release: prepare Meldwork V1.0.6

What changed
- Package version 0.1.5 -> 0.1.6 in desktop and frontend (package.json and the
  lock files).
- CHANGELOG entry for 0.1.6 (Meldwork-V1.0.6 prerelease).

Why
V1.0.5 was published on 2026-09-17 from the tag Meldwork-V1.0.5 (bb55fad).
Everything after it, 38 commits of Agent reliability, group conversation and
interface work, is version 1.0.6. This line was developed under the name
feature/1.0.5 by mistake. It is feature/1.0.6 now, and feature/1.0.5 is back at
the published V1.0.5.

Checks
- Only version fields and the changelog change, and no test pins the version.
- The code under test is that of ce995cb: frontend 357/357, desktop 1603/1603,
  deterministic evaluation 6 cases / 18 results.

AGENTS.md

@@ -11,7 +11,7 @@
 
 - Meldwork is a local-first desktop application. Do not introduce a required server, container, or cluster dependency unless the task explicitly calls for it.
 - Keep Agent execution in the Electron main process. Renderer access must go through narrow, validated preload APIs.
-- Never expose executable paths, tokens, provider credentials, or unrestricted shell execution to renderer code.
+- Never expose executable paths, tokens, credentials, or unrestricted shell execution to renderer code.
 - Store secrets only through operating-system-backed secure storage. Do not add plaintext fallbacks.
 - Keep workspace write access opt-in. Conversations and groups stay local; do not add remote participants, cloud channels, or remote conversation storage.
 

CHANGELOG.md

@@ -1,5 +1,90 @@
 # Changelog
 
+## 0.1.6 - 2026-10-17 (Meldwork-V1.0.6 prerelease)
+
+### Agent Reliability
+
+- Tier authentication failures instead of treating any authentication-looking diagnostic as a sign-in problem, and retry transient provider errors.
+- Keep long-lived ACP Agents (Hermes, Kimi, MiMo, OpenCode) usable across stops, restarts, and hangs, including when a group alternates between read-only and write turns.
+- Stop the watchdog from killing Agents that are still thinking or building, and keep an ACP turn alive through traffic this client does not recognise.
+- Keep an Agent available through an inconclusive readiness probe and preserve its shell environment.
+- Let MiMo and OpenCode turns that call tools finish instead of returning a partial answer, and report a CLI's own error instead of EPIPE.
+- Launch WorkBuddy's app-bundled CLI again (it now ships only the lite bundle) and stop skipping every WorkBuddy permission check in write mode.
+- Refuse unsafe ACP permission requests during read-only turns.
+- Give direct chats the current request and the delivery contract again after a native session is rebuilt.
+- Compare installed Agent CLI versions as releases, so a newer installed CLI is not offered a downgrade.
+- Write workspace state compactly and durably, and record unhandled main-process faults in the diagnostics log.
+- Internal: split the auto-runner and the Agent invocation into modules grouped by purpose.
+
+### Group Conversations
+
+- Show group replies in the order Agents were dispatched.
+- Keep natural discussions going instead of ending on one slip, ask Agents to engage with each other, and let peers see workspace file paths.
+- Keep group members after an Agent failure, so one proxy 403 or an expired sign-in no longer removes an Agent from later turns, and recover groups that were emptied.
+- A member that is not ready no longer blocks a message to the whole group; skipped members are named.
+- Deliver attachments and knowledge bases to group Agents, and send Hermes image attachments over ACP.
+- Conversations no longer wait for each other, even in the same folder; writers inside one conversation still take turns.
+- Import only files an Agent wrote, including paths with spaces.
+- Explain run failures in Chinese and English instead of showing raw codes.
+
+### Interface
+
+- Start without waiting for Agent detection: the sidebar, home page, and composer show that detection is running.
+- Mark conversations that wait for your decision in the sidebar and list them on the home page; mark group runs that finished while you were in another conversation.
+- Add a General settings tab with language and a System, Light, or Dark appearance; System follows the operating system until you choose.
+- Explain on each Agent card why an installed Agent is unavailable: its installed version, the verified range, and the failed check.
+- Use a shared type scale and semantic status colors, keep secondary text within contrast guidelines, and keep long notices on screen long enough to read.
+- Keep unsent text per conversation, add Skip and arrow-key navigation to first-run onboarding, ask before deleting a cloud server connection, and show Gemini CLI's own logo.
+
+### Verification And Distribution
+
+- Frontend 357/357, desktop 1603/1603, and deterministic evaluation 6 cases / 18 results passed on a clean checkout.
+- DMG and ZIP are ad-hoc signed, without Apple Developer ID signing or notarization. No installed user application is replaced automatically.
+- Live coverage is limited to observed CLI versions and configurations; this release did not repeat the live Agent acceptance run.
+
+## 0.1.5 - 2026-09-09 (Meldwork-V1.0.5 prerelease)
+
+### Local Agent Reliability
+
+- Accept validated OpenClaw gateway configuration migrations without disabling path, permission, or credential isolation checks.
+- Keep installed Agents visible when credential or capability checks fail; isolate failed probes and refresh the catalog after current readiness settles.
+- Refresh native shell settings before manual detection and preserve supported proxy, CA, cloud Provider, and model-region settings without silently replacing a selected Provider.
+
+### Group Conversations
+
+- Route peer requests through explicit Agent decisions instead of interpreting prose mentions or media keywords as commands.
+- Persist AI task completion separately from individual invocation success and human adoption; support single-Agent automatic work and exact visible answers.
+- Resume human clarification gates after restart, reject conflicting responses, and prevent repeated responses from replaying calls.
+- Serialize workspace writers, recover interrupted writes through bound approvals, retain partial results, and send acknowledged session context incrementally.
+- Preserve completed and stopped task states after restart and show healthy contributions alongside failed participants.
+
+### Verification And Distribution
+
+- Frontend 343/343, desktop 1562/1562, and deterministic evaluation 6 cases / 18 results passed.
+- Web and desktop builds, packaging, real packaged Codex/OpenClaw group execution, cancellation, and restart acceptance passed on Apple silicon macOS.
+- DMG and ZIP are ad-hoc signed, without Apple Developer ID signing or notarization. No installed user application is replaced automatically.
+- Live coverage is limited to observed CLI versions and configurations; cloud credential forwarding is not cloud-account certification. Full evidence: `docs/releases/Meldwork-V1.0.5.md`.
+
+## 0.1.4 - 2026-09-01 (Meldwork-V1.0.4 prerelease)
+
+### Natural Group Discussions
+
+- Sequential discussion mode now runs one Agent at a time in the configured CLI order and records every round visibly.
+- Auto Discussion starts with concurrent proposals, then lets Agent responses route the next turn through exact final-line `@Agent` mentions. One mention runs one Agent; multiple mentions run concurrently.
+- Natural discussion prompts allow Agents to exchange views before work and review instead of forcing every turn into a review-only role.
+
+### Session And Recovery
+
+- Reuses one native Agent session per Agent task across discussion rounds and restarts.
+- Persists round messages and checkpoints before and after Agent execution so completed turns are recovered without duplicate reruns.
+- Rejects stale operation and snapshot bindings before late results can be committed.
+
+### Verification And Distribution
+
+- Focused natural discussion coverage: 17/17 tests passed.
+- Desktop syntax checks and `git diff --check` passed.
+- The Apple silicon DMG and ZIP are ad-hoc signed and not notarized. Live CLI behavior still depends on each installed Agent's version, authentication, Provider, and capabilities.
+
 ## 0.1.3 - 2026-08-18 (Meldwork-V1.0.3 prerelease)
 
 ### Concurrent Conversations

CODE_OF_CONDUCT.md

@@ -0,0 +1,51 @@
+# Code of Conduct
+
+Meldwork is a workspace for people and Agents to build together. We want issues,
+pull requests, discussions, and community spaces to stay useful, direct, and
+respectful.
+
+## Expected Behavior
+
+- Be clear, constructive, and respectful when giving feedback.
+- Assume good intent while keeping discussions focused on facts, tradeoffs, and outcomes.
+- Respect different experience levels, backgrounds, languages, and working styles.
+- Keep disagreements technical, specific, and connected to the project.
+- Help maintainers keep issues and pull requests focused.
+- Protect private prompts, credentials, customer information, and other sensitive data.
+
+## Unacceptable Behavior
+
+The following behavior is not acceptable in Meldwork project spaces:
+
+- Personal attacks, insults, harassment, or sustained disruption.
+- Demeaning comments about identity, background, language, ability, or experience.
+- Threats, intimidation, stalking, or unwanted contact.
+- Publishing another person's private information without consent.
+- Spam, trolling, discrimination, or repeated off-topic posting.
+- Submitting malicious code, secrets, private data, or intentionally misleading reports.
+- Any behavior maintainers reasonably judge to make the project unsafe or unproductive.
+
+## Maintainer Responsibilities
+
+Maintainers may edit or remove comments, close issues, limit participation, or
+block participants when needed to protect the project community. Enforcement
+should be proportional, documented when practical, and focused on keeping
+collaboration healthy.
+
+## Reporting
+
+Report conduct concerns privately through the repository's GitHub contact or
+security reporting channel. Do not open a public issue for sensitive reports.
+
+Please include:
+
+- a short description of what happened;
+- links, screenshots, or message references when available;
+- whether you want follow-up, confidentiality, or no direct reply.
+
+Reports will be reviewed by project maintainers. We will handle reports with
+care and limit access to people who need the information to respond.
+
+This Code of Conduct applies wherever the project or its maintainers represent
+the Meldwork community. It is adapted from the Contributor Covenant and the
+community standards used by comparable open-source projects.

COMMERCIAL_USE.md

@@ -1,17 +1,14 @@
 # Commercial Use And Licensing
 
-Meldwork is source-available under the [Meldwork Non-Commercial Source License](LICENSE). The public repository can be inspected, evaluated, run, and modified for non-commercial purposes, but it is not licensed for commercial use.
+Meldwork is licensed under the [Apache License 2.0](LICENSE). Commercial use,
+private use, modification, redistribution, and use inside a paid service are
+permitted under the license terms.
 
-Any commercial use requires prior written permission from the copyright holder. This includes monetizing Meldwork, reselling or hosting it, embedding it in a proprietary product, using it to provide paid services, or using it as part of a for-profit organization's business operations.
+This page is a practical summary and does not replace the license text. Apache
+License 2.0 does not grant trademark rights to the Meldwork name, logo,
+screenshots, or other brand assets. Third-party software, Agent products,
+models, names, logos, and other materials remain subject to their own licenses
+and terms.
 
-The copyright holder may offer a separate commercial license for organizations that need permission for:
-
-- proprietary redistribution or resale of Meldwork or a modified version;
-- embedding Meldwork code in a closed-source product;
-- offering a hosted, managed, or network-accessible version;
-- internal business operations at a for-profit organization;
-- separately licensed commercial features, enterprise integrations, support, or managed deployment.
-
-A commercial license does not change the licenses of third-party dependencies, Agent products, names, models, or logos. The repository license also does not grant trademark rights to the Meldwork name, logo, screenshots, or other brand assets.
-
-This page is a practical summary and does not replace the license text or legal advice. When a planned use may be commercial, contact the repository owner through GitHub before using, distributing, hosting, reselling, or integrating the product.
+Before distributing a product that includes third-party materials, review the
+relevant notices and preserve the required attribution and license information.

CONTRIBUTING.md

@@ -1,18 +1,19 @@
 # Contributing
 
-Meldwork is source-available under the [Meldwork Non-Commercial Source License](LICENSE). Contributions are welcome when they improve the local-first desktop product, documentation, tests, or security posture.
+Meldwork is licensed under the [Apache License 2.0](LICENSE). Contributions are welcome when they improve the local-first desktop product, documentation, tests, or security posture.
 
-Unless a separate written agreement applies, a contribution accepted into this repository is available under the repository license. The project may also offer separately licensed commercial editions. Before external code is incorporated into a separately licensed commercial build, the maintainer may require a reviewed contributor-license agreement that grants the necessary relicensing rights. Opening a pull request by itself does not transfer copyright or silently grant those additional rights.
+Unless a separate written agreement applies, a contribution accepted into this repository is available under the repository license. Opening a pull request does not transfer copyright beyond the rights granted by the Apache License 2.0.
 
 ## Ground Rules
 
 - Keep changes small and tied to a clear product or engineering need.
 - Do not submit secrets, tokens, private logs, private prompts, or user data.
 - Do not add remote services, telemetry, credential collection, or server dependencies without explicit project approval.
-- Preserve the local-first boundary: renderer code must not gain unrestricted filesystem, shell, Provider credential, or executable-path access.
+- Preserve the local-first boundary: renderer code must not gain unrestricted filesystem, shell, credential, or executable-path access.
 - Include focused tests for behavior, security boundary, storage, or IPC changes.
-- Confirm that all submitted code and assets are yours to license and do not introduce incompatible license obligations.
+- Confirm that all submitted code and assets are yours to contribute and do not introduce incompatible license obligations.
 
-## Commercial Editions
+## Community Standards
 
-Commercial use of Meldwork requires prior written permission. Commercial editions, proprietary integrations, support, and other paid offerings may use separate terms; see [COMMERCIAL_USE.md](COMMERCIAL_USE.md).
+Please follow the [Code of Conduct](CODE_OF_CONDUCT.md) in issues, pull
+requests, discussions, and other project spaces.

LICENSE

@@ -1,82 +1,201 @@
-Meldwork Non-Commercial Source License 1.0
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
 
-Copyright (c) 2026 Ryder-MHumble.
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
 
-This license applies to Meldwork and the source code, documentation, assets, and
-other materials in this repository, except where a file or third-party notice
-states a different license.
+   1. Definitions.
 
-Meldwork is source-available, not open source in the OSI sense. Commercial use is
-not permitted without prior written permission from the copyright holder.
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
 
-1. Allowed non-commercial use
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
 
-You may download, inspect, run, copy, and modify Meldwork for personal learning,
-research, evaluation, security review, and other non-commercial purposes.
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
 
-You may share unmodified or modified copies for non-commercial purposes if you:
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
 
-- keep this license and all copyright, attribution, and trademark notices;
-- clearly mark your changes when you distribute a modified copy;
-- do not remove or obscure third-party notices; and
-- do not imply that the copyright holder endorses your copy or service.
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
 
-2. Commercial use requires permission
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
 
-You may not use Meldwork for a commercial purpose without a separate written
-license or other written permission from the copyright holder.
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
 
-Commercial purpose includes, without limitation:
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
 
-- selling, renting, sublicensing, or otherwise monetizing Meldwork or a modified
-  version;
-- offering Meldwork or a modified version as a hosted, managed, consulting,
-  outsourcing, integration, or support service;
-- embedding Meldwork code, UI, workflow, or packaged binaries in a proprietary,
-  paid, or revenue-generating product;
-- using Meldwork to provide paid deliverables, paid automation, paid Agent
-  operations, or other business services to a customer; and
-- use by or for a for-profit company, commercial product team, or revenue-driven
-  organization as part of its business operations.
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
 
-If you want to use Meldwork commercially, contact the repository owner and obtain
-written permission before using, distributing, hosting, reselling, or integrating
-the product.
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
 
-3. No trademark license
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
 
-This license does not grant rights to use the Meldwork name, logo, product
-identity, screenshots, or other brand assets except as needed to describe an
-unaltered copy of the software in a truthful, non-confusing way.
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
 
-4. Third-party materials
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
 
-Third-party software, Agent products, models, names, logos, dependencies, and
-other materials remain subject to their own licenses and terms. This license does
-not give you rights that the copyright holder does not own.
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
 
-5. Contributions
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
 
-Unless a separate written agreement says otherwise, contributions submitted to
-this repository may be used under this license. The copyright holder may require
-a separate contributor agreement before accepting contributions intended for a
-commercially licensed edition.
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
 
-6. Termination
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
 
-If you violate this license, your rights under it terminate automatically. The
-copyright holder may reinstate rights in writing after the violation is cured.
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
 
-7. No warranty
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
 
-Meldwork is provided "as is", without warranties or conditions of any kind,
-including implied warranties of merchantability, fitness for a particular
-purpose, title, and non-infringement. You are responsible for evaluating whether
-the software is suitable and lawful for your use.
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
 
-8. Limitation of liability
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
 
-To the maximum extent permitted by law, the copyright holder and contributors
-will not be liable for any damages arising from or related to the software or
-this license, including direct, indirect, incidental, consequential, special,
-exemplary, punitive, or business-interruption damages.
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -1,9 +1,7 @@
 Meldwork
 Copyright 2026 Ryder-MHumble and Meldwork contributors.
 
-Meldwork is licensed under the Meldwork Non-Commercial Source License. See
-LICENSE. Commercial use requires separate written permission from the copyright
-holder.
+Meldwork is licensed under the Apache License 2.0. See LICENSE.
 
 Third-party Agent names, product names, and logos remain the property of their
 respective owners. They are used only to identify compatible integrations and

README.md

@@ -1,11 +1,3 @@
-<!-- <p align="center">
-  <picture>
-    <source media="(prefers-color-scheme: dark)" srcset="frontend/public/logos/meldwork-wordmark-v3-dark.svg">
-    <source media="(prefers-color-scheme: light)" srcset="frontend/public/logos/meldwork-wordmark-v3.svg">
-    <img src="frontend/public/logos/meldwork-wordmark-v3.svg" alt="Meldwork" width="360">
-  </picture>
-</p> -->
-
 <p align="center">
   <img src="frontend/public/logos/meldwork-readme-banner-en.png" alt="Meldwork README banner" width="100%">
 </p>
@@ -14,113 +6,129 @@
   <strong>English</strong> · <a href="README.zh-CN.md">简体中文</a>
 </p>
 
-# Meldwork builds the organization layer for AI agents.
+# Meldwork — A local-first workspace for General Agents
 
-**Most tools help humans run more agents. Meldwork is built for a different job: make multi-agent work legible, accountable, and reusable.**
+**Turn several General Agents into a reviewable working group. Meldwork gives every task a shared context, lets selected Agents investigate independently or discuss across rounds, and keeps the evidence and human decision visible before anything is adopted.**
 
-The current preview is a local work cell for supported Agent CLIs. It keeps participants explicit, context scoped, collaboration bounded, run state inspectable, and human review in the loop. Selected Agents can now respond concurrently or move through independent proposals, peer negotiation, responsibility-based work, single-writer integration, and independent verification around one goal.
+Meldwork is a local-first Electron workspace for using the Agent tools you already have on your Mac. It supports focused conversations, independent concurrent responses, and Auto Discussion V4 for proposal, challenge, negotiated work, synthesis, and verification. Agents can work with text, files, images, media, Skills, and selected knowledge sources, so the same workspace can support research, analysis, writing, planning, review, and implementation.
+
+The product is organized around a durable review record: **Case → Finding → Evidence → Decision → Disposition**. Meldwork keeps the task snapshot, Agent contributions, run state, artifacts, Human Gates, and adoption decision together in one local work cell. Workspace writes are opt-in and remain under the user's control.
 
 <p align="center">
-  <a href="https://github.com/Ryder-MHumble/Meldwork/releases/download/Meldwork-V1.0.3/Meldwork-0.1.3-arm64.dmg"><strong>Download Meldwork V1.0.3 for Apple silicon macOS</strong></a>
+  <a href="https://github.com/Ryder-MHumble/Meldwork/releases/download/Meldwork-V1.0.5/Meldwork-0.1.5-arm64.dmg"><strong>Download Meldwork V1.0.5 for Apple silicon macOS</strong></a>
   · <a href="architecture.md">Architecture</a>
   · <a href="LICENSE">License</a>
 </p>
 
-## The missing layer in the Agent workforce
+## What Meldwork is for
 
-Agent supply is growing quickly. The organization around it is not.
+Meldwork is useful when a decision benefits from more than one perspective and a clear record of why a result was adopted. Typical work includes:
 
-Today, people still choose every Agent, move every piece of context, explain every role, reconcile conflicting answers, and carry the final accountability. Parallel managers improve throughput. Agent frameworks let developers program workflows. Single-Agent products improve one worker.
+- research and information synthesis across local Agent tools and selected knowledge sources;
+- product, market, operational, and technical analysis;
+- writing, planning, review, and structured document work;
+- multimodal work with local documents, images, audio, video, PDFs, and code or configuration files;
+- implementation and review tasks where findings, evidence, permissions, and the final adoption decision should remain inspectable.
 
-Meldwork is focused on the work between them:
+## How a task moves through Meldwork
 
-- one shared goal and acceptance boundary;
-- independent viewpoints instead of hidden routing;
-- explicit authority, responsibility, handoffs, and stopping conditions;
-- artifacts and evidence that survive Agent changes;
-- a human-controlled adoption decision.
+1. **Select** the Agents and participants for this Case.
+2. **Scope** the goal, working directory, context, attachments, Skills, knowledge sources, and permissions.
+3. **Run** a direct conversation, Concurrent Responses, or Auto Discussion V4.
+4. **Review** the findings, evidence, artifacts, trace, and any Human Gate.
+5. **Adopt** the approved result or leave the Case unresolved for further work.
 
-## What Meldwork is building
+## Collaboration modes
 
-The collaboration mechanism is not a hidden preset A → B → C assignment. The user selects the participants; those Agents expose the solution space, negotiate responsibilities, and execute inside Harness-enforced context, permission, budget, recovery, and evidence boundaries.
-
-The selected-Agent proposal, challenge, responsibility, work, integration, and verification loop is available today. Automatic participant selection from a broader roster, remote Agents, enterprise governance, and outcome-driven reputation remain future direction.
-
-### Demo
-
-| Local Agent discovery | Direct multimodal work | Concurrent multi-Agent work |
+| Mode | What happens | Best for |
 | --- | --- | --- |
-| [![Meldwork detecting supported local Agent CLIs](assets/meldwork-agent-discovery.png)](assets/meldwork-agent-discovery.png) | [![A direct Agent conversation returning local files and media](assets/meldwork-direct-multimodal.png)](assets/meldwork-direct-multimodal.png) | [![A group conversation with user-selected Agent replies](assets/meldwork-group-collaboration.png)](assets/meldwork-group-collaboration.png) |
-| See which supported Agents are ready. | Keep prompts, files, permissions, and compatible sessions together. | Run concurrent replies or an Agent-negotiated collaboration with the participants you selected. |
+| **Direct** | One selected Agent keeps its conversation and native session when supported. | Focused research, writing, analysis, or execution. |
+| **Concurrent Responses** | Selected Agents receive the same frozen task snapshot and return independent responses in stable order. | Comparing perspectives before choosing a path. |
+| **Auto Discussion V4** | Agents propose, challenge, negotiate responsibilities, execute agreed work, synthesize a candidate result, and verify it across bounded rounds. Native Agent sessions continue when supported. | Multi-step work that needs discussion, division of responsibility, and review. |
 
-## What makes Meldwork different
+Participants are selected by the user. Automatic selection from a larger roster is outside the current preview.
 
-### Harness design
+## The review record
 
-The Harness is the control plane that keeps conversations, context, adapters, streams, and durable state aligned in one workspace. It is what makes a run continuous across sessions, timeouts, recovery, and Agent changes.
+Meldwork treats collaboration as a decision process rather than a stream of chat messages:
 
-<p align="center">
-  <img src="frontend/public/logos/Harness-readme.png" alt="Meldwork Harness" width="100%">
-</p>
+- **Case** defines the question, scope, context, and permissions.
+- **Finding** records an Agent's claim, observation, or proposed action.
+- **Evidence** links the finding to a response, artifact, file, knowledge result, or other bounded reference.
+- **Decision** records what the evidence supports, what remains uncertain, and which path is selected.
+- **Disposition** records whether the result was accepted, revised, rejected, superseded, or left unresolved.
 
-This is a comparison of working models, not a feature-scorecard claiming Meldwork should replace every tool.
+The Run Ledger preserves phase, participant, attempt, receipt, artifact, recovery, and Human Gate state. Diagnostic tool output and sensitive runtime details stay behind the Electron main-process boundary.
 
-| Approach | Examples | Best at | Main tradeoff | Meldwork's choice |
-| --- | --- | --- | --- | --- |
-| One Agent CLI or AI IDE | Claude Code, Codex, Cursor, Windsurf | Deep work inside one Agent or product ecosystem | Moving to another Agent still requires a manual handoff | Preserve work across heterogeneous Agents |
-| Parallel coding-Agent managers | Superset, Vibe Kanban, Nimbalyst, Conductor | Running multiple coding tasks and isolated workspaces in parallel | Primarily optimized for throughput and code delivery | Optimize proposal quality, responsibility, evidence, and adoption |
-| Cloud Agent control planes | GitHub Agent HQ, Warp Oz, Devin, Factory, OpenHands Cloud | Dispatching and observing remote jobs, sandboxes, and Agent fleets | Work is organized around cloud or platform infrastructure | Start local-first and bring your existing Agents |
-| Programmable Agent frameworks | LangGraph, CrewAI, AutoGen, Agents SDK, ADK | Building custom graphs, roles, routing, and automation | Developers must design and maintain the orchestration | Deliver an organization and acceptance experience to users |
-| Multiple terminals and scripts | tmux, shell scripts, manual copy and paste | Maximum access to native Agent capabilities | The user becomes the router, context bus, and final judge | Make the work, boundaries, and decisions durable |
-| **Meldwork** | Supported local Agent CLIs; remote supply is direction | Concurrent independent responses and selected-Agent proposal, negotiation, responsibility, evidence, and verification | Local single-user preview today; participants are still selected by the user | Build the organization layer for the Agent workforce |
+## General Agent catalog
 
-If one Agent and one work surface are enough, use the native tool. Use Meldwork when handoff, independent judgment, authority, evidence, and acceptance become part of the problem.
+Meldwork detects and invokes compatible locally installed Agent commands. The catalog currently includes:
 
-## Install
+<table>
+  <tr>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/codex.svg" alt="Codex logo" width="32" height="32"><br><strong>Codex</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/hermes.png" alt="Hermes logo" width="32" height="32"><br><strong>Hermes</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/openclaw-transparent.png" alt="OpenClaw logo" width="32" height="32"><br><strong>OpenClaw</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/workbuddy.png" alt="WorkBuddy logo" width="32" height="32"><br><strong>WorkBuddy</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/pi.svg" alt="Pi Agent logo" width="32" height="32"><br><strong>Pi Agent</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/kimi.png" alt="Kimi Code logo" width="32" height="32"><br><strong>Kimi Code</strong></td>
+  </tr>
+  <tr>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/mimo.svg" alt="MiMo Code logo" width="32" height="32"><br><strong>MiMo Code</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/claude.png" alt="Claude Code logo" width="32" height="32"><br><strong>Claude Code</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/gemini.svg" alt="Gemini CLI logo" width="32" height="32"><br><strong>Gemini CLI</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/opencode.svg" alt="OpenCode logo" width="32" height="32"><br><strong>OpenCode</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/qwen.svg" alt="Qwen Code logo" width="32" height="32"><br><strong>Qwen Code</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/opencodereview.svg" alt="OpenCodeReview logo" width="32" height="32"><br><strong>OpenCodeReview</strong></td>
+  </tr>
+</table>
 
-The current release-validated target is Apple silicon macOS.
+Approved Agent Connectors and custom executable Agents use the desktop connector paths. The [desktop guide](desktop/README.md) contains the current adapter, installation, provider, and capability matrix.
 
-1. Open the [Meldwork V1.0.3 prerelease](https://github.com/Ryder-MHumble/Meldwork/releases/tag/Meldwork-V1.0.3).
-2. Download Meldwork-0.1.3-arm64.dmg.
-3. Drag Meldwork into Applications and try to open it once.
-4. This prerelease is ad-hoc signed, not signed with an Apple Developer ID, and not Apple-notarized. If macOS blocks it, open **System Settings → Privacy & Security**, find the Meldwork warning, choose **Open Anyway**, and confirm.
-5. Connect a supported Agent CLI already installed on the computer, or configure an independent Provider profile where supported.
+## Context and local boundaries
 
-Use only artifacts downloaded from the official Release.
+Meldwork runs as a local Electron application. Conversations, group configuration, run records, and app-owned attachments stay in the local user data directory. The renderer receives only validated snapshots through a narrow preload API; executable paths, credentials, native session references, Skill paths, and unrestricted shell access remain in the main process.
 
-<details>
-<summary><strong>Build from source</strong></summary>
+Selected Agents may send prompts, attachments, or Skills to the model provider configured in their own CLI. Knowledge access is explicit and bounded: the current implementation supports local Obsidian retrieval and CLI-owned Feishu or DingTalk access modes. Local-first describes where Meldwork stores and coordinates work; it does not promise that an Agent's model provider is offline.
 
-Requires Node.js 22.12 or newer and npm.
+## Download V1.0.5
 
-~~~bash
+For Apple silicon macOS, download the [Meldwork V1.0.5 prerelease](https://github.com/Ryder-MHumble/Meldwork/releases/tag/Meldwork-V1.0.5) and choose [`Meldwork-0.1.5-arm64.dmg`](https://github.com/Ryder-MHumble/Meldwork/releases/download/Meldwork-V1.0.5/Meldwork-0.1.5-arm64.dmg). The release also includes a ZIP archive and SHA-256 manifest. Install at least one supported local Agent CLI before opening the app.
+
+The preview artifacts are ad-hoc signed and not notarized with an Apple Developer ID. macOS may require **Open Anyway** under **System Settings → Privacy & Security** on first launch.
+
+## Run from source
+
+Prerequisites: Node.js `22.12+` and npm.
+
+```bash
 npm --prefix frontend ci
 npm --prefix desktop ci
 npm --prefix desktop run dev
-~~~
+```
 
-Run the repository checks before distributing a build:
+Before distributing a build:
 
-~~~bash
+```bash
 npm --prefix frontend test
 npm --prefix frontend run build
 npm --prefix frontend run build:desktop
 npm --prefix desktop test
-~~~
+```
 
-</details>
+## V1.0.5
 
-## Trust boundaries
+V1.0.5 improves local Agent readiness and recovery, preserves healthy participants when another Agent fails, keeps completed and stopped work visible after restart, and tightens group execution and output handling. It also updates the workspace preference flow, sidebar and titlebar details, and the Pi capability probe contract.
 
-- Conversations and Meldwork orchestration state stay local.
-- Model requests still follow the Agent and Provider you choose; local-first does not mean every model runs offline.
-- Credentials use operating-system-backed secure storage where supported.
-- Workspace writes are explicit and opt-in.
-- Raw chain-of-thought, executable paths, secrets, and unrestricted command output are not exposed to the renderer.
-- Cloud, Channel, automatic participant selection, enterprise governance, and an Outcome Network are not current release promises.
+Verification for this prerelease includes 343/343 frontend tests, 1,562/1,562 desktop tests, six deterministic evaluation cases with 18 results, web and desktop builds, packaging, ZIP integrity, deep code-signature verification, and packaged macOS acceptance for group execution, cancellation, and restart recovery. The artifacts remain ad-hoc signed and unnotarized; live behavior still depends on the installed Agent CLI version, authentication, Provider, and capabilities.
 
-Read the [architecture](architecture.md) and [Agent Connector contract](docs/agent-connector-sdk.md) for the public technical boundary.
+## Documentation
+
+- [Architecture and product boundary](architecture.md)
+- [Desktop setup and Agent matrix](desktop/README.md)
+- [Contributing](CONTRIBUTING.md)
+- [Security policy](SECURITY.md)
 
 ## License
 
-Meldwork is source-available under the [Meldwork Non-Commercial Source License](LICENSE). You may inspect, run, modify, and share it for non-commercial purposes. Commercial use requires prior written permission; see [COMMERCIAL_USE.md](COMMERCIAL_USE.md).
+Meldwork is source-available under the [Meldwork Non-Commercial Source License 1.0](LICENSE). You may inspect, run, modify, and share it for non-commercial purposes. Commercial use requires prior written permission; see [COMMERCIAL_USE.md](COMMERCIAL_USE.md).

README.zh-CN.md

@@ -1,11 +1,3 @@
-<!-- <p align="center">
-  <picture>
-    <source media="(prefers-color-scheme: dark)" srcset="frontend/public/logos/meldwork-wordmark-v3-dark.svg">
-    <source media="(prefers-color-scheme: light)" srcset="frontend/public/logos/meldwork-wordmark-v3.svg">
-    <img src="frontend/public/logos/meldwork-wordmark-v3.svg" alt="Meldwork" width="360">
-  </picture>
-</p> -->
-
 <p align="center">
   <img src="frontend/public/logos/meldwork-readme-banner-cn.png" alt="Meldwork README banner" width="100%">
 </p>
@@ -14,115 +6,129 @@
   <a href="README.md">English</a> · <strong>简体中文</strong>
 </p>
 
-# Meldwork 为 AI Agent 构建组织层
+# Meldwork：本地优先的 General Agent 工作空间
 
-**多数工具是在帮人类管更多 Agent；Meldwork 解决的是另一件事：让多 Agent 协作变得可见、可追责、可复用。**
+**把多个 General Agent 组织成一个可复核的工作小组。Meldwork 为每个任务建立共享上下文，让选定的 Agent 独立调查或分轮讨论，并在采用结果前保留完整的证据和人工决策。**
 
-当前预览版是一个面向已支持本地 Agent CLI 的本地工作单元。它把参与者、上下文、协作边界、运行状态和人类复核放在同一个工作空间里。已选 Agent 现在可以并发回复，也可以围绕同一目标先独立提案、再协商职责、分工执行、单写者整合并由其他 Agent 独立复核。
+Meldwork 是一个本地优先的 Electron 工作空间，用来连接你已经在电脑上使用的 Agent 工具。它支持直接会话、独立并发回复，以及包含提案、质询、职责协商、分工执行、整合和复核的 Auto Discussion V4。Agent 可以处理文字、文件、图片、音视频、Skill 和选定的知识源，因此同一个工作空间可以覆盖研究、分析、写作、规划、评审和实施。
+
+产品围绕一条可持续复核的记录组织工作：**Case → Finding → Evidence → Decision → Disposition**。Meldwork 将任务快照、Agent 贡献、运行状态、产物、人工审批门和采用决定保存在同一个本地工作单元中。工作区写入默认受控，并由用户明确授权。
 
 <p align="center">
-  <a href="https://github.com/Ryder-MHumble/Meldwork/releases/download/Meldwork-V1.0.3/Meldwork-0.1.3-arm64.dmg"><strong>下载 Meldwork V1.0.3 Apple 芯片 macOS 版</strong></a>
+  <a href="https://github.com/Ryder-MHumble/Meldwork/releases/download/Meldwork-V1.0.5/Meldwork-0.1.5-arm64.dmg"><strong>下载 Meldwork V1.0.5 Apple 芯片 macOS 版</strong></a>
   · <a href="architecture.md">架构</a>
   · <a href="LICENSE">许可证</a>
 </p>
 
-## Agent workforce 缺少的那一层
+## Meldwork 适合什么工作
 
-Agent 的供给正在快速增长，但围绕它们工作的组织能力并没有跟上。
+当一个决定需要多个视角，并且需要留下清晰的采用依据时，Meldwork 更适合：
 
-今天，人类仍然需要自己选 Agent、搬运上下文、解释角色、处理冲突答案，并承担最终责任。并行管理器优化吞吐量，多 Agent 框架让开发者编写 Workflow，单 Agent 产品优化一个 Worker。
+- 研究与信息综合：结合本地 Agent 工具和明确选择的知识源；
+- 产品、市场、运营和技术分析；
+- 写作、规划、评审和结构化文档工作；
+- 处理本地文档、图片、音频、视频、PDF、代码和配置文件的多模态工作；
+- 需要保留发现、证据、权限、运行轨迹和最终采用决定的实施工作。
 
-Meldwork 关注的是中间那层工作：
+## 一个任务如何在 Meldwork 中推进
 
-- 一个共享 Goal 和明确验收边界；
-- 独立判断，而不是隐藏路由；
-- 明确的权限、责任、交接和停止条件；
-- 不随 Agent 切换而丢失的 Artifact 与 Evidence；
-- 由人类掌握的最终采用决定。
+1. **选择**本 Case 的 Agent 和参与者。
+2. **限定**目标、工作目录、上下文、附件、Skill、知识源和权限。
+3. **运行**直接会话、并发回复或 Auto Discussion V4。
+4. **复核**发现、证据、产物、运行轨迹和人工审批门。
+5. **采用**批准的结果，或保留为待解决状态继续处理。
 
+## 协作模式
 
-## Meldwork 正在构建什么
-
-当前协作机制不是 Harness 暗中预设的 A → B → C 分配链。用户选定参与者后，Agent 先展开解空间、自主协商职责，再在 Harness 管理的上下文、权限、预算、恢复和证据边界内执行。
-
-已选 Agent 的独立提案、交叉质询、职责协商、分工执行、整合和复核现在已经可用。从更大候选池自动选人组队、远程 Agent、企业治理和 Outcome 驱动的声誉系统仍是未来方向。
-
-### 效果演示
-
-| 本地 Agent 检测 | 单聊中的多模态工作 | 并发多 Agent 协作 |
+| 模式 | 运行方式 | 适合场景 |
 | --- | --- | --- |
-| [![Meldwork 检测已支持本地 Agent CLI](assets/meldwork-agent-discovery.png)](assets/meldwork-agent-discovery.png) | [![Agent 在直接会话中返回本地文件和媒体](assets/meldwork-direct-multimodal.png)](assets/meldwork-direct-multimodal.png) | [![用户选择多个 Agent 并分别检查结果](assets/meldwork-group-collaboration.png)](assets/meldwork-group-collaboration.png) |
-| 查看哪些受支持 Agent 已就绪。 | Prompt、文件、权限和兼容 Session 留在一起。 | 让已选 Agent 并发回复，或进入由 Agent 自主协商职责的协作流程。 |
+| **直接会话** | 一个已选 Agent 在适配器支持时保持对话和原生会话。 | 聚焦研究、写作、分析或执行。 |
+| **并发回复** | 已选 Agent 收到同一份冻结任务快照，并按稳定顺序返回独立回复。 | 在选择方案前比较不同视角。 |
+| **Auto Discussion V4** | Agent 先提出方案，再进行质询、职责协商、分工执行、候选结果整合和独立复核；支持时会在多轮中保持原生会话。 | 需要讨论、分工和复核的多步骤工作。 |
 
+参与者始终由用户选择。从更大候选池自动选人组队不属于当前预览版。
 
-## Meldwork 和别的产品有什么区别
+## 复核记录
 
-### Harness设计
+Meldwork 将协作组织成一套决策过程，而不是一串聊天消息：
 
-Harness 是把会话、上下文、适配器、流和持久化状态拴在一起的控制平面。它让一次任务可以跨会话、超时、恢复和 Agent 切换继续下去。
+- **Case**：定义问题、范围、上下文和权限；
+- **Finding**：记录 Agent 的判断、观察或行动建议；
+- **Evidence**：将 Finding 连接到回复、产物、文件、知识结果或其他受约束的依据；
+- **Decision**：记录证据支持的结论、未决问题和选定路径；
+- **Disposition**：记录结果已接受、修订、拒绝、被替代或仍未解决。
 
-<p align="center">
-  <img src="frontend/public/logos/Harness-readme.png" alt="Meldwork Harness" width="100%">
-</p>
+Run Ledger 会保留阶段、参与者、尝试、回执、产物、恢复状态和人工审批门。诊断工具输出与敏感运行信息留在 Electron 主进程边界内。
 
-这张表比较的是不同工作范式，不是一张声称 Meldwork 应当替代所有工具的功能勾选表。
+## General Agent 目录
 
-| 范式 | 例子 | 最擅长什么 | 主要代价 | Meldwork 的选择 |
-| --- | --- | --- | --- | --- |
-| 单 Agent CLI / AI IDE | Claude Code、Codex、Cursor、Windsurf | 在一个 Agent 或产品生态内完成深度工作 | 换 Agent 时仍需手工交接 | 保留跨异构 Agent 的工作 |
-| 并行 Coding-Agent 工作台 | Superset、Vibe Kanban、Nimbalyst、Conductor | 并行运行多个 Coding 任务和隔离工作区 | 主要优化吞吐量和代码交付 | 优化方案质量、责任、证据和采用 |
-| Cloud Agent 控制面 | GitHub Agent HQ、Warp Oz、Devin、Factory、OpenHands Cloud | 派发和观察远程 Job、Sandbox 与 Agent Fleet | 工作围绕云基础设施和平台边界 | 先从 Local-first 和 BYO Agent 切入 |
-| 可编程多 Agent 框架 | LangGraph、CrewAI、AutoGen、Agents SDK、ADK | 开发者自定义 Graph、角色、路由和自动化 | 需要开发者设计并维护编排 | 面向使用者交付组织与验收体验 |
-| 多终端与脚本 | tmux、Shell、手工复制粘贴 | 最大化原生 Agent 访问权 | 人类成为 Router、Context Bus 和最终裁决者 | 让工作、边界和决定保持持续 |
-| **Meldwork** | 已支持本地 Agent CLI；远程供给属于方向 | 并发独立回复，以及已选 Agent 之间的提案、协商、责任、Evidence 和复核 | 当前仍是本地单用户预览；参与者仍由用户选定 | 构建 Agent workforce 的组织层 |
+当适配器和 CLI 版本兼容时，Meldwork 会检测并调用已安装的 Agent 命令。当前目录包括：
 
-如果一个 Agent 和一个工作面已经足够，直接使用原生工具。只有当交接、独立判断、权限、证据和验收成为问题时，才需要 Meldwork。
+<table>
+  <tr>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/codex.svg" alt="Codex logo" width="32" height="32"><br><strong>Codex</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/hermes.png" alt="Hermes logo" width="32" height="32"><br><strong>Hermes</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/openclaw-transparent.png" alt="OpenClaw logo" width="32" height="32"><br><strong>OpenClaw</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/workbuddy.png" alt="WorkBuddy logo" width="32" height="32"><br><strong>WorkBuddy</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/pi.svg" alt="Pi Agent logo" width="32" height="32"><br><strong>Pi Agent</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/kimi.png" alt="Kimi Code logo" width="32" height="32"><br><strong>Kimi Code</strong></td>
+  </tr>
+  <tr>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/mimo.svg" alt="MiMo Code logo" width="32" height="32"><br><strong>MiMo Code</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/claude.png" alt="Claude Code logo" width="32" height="32"><br><strong>Claude Code</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/gemini.svg" alt="Gemini CLI logo" width="32" height="32"><br><strong>Gemini CLI</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/opencode.svg" alt="OpenCode logo" width="32" height="32"><br><strong>OpenCode</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/qwen.svg" alt="Qwen Code logo" width="32" height="32"><br><strong>Qwen Code</strong></td>
+    <td align="center" valign="top" width="16.66%"><img src="frontend/public/agent-logos/opencodereview.svg" alt="OpenCodeReview logo" width="32" height="32"><br><strong>OpenCodeReview</strong></td>
+  </tr>
+</table>
 
-## 安装
+已批准的 Agent Connector 和自定义可执行 Agent 使用桌面端的连接器入口。完整的适配器、安装、Provider 和能力矩阵见[桌面端指南](desktop/README.md)。
 
-当前经过发布验证的目标是 Apple 芯片 macOS。
+## 上下文与本地边界
 
-1. 打开 [Meldwork V1.0.3 预发布页](https://github.com/Ryder-MHumble/Meldwork/releases/tag/Meldwork-V1.0.3)。
-2. 下载 Meldwork-0.1.3-arm64.dmg。
-3. 将 Meldwork 拖入“应用程序”，然后先尝试打开一次。
-4. 当前预发布版使用临时签名，没有 Apple Developer ID 签名，也尚未经过 Apple 公证。如果 macOS 阻止启动，请打开“系统设置 → 隐私与安全性”，找到 Meldwork 提示，点击“仍要打开”并确认。
-5. 连接电脑上已安装的受支持 Agent CLI，或在兼容情况下单独配置 Provider。
+Meldwork 作为本地 Electron 应用运行。对话、群组配置、运行记录和应用管理的附件保存在本地用户数据目录。渲染进程只通过受约束的 preload API 获取校验后的状态；可执行路径、凭据、原生会话引用、Skill 路径和任意 Shell 权限都留在主进程中。
 
-请只使用官方 Release 下载的产物。
+选定的 Agent 仍可能把 Prompt、附件或 Skill 发送到其自身 CLI 凭据对应的模型服务商。知识访问必须明确选择并受边界约束：当前实现支持本地 Obsidian 检索，以及由本地 CLI 管理的飞书或钉钉访问模式。Local-first 描述的是 Meldwork 的数据和协作位置，不代表 Agent 使用的模型服务商一定离线。
 
-<details>
-<summary><strong>从源码运行</strong></summary>
+## 下载 V1.0.5
 
-需要 Node.js 22.12 或更高版本以及 npm。
+Apple 芯片 macOS 用户请前往 [Meldwork V1.0.5 预发布页](https://github.com/Ryder-MHumble/Meldwork/releases/tag/Meldwork-V1.0.5)，下载 [`Meldwork-0.1.5-arm64.dmg`](https://github.com/Ryder-MHumble/Meldwork/releases/download/Meldwork-V1.0.5/Meldwork-0.1.5-arm64.dmg)。发布页同时提供 ZIP 压缩包和 SHA-256 校验文件。启动应用前，请至少安装一个受支持的本地 Agent CLI。
 
-~~~bash
+该预览版安装包使用 ad-hoc 临时签名，未使用 Apple Developer ID 签名，也未进行公证。macOS 首次启动时可能需要在“系统设置 → 隐私与安全性”中选择“仍要打开 / Open Anyway”。
+
+## 从源码运行
+
+前置条件：Node.js `22.12+` 和 npm。
+
+```bash
 npm --prefix frontend ci
 npm --prefix desktop ci
 npm --prefix desktop run dev
-~~~
+```
 
-分发前运行：
+分发构建前运行：
 
-~~~bash
+```bash
 npm --prefix frontend test
 npm --prefix frontend run build
 npm --prefix frontend run build:desktop
 npm --prefix desktop test
-~~~
+```
 
-</details>
+## V1.0.5
 
-## 信任边界
+V1.0.5 改进了本地 Agent 的就绪检测与恢复，在单个 Agent 失败时保留健康参与者，重启后继续显示已完成和已停止的工作，并加强了群组执行与结果导入。同时更新了工作区偏好、侧栏和标题栏细节，以及 Pi 能力探测契约。
 
-- Meldwork 会话和编排状态保存在本地。
-- 模型请求仍然按照你选择的 Agent 和 Provider 发出；Local-first 不等于所有模型离线运行。
-- 兼容环境使用操作系统安全存储保存凭据。
-- 工作目录写入必须显式开启。
-- Renderer 不暴露 raw chain-of-thought、可执行路径、Secret 或不受限命令输出。
-- Cloud、Channel、自动选人组队、企业治理和 Outcome Network 不是当前发布承诺。
+本预览版已完成 343/343 前端测试、1,562/1,562 桌面端测试、6 个确定性评估场景共 18 条结果验证，并通过 Web 与桌面构建、打包、ZIP 完整性、深度代码签名校验，以及 Apple 芯片 macOS 上的群组执行、取消和重启恢复验收。安装包仍为 ad-hoc 临时签名且未公证；实际运行结果仍取决于本机 Agent CLI 版本、鉴权、Provider 和能力。
 
-公开技术边界见[架构](architecture.md)和 [Agent Connector 契约](docs/agent-connector-sdk.md)。
+## 文档
+
+- [架构与产品边界](architecture.md)
+- [桌面端设置与 Agent 矩阵](desktop/README.md)
+- [贡献指南](CONTRIBUTING.md)
+- [安全策略](SECURITY.md)
 
 ## 许可证
 
-Meldwork 采用 [Meldwork 非商用源码许可](LICENSE)。你可以为非商用目的查看、运行、修改和分享；商业使用需要事先书面许可，详见 [COMMERCIAL_USE.md](COMMERCIAL_USE.md)。
+Meldwork 采用 [Meldwork 非商用源码许可 1.0](LICENSE)。你可以出于非商业目的查看、运行、修改和分享。商业使用需要事先书面许可，详见 [COMMERCIAL_USE.md](COMMERCIAL_USE.md)。

SECURITY.md

@@ -1,6 +1,6 @@
 # Security Policy
 
-Meldwork is a local-first desktop application that invokes local Agent CLIs, stores local conversations, imports validated image, media, document, code, and archive attachments, and manages Provider credentials through the Electron main process.
+Meldwork is a local-first desktop application that invokes local Agent CLIs, stores local conversations, imports validated image, media, document, code, and archive attachments, and runs each Agent with its own native CLI credentials through the Electron main process.
 
 ## Reporting
 
@@ -11,10 +11,10 @@ Do not open a public issue with exploit details, secrets, local file paths, or p
 ## In Scope
 
 - Renderer-to-main IPC authorization bypasses.
-- Exposure of executable paths, attachment paths, native session references, Provider API keys, or Skill file paths to renderer code.
+- Exposure of executable paths, attachment paths, native session references, CLI API keys, or Skill file paths to renderer code.
 - Arbitrary shell execution, arbitrary file read/write, unsafe navigation, or unsafe installer behavior.
 - Attachment validation, private storage, checksum, or preview isolation failures.
-- Credential storage or Provider configuration failures that expose secrets.
+- Credential storage failures that expose secrets.
 
 ## Out of Scope
 

architecture.md

@@ -2,7 +2,7 @@
 
 ## Product Boundary
 
-Meldwork is a local Electron desktop product for direct and multi-Agent work with explicitly selected Skill, image, and knowledge-source context. It has no account system, tenant model, required application server, remote conversation store, email delivery, scheduled jobs, or public SEO surface. Optional Agent, Cloud, and Channel Connectors run behind the Electron main-process trust boundary and are disabled until explicitly configured.
+Meldwork is a local Electron desktop product for multi-agent orchestration and review with explicitly selected Skill, image, and knowledge-source context. It has no account system, tenant model, required application server, remote conversation store, email delivery, scheduled jobs, or hosted web surface. Optional Agent, Cloud, and Channel Connectors run behind the Electron main-process trust boundary and are disabled until explicitly configured. The landing page at https://ryder-mhumble.github.io/Meldwork/ is the public marketing surface for search engine and AI assistant discoverability.
 
 The supported product surface is the packaged Electron application. The frontend is a renderer bundle; without the local preload bridge it shows a desktop-required state and cannot operate the workspace.
 
@@ -11,15 +11,15 @@ The supported product surface is the packaged Electron application. The frontend
 | Component | Responsibility | Primary evidence |
 | --- | --- | --- |
 | Vue renderer | Agent-first multi-session UI, direct-session rename/delete, live direct-chat conclusions and inline execution summaries, group-chat result streaming with an expandable per-Agent/right-side trace panel, per-round navigation, background completion markers, first-run carousel, conversation/Skill/image/knowledge composer UI, bounded attachment-preview cache, theme/i18n, user confirmation and input collection | `frontend/src/App.vue`, `frontend/src/components/RunTracePanel.vue`, `frontend/src/composables/useAttachmentPreviews.js`, `frontend/src/desktop.js`, `frontend/src/knowledgeBaseCatalog.js` |
-| Electron preload | Exposes narrow workspace, Human Gate, Cloud control, installer/Skill, attachment, Provider, and knowledge-source methods through `window.meldworkDesktop` only to a local `file:` document; Channel ingress has no renderer API | `desktop/src/preload.cjs` |
-| Electron main | Trust boundary for IPC authorization and shutdown gating, persistence, attachment import/preview, Connector registration, Skill and knowledge-source validation, local/Cloud execution, Channel receivers, navigation, sanitized run events, operating-system notifications, Provider storage, and installer control | `desktop/src/main.cjs` |
+| Electron preload | Exposes narrow workspace, Human Gate, Cloud control, installer/Skill, attachment, and knowledge-source methods through `window.meldworkDesktop` only to a local `file:` document; Channel ingress has no renderer API | `desktop/src/preload.cjs` |
+| Electron main | Trust boundary for IPC authorization and shutdown gating, persistence, attachment import/preview, Connector registration, Skill and knowledge-source validation, local/Cloud execution, Channel receivers, navigation, sanitized run events, operating-system notifications, and installer control | `desktop/src/main.cjs` |
 | Local workspace | Persists Meldwork conversations, final replies, compact evidence capsules, safe message metadata, task-scoped group Sessions, conversation-scoped direct Sessions, legacy Session migration, Agent-specific transcript deltas, Agent preferences/runtime evidence, and active run coordination | `desktop/src/workspace/local-workspace.cjs` |
-| V4 orchestration | Validates frozen snapshots, phase-aware slots, typed collaboration receipts, negotiated responsibility graphs, work receipts, synthesis recovery, verification, and idempotent commits | `desktop/src/collaboration/orchestration-v4-records.cjs`, `desktop/src/workspace/local-workspace-message-submission.cjs`, `desktop/src/workspace/local-workspace-auto-runner.cjs` |
+| V4 orchestration | Validates frozen snapshots, phase-aware slots, typed collaboration receipts, negotiated responsibility graphs, work receipts, synthesis recovery, verification, and idempotent commits | `desktop/src/collaboration/orchestration-v4-records.cjs`, `desktop/src/workspace/local-workspace-message-submission.cjs`, `desktop/src/workspace/local-workspace-auto-runner.cjs` and its `local-workspace-auto-runner-*.cjs` parts (see Code Map) |
 | Run Ledger | Atomically checkpoints bounded, sanitized per-run and per-Agent state for stop/crash recovery, marks nonterminal runs and Agent attempts interrupted after restart, supports idempotent reconciliation of missing terminal Agent messages, and purges records with their conversation | `desktop/src/runs/run-ledger.cjs` |
 | Context Pack and delivery stores | Capture immutable, content-addressed Task inputs, attempt-specific approved prompts, exact outbound fingerprints, source hashes, target scope, and native Session provenance | `desktop/src/collaboration/context-pack-store.cjs`, `desktop/src/workspace/local-workspace-context-packs.cjs` |
 | Outcome store | Persists typed Artifacts, Evidence, Findings, and Adoption records separately from diagnostic reasoning and tool summaries | `desktop/src/collaboration/outcome-store.cjs`, `desktop/src/collaboration/outcome-records.cjs` |
 | Human Gate store | Persists permission, budget, and decision requests and their terminal user/system decisions without exposing raw request payloads to the renderer | `desktop/src/gates/human-gate-store.cjs`, `desktop/src/gates/human-gate-coordinator.cjs` |
-| Scheduling and failure policy | Enforces run budgets, retry policy, per-Agent controls, and global/workspace scheduling | `desktop/src/runs/run-scheduler.cjs`, `desktop/src/runs/failure-policy.cjs` |
+| Scheduling and failure policy | Enforces run budgets, per-Agent controls, and one writer at a time within a conversation (conversations never wait for each other); classifies failures (sign-in, rate limit, network, provider, context overflow) and decides retries with `Retry-After`; stops a silent Agent through the invocation watchdog | `desktop/src/runs/run-scheduler.cjs`, `desktop/src/runs/failure-policy.cjs`, `desktop/src/runs/failure-signals.cjs`, `desktop/src/runs/invocation-watchdog.cjs` |
 | Agent Connector registry/runtime | Validates approved manifests and instances, separates CredentialRefs, records Connector/upstream versions, and reduces versioned idempotent Run Events | `desktop/src/agents/connectors/agent-connector-manifest.cjs`, `desktop/src/agents/connectors/agent-connector-registry.cjs`, `desktop/src/agents/connectors/agent-connector-runtime.cjs` |
 | Attachment store | Imports validated local images, media, documents, code/configuration files, Office files, and archives into private app-owned storage, verifies metadata/checksums on read, and removes unreferenced entries | `desktop/src/attachments/attachment-store.cjs`, `desktop/src/attachments/image-dimensions.cjs` |
 | Local Skill catalog | Scans bounded known roots per Agent and returns sanitized Skill coordinates without exposing paths or contents | `desktop/src/skills/local-skill-catalog.cjs` |
@@ -30,9 +30,53 @@ The supported product surface is the packaged Electron application. The frontend
 | Knowledge-source store | Persists only the user-selected Obsidian Vault path; Feishu and DingTalk credentials remain owned by their local CLIs | `desktop/src/knowledge/knowledge-base-store.cjs` |
 | CLI adapters | Detect and invoke supported local Agent executables with per-Agent arguments and constrained child environments | `desktop/src/agents/cli/cli-adapters.cjs` |
 | Agent installer | Detects Agents and runs allowlisted npm packages or official installer scripts after user confirmation | `desktop/src/agents/installer/agent-installer.cjs` |
-| Provider store | Validates Provider metadata and encrypts the API key with Electron `safeStorage` | `desktop/src/providers/provider-store.cjs` |
 | Managed OpenClaw runtime | Creates isolated local configuration and a restricted tool policy for OpenClaw | `desktop/src/agents/cli/openclaw-runtime.cjs` |
 
+## Code Map
+
+The workspace is composed of small modules around a thin coordinator. A module either owns one piece of state or is a pure function of its inputs; each has a test file with the same name under `desktop/test/`.
+
+**One Agent invocation** (`desktop/src/workspace/local-workspace-agent-invocation.cjs`) is the run lifecycle: lease, run, race against stop and timeout, record the result. It delegates to:
+
+| Module | Owns |
+| --- | --- |
+| `invocation-session.cjs` | The native Agent session of one invocation: resume, adopt a new id, discard and rebuild, record the turn |
+| `invocation-outputs.cjs` | Workspace snapshot before a writable turn, media import and artifact descriptors after it |
+| `invocation-prompt.cjs` | The prompt one Agent receives (frozen, direct chat, group) |
+| `invocation-budget.cjs` | Usage recorded against the run budget |
+| `invocation-failure.cjs` | Why an invocation ended (stopped, timed out, interrupted, cancelled, failed) and the hidden error metadata |
+| `desktop/src/runs/invocation-watchdog.cjs` | Startup, inactivity and stall windows; liveness versus substantive progress |
+| `desktop/src/agents/agent-permission-policy.cjs` | Automatic answers to permission requests; a read-only turn only approves side-effect-free tools |
+| `desktop/src/collaboration/collaboration-receipt.cjs` | Parsing and normalising the control receipt at the end of a turn (lenient for natural discussion, strict for phased V4) |
+| `desktop/src/collaboration/answer-marker-filter.cjs` | Hides control markers while an answer streams |
+
+**Prompts** live in `desktop/src/workspace/prompts/`: `natural-discussion-prompt.cjs` (named sections: brief, writer, owner, routing, receipt), `manual-reply-prompt.cjs` (concurrent replies in a group), and `prompt-contracts.cjs` (the language and deliverable contracts shared by all of them). Direct chats are built by `directPromptFor` in `local-workspace-context.cjs`.
+
+**The automatic run** is `LocalWorkspaceAutoRunner`; its methods are grouped into files by what they do and assembled with `Object.assign` onto the class prototype:
+
+| File | Responsibility |
+| --- | --- |
+| `local-workspace-auto-runner.cjs` | Constructor, `start` and `resume`, run-level helpers |
+| `…-natural-discussion.cjs`, `…-natural-decisions.cjs` | The production path: transcript, prompts, the sequential and agent-led loops, and what happens after each round (owner takeover, ignored routes, peer blockers, missing decisions) |
+| `…-v4-discussion.cjs`, `…-v4-phase.cjs`, `…-v4-results.cjs` | The phased V4 run, one phase, and turning results into receipts and commits |
+| `…-v4-context.cjs`, `…-v4-delivery.cjs`, `…-v4-synthesis.cjs`, `…-v4-outcomes.cjs` | What a run freezes and shares, how context reaches each Agent, synthesis recovery, and the receipts, review issues, convergence and acceptance checks |
+| `…-recovery.cjs` | Retries, pausing a failed Agent for the run, unauthorized-session recovery |
+| `…-collaboration.cjs` | Collaboration records written as a run progresses |
+| `…-task-graph.cjs`, `…-rounds.cjs` | V3 task-graph and V2 round runs, kept only to resume older runs |
+
+Splitting by file did not change any method; the class still shares one `this`. New behaviour should go into a small module with its own tests and be called from the group that owns the topic, not appended to a large method.
+
+## Reliability Rules
+
+- Conversations are independent. A running chat, direct or group, never makes an Agent in another conversation wait, even when both use the same folder and the same Agent. The writer lease is scoped to the conversation; inside one conversation writers still take turns.
+- A failure pauses an Agent for the current run only. Group membership is never rewritten because an Agent failed; a group whose members were all removed by an older build is rebuilt from the Agents that replied in it.
+- Only a definitive sign-in failure marks credentials missing. A refusal that could be a proxy or a tool's own output is retried once and reported as a provider failure.
+- An Agent that was available stays available through an inconclusive readiness probe for ten minutes. A definitive answer applies immediately.
+- A stopped or failed turn keeps a healthy long-lived ACP process and its native session. The process is replaced only when it ignores cancellation or the protocol breaks.
+- Unknown ACP traffic (banners, new update kinds, unknown notifications) is ignored; unknown requests are answered with "method not found". Malformed JSON, invalid responses and invalid permission requests still fail closed.
+- Messages in a group topic carry a dispatch number (`turnSeq`) assigned before the Agent starts, so live and saved replies share one order. Failure and stop notices take the turn's number.
+- A discussion does not end on a malformed receipt, an absent recipient, a peer reporting a blocker or an owner who forgot the decision: those are discarded, handed to the owner, or asked for once.
+
 ## Concurrent Collaboration V4
 
 - A group with multiple selected Agents in `manual` mode runs as one concurrent batch. Harness freezes every selected Agent's task snapshot and delivery plan before any Scheduler lease, preserves all selected members, and publishes the batch in stable member order after the barrier. Direct conversations, one-Agent targets, and regeneration keep their existing execution paths.
@@ -45,11 +89,11 @@ Automatic participant selection from a larger roster, production remote Agents,
 
 ## Trust Boundaries
 
-1. The renderer is not trusted with Node.js, Agent executable paths, native Agent session IDs or storage paths, attachment paths, Skill paths/contents, raw Provider credentials after saving, arbitrary filesystem reads, or arbitrary process execution. User-selected conversation workdirs and the user-selected Obsidian Vault path are intentional UI-visible paths; pasted image bytes use one exact import payload.
+1. The renderer is not trusted with Node.js, Agent executable paths, native Agent session IDs or storage paths, attachment paths, Skill paths/contents, raw CLI credentials, arbitrary filesystem reads, or arbitrary process execution. User-selected conversation workdirs and the user-selected Obsidian Vault path are intentional UI-visible paths; pasted image bytes use one exact import payload.
 2. Preload exposes only named IPC methods. Main validates that every request comes from the current main frame and exact bundled frontend file.
 3. Main owns local files, OS dialogs, attachment validation/storage, Skill and knowledge-source discovery/revalidation, Connector approval/registration, Agent discovery, child processes, Cloud observation, Channel authentication, installer downloads, and credential encryption.
 4. Agent CLIs are external processes. They receive a selected working directory, a constrained environment, permission flags, main-resolved image paths, target-scoped immutable Skill snapshots, and validated knowledge snapshots or read-only hints after capability checks.
-5. Provider, Agent Connector, Cloud Connector, Channel Connector, and installer destinations are external trust boundaries. Provider URLs are user supplied and validated; manifests and Connector instances fail closed unless explicitly approved; Channel signatures and Workspace scope are verified before Task delivery.
+5. Agent Connector, Cloud Connector, Channel Connector, and installer destinations are external trust boundaries. Manifests and Connector instances fail closed unless explicitly approved; Channel signatures and Workspace scope are verified before Task delivery.
 6. Harness events expose only allowlisted status, conclusion deltas, reasoning summaries, plans, tool lifecycle summaries, bounded result metadata, and source message IDs. Raw chain-of-thought, commands, executable paths, native session references, credentials, unrestricted stdout/stderr, and arbitrary tool payloads remain outside the renderer and shared Agent context.
 
 ## Local Data
@@ -61,7 +105,6 @@ Automatic participant selection from a larger roster, production remote Agents,
 | Immutable Context Packs, delivery records, content blobs, typed outcomes, and Human Gates | Electron user data under `meldwork-private/` | Content-addressed or atomically written main-only records; renderer snapshots expose only bounded IDs, provenance summaries, budgets, and pending Gate choices |
 | Cloud operations and Channel Inbox | Electron user data under `meldwork-private/` | Private `0600` stores keep opaque remote references, idempotency state, signed-event-derived Task data, delivery cursors, Task mappings, and outbound audit records; raw signatures, headers, credentials, and complete IM history are not exposed to the renderer |
 | Imported attachments | Electron user data under `attachments/` | App-owned copy; `0700` directories, `0600` files, type/MIME/size validation, image-dimension checks where applicable, metadata checksum, ID-only renderer access |
-| Provider metadata and encrypted key | Electron user data: `meldwork-provider.json` | API key encrypted with `safeStorage`; file written atomically with mode `0600` |
 | Knowledge-source preference | Electron user data: `meldwork-knowledge-base.json` | Stores the selected absolute Obsidian Vault path only; written atomically as a private file |
 | Managed OpenClaw state/config | Electron user data under `openclaw-managed/` | Per-scope directories, mode `0700`; config mode `0600`; API key passed by child environment |
 | Theme, locale, and onboarding completion | Renderer `localStorage` | Non-secret UI preferences |
@@ -74,7 +117,6 @@ There is no Meldwork-hosted remote data store or application account namespace.
 
 - Agent CLIs may call their own configured model providers.
 - Selected attachments and Skill-influenced prompts may be sent by an Agent CLI to its configured model provider after local validation and capability checks.
-- Compatible Agents may receive the user-configured Provider URL/model/key from main.
 - Feishu and DingTalk readiness probes may invoke their configured local CLIs; a selected Agent may then use those CLI connections with read-only instructions.
 - An explicit Obsidian selection is searched and fetched through the main-owned read-only filesystem Connector, then captured as an immutable snapshot or explicit live reference before Agent invocation.
 - Approved Agent Connectors may contact only their declared outbound destinations and receive main-owned CredentialRefs rather than renderer-provided secrets.
@@ -86,6 +128,7 @@ There is no Meldwork-hosted remote data store or application account namespace.
 
 ## Known Risks And Assumptions
 
+- Writable conversations that share a folder are not coordinated. Two Agents in different conversations can edit the same files at the same time, as two terminals in one directory can. Output capture compares `.meldwork-output/` before and after a run, so when two conversations produce deliverables in the same folder at the same time, one may attach or record a file the other wrote.
 - Installer recipes use mutable upstream packages or scripts and currently do not verify a pinned digest or signature (`desktop/src/agents/installer/agent-installer.cjs`).
 - Conversation history is local but not application-encrypted (`desktop/src/workspace/local-workspace.cjs`).
 - Sanitized Run Ledger checkpoints are local and private but not application-encrypted. The store assumes one Electron main-process writer and recovers only the last persisted bounded checkpoint, not every in-memory event (`desktop/src/runs/run-ledger.cjs`).
@@ -95,7 +138,7 @@ There is no Meldwork-hosted remote data store or application account namespace.
 - Obsidian selections use a main-owned read-only Connector with bounded retrieval, immutable snapshots, and citation hash verification. Generic Feishu and DingTalk access remains CLI-owned and prompt-mediated; their authentication and document permissions are not controlled by Meldwork. Notion, Confluence, Google Drive, and SharePoint are reference-only planned sources and are not active Connectors.
 - Native image support is currently limited to Codex, Hermes, and OpenCode, with different per-run limits; mixed groups fail before execution when they cannot receive equal image context (`desktop/src/agents/cli/cli-adapters.cjs`).
 - Read-only enforcement is owned by the Obsidian Connector for captured selections. Agent workspace and CLI-mediated remote access still depends partly on each upstream CLI honoring its permission flags; Meldwork adds per-CLI hardening where available (`desktop/src/agents/cli/cli-adapters.cjs`).
-- Provider requests and native Agent authentication are external dependencies; Meldwork cannot guarantee their availability or data handling.
+- Native Agent authentication and model requests are external dependencies; Meldwork cannot guarantee their availability or data handling.
 - Agent/Cloud/Channel Connector contracts validate provenance, scope, and idempotency, but the external implementation must honor its CredentialRef, outbound destination, cancellation, and signature-verification obligations. The production build registers no Cloud or Channel Connector by default.
 - macOS packaging uses ad-hoc signing for local validation, not a notarized public release (`desktop/scripts/after-pack.cjs`).
 
@@ -103,14 +146,11 @@ There is no Meldwork-hosted remote data store or application account namespace.
 
 - No email sending, so there is no `emails.md`.
 - No cron or scheduled background work, so there is no `cron.md`.
-- No supported public or indexable routes, so there is no `seo.md`.
+- No supported public or indexable routes in the app itself, so there is no `seo.md`. The landing page at https://ryder-mhumble.github.io/Meldwork/ serves as the public discoverability surface for search engines and AI assistants.
 
 ## Related Documents
 
-- [Flows](docs/flows.md)
-- [Permissions](docs/permissions.md)
-- [Variables And Secrets](docs/variables.md)
-- [Test Coverage](docs/tests.md)
-- [Automation](docs/automation.md)
-- [Public MVP Release Checklist](docs/public-mvp-release.md)
-- [macOS Signing And Notarization](docs/macos-signing.md)
+- [Contributing](CONTRIBUTING.md)
+- [Security Policy](SECURITY.md)
+- [Repository Guidelines For Agents](AGENTS.md)
+- [Changelog](CHANGELOG.md)

desktop/package-lock.json

@@ -1,13 +1,13 @@
 {
   "name": "meldwork-desktop",
-  "version": "0.1.3",
+  "version": "0.1.6",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "meldwork-desktop",
-      "version": "0.1.3",
-      "license": "SEE LICENSE IN LICENSE",
+      "version": "0.1.6",
+      "license": "Apache-2.0",
       "dependencies": {
         "@agentclientprotocol/sdk": "0.23.0",
         "@streamparser/json": "0.0.22",

desktop/package.json

@@ -1,10 +1,10 @@
 {
   "name": "meldwork-desktop",
-  "version": "0.1.3",
+  "version": "0.1.6",
   "private": true,
   "description": "A local-first desktop workspace for multi-Agent collaboration",
   "author": "Ryder-MHumble",
-  "license": "SEE LICENSE IN LICENSE",
+  "license": "Apache-2.0",
   "engines": {
     "node": ">=22.12.0"
   },
@@ -38,6 +38,7 @@
     "productName": "Meldwork",
     "artifactName": "Meldwork-${version}-${arch}.${ext}",
     "afterPack": "scripts/after-pack.cjs",
+    "afterAllArtifactBuild": "scripts/after-all-artifact-build.cjs",
     "files": [
       "src/**/*",
       "samples/**/*",
@@ -59,6 +60,10 @@
       {
         "from": "../COMMERCIAL_USE.md",
         "to": "COMMERCIAL_USE.md"
+      },
+      {
+        "from": "build/dmg-background-ai-v4.tiff",
+        "to": "dmg-background.tiff"
       }
     ],
     "mac": {
@@ -70,21 +75,23 @@
       ]
     },
     "dmg": {
-      "background": "build/dmg-background-ai-v2.png",
+      "background": "build/dmg-background-ai-v4.png",
+      "icon": null,
+      "writeUpdateInfo": false,
       "window": {
-        "width": 540,
-        "height": 380
+        "width": 640,
+        "height": 420
       },
-      "iconSize": 128,
+      "iconSize": 112,
       "contents": [
         {
-          "x": 145,
-          "y": 210,
+          "x": 110,
+          "y": 205,
           "type": "file"
         },
         {
-          "x": 395,
-          "y": 210,
+          "x": 520,
+          "y": 205,
           "type": "link",
           "path": "/Applications"
         }

desktop/samples/local-echo.connector.json

@@ -1 +1 @@
-{"manifest":{"connectorId":"external.local-echo-sample","connectorVersion":"1.0.0","credentials":{"mode":"none","slots":[]},"description":"Local SDK sample that returns text without delegating to a built-in Agent.","domains":["general"],"eventProtocolVersion":1,"eventTypes":["Completed","Failed","Cancelled"],"inputTypes":["text"],"invocation":{"recipeId":"sdk.local-echo.v1"},"kind":"agent","label":"Local Echo Connector Sample","license":"LicenseRef-Meldwork-NC","manifestId":"connector-manifest-6ce79bf9cbe282e9f537a1b45d742cc507e21ade5ea2859dff9e007bd4f850db","outboundDestinations":[],"permissionModes":["read-only"],"recordType":"agent-connector-manifest","schemaVersion":1,"session":{"cancel":true,"checkpoint":false,"resume":true,"supported":true},"transport":{"protocol":"json","type":"cli"},"upstream":{"id":"meldwork-sdk","maxVersion":"1.0.0","minVersion":"1.0.0"},"usage":{"costMicros":false,"elapsedMs":false,"inputTokens":false,"outboundBytes":false,"outputTokens":false,"toolCalls":false}},"packageId":"connector-package-f4908cee8fa7acf5d17a403b39bd58cb8d4a1dfec3fbfb43c62c5727ee793cd4","provider":{"config":{},"id":"sdk.local-echo.v1"},"publisher":{"id":"meldwork","name":"Meldwork"},"recordType":"agent-connector-package","schemaVersion":1}
+{"manifest":{"connectorId":"external.local-echo-sample","connectorVersion":"1.0.0","credentials":{"mode":"none","slots":[]},"description":"Local SDK sample that returns text without delegating to a built-in Agent.","domains":["general"],"eventProtocolVersion":1,"eventTypes":["Completed","Failed","Cancelled"],"inputTypes":["text"],"invocation":{"recipeId":"sdk.local-echo.v1"},"kind":"agent","label":"Local Echo Connector Sample","license":"Apache-2.0","manifestId":"connector-manifest-6ce79bf9cbe282e9f537a1b45d742cc507e21ade5ea2859dff9e007bd4f850db","outboundDestinations":[],"permissionModes":["read-only"],"recordType":"agent-connector-manifest","schemaVersion":1,"session":{"cancel":true,"checkpoint":false,"resume":true,"supported":true},"transport":{"protocol":"json","type":"cli"},"upstream":{"id":"meldwork-sdk","maxVersion":"1.0.0","minVersion":"1.0.0"},"usage":{"costMicros":false,"elapsedMs":false,"inputTokens":false,"outboundBytes":false,"outputTokens":false,"toolCalls":false}},"packageId":"connector-package-f4908cee8fa7acf5d17a403b39bd58cb8d4a1dfec3fbfb43c62c5727ee793cd4","provider":{"config":{},"id":"sdk.local-echo.v1"},"publisher":{"id":"meldwork","name":"Meldwork"},"recordType":"agent-connector-package","schemaVersion":1}

desktop/scripts/after-all-artifact-build.cjs

@@ -0,0 +1,126 @@
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+const { execFileSync } = require('node:child_process')
+
+const ROOT_HIDDEN_RESOURCES = ['.background.tiff', '.VolumeIcon.icns']
+
+function selectDmgArtifacts(artifactPaths) {
+  return artifactPaths.filter(artifactPath => artifactPath.endsWith('.dmg'))
+}
+
+function findBundledPython(cacheRoot = process.env.ELECTRON_BUILDER_CACHE) {
+  const root = cacheRoot || path.join(os.homedir(), 'Library', 'Caches', 'electron-builder')
+  const pending = [root]
+
+  while (pending.length > 0) {
+    const current = pending.pop()
+    let entries
+    try {
+      entries = fs.readdirSync(current, { withFileTypes: true })
+    } catch {
+      continue
+    }
+
+    for (const entry of entries) {
+      const candidate = path.join(current, entry.name)
+      if (entry.isDirectory()) {
+        pending.push(candidate)
+      } else if (
+        entry.isFile()
+        && /^python3(?:\.\d+)?$/.test(entry.name)
+        && path.basename(current) === 'bin'
+        && candidate.includes(`${path.sep}dmg-builder@`)
+      ) {
+        return candidate
+      }
+    }
+  }
+
+  throw new Error(`Unable to locate dmg-builder's bundled Python under ${root}`)
+}
+
+function parseAttachedVolume(plist, workDir, run = execFileSync) {
+  const plistPath = path.join(workDir, 'attach.plist')
+  fs.writeFileSync(plistPath, plist)
+  const parsed = JSON.parse(run('plutil', ['-convert', 'json', '-o', '-', plistPath], {
+    encoding: 'utf8',
+  }))
+  const mounted = parsed['system-entities'].find(entity => entity['mount-point'])
+  if (!mounted) {
+    throw new Error('hdiutil did not return a mounted DMG volume')
+  }
+  return { device: mounted['dev-entry'], mountPath: mounted['mount-point'] }
+}
+
+function repointDmgBackground(dmgPath, options = {}) {
+  const run = options.execFile || execFileSync
+  const pythonPath = options.pythonPath || findBundledPython(options.cacheRoot)
+  const workDir = fs.mkdtempSync(path.join(os.tmpdir(), 'meldwork-dmg-'))
+  const readWriteDmg = path.join(workDir, 'writable.dmg')
+  const repackedDmg = `${dmgPath}.repacked-${process.pid}.dmg`
+  let device
+
+  try {
+    run('hdiutil', ['convert', dmgPath, '-format', 'UDRW', '-o', readWriteDmg], {
+      stdio: 'pipe',
+    })
+    const attached = parseAttachedVolume(run('hdiutil', [
+      'attach', readWriteDmg, '-nobrowse', '-readwrite', '-plist',
+    ], { encoding: 'utf8' }), workDir, run)
+    device = attached.device
+
+    const backgroundPath = path.join(
+      attached.mountPath,
+      'Meldwork.app',
+      'Contents',
+      'Resources',
+      'dmg-background.tiff',
+    )
+    if (!fs.existsSync(backgroundPath)) {
+      throw new Error(`Packaged DMG background is missing: ${backgroundPath}`)
+    }
+
+    run(pythonPath, [
+      path.join(__dirname, 'repoint-dmg-background.py'),
+      path.join(attached.mountPath, '.DS_Store'),
+      backgroundPath,
+    ], { stdio: 'pipe' })
+
+    for (const resource of ROOT_HIDDEN_RESOURCES) {
+      fs.rmSync(path.join(attached.mountPath, resource), { force: true })
+    }
+    run('/usr/bin/SetFile', ['-a', 'c', attached.mountPath], { stdio: 'pipe' })
+    run('/bin/sync', [], { stdio: 'pipe' })
+    run('hdiutil', ['detach', device], { stdio: 'pipe' })
+    device = null
+
+    run('hdiutil', [
+      'convert', readWriteDmg, '-format', 'UDZO', '-imagekey', 'zlib-level=9', '-o', repackedDmg,
+    ], { stdio: 'pipe' })
+    fs.renameSync(repackedDmg, dmgPath)
+  } finally {
+    if (device) {
+      try {
+        run('hdiutil', ['detach', '-force', device], { stdio: 'pipe' })
+      } catch {}
+    }
+    fs.rmSync(repackedDmg, { force: true })
+    fs.rmSync(workDir, { recursive: true, force: true })
+  }
+}
+
+async function afterAllArtifactBuild(buildResult) {
+  for (const dmgPath of selectDmgArtifacts(buildResult.artifactPaths)) {
+    repointDmgBackground(dmgPath)
+  }
+  return []
+}
+
+afterAllArtifactBuild.ROOT_HIDDEN_RESOURCES = ROOT_HIDDEN_RESOURCES
+afterAllArtifactBuild.selectDmgArtifacts = selectDmgArtifacts
+afterAllArtifactBuild.findBundledPython = findBundledPython
+afterAllArtifactBuild.parseAttachedVolume = parseAttachedVolume
+afterAllArtifactBuild.repointDmgBackground = repointDmgBackground
+
+module.exports = afterAllArtifactBuild

desktop/scripts/render-dmg-background.swift

@@ -0,0 +1,72 @@
+#!/usr/bin/env swift
+import AppKit
+import Foundation
+
+let arguments = CommandLine.arguments
+guard arguments.count == 3 else {
+    fputs("Usage: render-dmg-background.swift <source.png> <output.png>\n", stderr)
+    exit(1)
+}
+
+let sourceURL = URL(fileURLWithPath: arguments[1])
+let outputURL = URL(fileURLWithPath: arguments[2])
+guard let source = NSImage(contentsOf: sourceURL) else {
+    fputs("Unable to read source image\n", stderr)
+    exit(1)
+}
+guard let bitmap = NSBitmapImageRep(
+    bitmapDataPlanes: nil,
+    pixelsWide: 1280,
+    pixelsHigh: 840,
+    bitsPerSample: 8,
+    samplesPerPixel: 4,
+    hasAlpha: true,
+    isPlanar: false,
+    colorSpaceName: .deviceRGB,
+    bytesPerRow: 0,
+    bitsPerPixel: 0
+) else {
+    fputs("Unable to create output bitmap\n", stderr)
+    exit(1)
+}
+
+NSGraphicsContext.saveGraphicsState()
+NSGraphicsContext.current = NSGraphicsContext(bitmapImageRep: bitmap)
+source.draw(
+    in: NSRect(x: 0, y: 0, width: 1280, height: 840),
+    from: NSRect(x: 0, y: 4, width: 1280, height: 840),
+    operation: .copy,
+    fraction: 1
+)
+
+let centered = NSMutableParagraphStyle()
+centered.alignment = .center
+
+let title = "INSTALL MELDWORK" as NSString
+title.draw(
+    in: NSRect(x: 0, y: 712, width: 1280, height: 52),
+    withAttributes: [
+        .font: NSFont(name: "DIN Alternate Bold", size: 38) ?? NSFont.boldSystemFont(ofSize: 38),
+        .foregroundColor: NSColor(white: 0.96, alpha: 0.94),
+        .kern: 4.5,
+        .paragraphStyle: centered,
+    ]
+)
+
+let explanation = "DRAG MELDWORK INTO APPLICATIONS  /  YOUR WORKSPACE STAYS LOCAL" as NSString
+explanation.draw(
+    in: NSRect(x: 0, y: 675, width: 1280, height: 30),
+    withAttributes: [
+        .font: NSFont(name: "Menlo-Regular", size: 14) ?? NSFont.monospacedSystemFont(ofSize: 14, weight: .regular),
+        .foregroundColor: NSColor(white: 0.82, alpha: 0.74),
+        .kern: 1.5,
+        .paragraphStyle: centered,
+    ]
+)
+NSGraphicsContext.restoreGraphicsState()
+
+guard let png = bitmap.representation(using: .png, properties: [:]) else {
+    fputs("Unable to encode PNG\n", stderr)
+    exit(1)
+}
+try png.write(to: outputURL)

desktop/scripts/repoint-dmg-background.py

@@ -0,0 +1,17 @@
+#!/usr/bin/env python3
+import sys
+
+from ds_store import DSStore
+from mac_alias import Alias
+
+
+def main():
+    ds_store_path, background_path = sys.argv[1:3]
+    with DSStore.open(ds_store_path, "r+") as store:
+        icon_view = store["."]["icvp"]
+        icon_view["backgroundImageAlias"] = Alias.for_file(background_path).to_bytes()
+        store["."]["icvp"] = icon_view
+
+
+if __name__ == "__main__":
+    main()

desktop/scripts/test-discovery.cjs

@@ -5,7 +5,7 @@ const TEST_SUITES = Object.freeze({
   agents: ['agents'],
   attachments: ['attachments'],
   collaboration: ['collaboration'],
-  core: ['channels', 'knowledge', 'media', 'providers', 'skills'],
+  core: ['channels', 'knowledge', 'media', 'skills'],
   runs: ['runs'],
   security: ['security', 'shell'],
   workspace: ['workspace', 'gates'],

desktop/src/agents/agent-compatibility.cjs

@@ -57,7 +57,9 @@ const AGENT_COMPATIBILITY = Object.freeze({
     ]),
   ]),
   pi: exactProfile('0.84.2', [
-    probe('pi-cli', ['--help'], []),
+    probe('pi-cli', ['--help'], [
+      '--mode', '--print', '--session', '--tools', '--approve', '--no-approve', 'json',
+    ]),
   ]),
   kimi: profile('0.19.2', '0.32.0', [
     probe('kimi-stream', ['--help'], [
@@ -99,11 +101,13 @@ const AGENT_COMPATIBILITY = Object.freeze({
   ]),
 })
 
-function probe(id, args, requiredText) {
+function probe(id, args, requiredText, validateOutput = null, timeout = null) {
   return Object.freeze({
     id,
     args: Object.freeze([...args]),
     requiredText: Object.freeze([...requiredText]),
+    validateOutput,
+    timeout,
   })
 }
 
@@ -261,6 +265,7 @@ module.exports = {
   AGENT_COMPATIBILITY,
   assessAgentVersion,
   capabilityProbes,
+  compareReleaseVersions,
   extractAgentVersion,
   identifyAgentVersion,
 }

desktop/src/agents/agent-output-importer.cjs

@@ -426,7 +426,7 @@ function importAgentOutputs(input, attachmentStore) {
     && !Array.isArray(input.baseline.files)
     ? input.baseline.files
     : {}
-  const candidates = outputFiles(safeOutputDirectory(workdirRealPath))
+  const controlledCandidates = outputFiles(safeOutputDirectory(workdirRealPath))
     .filter(file => !sameFileState(baselineFiles[file.name], file))
     .filter(file => Math.max(file.mtimeMs, file.ctimeMs) >= startedAt - 2000)
     .sort((left, right) => (
@@ -434,6 +434,30 @@ function importAgentOutputs(input, attachmentStore) {
       || left.name.localeCompare(right.name)
     ))
 
+  const reportedCandidates = []
+  const reportedPaths = Array.isArray(input.reportedPaths) ? input.reportedPaths : []
+  for (const reportedPath of reportedPaths.slice(0, MAX_SCANNED_ENTRIES)) {
+    if (typeof reportedPath !== 'string' || !path.isAbsolute(reportedPath)) continue
+    try {
+      const realPath = fs.realpathSync(reportedPath)
+      const stat = fs.statSync(realPath)
+      const extension = path.extname(realPath).toLowerCase()
+      // A path the Agent merely mentioned (a file it read, the user's input,
+      // a staged attachment) is not an output. Only files written during this
+      // run outside the input staging folder are imported.
+      if (!stat.isFile() || !SUPPORTED_EXTENSIONS.has(extension)
+          || stat.size > 64 * 1024 * 1024 || !isInside(workdirRealPath, realPath)
+          || realPath.split(path.sep).includes('.meldwork-input')
+          || stat.mtimeMs < startedAt - 2000) continue
+      reportedCandidates.push({
+        name: path.basename(realPath), path: realPath, size: stat.size,
+        mtimeMs: stat.mtimeMs, ctimeMs: stat.ctimeMs,
+      })
+    } catch { /* reported paths are best effort */ }
+  }
+  const candidates = [...controlledCandidates, ...reportedCandidates]
+    .filter((file, index, all) => all.findIndex(candidate => candidate.path === file.path) === index)
+
   const imported = []
   if (input.signal?.aborted) return imported
   for (const file of candidates) {

desktop/src/agents/agent-permission-policy.cjs

@@ -0,0 +1,77 @@
+const { createHash } = require('node:crypto')
+const { canonicalJson } = require('../collaboration/context-pack-records.cjs')
+
+function sha256(value) {
+  return createHash('sha256').update(String(value || '')).digest('hex')
+}
+
+// Tool kinds an Agent may use without approval during a read-only turn.
+const READ_ONLY_SAFE_TOOL_KINDS = new Set(['read', 'search', 'fetch', 'think'])
+
+// Automatic answer to an Agent's permission request when the run is not
+// waiting for a person. Returns the response to send, `null` when the request
+// is refused but the Agent offered nothing to refuse with, or `undefined`
+// when no automatic answer exists and the caller must ask the user.
+//
+// A read-only turn stays read-only even when the Agent asks: only
+// side-effect-free tools are approved, while edits, deletes, moves, commands
+// and mode switches are refused.
+function automaticPermissionResponse(request, { allowWrite }) {
+  const options = Array.isArray(request?.options) ? request.options : []
+  if (!allowWrite && !READ_ONLY_SAFE_TOOL_KINDS.has(request?.toolCall?.kind)) {
+    const rejection = options.find(candidate => candidate?.kind === 'reject_once')
+      || options.find(candidate => candidate?.kind === 'reject_always')
+    return rejection?.optionId ? { status: 'rejected', optionId: rejection.optionId } : null
+  }
+  const option = options.find(candidate => candidate?.kind === 'allow_once')
+    || options.find(candidate => candidate?.kind === 'allow_always')
+    || options[0]
+  return option?.optionId ? { status: 'approved', optionId: option.optionId } : undefined
+}
+
+// A person's decision applies only to the exact request it was made for.
+function matchesResumedPermission(request, resumedGate) {
+  if (!request || !resumedGate?.request) return false
+  try {
+    if (canonicalJson(request) !== canonicalJson(resumedGate.request)) return false
+  } catch {
+    return false
+  }
+  const selected = request.options?.find(option => option.optionId === resumedGate.optionId)
+  if (!selected) return false
+  return resumedGate.status === 'approved'
+    ? selected.kind.startsWith('allow_')
+    : selected.kind.startsWith('reject_')
+}
+
+function permissionContinuationRequestId(requestHash, binding) {
+  return sha256(canonicalJson({
+    requestHash,
+    sessionRefHash: binding.sessionRefHash,
+    sessionProvenanceHash: binding.sessionProvenanceHash,
+  }))
+}
+
+// A resumed permission decision must belong to the native session the
+// request came from; anything else is refused before the Agent runs.
+function assertResumedPermissionBinding(resumed, currentBinding, sessionRef) {
+  const requestHash = sha256(canonicalJson(resumed.request))
+  const persistedBinding = {
+    sessionRefHash: resumed.sessionRefHash,
+    sessionProvenanceHash: resumed.sessionProvenanceHash,
+  }
+  if (resumed.requestId !== permissionContinuationRequestId(requestHash, persistedBinding)
+      || resumed.requestHash !== requestHash
+      || resumed.sessionProvenanceHash !== currentBinding.sessionProvenanceHash
+      || (sessionRef && resumed.sessionRefHash !== currentBinding.sessionRefHash)) {
+    throw new Error('LOCAL_RUN_PERMISSION_REQUEST_MISMATCH')
+  }
+}
+
+module.exports = {
+  READ_ONLY_SAFE_TOOL_KINDS,
+  assertResumedPermissionBinding,
+  automaticPermissionResponse,
+  matchesResumedPermission,
+  permissionContinuationRequestId,
+}

desktop/src/agents/agent-runtime-contract.cjs

@@ -1,3 +1,4 @@
+const { authFailureCertainty } = require('../runs/failure-signals.cjs')
 const AGENT_OUTCOMES = Object.freeze([
   'accepted',
   'running',
@@ -210,10 +211,34 @@ function normalizeExternalRunRef(value) {
   return EXTERNAL_RUN_REF.test(ref) ? ref : ''
 }
 
+function terminalAuthenticationDiagnostic(value) {
+  const text = String(value || '').replace(/\s+/g, ' ').trim()
+  if (!text || text.length > 2000) return ''
+  return /^HTTP\s+(?:401|403)\s*:\s*(?:unauthorized|forbidden|(?:invalid|expired|missing)\s+(?:access\s+token|token|credentials?|api[ _-]?key|authentication))(?:\s*\(request[ _-]?id:\s*[^()]{1,160}\))?\.?$/i
+    .test(text)
+    ? text
+    : ''
+}
+
+// True only for explicit sign-in or API-key statements. Ambiguous refusals
+// (HTTP 403, proxy or region blocks, "provider rejected") are classified as
+// LOCAL_AGENT_ACCESS_DENIED so they are retried instead of signing the Agent out.
+function authConfigurationFailure(detail) {
+  return authFailureCertainty(detail) === 'definitive'
+}
+
+function authFailureCode(detail) {
+  const certainty = authFailureCertainty(detail)
+  if (certainty === 'definitive') return 'LOCAL_AGENT_AUTH_REQUIRED'
+  if (certainty === 'ambiguous') return 'LOCAL_AGENT_ACCESS_DENIED'
+  return ''
+}
+
 function failureCategory(code) {
   const value = String(code || '')
   if (value === 'LOCAL_AGENT_SESSION_INVALID') return 'session'
   if (value === 'LOCAL_AGENT_AUTH_REQUIRED') return 'authentication'
+  if (value === 'LOCAL_AGENT_ACCESS_DENIED') return 'provider'
   if (value === 'LOCAL_AGENT_TIMEOUT') return 'timeout'
   if (value === 'LOCAL_AGENT_EXECUTION_STOPPED') return 'cancellation'
   if (value === 'LOCAL_AGENT_REFUSED') return 'refusal'
@@ -290,15 +315,32 @@ function requireTerminalAgentResult(result) {
     throw agentRuntimeError('LOCAL_AGENT_EXECUTION_STOPPED')
   }
   if (normalized.outcome === 'failed') {
-    const failure = normalized.failure?.code
+    let failure = normalized.failure?.code
       ? normalized.failure
       : agentRuntimeFailure('LOCAL_AGENT_PROCESS_FAILED')
+    const diagnostic = String(normalized.diagnostic || normalized.text || '')
+    // A structured failure whose diagnostic names an authentication problem is
+    // an authentication failure even when the CLI reported a generic process
+    // error, so the catalog can mark the Agent as needing a fresh login.
+    if (failure.category !== 'authentication' && authConfigurationFailure(diagnostic)) {
+      failure = agentRuntimeFailure('LOCAL_AGENT_AUTH_REQUIRED')
+    } else if (failure.code === 'LOCAL_AGENT_PROCESS_FAILED'
+        && authFailureCertainty(normalized.diagnostic) === 'ambiguous') {
+      failure = agentRuntimeFailure('LOCAL_AGENT_ACCESS_DENIED')
+    }
     throw agentRuntimeError(
       failure.code,
       normalized.diagnostic || normalized.text,
       failure,
     )
   }
+  const authenticationDiagnostic = terminalAuthenticationDiagnostic(normalized.text)
+  if (authenticationDiagnostic) {
+    throw agentRuntimeError(
+      authFailureCode(authenticationDiagnostic) || 'LOCAL_AGENT_ACCESS_DENIED',
+      authenticationDiagnostic,
+    )
+  }
   return normalized
 }
 
@@ -325,6 +367,8 @@ module.exports = {
   agentRuntimeCapabilities,
   agentRuntimeError,
   agentRuntimeFailure,
+  authConfigurationFailure,
+  authFailureCode,
   isCodeReviewAgentKind,
   isReviewOnlyAgentKind,
   normalizeAgentResult,
@@ -332,4 +376,5 @@ module.exports = {
   normalizeOutcome,
   outcomeForAcpStopReason,
   requireTerminalAgentResult,
+  terminalAuthenticationDiagnostic,
 }

desktop/src/agents/cli/cli-acp-runner.cjs

@@ -1,5 +1,7 @@
 const { spawn } = require('node:child_process')
 const { createHash } = require('node:crypto')
+const fs = require('node:fs')
+const path = require('node:path')
 const { PassThrough, Readable, Writable } = require('node:stream')
 const { AGENT_PROFILES, prepareCommand } = require('./cli-discovery.cjs')
 const {
@@ -23,8 +25,24 @@ const { createAcpOutboundPayload } = require('../../collaboration/outbound-paylo
 const { redactSecrets } = require('../../security/secret-redaction.cjs')
 
 const ACP_CANCEL_GRACE_MS = 250
-const ACP_MAX_LINE_BYTES = 1024 * 1024
+// After a stop or timeout the Agent gets this long to answer the cancelled
+// prompt. A clean cancellation keeps the long-lived process and its native
+// session; only an Agent that ignores cancel is torn down.
+const ACP_CANCEL_SETTLE_MS = 3000
+const ACP_INITIALIZE_TIMEOUT_MS = 60 * 1000
+const ACP_STDERR_TAIL_BYTES = 64 * 1024
+// Process identity must not change when only volatile shell values change.
+// Version managers such as fnm put a per-shell directory on PATH, so a fresh
+// login-shell capture used to restart the ACP process and drop its session.
+const ACP_VOLATILE_ENV_KEYS = new Set([
+  'PATH', 'Path', 'TERM', 'TERM_PROGRAM', 'TERM_PROGRAM_VERSION', 'TERM_SESSION_ID',
+  'SHLVL', 'PWD', 'OLDPWD', '_', 'TMPDIR', 'COLUMNS', 'LINES', 'SHELL',
+  'FNM_MULTISHELL_PATH', 'FNM_DIR', 'FNM_VERSION_FILE_STRATEGY', 'FNM_LOGLEVEL',
+  'FNM_NODE_DIST_MIRROR', 'FNM_ARCH', 'FNM_COREPACK_ENABLED', 'FNM_RESOLVE_ENGINES',
+])
+const ACP_MAX_LINE_BYTES = 4 * 1024 * 1024
 const ACP_MAX_INPUT_BYTES = 16 * 1024 * 1024
+const ACP_UNKNOWN_METHOD_NAME_LIMIT = 120
 const ACP_MAX_REPLY_BYTES = 10 * 1024 * 1024
 const ACP_PERSISTENT_RUNTIME_LIMIT = 16
 const ACP_PERMISSION_OPTION_LIMIT = 16
@@ -88,6 +106,10 @@ function acpTransportError(diagnostic) {
   return agentExecutionError('LOCAL_AGENT_PROCESS_FAILED', diagnostic)
 }
 
+// Returns the message to deliver, or null for traffic that is safe to ignore.
+// A newer Agent can send updates and notifications this client does not know;
+// failing the whole turn over them would break Meldwork on every Agent upgrade,
+// so only messages that decide what is executed or approved stay strict.
 function validateAcpInboundMessage(message, validators, replyState, protocolLabel) {
   if (!message || typeof message !== 'object' || Array.isArray(message)
       || message.jsonrpc !== '2.0') {
@@ -99,7 +121,8 @@ function validateAcpInboundMessage(message, validators, replyState, protocolLabe
     if (message.method === 'session/update' && !hasId) {
       const parsed = validators.zSessionNotification.safeParse(message.params)
       if (!parsed.success) {
-        throw acpTransportError(`${protocolLabel} returned invalid session update parameters.`)
+        replyState.ignoredMessages = (replyState.ignoredMessages || 0) + 1
+        return null
       }
       const update = parsed.data.update
       if (replyState.collecting !== false
@@ -119,7 +142,23 @@ function validateAcpInboundMessage(message, validators, replyState, protocolLabe
       }
       return message
     }
-    throw acpTransportError(`${protocolLabel} requested an unsupported client method.`)
+    // An unknown request still gets the connection's "method not found"
+    // answer, so the Agent is not left waiting. The SDK logs every request it
+    // rejects together with its parameters, so only the method name is
+    // forwarded. An unknown notification needs no answer and is dropped
+    // before it can reach the SDK's logging.
+    const validId = typeof message.id === 'string'
+      || (Number.isInteger(message.id) && message.id >= 0)
+    if (hasId && validId) {
+      return {
+        jsonrpc: '2.0',
+        id: message.id,
+        method: message.method.slice(0, ACP_UNKNOWN_METHOD_NAME_LIMIT),
+        params: {},
+      }
+    }
+    replyState.ignoredMessages = (replyState.ignoredMessages || 0) + 1
+    return null
   }
 
   if (!Number.isInteger(message.id) || message.id < 0) {
@@ -215,6 +254,12 @@ function boundedAcpStream(
       const parseLine = (line) => {
         const text = line.toString('utf8').trim()
         if (!text) return
+        // Some CLIs print banners or warnings on stdout next to the protocol.
+        // Anything that is not an object is not protocol traffic.
+        if (!text.startsWith('{')) {
+          replyState.ignoredMessages = (replyState.ignoredMessages || 0) + 1
+          return
+        }
         let message
         try {
           message = JSON.parse(text)
@@ -224,6 +269,7 @@ function boundedAcpStream(
         const validated = validateAcpInboundMessage(
           message, validators, replyState, protocolLabel,
         )
+        if (!validated) return
         if (message.method === 'session/request_permission') {
           replyState.pendingPermissionRequestIds.add(message.id)
         } else if (message.method === 'session/update' && message.params?.update) {
@@ -529,6 +575,32 @@ function persistenceKey(agent, options) {
   return `${agent.kind}:${value}`
 }
 
+const ACP_IMAGE_MIME_TYPES = new Map([
+  ['.png', 'image/png'], ['.jpg', 'image/jpeg'], ['.jpeg', 'image/jpeg'],
+  ['.gif', 'image/gif'], ['.webp', 'image/webp'],
+])
+const ACP_MAX_IMAGE_BYTES = 10 * 1024 * 1024
+
+// Images travel as ACP content blocks only when the Agent advertises image
+// prompts during initialize. Otherwise the turn fails before prompting so the
+// adapter can fall back to a transport that accepts image arguments.
+function acpImageBlocks(runtime, imageAttachments) {
+  const images = Array.isArray(imageAttachments) ? imageAttachments : []
+  if (!images.length) return []
+  if (runtime.agentCapabilities?.promptCapabilities?.image !== true) {
+    throw allowAcpSetupFallback(agentExecutionError('LOCAL_AGENT_IMAGE_UNSUPPORTED'))
+  }
+  return images.map((filename) => {
+    const mimeType = ACP_IMAGE_MIME_TYPES.get(path.extname(filename).toLowerCase())
+    if (!mimeType) throw agentExecutionError('LOCAL_AGENT_IMAGE_UNSUPPORTED')
+    const stat = fs.statSync(filename)
+    if (!stat.isFile() || stat.size > ACP_MAX_IMAGE_BYTES) {
+      throw agentExecutionError('LOCAL_AGENT_IMAGE_LIMIT')
+    }
+    return { type: 'image', mimeType, data: fs.readFileSync(filename).toString('base64') }
+  })
+}
+
 function withAcpRuntimeLock(key, callback) {
   const previous = acpRuntimeLocks.get(key) || Promise.resolve()
   const current = previous.catch(() => {}).then(callback)
@@ -540,14 +612,19 @@ function withAcpRuntimeLock(key, callback) {
 }
 
 function runtimeSignature(agent, workdir, spec, profile, childEnv) {
-  const environment = Object.keys(childEnv).sort().map(key => [key, childEnv[key]])
+  const environment = Object.keys(childEnv).sort()
+    .filter(key => !ACP_VOLATILE_ENV_KEYS.has(key))
+    .map(key => [key, childEnv[key]])
   return createHash('sha256').update(JSON.stringify({
     kind: agent.kind,
     executable: agent.executable,
     workdir,
     command: spec.command,
     args: spec.args,
-    acpMode: spec.acpMode || '',
+    // The session mode (read-only plan versus write) is applied on every turn
+    // with session/set_mode, so it does not identify a process. Anything that
+    // really needs a separate process (a permission policy in the environment)
+    // is already part of the environment below.
     profileId: profile.profileId,
     environment,
   })).digest('hex')
@@ -653,8 +730,14 @@ async function createAcpRuntime(input) {
   })
   child.stderr.on('data', (chunk) => {
     try { runtime.activeTurn?.noteActivity?.() } catch { /* activity is best-effort */ }
+    // Keep only recent stderr. A long-lived process used to keep the first
+    // megabyte of its whole lifetime, so a new failure was classified from
+    // stale early lines (for example an old auth warning).
+    runtime.stderr.push(chunk)
     runtime.stderrBytes += chunk.length
-    if (runtime.stderrBytes <= 1024 * 1024) runtime.stderr.push(chunk)
+    while (runtime.stderrBytes > ACP_STDERR_TAIL_BYTES && runtime.stderr.length > 1) {
+      runtime.stderrBytes -= runtime.stderr.shift().length
+    }
   })
   runtime.observedStdout = createObservedAcpInput(
     child.stdout, runtime.replyState, protocolLabel,
@@ -713,15 +796,42 @@ async function createAcpRuntime(input) {
   )
   runtime.connection = new ClientSideConnection(() => client, stream)
   acpInitializingRuntimes.add(runtime)
+  const abortSignal = input.options?.signal
+  const initializeTimeoutMs = Number.isFinite(input.options?.acpInitializeTimeoutMs)
+    && input.options.acpInitializeTimeoutMs > 0
+    ? input.options.acpInitializeTimeoutMs
+    : ACP_INITIALIZE_TIMEOUT_MS
+  let initializeTimer = null
+  let abortListener = null
   try {
-    await Promise.race([
+    // initialize used to wait forever while holding the per-conversation
+    // lock, leaving the Agent unusable in that conversation until restart.
+    const initialized = await Promise.race([
       runtime.connection.initialize({ protocolVersion: PROTOCOL_VERSION, clientCapabilities: {} }),
       runtime.failurePromise,
+      new Promise((_, reject) => {
+        initializeTimer = setTimeout(() => reject(agentExecutionError(
+          'LOCAL_AGENT_PROTOCOL_UNAVAILABLE',
+          `${protocolLabel} did not answer the ACP initialize request.`,
+        )), initializeTimeoutMs)
+      }),
+      new Promise((_, reject) => {
+        abortListener = () => reject(agentExecutionError('LOCAL_AGENT_EXECUTION_STOPPED'))
+        if (abortSignal?.aborted) abortListener()
+        else abortSignal?.addEventListener('abort', abortListener, { once: true })
+      }),
     ])
+    runtime.agentCapabilities = initialized?.agentCapabilities
+      && typeof initialized.agentCapabilities === 'object'
+      ? initialized.agentCapabilities
+      : {}
   } catch (error) {
     await closeAcpRuntime(runtime)
+    if (abortSignal?.aborted) throw agentExecutionError('LOCAL_AGENT_EXECUTION_STOPPED')
     throw allowAcpSetupFallback(acpProtocolError(error, childEnv))
   } finally {
+    clearTimeout(initializeTimer)
+    if (abortListener) abortSignal?.removeEventListener('abort', abortListener)
     acpInitializingRuntimes.delete(runtime)
   }
   return runtime
@@ -745,7 +855,7 @@ function publicSessionRef(value, childEnv) {
   return safe.includes('[redacted]') ? '' : safe
 }
 
-async function runAcpTurn(runtime, prompt, options, spec, profile, persistent) {
+async function runAcpTurn(runtime, prompt, options, spec, profile) {
   if (runtime.closed || runtime.failureError) {
     throw runtime.failureError || agentExecutionError('LOCAL_AGENT_PROCESS_FAILED')
   }
@@ -769,6 +879,8 @@ async function runAcpTurn(runtime, prompt, options, spec, profile, persistent) {
     pendingPermissionCallbacks: new Set(),
   }
   runtime.activeTurn = turn
+  runtime.stderr = []
+  runtime.stderrBytes = 0
   runtime.replyState.bytes = 0
   runtime.replyState.inputBytes = 0
   runtime.replyState.observedInputBytes = 0
@@ -785,6 +897,8 @@ async function runAcpTurn(runtime, prompt, options, spec, profile, persistent) {
     runtime.publicSessionRef = visibleSessionRef
   }
   let promptStarted = false
+  let promptRequest = null
+  let caughtError = null
   let abortRequested = false
   let timeoutRequested = false
   let cancelPromise = Promise.resolve()
@@ -812,9 +926,6 @@ async function runAcpTurn(runtime, prompt, options, spec, profile, persistent) {
 
   try {
     if (!protocolSessionRef) {
-      if (persistent && options.sessionRef) {
-        throw agentExecutionError('LOCAL_AGENT_SESSION_INVALID')
-      }
       if (spec.acpSessionStrategy !== 'new' && options.sessionRef) {
         protocolSessionRef = String(options.sessionRef)
         await operation(runtime.connection.resumeSession({
@@ -847,14 +958,17 @@ async function runAcpTurn(runtime, prompt, options, spec, profile, persistent) {
     if (visibleSessionRef && typeof options.onSessionRef === 'function') {
       await operation(options.onSessionRef(visibleSessionRef, { transport: 'acp' }))
     }
+    const imageBlocks = acpImageBlocks(runtime, spec.imageAttachments)
     runtime.replyState.collecting = true
     let promptResult
     try {
       promptStarted = true
-      promptResult = await operation(runtime.connection.prompt({
+      promptRequest = Promise.resolve(runtime.connection.prompt({
         sessionId: protocolSessionRef,
-        prompt: [{ type: 'text', text: prompt }],
+        prompt: [{ type: 'text', text: prompt }, ...imageBlocks],
       }))
+      promptRequest.catch(() => {})
+      promptResult = await operation(promptRequest)
     } finally {
       runtime.replyState.collecting = false
     }
@@ -889,9 +1003,15 @@ async function runAcpTurn(runtime, prompt, options, spec, profile, persistent) {
         ? 'timeout'
         : normalized.failure?.category === 'cancellation' ? 'stopped' : 'failed',
     })
+    if (promptStarted && options.sessionRef
+        && runtime.replyState.bytes === 0
+        && normalized.message === 'LOCAL_AGENT_REFUSED') {
+      throw agentExecutionError('LOCAL_AGENT_SESSION_INVALID')
+    }
     if (!promptStarted && normalized.message !== 'LOCAL_AGENT_SESSION_INVALID') {
       allowAcpSetupFallback(normalized)
     }
+    caughtError = normalized
     throw normalized
   } finally {
     clearTimeout(timeout)
@@ -904,7 +1024,25 @@ async function runAcpTurn(runtime, prompt, options, spec, profile, persistent) {
     turn.pendingPermissionCallbacks.clear()
     runtime.activeTurn = null
     runtime.lastUsedAt = Date.now()
-    if (abortRequested) await settleWithin(cancelPromise, ACP_CANCEL_GRACE_MS)
+    if (abortRequested) {
+      await settleWithin(cancelPromise, ACP_CANCEL_GRACE_MS)
+      // A user stop or watchdog timeout used to destroy the process and its
+      // native session. Keep both when the Agent honours the cancellation.
+      let promptSettled = false
+      // Disposable runtimes are always closed after the turn; only a
+      // persistent runtime is worth waiting for.
+      if (promptRequest && !acpDisposableRuntimes.has(runtime)) {
+        await settleWithin(
+          promptRequest.then(() => { promptSettled = true }, () => { promptSettled = true }),
+          ACP_CANCEL_SETTLE_MS,
+        )
+      }
+      if (caughtError && promptSettled && !runtime.closed && !runtime.failureError) {
+        Object.defineProperty(caughtError, 'acpRuntimeReusable', {
+          value: true, enumerable: false, configurable: true,
+        })
+      }
+    }
   }
 }
 
@@ -955,9 +1093,6 @@ async function runPersistentAcpAgent(agent, prompt, workdir, options, spec, prof
       runtime = null
       if (options.sessionRef) throw agentExecutionError('LOCAL_AGENT_SESSION_INVALID')
     }
-    if (!runtime && options.sessionRef) {
-      throw agentExecutionError('LOCAL_AGENT_SESSION_INVALID')
-    }
     if (!runtime) {
       if (shutdownGeneration !== acpShutdownGeneration) {
         throw agentExecutionError('LOCAL_AGENT_EXECUTION_STOPPED')
@@ -969,10 +1104,12 @@ async function runPersistentAcpAgent(agent, prompt, workdir, options, spec, prof
     }
     runtime.active = true
     try {
-      return await runAcpTurn(runtime, prompt, options, spec, profile, true)
+      return await runAcpTurn(runtime, prompt, options, spec, profile)
     } catch (error) {
-      acpSessionRuntimes.delete(key)
-      await closeAcpRuntime(runtime)
+      if (error?.acpRuntimeReusable !== true) {
+        acpSessionRuntimes.delete(key)
+        await closeAcpRuntime(runtime)
+      }
       throw error
     } finally {
       runtime.active = false
@@ -1003,7 +1140,7 @@ async function runDisposableAcpAgent(agent, prompt, workdir, options, spec, prof
       throw agentExecutionError('LOCAL_AGENT_EXECUTION_STOPPED')
     }
     acpDisposableRuntimes.add(runtime)
-    return await runAcpTurn(runtime, prompt, options, spec, profile, false)
+    return await runAcpTurn(runtime, prompt, options, spec, profile)
   } finally {
     if (runtime) {
       acpDisposableRuntimes.delete(runtime)

desktop/src/agents/cli/cli-adapters.cjs

@@ -15,6 +15,7 @@ const {
   parseMimoOutput,
   parseOpenCodeOutput,
   parseOpenCodeReviewOutput,
+  parsePiOutput,
   parseWorkBuddyOutput,
   readHermesFinalResponse,
   readHermesMessageWatermark,
@@ -52,6 +53,17 @@ const MAX_STDOUT_CAPTURE_BYTES = 10 * 1024 * 1024
 const MAX_STDERR_CAPTURE_BYTES = 1024 * 1024
 const OUTPUT_TRUNCATION_MARKER = Buffer.from('\n[output truncated]\n')
 
+const BROKEN_INPUT_GRACE_MS = 5000
+
+// A step that ends because the model asked for tools is followed by more
+// steps (the tools run, then the model continues); only the last step's finish
+// ends the run. Ending the process after the first one cut every tool-using
+// MiMo or OpenCode turn short.
+function stepFinishContinues(event) {
+  const reason = String(event?.part?.reason || event?.reason || '').toLowerCase()
+  return /^tool[-_ ]?(calls?|use)$/.test(reason)
+}
+
 function createBoundedOutputCapture(maxBytes) {
   const headLimit = Math.floor(maxBytes / 2)
   const tailLimit = maxBytes - headLimit
@@ -122,8 +134,8 @@ async function runAgent(agent, prompt, workdir, options = {}) {
     ? options.hermesAcpAvailable
     : agent.acpAvailable
   const spec = invocation(agent.kind, agent.executable, workdir, sessionRef, {
+    platform,
     sandbox: options.sandbox,
-    provider: options.provider,
     attachments: options.attachments,
     hermesAcpAvailable,
     invocationTransport: options.invocationTransport,
@@ -244,6 +256,7 @@ async function runAgent(agent, prompt, workdir, options = {}) {
     let stopRequested = false
     let timeoutRequested = false
     let inputError = null
+    let brokenInputTimer = null
     let timeout
     let forceKillTimeout
     let forceSettleTimeout
@@ -293,6 +306,7 @@ async function runAgent(agent, prompt, workdir, options = {}) {
       clearTimeout(forceSettleTimeout)
       clearTimeout(terminalKillTimer)
       clearTimeout(terminalForceKillTimer)
+      clearTimeout(brokenInputTimer)
       options.signal?.removeEventListener('abort', abort)
       callback()
     }
@@ -312,9 +326,10 @@ async function runAgent(agent, prompt, workdir, options = {}) {
       if (terminalObserved || !event || typeof event !== 'object') return
       const terminal = (
         (['workbuddy', 'claude', 'qwen'].includes(agent.kind) && event.type === 'result')
-        || (agent.kind === 'pi' && ['turn_end', 'agent_end'].includes(event.type))
+        || (agent.kind === 'pi' && event.type === 'agent_end')
         || (agent.kind === 'codex' && event.type === 'turn.completed')
-        || (['mimo', 'opencode'].includes(agent.kind) && event.type === 'step_finish')
+        || (['mimo', 'opencode'].includes(agent.kind) && event.type === 'step_finish'
+          && !stepFinishContinues(event))
         || (agent.kind === 'gemini' && event.type === 'result')
         || (agent.kind === 'kimi' && event.type === 'session.resume_hint')
       )
@@ -383,10 +398,23 @@ async function runAgent(agent, prompt, workdir, options = {}) {
     })
     child.stdin.on('error', (error) => {
       if (settled || inputError) return
-      inputError = agentExecutionError(
+      const sendFailure = () => agentExecutionError(
         'LOCAL_AGENT_SEND_FAILED',
         redactChildSecrets(error?.message || error, childEnv),
       )
+      // A closed pipe usually means the CLI exited before it read the whole
+      // prompt (expired sign-in, rejected flag). Its own output says why, so
+      // let the process finish and report that. Only a process that stays
+      // alive without reading its input is a send failure.
+      if (['EPIPE', 'ERR_STREAM_DESTROYED', 'ECONNRESET'].includes(error?.code)) {
+        brokenInputTimer ||= setTimeout(() => {
+          if (settled || inputError) return
+          inputError = sendFailure()
+          abort()
+        }, BROKEN_INPUT_GRACE_MS)
+        return
+      }
+      inputError = sendFailure()
       abort()
     })
     child.on('error', error => finish(() => rejectTerminal(
@@ -500,6 +528,7 @@ module.exports = {
   parseMimoOutput,
   parseOpenCodeOutput,
   parseOpenCodeReviewOutput,
+  parsePiOutput,
   parseWorkBuddyOutput,
   prepareCommand,
   readHermesFinalResponse,

desktop/src/agents/cli/cli-discovery.cjs

@@ -3,6 +3,7 @@ const fs = require('node:fs')
 const os = require('node:os')
 const path = require('node:path')
 const { promisify } = require('node:util')
+const { networkEnvironment } = require('./cli-network-environment.cjs')
 const {
   assessAgentVersion,
   capabilityProbes,
@@ -26,6 +27,8 @@ const AGENT_PROFILES = {
 }
 const ALLOWED_KINDS = Object.keys(AGENT_PROFILES)
 const DEFAULT_WINDOWS_PATHEXT = ['.COM', '.EXE', '.BAT', '.CMD']
+const CAPABILITY_CACHE_TTL_MS = 5 * 60 * 1000
+const capabilityCache = new Map()
 const SYSTEM_CHILD_ENV_KEYS = Object.freeze([
   'HOME', 'USER', 'LOGNAME', 'SHELL',
   'TMPDIR', 'TMP', 'TEMP',
@@ -38,7 +41,7 @@ const SYSTEM_CHILD_ENV_KEYS = Object.freeze([
   'APPDATA', 'LOCALAPPDATA', 'PROGRAMDATA',
   'PROGRAMFILES', 'PROGRAMFILES(X86)', 'PROGRAMW6432',
   'OS', 'PROCESSOR_ARCHITECTURE', 'PROCESSOR_IDENTIFIER', 'NUMBER_OF_PROCESSORS',
-  'SSL_CERT_FILE', 'SSL_CERT_DIR', 'NODE_EXTRA_CA_CERTS',
+  'CODEX_HOME', 'CLAUDE_CONFIG_DIR', 'PI_CODING_AGENT_DIR',
 ])
 
 function envValue(env, name) {
@@ -47,7 +50,7 @@ function envValue(env, name) {
 }
 
 function systemChildEnvironment(sourceEnv = process.env, platform = process.platform) {
-  const env = {}
+  const env = networkEnvironment(sourceEnv, platform)
   for (const name of SYSTEM_CHILD_ENV_KEYS) {
     const value = envValue(sourceEnv, name)
     if (typeof value === 'string' && value) env[name] = value
@@ -206,6 +209,23 @@ function agentCommands(kind, options = {}) {
   return commands
 }
 
+// The WorkBuddy app ships its CLI with only the lite bundle, which the
+// launcher loads solely when this variable is set. Without it every headless
+// call, `--help` included, exits with MODULE_NOT_FOUND. Installs that carry a
+// regular bundle launch exactly as before.
+function launchEnvironment(kind, executable, options = {}) {
+  if (kind !== 'workbuddy' || !executable) return {}
+  const { pathApi } = runtimeOptions(options)
+  try {
+    const dist = pathApi.join(pathApi.dirname(fs.realpathSync(executable)), '..', 'dist')
+    const shipped = name => fs.existsSync(pathApi.join(dist, name))
+    if (shipped('codebuddy-lite-wb.mjs') && !shipped('codebuddy-headless.js') && !shipped('codebuddy.js')) {
+      return { CODEBUDDY_FORCE_LITE_WB_BUNDLE: '1' }
+    }
+  } catch { /* an unreadable layout keeps the default launch */ }
+  return {}
+}
+
 function executableExtensions(options = {}) {
   const { platform, env } = runtimeOptions(options)
   if (platform !== 'win32') return ['']
@@ -330,31 +350,57 @@ async function probeAgentCapabilities(kind, executable, options = {}) {
   const runtime = runtimeOptions(options)
   const { env, execFileFn, platform } = runtime
   const prepareCommandFn = options.prepareCommandFn || prepareCommand
-  const childEnv = options.childEnv || {
-    ...systemChildEnvironment(env, platform),
-    PATH: searchPath(options),
+  const childEnv = {
+    ...(options.childEnv || {
+      ...systemChildEnvironment(env, platform),
+      PATH: searchPath(options),
+    }),
+    ...launchEnvironment(kind, executable, options),
   }
+  let identity = ''
+  try {
+    const stat = (options.statFn || fs.statSync)(executable)
+    identity = JSON.stringify([
+      kind, executable, options.version || '', stat.dev, stat.ino,
+      stat.size, stat.mtimeMs, stat.ctimeMs,
+    ])
+  } catch { /* caching requires a verifiable executable identity */ }
+  const cacheKey = `${kind}:${executable}`
+  const prior = capabilityCache.get(cacheKey)
+  const recentCompatible = identity && prior?.identity === identity
+    && Date.now() - prior.checkedAt < CAPABILITY_CACHE_TTL_MS
   const results = await Promise.all(capabilityProbes(kind).map(async (probe) => {
-    try {
-      const command = prepareCommandFn(executable, probe.args, options)
-      const result = await execFileFn(command.command, command.args, {
-        timeout: 8000,
-        windowsHide: true,
-        env: childEnv,
-      })
-      return {
-        probe,
-        output: `${result.stdout || ''}\n${result.stderr || ''}`,
-      }
-    } catch {
-      return {
-        probe,
-        error: true,
+    for (const timeout of probe.timeout ? [probe.timeout] : [8000, 16000]) {
+      try {
+        const command = prepareCommandFn(executable, probe.args, options)
+        const result = await execFileFn(command.command, command.args, {
+          timeout,
+          windowsHide: true,
+          env: childEnv,
+        })
+        return { probe, output: `${result.stdout || ''}\n${result.stderr || ''}` }
+      } catch (error) {
+        const timedOut = error?.killed === true || error?.code === 'ETIMEDOUT'
+        if (timedOut && timeout === 8000) continue
+        return { probe, error: true, timedOut }
       }
     }
   }))
+  const failed = results.find(result => !result.timedOut && (result.error
+    || result.probe.requiredText.some(value => !result.output.includes(value))
+    || (result.probe.validateOutput && !result.probe.validateOutput(result.output))))
+  if (failed) capabilityCache.delete(cacheKey)
   for (const result of results) {
     const { probe } = result
+    if (result.timedOut && !failed) {
+      if (recentCompatible) continue
+      return {
+        compatibilityState: 'unknown',
+        incompatibilityReason: 'LOCAL_AGENT_CAPABILITY_PROBE_TIMEOUT',
+        incompatibilityProbe: probe.id,
+      }
+    }
+    if (result.timedOut) continue
     if (result.error) {
       return {
         compatibilityState: 'incompatible',
@@ -372,6 +418,9 @@ async function probeAgentCapabilities(kind, executable, options = {}) {
       }
     }
   }
+  if (identity && !results.some(result => result.timedOut)) {
+    capabilityCache.set(cacheKey, { identity, checkedAt: Date.now() })
+  }
   return {
     compatibilityState: 'compatible',
     incompatibilityReason: '',
@@ -402,18 +451,17 @@ async function inspectAgentCandidate(kind, executable, options, childEnv) {
         version: identified?.line || '',
         versionIdentified: Boolean(identified),
       }
-    } catch { /* a broken shim is not a usable CLI */ }
+    } catch { /* capability probes can still establish whether this installation works */ }
     return { succeeded: false, version: '', versionIdentified: false }
   })()
 
-  const { succeeded: versionCommandSucceeded, version } = await versionTask
-  if (!versionCommandSucceeded) return null
+  const { version } = await versionTask
   const versionCompatibility = assessAgentVersion(kind, version)
   const capabilityTask = Promise.resolve().then(async () => {
     const result = await (options.probeAgentCapabilitiesFn || probeAgentCapabilities)(
       kind,
       executable,
-      { ...options, childEnv },
+      { ...options, childEnv, version },
     )
     return result && typeof result === 'object' ? result : incompatibleCapabilities
   }).catch(() => incompatibleCapabilities)
@@ -433,7 +481,7 @@ async function inspectAgentCandidate(kind, executable, options, childEnv) {
       })()
     : Promise.resolve(undefined)
   const [capabilityCompatibility, acpAvailable] = await Promise.all([capabilityTask, acpTask])
-  const compatibility = capabilityCompatibility.compatibilityState === 'incompatible'
+  const compatibility = capabilityCompatibility.compatibilityState !== 'compatible'
     ? { ...versionCompatibility, ...capabilityCompatibility }
     : {
         ...versionCompatibility,
@@ -485,6 +533,7 @@ module.exports = {
   AGENT_PROFILES,
   ALLOWED_KINDS,
   detectAgents,
+  launchEnvironment,
   prepareCommand,
   probeAgentCapabilities,
   resolveExecutable,

desktop/src/agents/cli/cli-invocations.cjs

@@ -73,9 +73,10 @@ function invocation(kind, executable, workdir, sessionRef = '', options = {}) {
     }
   }
   if (kind === 'hermes') {
-    const legacySessionRef = options.sessionTransport === 'acp' ? '' : sessionRef
+    // Hermes uses the same native conversation identifier across its ACP and
+    // legacy transports. Keep it when an attachment requires the legacy path.
+    const legacySessionRef = sessionRef
     const useAcp = options.invocationTransport !== 'legacy'
-      && !attachments.length
       && options.hermesAcpAvailable !== false
       && (!sessionRef || options.sessionTransport === 'acp')
     if (useAcp) {
@@ -86,6 +87,9 @@ function invocation(kind, executable, workdir, sessionRef = '', options = {}) {
         eventTransport: 'acp',
         fallbackTransport: 'legacy',
         fallbackSessionPolicy: 'invalidate',
+        // Hermes ACP advertises image prompts; send images as content blocks
+        // instead of silently dropping them from a text-only prompt.
+        ...(attachments.length ? { imageAttachments: attachments } : {}),
       }
     }
     return {
@@ -94,8 +98,6 @@ function invocation(kind, executable, workdir, sessionRef = '', options = {}) {
         'chat',
         '--quiet',
         ...(options.sandbox === 'workspace-write' ? ['--yolo'] : []),
-        ...(options.provider?.id ? ['--provider', options.provider.id] : []),
-        ...(options.provider?.model ? ['--model', options.provider.model] : []),
         ...(legacySessionRef ? ['--resume', legacySessionRef] : []),
         ...(attachments[0] ? ['--image', attachments[0]] : []),
         '--query',
@@ -143,6 +145,9 @@ function invocation(kind, executable, workdir, sessionRef = '', options = {}) {
         '--print',
         '--output-format', 'stream-json',
         '--include-partial-messages',
+        // Workspace write access means accepting file edits in the working
+        // directory. Skipping every permission check would also let the CLI
+        // run any command anywhere on the machine without confirmation.
         '--permission-mode', options.sandbox === 'workspace-write' ? 'acceptEdits' : 'plan',
         ...(maxTurns ? ['--max-turns', String(maxTurns)] : []),
         ...(sessionRef ? ['--resume', sessionRef] : []),
@@ -156,7 +161,9 @@ function invocation(kind, executable, workdir, sessionRef = '', options = {}) {
       command: executable,
       args: [
         '--mode', 'json', '--print',
-        ...(options.sandbox === 'workspace-write' ? ['--approve'] : ['--no-approve']),
+        ...(options.sandbox === 'workspace-write'
+          ? ['--approve']
+          : ['--no-approve', '--tools', 'read,grep,find,ls']),
         ...(sessionRef ? ['--session', sessionRef] : []),
       ],
       promptArg: true,
@@ -185,7 +192,7 @@ function invocation(kind, executable, workdir, sessionRef = '', options = {}) {
     }
   }
   if (kind === 'mimo') {
-    if (options.invocationTransport !== 'json') {
+    if (!attachments.length && options.invocationTransport !== 'json') {
       return {
         command: executable,
         args: ['acp', '--pure', '--cwd', workdir],
@@ -226,8 +233,6 @@ function invocation(kind, executable, workdir, sessionRef = '', options = {}) {
         '--output-format', 'stream-json',
         '--include-partial-messages',
         '--approval-mode', options.sandbox === 'workspace-write' ? 'auto-edit' : 'plan',
-        ...(options.provider?.id === 'openai' ? ['--auth-type', 'openai'] : []),
-        ...(options.provider?.model ? ['--model', options.provider.model] : []),
         ...(sessionRef ? ['--resume', sessionRef] : []),
       ],
       promptArg: true,

desktop/src/agents/cli/cli-network-environment.cjs

@@ -0,0 +1,19 @@
+const NETWORK_ENV_KEYS = Object.freeze([
+  'HTTP_PROXY', 'HTTPS_PROXY', 'ALL_PROXY', 'NO_PROXY',
+  'http_proxy', 'https_proxy', 'all_proxy', 'no_proxy',
+  'SSL_CERT_FILE', 'SSL_CERT_DIR', 'NODE_EXTRA_CA_CERTS',
+  'REQUESTS_CA_BUNDLE', 'CURL_CA_BUNDLE', 'NODE_USE_ENV_PROXY',
+])
+
+function networkEnvironment(source = {}, platform = process.platform) {
+  const result = {}
+  for (const [key, value] of Object.entries(source)) {
+    const allowed = platform === 'win32'
+      ? NETWORK_ENV_KEYS.find(name => name.toLowerCase() === key.toLowerCase())
+      : NETWORK_ENV_KEYS.find(name => name === key)
+    if (allowed && typeof value === 'string') result[allowed] = value
+  }
+  return result
+}
+
+module.exports = { NETWORK_ENV_KEYS, networkEnvironment }

desktop/src/agents/cli/cli-openclaw-gateway.cjs

@@ -7,7 +7,10 @@ const {
   TERMINATE_GRACE_MS,
   childEnvironment,
 } = require('./cli-process-support.cjs')
-const { configureOpenClawGatewayRuntime } = require('./openclaw-runtime.cjs')
+const {
+  configureOpenClawGatewayRuntime,
+  refreshOpenClawRuntimeAfterStartup,
+} = require('./openclaw-runtime.cjs')
 
 const GATEWAY_READY_MARKER = '[gateway] ready'
 const GATEWAY_READY_TIMEOUT_MS = 15_000
@@ -286,8 +289,11 @@ async function withOpenClawGateway(options, callback) {
         signal: options.signal,
         workdir: options.workdir,
       })
+      runtimeOptions = refreshOpenClawRuntimeAfterStartup(runtimeOptions)
     } catch (error) {
       await terminateGateway(child)
+      // Startup can migrate the config before a port conflict or health failure.
+      runtimeOptions = refreshOpenClawRuntimeAfterStartup(runtimeOptions)
       if (error?.retryable && attempt + 1 < GATEWAY_SETUP_ATTEMPTS) continue
       throw attachRuntimeOptions(error, runtimeOptions)
     }

desktop/src/agents/cli/cli-output-parsers.cjs

@@ -172,17 +172,38 @@ function stepFinishOutcome(reason) {
   return 'partial'
 }
 
+function structuredResultDiagnostic(event) {
+  const parts = []
+  if (typeof event?.error === 'string' && event.error.trim()) parts.push(event.error.trim())
+  else if (event?.error?.message) parts.push(String(event.error.message))
+  if (typeof event?.result === 'string' && event.result.trim()) parts.push(event.result.trim())
+  if (Array.isArray(event?.errors)) {
+    parts.push(...event.errors.map(value => String(value || '').trim()).filter(Boolean))
+  }
+  return parts.join('\n')
+}
+
 function ingestStructuredEvent(state, event) {
   if (!event || typeof event !== 'object' || Array.isArray(event)) return
   state.seen = true
   if (['workbuddy', 'claude', 'qwen'].includes(state.kind)
-      && event.type === 'result' && typeof event.result === 'string') {
+      && event.type === 'result') {
     if (typeof event.session_id === 'string') state.sessionRef = event.session_id
-    state.failure = null
-    state.diagnostic = ''
-    state.text = ''
-    appendStructuredText(state, event.result, '')
-    state.outcome = event.is_error === true ? 'partial' : 'completed'
+    if (typeof event.result === 'string') {
+      state.failure = null
+      state.diagnostic = ''
+      state.text = ''
+      appendStructuredText(state, event.result, '')
+      state.outcome = event.is_error === true ? 'partial' : 'completed'
+      return
+    }
+    // An error result without a usable result string still carries the CLI's
+    // error list; keep it as the failure diagnostic so authentication and
+    // session problems are classified instead of surfacing as a bare
+    // LOCAL_AGENT_PROCESS_FAILED.
+    if (event.is_error === true || String(event.subtype || '').startsWith('error')) {
+      markStructuredFailure(state, 'LOCAL_AGENT_PROCESS_FAILED', structuredResultDiagnostic(event))
+    }
     return
   }
   if (structuredEventFailed(event)) {
@@ -216,13 +237,14 @@ function ingestStructuredEvent(state, event) {
   }
 
   if (state.kind === 'pi') {
-    if (typeof event.id === 'string') state.sessionRef = event.id
+    if (event.type === 'session' && typeof event.id === 'string') state.sessionRef = event.id
     const update = event.assistantMessageEvent
     if (event.type === 'message_update' && update?.type === 'text_delta'
         && typeof update.delta === 'string') {
       appendStructuredText(state, update.delta, '')
     }
-    if (event.type === 'message_end' && Array.isArray(event.message?.content)) {
+    if (event.type === 'message_end' && event.message?.role === 'assistant'
+        && Array.isArray(event.message?.content)) {
       for (const block of event.message.content) {
         if (block?.type === 'text' && typeof block.text === 'string') {
           if (!state.text) appendStructuredText(state, block.text, '')
@@ -234,7 +256,15 @@ function ingestStructuredEvent(state, event) {
     if (event.type === 'turn_end') {
       state.outcome = stepFinishOutcome(event.message?.stopReason || event.stopReason || 'stop')
     }
-    if (event.type === 'agent_end') state.outcome = 'completed'
+    if (event.type === 'agent_end') {
+      const lastAssistant = event.messages?.findLast(message => message?.role === 'assistant')
+      if (lastAssistant?.stopReason) state.outcome = stepFinishOutcome(lastAssistant.stopReason)
+      else if (!state.outcome) state.outcome = 'completed'
+      if (state.outcome === 'completed' && event.willRetry !== true) {
+        state.failure = null
+        state.diagnostic = ''
+      }
+    }
     if (state.outcome === 'failed') markStructuredFailure(state)
     return
   }
@@ -709,7 +739,8 @@ function parsePiOutput(stdout) {
         && typeof event.assistantMessageEvent.delta === 'string') {
       text += event.assistantMessageEvent.delta
     }
-    if (!text && event.type === 'message_end' && Array.isArray(event.message?.content)) {
+    if (!text && event.type === 'message_end' && event.message?.role === 'assistant'
+        && Array.isArray(event.message?.content)) {
       text = event.message.content
         .filter(block => block?.type === 'text' && typeof block.text === 'string')
         .map(block => block.text)
@@ -718,7 +749,7 @@ function parsePiOutput(stdout) {
   }
   const parsedText = text.trim()
   const answer = stripPiInternalPromptEcho(parsedText)
-  return { text: parsedText ? answer : String(stdout || '').trim(), sessionRef }
+  return { text: parsedText ? answer : events.length ? '' : String(stdout || '').trim(), sessionRef }
 }
 
 function openCodeReviewText(output) {
@@ -833,8 +864,11 @@ function classifyCliOutcome(kind, stdout) {
   }
   if (kind === 'pi') {
     const terminal = events.findLast(event => event?.type === 'turn_end' || event?.type === 'agent_end')
-    const reason = terminal?.message?.stopReason || terminal?.stopReason || 'stop'
-    return { outcome: stepFinishOutcome(reason) }
+    const lastAssistant = terminal?.messages?.findLast(message => message?.role === 'assistant')
+    const previousTurn = events.findLast(event => event?.type === 'turn_end')
+    const reason = terminal?.message?.stopReason || terminal?.stopReason || lastAssistant?.stopReason
+      || previousTurn?.message?.stopReason || previousTurn?.stopReason
+    return { outcome: reason ? stepFinishOutcome(reason) : terminal ? 'completed' : 'partial' }
   }
   if (['mimo', 'opencode'].includes(kind)) {
     const finish = events.findLast(event => (
@@ -872,7 +906,8 @@ function classifyCliOutcome(kind, stdout) {
 function structuredCliError(stdout) {
   const failed = parseJsonOutputEvents(stdout)
     .findLast(event => event?.is_error || event?.subtype?.startsWith('error'))
-  return String(failed?.error?.message || failed?.result || '').trim()
+  if (!failed) return ''
+  return structuredResultDiagnostic(failed)
 }
 
 function hermesSessionRef(stderr) {

desktop/src/agents/cli/cli-process-support.cjs

@@ -1,7 +1,12 @@
 const path = require('node:path')
-const { searchPath, systemChildEnvironment } = require('./cli-discovery.cjs')
+const { launchEnvironment, searchPath, systemChildEnvironment } = require('./cli-discovery.cjs')
+const { NETWORK_ENV_KEYS, networkEnvironment } = require('./cli-network-environment.cjs')
 const { redactChildSecrets } = require('./cli-runtime-events.cjs')
-const { agentRuntimeError } = require('../agent-runtime-contract.cjs')
+const {
+  agentRuntimeError,
+  authFailureCode,
+} = require('../agent-runtime-contract.cjs')
+const { authFailureCertainty, isContextOverflow } = require('../../runs/failure-signals.cjs')
 const { validateOpenClawRuntimeGuard } = require('./openclaw-runtime.cjs')
 
 const TERMINATE_GRACE_MS = 500
@@ -21,20 +26,26 @@ const OPENCLAW_RUNTIME_PATH_KEYS = Object.freeze([
   'OPENCLAW_CONFIG_PATH',
   'OPENCLAW_WORKSPACE_DIR',
 ])
+const CONTROLLED_RUNTIME_ENV = /^(?:AWS_|AZURE_|GOOGLE_|GEMINI_|ANTHROPIC_|OPENAI_|OPENROUTER_|CODEBUDDY_|HERMES_|PI_|KIMI_|MIMO_|DASHSCOPE_|CLAUDE_|OCR_|CLOUDSDK_|VERTEX_|CLOUD_ML_|MOONSHOT_|CONNECTOR_TEST_|MELDWORK_TEST_)/
+const BLOCKED_RUNTIME_ENV = /^(?:NODE_OPTIONS|NODE_PATH|LD_PRELOAD|LD_LIBRARY_PATH|DYLD_|ELECTRON_RUN_AS_NODE)$/
 
 function agentExecutionError(code, diagnostic = '') {
   return agentRuntimeError(code, diagnostic)
 }
 
-function authConfigurationFailure(detail) {
-  return /api[ _-]?key|access[ _-]?token|refresh[ _-]?token|auth[ _-]?token|credential|auth(?:entication|orization)?|log(?:ged)?[ -]?in|sign(?:ed)?[ -]?in|unauthorized|forbidden|\b(?:401|403)\b|select an auth type|(?:provider|model).{0,80}(?:reject|configur|missing|invalid)|(?:reject|configur|missing|invalid).{0,80}(?:provider|model)|令牌|凭据|登录|认证|鉴权|(?:提供商|供应商|模型).{0,40}(?:配置|拒绝|缺失|无效)|(?:配置|拒绝|缺失|无效).{0,40}(?:提供商|供应商|模型)/i
-    .test(String(detail || ''))
-}
-
 function invalidSessionFailure(detail) {
   const text = String(detail || '').trim()
-  if (!text || authConfigurationFailure(text)
+  if (!text || authFailureCertainty(text)
       || /\b(?:api|access|refresh|auth|session)[ _-]?token\b/i.test(text)) return false
+  // A resumed native session that no longer fits the model context cannot be
+  // continued; rebuilding it with compact continuation context is the fix.
+  if (isContextOverflow(text)) return true
+  // Runtime crashes ("thread 'main' panicked ... invalid ...") and parser
+  // errors mention threads or invalid input without saying anything about the
+  // stored session. Discarding a healthy session for them loses context.
+  if (/\bpanicked\b|\bSyntaxError\b|\binvalid JSON\b|\bUnexpected token\b|\bstack backtrace\b/i.test(text)) {
+    return false
+  }
   return /\b(?:no|unknown|invalid|expired|missing|stale)\s+(?:saved\s+)?(?:session|conversation|thread)\b|\b(?:session|conversation|thread)(?:\s+(?:id|key|reference))?\b[^\n]{0,100}\b(?:not found|does not exist|no longer exists|invalid|unknown|expired|has expired|was deleted|cannot be resumed)\b/i
     .test(text)
 }
@@ -44,9 +55,7 @@ function failedAgentProcessError(detail, options = {}) {
   return agentExecutionError(
     options.sessionRef && invalidSessionFailure(detail)
       ? 'LOCAL_AGENT_SESSION_INVALID'
-      : authConfigurationFailure(detail)
-        ? 'LOCAL_AGENT_AUTH_REQUIRED'
-        : 'LOCAL_AGENT_PROCESS_FAILED',
+      : authFailureCode(detail) || 'LOCAL_AGENT_PROCESS_FAILED',
     detail,
   )
 }
@@ -60,7 +69,10 @@ function openClawChildEnvironment(workdir, options, platform) {
     throw new Error('OPENCLAW_RUNTIME_UNSAFE_PATH')
   }
 
-  const env = systemChildEnvironment(process.env, platform)
+  const env = {
+    ...systemChildEnvironment(process.env, platform),
+    ...networkEnvironment(source, platform),
+  }
   for (const key of [
     'HOME', 'USERPROFILE', 'HOMEDRIVE', 'HOMEPATH',
     'XDG_CONFIG_HOME', 'XDG_DATA_HOME', 'XDG_STATE_HOME', 'XDG_CACHE_HOME',
@@ -103,11 +115,17 @@ function childEnvironment(agent, workdir, options, platform) {
     ? { OPENCODE_PERMISSION: OPENCODE_READ_ONLY_PERMISSION }
     : {}
   const source = { ...process.env, ...options.env }
+  const controlled = Object.fromEntries(Object.entries(options.env || {}).filter(([key, value]) => (
+    typeof value === 'string'
+      && (CONTROLLED_RUNTIME_ENV.test(key) || NETWORK_ENV_KEYS.includes(key))
+      && !BLOCKED_RUNTIME_ENV.test(key)
+  )))
   return {
     ...systemChildEnvironment(process.env, platform),
-    ...options.env,
+    ...controlled,
     ...hermesSafetyEnv,
     ...openCodeSafetyEnv,
+    ...launchEnvironment(agent.kind, agent.executable, { platform }),
     PATH: searchPath({ platform, env: source }),
   }
 }

desktop/src/agents/cli/cli-runtime-event-mappers.cjs

@@ -425,7 +425,7 @@ function acpPlanSummary(update) {
 }
 
 function createAcpRuntimeState() {
-  return { tools: new Map() }
+  return { tools: new Map(), reasoning: new Map() }
 }
 
 const ACP_TOOL_KINDS = new Set([
@@ -447,10 +447,22 @@ function acpToolOperation(update) {
 function acpRuntimeEvents(update, state) {
   if (!update || typeof update !== 'object') return []
   if (update.sessionUpdate === 'agent_message_chunk' && update.content?.type === 'text') {
-    return [{ type: 'answer_delta', status: 'running', delta: update.content.text }]
+    // Close an open reasoning phase once the visible answer starts.
+    const closed = state?.reasoning?.get('reasoning:acp')?.startsWith('running:')
+      ? reasoningLifecycleEvent(state, 'reasoning:acp', 'completed')
+      : null
+    return [
+      ...(closed ? [closed] : []),
+      { type: 'answer_delta', status: 'running', delta: update.content.text },
+    ]
   }
   if (update.sessionUpdate === 'agent_thought_chunk' && update.content?.type === 'text') {
-    return []
+    // Thinking is real progress: report a text-free reasoning status so the
+    // watchdog and the trace know the Agent is working. The private thought
+    // text itself never leaves the main process.
+    if (!state?.reasoning) return []
+    const event = reasoningLifecycleEvent(state, 'reasoning:acp', 'running')
+    return event ? [event] : []
   }
   if (update.sessionUpdate === 'tool_call' || update.sessionUpdate === 'tool_call_update') {
     const status = runtimeEventStatus(update.status)

desktop/src/agents/cli/cli-runtime-event-sanitizer.cjs

@@ -28,21 +28,38 @@ function utf8ByteLength(value) {
 
 function redactChildSecrets(value, env) {
   let result = String(value || '')
-  for (const [name, secret] of Object.entries(env)) {
-    if (!/(?:API[_-]?KEY|TOKEN|SECRET|PASSWORD|AUTHORIZATION)/i.test(name)
-        || typeof secret !== 'string' || utf8ByteLength(secret) < 8) continue
+  for (const secret of childSecrets(env)) {
     result = result.split(secret).join('[redacted]')
   }
   return result
 }
 
 function childSecrets(env) {
-  return Object.entries(env || {})
-    .filter(([name, secret]) => (
-      /(?:API[_-]?KEY|TOKEN|SECRET|PASSWORD|AUTHORIZATION)/i.test(name)
-      && typeof secret === 'string' && utf8ByteLength(secret) >= 8
-    ))
-    .map(([, secret]) => secret)
+  const secrets = new Set()
+  for (const [name, value] of Object.entries(env || {})) {
+    if (typeof value !== 'string' || !value) continue
+    if (/(?:API[_-]?KEY|ACCESS[_-]?KEY|TOKEN|SECRET|PASSWORD|AUTHORIZATION)/i.test(name)
+        && utf8ByteLength(value) >= 8) secrets.add(value)
+    if (!/^(?:https?|all)_proxy$/i.test(name)) continue
+    if (value.includes('@')) secrets.add(value)
+    try {
+      const url = new URL(value)
+      if (!url.username && !url.password) continue
+      secrets.add(value)
+      secrets.add(url.href)
+      const userinfo = `${url.username}:${url.password}`
+      secrets.add(userinfo)
+      const decoded = decodeURIComponent(userinfo)
+      secrets.add(decoded)
+      secrets.add(Buffer.from(decoded).toString('base64'))
+      for (const part of [url.username, url.password]) {
+        if (utf8ByteLength(part) >= 8) secrets.add(part)
+        const literal = decodeURIComponent(part)
+        if (utf8ByteLength(literal) >= 8) secrets.add(literal)
+      }
+    } catch { /* malformed proxy URLs are diagnosed by the native runtime */ }
+  }
+  return [...secrets].sort((left, right) => right.length - left.length)
 }
 
 function assignmentNameVariants(parts) {

desktop/src/agents/cli/cli-runtime-events.cjs

@@ -13,6 +13,7 @@ const {
   parseMimoOutput,
   parseOpenCodeOutput,
   parseOpenCodeReviewOutput,
+  parsePiOutput,
   parseWorkBuddyOutput,
   readHermesFinalResponse,
   readHermesMessageWatermark,
@@ -54,6 +55,7 @@ module.exports = {
   parseMimoOutput,
   parseOpenCodeOutput,
   parseOpenCodeReviewOutput,
+  parsePiOutput,
   parseWorkBuddyOutput,
   readHermesFinalResponse,
   readHermesMessageWatermark,

desktop/src/agents/cli/openclaw-runtime.cjs

@@ -1,10 +1,10 @@
 const crypto = require('node:crypto')
 const fs = require('node:fs')
 const path = require('node:path')
+const { isDeepStrictEqual } = require('node:util')
 
 const { atomicWritePrivateFile } = require('../../security/private-file.cjs')
 
-const MANAGED_OPENCLAW_PROVIDER_ID = 'meldwork-desktop'
 const READ_ONLY_TOOLS = [
   'read', 'web_search', 'web_fetch', 'memory_search', 'memory_get', 'session_status',
 ]
@@ -36,12 +36,12 @@ const OPENCLAW_RUNTIME_PATH_ENV_KEYS = Object.freeze([
   'OPENCLAW_WORKSPACE_DIR',
 ])
 const OPENCLAW_RUNTIME_CREDENTIAL_KEYS = new Set([
-  'MELDWORK_OPENCLAW_API_KEY',
   'MELDWORK_OPENCLAW_NATIVE_API_KEY',
   'OPENCLAW_GATEWAY_TOKEN',
 ])
 const issuedOpenClawRuntimeGuards = new WeakSet()
 const openClawRuntimeCredentialDigests = new WeakMap()
+const openClawRuntimeConfigs = new WeakMap()
 const OPENCLAW_ENV_SECRET_NAME = /^[A-Z][A-Z0-9_]{0,127}$/
 const OPENCLAW_SECRET_MARKERS = new Set([
   'secretref-managed',
@@ -89,7 +89,7 @@ function directoryIdentity(filename) {
   })
 }
 
-function fileIdentity(filename) {
+function readRuntimeConfigFile(filename) {
   const resolved = path.resolve(filename)
   const entry = fs.lstatSync(resolved)
   if (!entry.isFile() || entry.isSymbolicLink() || fs.realpathSync(resolved) !== resolved) {
@@ -99,14 +99,22 @@ function fileIdentity(filename) {
   const descriptor = fs.openSync(resolved, fs.constants.O_RDONLY | noFollow)
   try {
     const before = fs.fstatSync(descriptor)
-    if (!before.isFile() || !sameFilesystemIdentity(before, {
+    if (!before.isFile() || before.size > 256 * 1024 || !sameFilesystemIdentity(before, {
       dev: identityValue(entry.dev),
       ino: identityValue(entry.ino),
       birthtime: identityValue(entry.birthtimeMs),
     })) {
       throw new Error('OPENCLAW_RUNTIME_UNSAFE_PATH')
     }
-    const contents = fs.readFileSync(descriptor)
+    const buffer = Buffer.alloc(256 * 1024 + 1)
+    let length = 0
+    while (length < buffer.length) {
+      const count = fs.readSync(descriptor, buffer, length, buffer.length - length, null)
+      if (!count) break
+      length += count
+    }
+    if (length === buffer.length) throw new Error('OPENCLAW_RUNTIME_UNSAFE_PATH')
+    const contents = buffer.subarray(0, length)
     const after = fs.fstatSync(descriptor)
     const current = fs.lstatSync(resolved)
     if (!sameFilesystemIdentity(after, {
@@ -121,7 +129,7 @@ function fileIdentity(filename) {
         }) || fs.realpathSync(resolved) !== resolved) {
       throw new Error('OPENCLAW_RUNTIME_UNSAFE_PATH')
     }
-    return Object.freeze({
+    const identity = Object.freeze({
       path: resolved,
       dev: identityValue(before.dev),
       ino: identityValue(before.ino),
@@ -130,11 +138,16 @@ function fileIdentity(filename) {
       size: contents.length,
       digest: crypto.createHash('sha256').update(contents).digest('hex'),
     })
+    return { identity, contents: contents.toString('utf8') }
   } finally {
     fs.closeSync(descriptor)
   }
 }
 
+function fileIdentity(filename) {
+  return readRuntimeConfigFile(filename).identity
+}
+
 function validateDirectoryIdentities(identities) {
   for (const expected of identities) {
     const current = directoryIdentity(expected.path)
@@ -158,26 +171,7 @@ function validateFileIdentity(expected) {
   }
 }
 
-function normalizeProvider(input) {
-  const apiKey = typeof input?.OPENAI_API_KEY === 'string' ? input.OPENAI_API_KEY.trim() : ''
-  const baseUrl = String(input?.OPENAI_BASE_URL || '')
-    .trim()
-    .replace(/\/+$/, '')
-    .replace(/\/chat\/completions$/i, '')
-  const model = String(input?.OPENAI_MODEL || '').trim()
-  let parsed
-  try { parsed = new URL(baseUrl) } catch { throw new Error('OPENCLAW_PROVIDER_INVALID') }
-  const loopback = ['localhost', '127.0.0.1', '[::1]'].includes(parsed.hostname)
-  if (!apiKey || apiKey.length > 8192 || isOpenClawSecretReference(apiKey)
-      || !model || model.length > 120
-      || (parsed.protocol !== 'https:' && !(parsed.protocol === 'http:' && loopback))
-      || parsed.username || parsed.password || parsed.search || parsed.hash) {
-    throw new Error('OPENCLAW_PROVIDER_INVALID')
-  }
-  return { apiKey, baseUrl, model }
-}
-
-function runtimePaths({ storageRoot, workdir, sessionRef, configName }) {
+function runtimePaths({ storageRoot, workdir, sessionRef, configName, runtimeVersion }) {
   if (!path.isAbsolute(storageRoot) || !path.isAbsolute(workdir)) {
     throw new Error('OPENCLAW_RUNTIME_INVALID_SCOPE')
   }
@@ -189,8 +183,15 @@ function runtimePaths({ storageRoot, workdir, sessionRef, configName }) {
     throw new Error('OPENCLAW_RUNTIME_INVALID_SCOPE')
   }
   const resolvedWorkdir = path.resolve(workdir)
+  // The managed scope is keyed by session, workspace, and CLI version. OpenClaw
+  // refuses to start on state written by an older CLI ("Legacy workspace setup
+  // state requires migration"), so a version-scoped directory gives an upgraded
+  // CLI a fresh isolated scope instead of a guaranteed startup failure.
+  const versionSegment = String(runtimeVersion || '').trim()
   const scope = crypto.createHash('sha256')
-    .update(`${sessionRef || 'configure'}\0${resolvedWorkdir}`)
+    .update(versionSegment
+      ? `${sessionRef || 'configure'}\0${resolvedWorkdir}\0${versionSegment}`
+      : `${sessionRef || 'configure'}\0${resolvedWorkdir}`)
     .digest('hex')
     .slice(0, 24)
   const runtimeRoot = path.join(resolvedStorageRoot, 'openclaw-managed', scope)
@@ -268,6 +269,7 @@ function writeRuntimeConfig(runtime, config, credentials, credentialKey) {
     credentialKeys: Object.freeze(credentialEntries.map(([key]) => key)),
   })
   issuedOpenClawRuntimeGuards.add(guard)
+  openClawRuntimeConfigs.set(guard, contents)
   openClawRuntimeCredentialDigests.set(guard, Object.freeze(Object.fromEntries(
     credentialEntries.map(([key, value]) => [
       key,
@@ -277,7 +279,7 @@ function writeRuntimeConfig(runtime, config, credentials, credentialKey) {
   return guard
 }
 
-function validateOpenClawRuntimeGuard(guard, env = {}) {
+function validateRuntimeScope(guard, env = {}) {
   const credentialDigests = guard && typeof guard === 'object'
     ? openClawRuntimeCredentialDigests.get(guard)
     : undefined
@@ -301,10 +303,57 @@ function validateOpenClawRuntimeGuard(guard, env = {}) {
     }
   }
   validateDirectoryIdentities(guard.directories)
+}
+
+function validateOpenClawRuntimeGuard(guard, env = {}) {
+  validateRuntimeScope(guard, env)
   validateFileIdentity(guard.config)
   return true
 }
 
+function comparableMigratedConfig(config) {
+  // OpenClaw materializes the implicit main agent and stamps migrations at startup.
+  const result = JSON.parse(JSON.stringify(config))
+  const meta = result.meta
+  if (meta && typeof meta === 'object' && !Array.isArray(meta)
+      && Object.keys(meta).every(key => ['migrations', 'lastTouchedVersion', 'lastTouchedAt'].includes(key))
+      && (meta.lastTouchedVersion === undefined || /^\d{4}\.\d{1,2}\.\d{1,2}(?:[-+.][\w.-]+)?$/.test(meta.lastTouchedVersion))
+      && (meta.lastTouchedAt === undefined || (typeof meta.lastTouchedAt === 'string' && Number.isFinite(Date.parse(meta.lastTouchedAt))))
+      && (meta.migrations === undefined || isDeepStrictEqual(meta.migrations, { modelPolicyAllowlist: true }))) {
+    delete result.meta
+  }
+  if (isDeepStrictEqual(result.agents?.entries, { main: {} })) delete result.agents.entries
+  if (isDeepStrictEqual(result.discovery?.wideArea, { enabled: false })) {
+    result.discovery.wideArea = {}
+  }
+  return result
+}
+
+function refreshOpenClawRuntimeAfterStartup(options) {
+  const guard = options?.openClawRuntimeGuard
+  const env = options?.env || {}
+  validateRuntimeScope(guard, env)
+  const { identity: config, contents } = readRuntimeConfigFile(guard.config.path)
+  if (process.platform !== 'win32' && (config.mode & 0o777) !== 0o600) {
+    throw new Error('OPENCLAW_RUNTIME_UNSAFE_PATH')
+  }
+  validateDirectoryIdentities(guard.directories)
+  validateFileIdentity(config)
+  let equivalent = false
+  try {
+    equivalent = isDeepStrictEqual(
+      comparableMigratedConfig(JSON.parse(openClawRuntimeConfigs.get(guard))),
+      comparableMigratedConfig(JSON.parse(contents)),
+    )
+  } catch { /* malformed config is rejected below */ }
+  if (!equivalent) throw new Error('OPENCLAW_RUNTIME_UNSAFE_PATH')
+  const refreshed = Object.freeze({ ...guard, config })
+  issuedOpenClawRuntimeGuards.add(refreshed)
+  openClawRuntimeConfigs.set(refreshed, contents)
+  openClawRuntimeCredentialDigests.set(refreshed, openClawRuntimeCredentialDigests.get(guard))
+  return { ...options, openClawRuntimeGuard: refreshed }
+}
+
 function configureOpenClawGatewayRuntime(options, port) {
   if (!Number.isInteger(port) || port < 1 || port > 65535) {
     throw new Error('OPENCLAW_GATEWAY_PORT_INVALID')
@@ -313,7 +362,7 @@ function configureOpenClawGatewayRuntime(options, port) {
   const env = options?.env || {}
   validateOpenClawRuntimeGuard(guard, env)
 
-  const config = JSON.parse(fs.readFileSync(guard.config.path, 'utf8'))
+  const config = JSON.parse(openClawRuntimeConfigs.get(guard))
   const gatewayCredentialKey = 'OPENCLAW_GATEWAY_TOKEN'
   const nextConfig = {
     ...config,
@@ -395,74 +444,8 @@ function normalizedNativeModel(input) {
   return model
 }
 
-function managedOpenClawOptions({
-  storageRoot, workdir, sessionRef = '', allowWrite = false, provider,
-}) {
-  const normalized = normalizeProvider(provider)
-  const configName = allowWrite
-    ? 'openclaw.provider.workspace-write.json'
-    : 'openclaw.provider.read-only.json'
-  const runtime = runtimePaths({
-    storageRoot, workdir, sessionRef, configName,
-  })
-  const { resolvedWorkdir, home, state, configPath } = runtime
-
-  const modelRef = `${MANAGED_OPENCLAW_PROVIDER_ID}/${normalized.model}`
-  const gatewayCredentialKey = 'OPENCLAW_GATEWAY_TOKEN'
-  const gatewayToken = crypto.randomBytes(32).toString('base64url')
-  const config = {
-    agents: {
-      defaults: {
-        workspace: resolvedWorkdir,
-        model: { primary: modelRef },
-        skipBootstrap: true,
-      },
-    },
-    models: {
-      mode: 'replace',
-      providers: {
-        [MANAGED_OPENCLAW_PROVIDER_ID]: {
-          baseUrl: normalized.baseUrl,
-          apiKey: {
-            source: 'env', provider: 'default', id: 'MELDWORK_OPENCLAW_API_KEY',
-          },
-          api: 'openai-completions',
-          models: [{ id: normalized.model, name: normalized.model, input: ['text'] }],
-        },
-      },
-    },
-    tools: toolPolicy(allowWrite),
-    gateway: {
-      mode: 'local',
-      bind: 'loopback',
-      auth: {
-        mode: 'token',
-        token: { source: 'env', provider: 'default', id: gatewayCredentialKey },
-      },
-      controlUi: { enabled: false },
-    },
-  }
-  const credentialKey = 'MELDWORK_OPENCLAW_API_KEY'
-  const openClawRuntimeGuard = writeRuntimeConfig(runtime, config, {
-    [credentialKey]: normalized.apiKey,
-    [gatewayCredentialKey]: gatewayToken,
-  }, credentialKey)
-
-  return {
-    openClawRuntimeGuard,
-    env: {
-      OPENCLAW_HOME: home,
-      OPENCLAW_STATE_DIR: state,
-      OPENCLAW_CONFIG_PATH: configPath,
-      OPENCLAW_WORKSPACE_DIR: resolvedWorkdir,
-      [credentialKey]: normalized.apiKey,
-      [gatewayCredentialKey]: gatewayToken,
-    },
-  }
-}
-
 function nativeOpenClawOptions({
-  storageRoot, workdir, sessionRef = '', allowWrite = false, runtime,
+  storageRoot, workdir, sessionRef = '', allowWrite = false, runtime, openclawVersion,
 }) {
   const modelRef = String(runtime?.model || '').trim()
   const provider = runtime?.provider
@@ -494,7 +477,7 @@ function nativeOpenClawOptions({
     ? 'openclaw.native.workspace-write.json'
     : 'openclaw.native.read-only.json'
   const runtimePathsResult = runtimePaths({
-    storageRoot, workdir, sessionRef, configName,
+    storageRoot, workdir, sessionRef, configName, runtimeVersion: openclawVersion,
   })
   const { resolvedWorkdir, home, state, configPath } = runtimePathsResult
   const modelConfig = { primary: modelRef }
@@ -555,7 +538,7 @@ function nativeOpenClawOptions({
 module.exports = {
   configureOpenClawGatewayRuntime,
   isOpenClawSecretReference,
-  managedOpenClawOptions,
   nativeOpenClawOptions,
+  refreshOpenClawRuntimeAfterStartup,
   validateOpenClawRuntimeGuard,
 }

desktop/src/agents/cloud/cloud-agent-bridge.cjs

@@ -632,8 +632,6 @@ class CloudAgentBridgeStore {
       installed: true,
       installSupported: false,
       installErrorCode: '',
-      providerCompatible: false,
-      providerMode: 'connector',
       imageAttachmentLimit: 0,
       custom: false,
       connector: true,

desktop/src/agents/connectors/agent-connector-local.cjs

@@ -49,7 +49,7 @@ const SAMPLE_AGENT_CONNECTOR_MANIFEST = createAgentConnectorManifest({
   },
   outboundDestinations: [],
   credentials: { mode: 'none', slots: [] },
-  license: 'LicenseRef-Meldwork-NC',
+  license: 'Apache-2.0',
 })
 const SAMPLE_CREDENTIAL_AGENT_CONNECTOR_MANIFEST = createAgentConnectorManifest({
   connectorId: 'external.codex-provider-sample',
@@ -82,7 +82,7 @@ const SAMPLE_CREDENTIAL_AGENT_CONNECTOR_MANIFEST = createAgentConnectorManifest(
     mode: 'credential-ref',
     slots: [{ slotId: 'openai-api-key', type: 'api-key', required: true }],
   },
-  license: 'LicenseRef-Meldwork-NC',
+  license: 'Apache-2.0',
 })
 const APPROVED_AGENT_CONNECTOR_MANIFESTS = Object.freeze([
   SAMPLE_AGENT_CONNECTOR_MANIFEST,
@@ -117,7 +117,7 @@ const SAMPLE_LOCAL_ECHO_AGENT_CONNECTOR_MANIFEST = createAgentConnectorManifest(
   },
   outboundDestinations: [],
   credentials: { mode: 'none', slots: [] },
-  license: 'LicenseRef-Meldwork-NC',
+  license: 'Apache-2.0',
 })
 const SAMPLE_LOCAL_ECHO_AGENT_CONNECTOR_PACKAGE = createAgentConnectorPackage({
   publisher: { id: 'meldwork', name: 'Meldwork' },
@@ -603,8 +603,6 @@ class LocalAgentConnectors {
         installed: true,
         installSupported: false,
         installErrorCode: '',
-        providerCompatible: false,
-        providerMode: 'connector',
         imageAttachmentLimit: manifest.inputTypes.includes('image') ? 32 : 0,
         custom: !cloud,
         connector: true,

desktop/src/agents/custom-agent-contract.cjs

@@ -127,9 +127,6 @@ function publicProfile(definition, version = '') {
     promptMode: definition.promptMode,
     custom: true,
     recommended: false,
-    providerCompatible: false,
-    providerSupport: 'custom-cli',
-    providerMode: 'custom',
     installed: true,
     installSupported: false,
     installErrorCode: '',

desktop/src/agents/installer/agent-installer-contract.cjs

@@ -4,6 +4,17 @@ const fs = require('node:fs')
 const path = require('node:path')
 const { promisify } = require('node:util')
 const { searchPath } = require('../cli/cli-adapters.cjs')
+const { compareReleaseVersions, extractAgentVersion } = require('../agent-compatibility.cjs')
+
+// True when the installed CLI is the recipe's version or newer. A substring
+// test treated "11.2.3" as "1.2.3" and offered downgrades for newer CLIs.
+function installedVersionCurrent(installedVersion, targetVersion) {
+  const installed = extractAgentVersion(installedVersion).split('-')[0]
+  const target = extractAgentVersion(targetVersion).split('-')[0]
+  const comparison = installed && target ? compareReleaseVersions(installed, target) : null
+  if (comparison != null) return comparison >= 0
+  return Boolean(targetVersion) && String(installedVersion || '').includes(targetVersion)
+}
 
 const execFileAsync = promisify(execFile)
 const DETECTION_CACHE_TTL_MS = 30000
@@ -99,50 +110,17 @@ const NPM_RELEASES = Object.freeze({
 const ALLOWED_SCRIPT_URLS = new Set(Object.values(HERMES_RECIPES).map(recipe => recipe.url))
 
 const AGENT_CATALOG = Object.freeze([
-  {
-    kind: 'codex', label: 'Codex', recommended: false, providerCompatible: false,
-    providerSupport: 'responses-required',
-  },
-  {
-    kind: 'claude', label: 'Claude Code', recommended: false, providerCompatible: false,
-    providerSupport: 'anthropic-required',
-  },
-  {
-    kind: 'hermes', label: 'Hermes', recommended: true, providerCompatible: true,
-    providerSupport: 'supported',
-  },
-  {
-    kind: 'workbuddy', label: 'WorkBuddy', recommended: true, providerCompatible: true,
-    providerSupport: 'experimental',
-  },
-  {
-    kind: 'openclaw', label: 'OpenClaw', recommended: true, providerCompatible: true,
-    providerSupport: 'supported',
-  },
-  {
-    kind: 'pi', label: 'Pi Agent', recommended: false, providerCompatible: false,
-    providerSupport: 'native-config',
-  },
-  {
-    kind: 'opencode', label: 'OpenCode', recommended: false, providerCompatible: false,
-    providerSupport: 'native-config',
-  },
-  {
-    kind: 'opencodereview', label: 'OpenCodeReview', recommended: false, providerCompatible: true,
-    providerSupport: 'supported',
-  },
-  {
-    kind: 'kimi', label: 'Kimi Code', recommended: false, providerCompatible: false,
-    providerSupport: 'native-config',
-  },
-  {
-    kind: 'mimo', label: 'MiMo Code', recommended: false, providerCompatible: false,
-    providerSupport: 'native-config',
-  },
-  {
-    kind: 'qwen', label: 'Qwen Code', recommended: false, providerCompatible: true,
-    providerSupport: 'supported',
-  },
+  { kind: 'codex', label: 'Codex', recommended: false },
+  { kind: 'claude', label: 'Claude Code', recommended: false },
+  { kind: 'hermes', label: 'Hermes', recommended: true },
+  { kind: 'workbuddy', label: 'WorkBuddy', recommended: true },
+  { kind: 'openclaw', label: 'OpenClaw', recommended: true },
+  { kind: 'pi', label: 'Pi Agent', recommended: false },
+  { kind: 'opencode', label: 'OpenCode', recommended: false },
+  { kind: 'opencodereview', label: 'OpenCodeReview', recommended: false },
+  { kind: 'kimi', label: 'Kimi Code', recommended: false },
+  { kind: 'mimo', label: 'MiMo Code', recommended: false },
+  { kind: 'qwen', label: 'Qwen Code', recommended: false },
 ])
 
 function installRecipe(kind, platform) {
@@ -372,6 +350,7 @@ function verifiedRecipeAgent(agent, recipe) {
 }
 
 module.exports = {
+  installedVersionCurrent,
   AGENT_CATALOG,
   DETECTION_CACHE_TTL_MS,
   abortError,

desktop/src/agents/installer/agent-installer.cjs

@@ -3,6 +3,7 @@ const fs = require('node:fs')
 const { randomUUID } = require('node:crypto')
 const { DISPLAY_LIMIT, listLocalAgentSkills } = require('../../skills/local-skill-catalog.cjs')
 const {
+  installedVersionCurrent,
   AGENT_CATALOG,
   DETECTION_CACHE_TTL_MS,
   abortError,
@@ -60,12 +61,15 @@ function publicCompatibility(agent) {
     ? agent.versionIdentified
     : Boolean(resolvedVersion)
   const compatible = compatibilityState === 'compatible'
-  const incompatibilityReason = compatibilityState === 'incompatible'
+  const incompatibilityReason = compatibilityState === 'unknown'
+      && agent.incompatibilityReason === 'LOCAL_AGENT_CAPABILITY_PROBE_TIMEOUT'
+    ? 'LOCAL_AGENT_CAPABILITY_PROBE_TIMEOUT'
+    : compatibilityState === 'incompatible'
     ? (COMPATIBILITY_REASONS.has(agent.incompatibilityReason)
         ? agent.incompatibilityReason
         : 'LOCAL_AGENT_VERSION_UNSUPPORTED')
     : ''
-  const incompatibilityProbe = compatibilityState === 'incompatible'
+  const incompatibilityProbe = Boolean(incompatibilityReason)
     && /^[a-z0-9-]{1,80}$/.test(String(agent.incompatibilityProbe || ''))
     ? agent.incompatibilityProbe
     : ''
@@ -164,6 +168,10 @@ class AgentInstaller extends EventEmitter {
       agents: AGENT_CATALOG.map(profile => {
         const agent = installed.get(profile.kind)
         const recipe = installRecipe(profile.kind, this.platform)
+        const targetVersion = recipe?.detectedVersion || recipe?.version || ''
+        const currentVersion = String(agent?.version || '')
+        const upToDate = Boolean(agent && targetVersion
+          && installedVersionCurrent(currentVersion, targetVersion))
         let installSupported = Boolean(recipe)
         let installErrorCode = ''
         if (!recipe) {
@@ -177,6 +185,8 @@ class AgentInstaller extends EventEmitter {
           ...profile,
           installed: Boolean(agent),
           version: agent?.version || '',
+          targetVersion,
+          installAction: agent ? (upToDate ? 'current' : 'update') : 'install',
           ...publicCompatibility(agent),
           installSupported,
           installErrorCode,
@@ -230,7 +240,10 @@ class AgentInstaller extends EventEmitter {
     this.invalidateDetectionCache()
     const installed = await abortable(this.detectedAgents(), signal)
     if (signal.aborted) throw abortError()
-    if (installed.some(agent => agent.kind === profile.kind)) {
+    const current = installed.find(agent => agent.kind === profile.kind)
+    const targetVersion = recipe.detectedVersion || recipe.version
+    if (current && (current.compatibilityState !== 'compatible'
+      || installedVersionCurrent(current.version, targetVersion))) {
       throw installerError('INSTALL_AGENT_ALREADY_INSTALLED')
     }
     let command = recipe.interpreter || ''

desktop/src/agents/local-agent-readiness.cjs

@@ -1,10 +1,12 @@
 const { execFile } = require('node:child_process')
+const { createHash } = require('node:crypto')
 const fs = require('node:fs')
 const os = require('node:os')
 const path = require('node:path')
 const { promisify } = require('node:util')
 const { prepareCommand, searchPath } = require('./cli/cli-adapters.cjs')
 const { isOpenClawSecretReference } = require('./cli/openclaw-runtime.cjs')
+const { NETWORK_ENV_KEYS, networkEnvironment } = require('./cli/cli-network-environment.cjs')
 
 const execFileAsync = promisify(execFile)
 
@@ -27,13 +29,20 @@ const OPENCLAW_MODEL_INPUTS = new Set(['text', 'image', 'audio', 'video'])
 const MAX_CREDENTIAL_FILE_BYTES = 2 * 1024 * 1024
 const MAX_SHELL_ENV_BYTES = 256 * 1024
 const SHELL_ENV_CACHE_TTL_MS = 30000
+// A login shell that loads nvm/conda/compinit can take several seconds,
+// especially right after boot or under load.
+const SHELL_ENV_TIMEOUT_MS = 10000
+// When the shell does not answer and nothing better is known, retry soon
+// instead of pinning the bare app environment for the whole cache window.
+const SHELL_ENV_FAILURE_TTL_MS = 5000
 const SHELL_ENV_MARKER = '__MELDWORK_NATIVE_ENV_V1__'
+const VERTEX_MODEL_REGION_KEY = /^VERTEX_REGION_[A-Z0-9_]{1,120}$/
 const CREDENTIAL_ENV_KEYS = Object.freeze({
   codex: ['OPENAI_API_KEY'],
   hermes: ['OPENAI_API_KEY', 'OPENROUTER_API_KEY', 'ANTHROPIC_API_KEY'],
   openclaw: ['OPENAI_API_KEY', 'OPENROUTER_API_KEY'],
   workbuddy: ['CODEBUDDY_API_KEY', 'OPENAI_API_KEY', 'ANTHROPIC_API_KEY'],
-  pi: ['PI_API_KEY', 'OPENAI_API_KEY', 'ANTHROPIC_API_KEY'],
+  pi: ['PI_API_KEY', 'OPENAI_API_KEY', 'ANTHROPIC_API_KEY', 'OPENROUTER_API_KEY', 'GEMINI_API_KEY'],
   kimi: ['MOONSHOT_API_KEY', 'KIMI_API_KEY', 'KIMI_MODEL_API_KEY', 'OPENAI_API_KEY'],
   mimo: ['MIMO_API_KEY'],
   claude: ['ANTHROPIC_API_KEY'],
@@ -42,8 +51,12 @@ const CREDENTIAL_ENV_KEYS = Object.freeze({
   opencode: ['OPENAI_API_KEY', 'ANTHROPIC_API_KEY', 'OPENROUTER_API_KEY'],
   opencodereview: ['OCR_LLM_TOKEN', 'OPENAI_API_KEY'],
 })
+const GOOGLE_CLOUD_ENV_KEYS = [
+  'GOOGLE_APPLICATION_CREDENTIALS', 'GOOGLE_CLOUD_PROJECT', 'GCLOUD_PROJECT',
+  'GOOGLE_CLOUD_LOCATION', 'GOOGLE_CLOUD_QUOTA_PROJECT', 'CLOUDSDK_CONFIG',
+]
 const RUNTIME_ENV_KEYS = Object.freeze({
-  codex: ['OPENAI_API_KEY', 'OPENAI_BASE_URL', 'OPENAI_MODEL'],
+  codex: ['OPENAI_API_KEY', 'OPENAI_BASE_URL', 'OPENAI_MODEL', 'CODEX_HOME'],
   hermes: [
     'OPENAI_API_KEY', 'OPENROUTER_API_KEY', 'ANTHROPIC_API_KEY',
     'OPENAI_BASE_URL', 'OPENAI_MODEL', 'HERMES_INFERENCE_PROVIDER', 'HERMES_INFERENCE_MODEL',
@@ -56,6 +69,7 @@ const RUNTIME_ENV_KEYS = Object.freeze({
   pi: [
     'PI_API_KEY', 'PI_BASE_URL', 'PI_MODEL',
     'OPENAI_API_KEY', 'OPENAI_BASE_URL', 'OPENAI_MODEL', 'ANTHROPIC_API_KEY',
+    'OPENROUTER_API_KEY', 'GEMINI_API_KEY', 'PI_CODING_AGENT_DIR',
   ],
   kimi: [
     'MOONSHOT_API_KEY', 'KIMI_API_KEY', 'KIMI_MODEL_API_KEY',
@@ -63,9 +77,25 @@ const RUNTIME_ENV_KEYS = Object.freeze({
     'OPENAI_API_KEY', 'OPENAI_BASE_URL', 'OPENAI_MODEL',
   ],
   mimo: ['MIMO_API_KEY', 'MIMO_BASE_URL', 'MIMO_MODEL'],
-  claude: ['ANTHROPIC_API_KEY', 'ANTHROPIC_BASE_URL', 'ANTHROPIC_MODEL'],
+  claude: [
+    'ANTHROPIC_API_KEY', 'ANTHROPIC_AUTH_TOKEN', 'CLAUDE_CODE_OAUTH_TOKEN',
+    'ANTHROPIC_BASE_URL', 'ANTHROPIC_MODEL', 'CLAUDE_CONFIG_DIR',
+    'ANTHROPIC_DEFAULT_OPUS_MODEL', 'ANTHROPIC_DEFAULT_SONNET_MODEL', 'ANTHROPIC_DEFAULT_HAIKU_MODEL',
+    'CLAUDE_CODE_USE_BEDROCK', 'CLAUDE_CODE_USE_VERTEX', 'CLAUDE_CODE_USE_FOUNDRY', 'CLAUDE_CODE_USE_MANTLE',
+    'AWS_ACCESS_KEY_ID', 'AWS_SECRET_ACCESS_KEY', 'AWS_SESSION_TOKEN', 'AWS_BEARER_TOKEN_BEDROCK',
+    'AWS_PROFILE', 'AWS_REGION', 'AWS_DEFAULT_REGION', 'AWS_CONFIG_FILE', 'AWS_SHARED_CREDENTIALS_FILE',
+    'AWS_ROLE_ARN', 'AWS_ROLE_SESSION_NAME', 'AWS_WEB_IDENTITY_TOKEN_FILE',
+    'ANTHROPIC_BEDROCK_BASE_URL', 'ANTHROPIC_BEDROCK_MANTLE_BASE_URL', 'ANTHROPIC_SMALL_FAST_MODEL_AWS_REGION',
+    ...GOOGLE_CLOUD_ENV_KEYS,
+    'CLOUD_ML_REGION', 'ANTHROPIC_VERTEX_PROJECT_ID', 'ANTHROPIC_VERTEX_BASE_URL',
+    'ANTHROPIC_FOUNDRY_RESOURCE', 'ANTHROPIC_FOUNDRY_BASE_URL',
+    'ANTHROPIC_FOUNDRY_API_KEY', 'ANTHROPIC_FOUNDRY_AUTH_TOKEN',
+    'AZURE_CLIENT_ID', 'AZURE_TENANT_ID', 'AZURE_CLIENT_SECRET', 'AZURE_AUTHORITY_HOST',
+    'AZURE_CLIENT_CERTIFICATE_PATH', 'AZURE_CLIENT_CERTIFICATE_PASSWORD', 'AZURE_FEDERATED_TOKEN_FILE',
+    'AZURE_CONFIG_DIR',
+  ],
   qwen: ['DASHSCOPE_API_KEY', 'OPENAI_API_KEY', 'OPENAI_BASE_URL', 'OPENAI_MODEL'],
-  gemini: ['GEMINI_API_KEY', 'GOOGLE_API_KEY', 'GEMINI_MODEL'],
+  gemini: ['GEMINI_API_KEY', 'GOOGLE_API_KEY', 'GEMINI_MODEL', ...GOOGLE_CLOUD_ENV_KEYS],
   opencode: [
     'OPENAI_API_KEY', 'ANTHROPIC_API_KEY', 'OPENROUTER_API_KEY',
     'OPENAI_BASE_URL', 'OPENAI_MODEL',
@@ -77,6 +107,7 @@ const RUNTIME_ENV_KEYS = Object.freeze({
 })
 const SHELL_ENV_KEYS = Object.freeze([...new Set([
   'PATH',
+  ...NETWORK_ENV_KEYS,
   ...Object.values(RUNTIME_ENV_KEYS).flat(),
 ])])
 let shellEnvironmentCache = null
@@ -94,6 +125,8 @@ function readCredentialFile(filename) {
 function containsCredential(value) {
   if (!value || typeof value !== 'object') return false
   if (Array.isArray(value)) return value.some(containsCredential)
+  if (value.type === 'api_key' && credentialLiteral(value.key)) return true
+  if (value.type === 'oauth' && (credentialLiteral(value.access) || credentialLiteral(value.refresh))) return true
   return Object.entries(value).some(([key, child]) => {
     if (CREDENTIAL_FIELD.test(key) && credentialLiteral(child)) return true
     return child && typeof child === 'object' && containsCredential(child)
@@ -133,8 +166,10 @@ function textContainsCredential(filename) {
 
 function nativeCredentialEnvironment(kind, env = process.env) {
   const result = {}
-  for (const key of RUNTIME_ENV_KEYS[kind] || []) {
-    if (typeof env[key] === 'string' && env[key].trim()) result[key] = env[key]
+  const keys = [...(RUNTIME_ENV_KEYS[kind] || []), ...(kind === 'claude'
+    ? Object.keys(env).filter(key => VERTEX_MODEL_REGION_KEY.test(key)) : [])]
+  for (const key of keys) {
+    if (typeof env[key] === 'string') result[key] = env[key]
   }
   return result
 }
@@ -147,23 +182,41 @@ function nativeCredentialKeyEnvironment(kind, env = process.env) {
   return result
 }
 
-function allowedShellEnvironment(env = process.env) {
+function allowedShellEnvironment(env = process.env, platform = process.platform) {
+  const source = { ...env, ...networkEnvironment(env, platform) }
   const result = {}
-  for (const key of SHELL_ENV_KEYS) {
-    const value = env[key]
-    if (typeof value !== 'string' || !value || value.length > MAX_SHELL_ENV_BYTES) continue
+  for (const key of [...SHELL_ENV_KEYS, ...Object.keys(source).filter(key => VERTEX_MODEL_REGION_KEY.test(key))]) {
+    const value = source[key]
+    if (typeof value !== 'string' || value.length > MAX_SHELL_ENV_BYTES) continue
     result[key] = value
   }
   return result
 }
 
-function nativeShellCommand() {
+function nativeShellCommand(shellName) {
   const keys = SHELL_ENV_KEYS.join(' ')
+  // Enumerate names only; do not dump unrelated values or depend on Electron RunAsNode.
+  const regionNames = shellName === 'zsh'
+    ? 'builtin print -rl -- ${(k)parameters[(I)VERTEX_REGION_*]}'
+    : [
+        'if [ -n "${BASH_VERSION-}" ]; then',
+        '  builtin compgen -e VERTEX_REGION_',
+        'elif [ -x /bin/bash ]; then',
+        "  BASH_ENV= ENV= /bin/bash --noprofile --norc -c 'builtin compgen -e VERTEX_REGION_'",
+        'fi',
+      ].join('\n')
   return [
     `printf '${SHELL_ENV_MARKER}\\0'`,
-    `for __meldwork_key in ${keys}; do`,
+    `for __meldwork_key in ${keys} $(${regionNames}); do`,
+    '  case "$__meldwork_key" in ""|*[!A-Za-z0-9_]*) continue ;; esac',
+    '  case "$__meldwork_key" in VERTEX_REGION_*)',
+    '    __meldwork_model=${__meldwork_key#VERTEX_REGION_}',
+    '    case "$__meldwork_model" in ""|*[!A-Z0-9_]*) continue ;; esac',
+    '    [ ${#__meldwork_model} -le 120 ] || continue ;;',
+    '  esac',
     '  eval "__meldwork_value=\\${$__meldwork_key-}"',
-    '  if [ -n "$__meldwork_value" ]; then',
+    '  eval "__meldwork_set=\\${$__meldwork_key+x}"',
+    '  if [ -n "$__meldwork_set" ]; then',
     "    printf '%s=%s\\0' \"$__meldwork_key\" \"$__meldwork_value\"",
     '  fi',
     'done',
@@ -182,7 +235,8 @@ function parseNativeShellEnvironment(output) {
     if (separator <= 0) continue
     const key = entry.slice(0, separator)
     const value = entry.slice(separator + 1)
-    if (!allowed.has(key) || !value || value.length > MAX_SHELL_ENV_BYTES) continue
+    if ((!allowed.has(key) && !VERTEX_MODEL_REGION_KEY.test(key))
+        || value.length > MAX_SHELL_ENV_BYTES) continue
     result[key] = value
   }
   return result
@@ -192,7 +246,7 @@ async function queryNativeShellEnvironment(options = {}) {
   const source = options.env || process.env
   const platform = options.platform || process.platform
   const home = options.home || os.homedir()
-  const fallback = allowedShellEnvironment(source)
+  const fallback = allowedShellEnvironment(source, platform)
   if (platform === 'win32') return { env: fallback, source: 'process' }
   const shell = String(options.shell || source.SHELL || (platform === 'darwin' ? '/bin/zsh' : '/bin/sh'))
   const shellName = path.basename(shell)
@@ -207,9 +261,9 @@ async function queryNativeShellEnvironment(options = {}) {
   try {
     const result = await execFileFn(
       shell,
-      [shellName === 'sh' ? '-lc' : '-lic', nativeShellCommand()],
+      [shellName === 'sh' ? '-lc' : '-lic', nativeShellCommand(shellName)],
       {
-        timeout: 5000,
+        timeout: SHELL_ENV_TIMEOUT_MS,
         maxBuffer: MAX_SHELL_ENV_BYTES,
         windowsHide: true,
         env: childEnv,
@@ -221,30 +275,60 @@ async function queryNativeShellEnvironment(options = {}) {
       source: Object.keys(loaded).some(key => key !== 'PATH') ? 'native-shell' : 'process',
     }
   } catch {
-    return { env: fallback, source: 'process' }
+    return { env: fallback, source: 'process', failed: true }
   }
 }
 
+function publicShellEnvironment({ env, source }) {
+  return { env, source }
+}
+
+let lastGoodShellEnvironment = null
+
 function resolveNativeShellEnvironment(options = {}) {
   const useCache = options.cache !== false && !options.execFileFn
-  if (!useCache) return queryNativeShellEnvironment(options)
+  if (!useCache) return queryNativeShellEnvironment(options).then(publicShellEnvironment)
   const source = options.env || process.env
   const platform = options.platform || process.platform
   const home = options.home || os.homedir()
   const shell = String(options.shell || source.SHELL || '')
-  const key = [platform, home, shell, String(source.PATH || '')].join('\u0000')
+  const environmentHash = createHash('sha256')
+    .update(JSON.stringify(allowedShellEnvironment(source, platform))).digest('hex')
+  const key = [platform, home, shell, environmentHash].join('\u0000')
   const now = Date.now()
-  if (shellEnvironmentCache?.key === key && shellEnvironmentCache.expiresAt > now) {
+  if (options.refresh !== true && shellEnvironmentCache?.key === key
+      && shellEnvironmentCache.expiresAt > now) {
     return shellEnvironmentCache.promise
   }
-  const promise = queryNativeShellEnvironment(options)
-  shellEnvironmentCache = { key, expiresAt: now + SHELL_ENV_CACHE_TTL_MS, promise }
-  return promise
+  // The window is measured from when the capture finishes, so a capture that
+  // used up its whole timeout is not already stale when it returns.
+  const entry = { key, promise: null, expiresAt: now + SHELL_ENV_CACHE_TTL_MS }
+  entry.promise = queryNativeShellEnvironment(options).then((result) => {
+    entry.expiresAt = Date.now() + SHELL_ENV_CACHE_TTL_MS
+    if (!result.failed) {
+      if (result.source === 'native-shell') lastGoodShellEnvironment = { key, result }
+      return publicShellEnvironment(result)
+    }
+    // The login shell did not answer (slow rc file, timeout, startup error).
+    // The bare app environment has none of the user's PATH entries or API
+    // keys, so an Agent started with it fails with misleading sign-in errors.
+    // Keep using what the shell last produced for this account instead.
+    if (lastGoodShellEnvironment?.key === key) {
+      return publicShellEnvironment(lastGoodShellEnvironment.result)
+    }
+    entry.expiresAt = Date.now() + SHELL_ENV_FAILURE_TTL_MS
+    return publicShellEnvironment(result)
+  })
+  shellEnvironmentCache = entry
+  return entry.promise
 }
 
 function nativeCredentialState(kind, options = {}) {
   const home = options.home || os.homedir()
   const env = options.env || process.env
+  const codexHome = env.CODEX_HOME || path.join(home, '.codex')
+  const claudeHome = env.CLAUDE_CONFIG_DIR || path.join(home, '.claude')
+  const piHome = env.PI_CODING_AGENT_DIR || path.join(home, '.pi', 'agent')
   if (Object.keys(nativeCredentialKeyEnvironment(kind, env)).length) {
     return {
       state: 'ready',
@@ -253,11 +337,13 @@ function nativeCredentialState(kind, options = {}) {
   }
 
   const jsonFiles = {
-    codex: [path.join(home, '.codex', 'auth.json'), path.join(home, '.codex', 'config.json')],
+    codex: [path.join(codexHome, 'auth.json'), path.join(codexHome, 'config.json')],
     hermes: [path.join(home, '.hermes', 'auth.json'), path.join(home, '.hermes', 'config.json')],
     openclaw: [path.join(home, '.openclaw', 'openclaw.json')],
     workbuddy: [path.join(home, '.workbuddy', 'models.json')],
     pi: [
+      path.join(piHome, 'auth.json'),
+      path.join(piHome, 'models.json'),
       path.join(home, '.pi', 'auth.json'),
       path.join(home, '.pi', 'config.json'),
       path.join(home, '.pi-agent', 'auth.json'),
@@ -265,16 +351,16 @@ function nativeCredentialState(kind, options = {}) {
     ],
     kimi: [path.join(home, '.kimi-code', 'credentials', 'kimi-code.json')],
     claude: [
-      path.join(home, '.claude', '.credentials.json'),
-      path.join(home, '.claude', 'settings.json'),
-      path.join(home, '.claude', 'settings.local.json'),
+      path.join(claudeHome, '.credentials.json'),
+      path.join(claudeHome, 'settings.json'),
+      path.join(claudeHome, 'settings.local.json'),
     ],
     qwen: [path.join(home, '.qwen', 'oauth_creds.json'), path.join(home, '.qwen', 'settings.json')],
     gemini: [
       path.join(home, '.gemini', 'oauth_creds.json'),
       path.join(home, '.gemini', 'settings.json'),
     ],
-    opencode: [path.join(home, '.local', 'share', 'opencode', 'auth.json')],
+    opencode: [path.join(env.XDG_DATA_HOME || path.join(home, '.local', 'share'), 'opencode', 'auth.json')],
     opencodereview: [path.join(home, '.opencodereview', 'config.json')],
   }[kind] || []
   if (jsonFiles.some(jsonContainsCredential)) {
@@ -282,7 +368,7 @@ function nativeCredentialState(kind, options = {}) {
   }
 
   const textFiles = {
-    codex: [path.join(home, '.codex', '.env'), path.join(home, '.codex', 'config.toml')],
+    codex: [path.join(codexHome, '.env'), path.join(codexHome, 'config.toml')],
     hermes: [
       path.join(home, '.hermes', '.env'),
       path.join(home, '.hermes', 'config.yaml'),
@@ -302,7 +388,7 @@ function nativeCredentialState(kind, options = {}) {
       path.join(home, '.kimi-code', '.env'),
     ],
     mimo: [path.join(home, '.mimo', '.env'), path.join(home, '.mimocode', '.env')],
-    claude: [path.join(home, '.claude', '.env')],
+    claude: [path.join(claudeHome, '.env')],
     qwen: [path.join(home, '.qwen', '.env')],
     gemini: [path.join(home, '.gemini', '.env')],
     opencode: [
@@ -322,13 +408,14 @@ function probeEnvironment(kind, options = {}) {
   const source = options.env || process.env
   const platform = options.platform || process.platform
   const home = options.home || os.homedir()
-  const env = {}
+  const env = networkEnvironment(source, platform)
   for (const key of [
     'USER', 'LOGNAME', 'TMPDIR', 'TMP', 'TEMP',
     'LANG', 'LANGUAGE', 'LC_ALL', 'LC_CTYPE',
     'USERPROFILE', 'HOMEDRIVE', 'HOMEPATH', 'USERNAME',
     'SYSTEMROOT', 'WINDIR', 'COMSPEC', 'PATHEXT',
     'APPDATA', 'LOCALAPPDATA',
+    'XDG_CONFIG_HOME', 'XDG_DATA_HOME', 'XDG_STATE_HOME', 'XDG_CACHE_HOME',
   ]) {
     if (typeof source[key] === 'string' && source[key]) env[key] = source[key]
   }
@@ -342,6 +429,13 @@ function probeEnvironment(kind, options = {}) {
 function claudeAuthState(output) {
   try {
     const status = JSON.parse(String(output || '').trim())
+    // Cloud status recognizes a provider selection without checking SDK credentials.
+    if (status?.loggedIn === true && status.authMethod === 'third_party') {
+      return { state: 'unknown', source: 'unverified' }
+    }
+    if (status?.loggedIn === true && status.authMethod === 'api_key') {
+      return { state: 'ready', source: 'native-credential' }
+    }
     if (status?.loggedIn === true) return { state: 'ready', source: 'native-auth-status' }
     if (status?.loggedIn === false) return { state: 'missing', source: 'native-auth-status' }
   } catch { /* keep the file-based result */ }
@@ -666,8 +760,7 @@ async function resolveNativeCredentialState(kind, options = {}) {
       ? status.credentialState
       : { state: 'unknown', source: 'native-runtime-unavailable' }
   }
-  if (kind !== 'claude' || !options.executable
-      || Object.keys(nativeCredentialKeyEnvironment(kind, options.env)).length) return current
+  if (kind !== 'claude' || !options.executable) return current
 
   const platform = options.platform || process.platform
   const prepareCommandFn = options.prepareCommandFn || prepareCommand

desktop/src/collaboration/answer-marker-filter.cjs

@@ -0,0 +1,81 @@
+const { V4_RECEIPT_MARKER_PREFIX } = require('./collaboration-receipt.cjs')
+
+// Control markers an Agent appends to its answer. They steer the harness and
+// must never reach the user's screen, even while the answer is still streaming.
+const CONSENSUS_MARKERS = Object.freeze([
+  '[[MELDWORK_CONSENSUS:agree]]',
+  '[[MELDWORK_CONSENSUS:continue]]',
+])
+
+function escapeRegExp(value) {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+}
+
+// Length of the longest buffer suffix that could still grow into a marker.
+function heldSuffixLength(buffer, markers) {
+  let hold = 0
+  for (const marker of markers) {
+    for (let size = Math.min(marker.length - 1, buffer.length); size > hold; size -= 1) {
+      if (buffer.slice(-size).toLowerCase() === marker.slice(0, size).toLowerCase()) {
+        hold = size
+        break
+      }
+    }
+  }
+  return hold
+}
+
+// Streams answer deltas through a marker filter.
+//   plain    nothing to hide: deltas pass through untouched.
+//   auto     hide the legacy consensus markers.
+//   receipt  hide the consensus markers and everything from the start of the
+//            collaboration receipt on (the receipt is parsed from the final text).
+// `push` returns the text that is safe to show now; text that might be the
+// start of a marker is held back until the next delta settles it, and
+// `flush` releases what is left once the answer has finished.
+function createAnswerMarkerFilter({ mode = 'plain' } = {}) {
+  if (mode === 'plain') {
+    return { push: delta => String(delta || ''), flush: () => '' }
+  }
+  const hideReceipt = mode === 'receipt'
+  const markers = hideReceipt
+    ? [...CONSENSUS_MARKERS, V4_RECEIPT_MARKER_PREFIX]
+    : [...CONSENSUS_MARKERS]
+  const pattern = new RegExp(markers.map(escapeRegExp).join('|'), 'i')
+  let buffer = ''
+  let receiptStarted = false
+  return {
+    push(delta) {
+      if (receiptStarted) return ''
+      buffer += String(delta || '')
+      let visible = ''
+      while (buffer && !receiptStarted) {
+        const match = pattern.exec(buffer)
+        if (match) {
+          visible += buffer.slice(0, match.index)
+          buffer = buffer.slice(match.index + match[0].length)
+          if (hideReceipt
+              && match[0].toLowerCase() === V4_RECEIPT_MARKER_PREFIX.toLowerCase()) {
+            receiptStarted = true
+          }
+          continue
+        }
+        const hold = heldSuffixLength(buffer, markers)
+        visible += hold ? buffer.slice(0, -hold) : buffer
+        buffer = hold ? buffer.slice(-hold) : ''
+        break
+      }
+      return visible
+    },
+    flush(finalText = '') {
+      const rest = buffer
+      buffer = ''
+      if (!hideReceipt) return rest
+      // A held fragment is real answer text only when the final answer ends
+      // with it; otherwise it was the start of the (hidden) receipt.
+      return !receiptStarted && rest && String(finalText).endsWith(rest) ? rest : ''
+    },
+  }
+}
+
+module.exports = { CONSENSUS_MARKERS, createAnswerMarkerFilter }

desktop/src/collaboration/collaboration-receipt.cjs

@@ -0,0 +1,295 @@
+const { canonicalJson } = require('./context-pack-records.cjs')
+const { parseLenientTaskDecision } = require('./task-decision.cjs')
+
+const SHA256 = /^[a-f0-9]{64}$/
+
+const V4_AGENT_PHASES = new Set([
+  'proposal', 'discussion', 'challenge', 'work', 'synthesis', 'verification',
+])
+const V4_RECEIPT_MARKER_PREFIX = '[[MELDWORK_COLLABORATION'
+const V4_RECEIPT_MARKER = /\s*\[\[MELDWORK_COLLABORATION(?:_JSON)?:([\s\S]*?)\]\]/i
+const V4_RECEIPT_BLOCK = /\s*\[\[MELDWORK_COLLABORATION(?:_JSON)?\]\]([\s\S]*?)\[\[\/MELDWORK_COLLABORATION(?:_JSON)?\]\]/i
+const LEGACY_CONSENSUS_MARKER = /^\s*\[\[MELDWORK_CONSENSUS:(?:agree|continue)\]\]\s*$/gim
+const V4_RECEIPT_MAX_TEXT = 800
+const V4_RECEIPT_MAX_ITEMS = 16
+const V4_RECEIPT_MAX_BYTES = 16 * 1024
+
+function isV4InternalProtocolCommentary(value) {
+  const text = String(value || '').replace(/\s+/g, ' ').trim()
+  if (!text || text.length > 4000 || /\[\[MELDWORK_COLLABORATION/i.test(text)) return false
+  const mentionsControl = /(?:\breceipt\b|structured marker|MELDWORK framework|ExitPlanMode|\u7ed3\u6784\u5316\s*receipt|receipt\s*\u6807\u8bb0)/i.test(text)
+  const describesInternalProcess = /(?:coordination\/self-introduction|not a code implementation task|no plan file|implementation planning|plan mode|waiting for (?:the )?(?:next|subsequent).*(?:phase|instruction)|\u534f\u4f5c\u534f\u8c03\u7c7b\u6587\u672c\u4ea7\u51fa|\u975e\u4ee3\u7801\u5b9e\u73b0\u4efb\u52a1|\u6267\u884c\u8ba1\u5212\u6d41\u7a0b|\u4e0d\u8c03\u7528\s*ExitPlanMode|\u672c\u8f6e\u4ea4\u4ed8|\u7b49\u5f85\u540e\u7eed.*(?:\u9636\u6bb5|\u6307\u4ee4))/i.test(text)
+  return mentionsControl && describesInternalProcess
+}
+
+function stripV4InternalProtocolTail(value) {
+  const paragraphs = String(value || '').trim().split(/\n{2,}/)
+  while (paragraphs.length && isV4InternalProtocolCommentary(paragraphs.at(-1))) {
+    paragraphs.pop()
+  }
+  return paragraphs.join('\n\n').trim()
+}
+
+function boundedReceiptText(value, limit = V4_RECEIPT_MAX_TEXT) {
+  if (typeof value !== 'string' || value.trim().length > limit) return ''
+  const text = cleanReceiptText(value, limit)
+  return text && !text.includes('\u0000') ? text : ''
+}
+
+function boundedReceiptList(value) {
+  if (value == null) return []
+  if (!Array.isArray(value) || value.length > V4_RECEIPT_MAX_ITEMS) {
+    throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  }
+  const result = value.map(item => boundedReceiptText(item))
+  if (result.some(item => !item)) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  return result
+}
+
+// A decision that is still unreadable after normalisation is treated as no
+// decision: the orchestrator then asks the owner again instead of failing
+// the Agent and removing it from the discussion.
+function leniencyTaskDecision(value, summary) {
+  try {
+    return { taskDecision: parseLenientTaskDecision(value, { fallbackReason: summary }) }
+  } catch {
+    return {}
+  }
+}
+
+function parseV4CollaborationReceiptStrict(result, required = false, expectedPhase = '', options = {}) {
+  const rawText = typeof result?.text === 'string' ? result.text : ''
+  if (!required) return { text: rawText, collaboration: null }
+  let text = rawText
+  let receipt = result?.collaboration
+  // Natural discussion turns are free-form answers; a slightly malformed
+  // receipt there must not fail the turn and end the discussion. Strict
+  // phased-V4 receipts keep failing closed.
+  const lenient = expectedPhase === 'discussion'
+  let trailingVisibleText = ''
+  let unreadableReceipt = false
+  const block = rawText.match(V4_RECEIPT_BLOCK) || rawText.match(V4_RECEIPT_MARKER)
+  if (block) {
+    const markerStart = Number(block.index) || 0
+    const suffix = rawText.slice(markerStart + block[0].length).trim()
+    // A closing code fence after the receipt is formatting, not content.
+    if (suffix && !isV4InternalProtocolCommentary(suffix) && !/^[\s`~]*$/.test(suffix)) {
+      if (!lenient) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+      trailingVisibleText = suffix
+    }
+    const encoded = String(block[1] || '').trim()
+    if (!encoded || Buffer.byteLength(encoded) > V4_RECEIPT_MAX_BYTES) {
+      if (!lenient) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+      unreadableReceipt = true
+    }
+    if (!unreadableReceipt) {
+      try {
+        let parsed
+        try {
+          parsed = JSON.parse(encoded)
+        } catch (error) {
+          if (!lenient) throw error
+          // Common hand-written slips: smart quotes and trailing commas.
+          parsed = JSON.parse(encoded
+            .replace(/[\u201c\u201d]/g, '"')
+            .replace(/[\u2018\u2019]/g, "'")
+            .replace(/,\s*([}\]])/g, '$1'))
+        }
+        if (receipt != null && canonicalJson(receipt) !== canonicalJson(parsed)) {
+          throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_MISMATCH')
+        }
+        receipt = parsed
+      } catch (error) {
+        if (error?.message === 'LOCAL_RUN_COLLABORATION_RECEIPT_MISMATCH') throw error
+        if (!lenient) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+        unreadableReceipt = true
+      }
+    }
+    text = rawText.slice(0, markerStart) + (trailingVisibleText ? `\n\n${trailingVisibleText}` : '')
+  } else if (/\[\[MELDWORK_COLLABORATION/i.test(rawText)) {
+    if (!lenient) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+    unreadableReceipt = true
+    text = rawText.slice(0, rawText.search(/\[\[MELDWORK_COLLABORATION/i))
+  }
+  if (unreadableReceipt) receipt = null
+  if (lenient && receipt && typeof receipt === 'object' && !Array.isArray(receipt)) {
+    // Keep only discussion fields and fill a missing summary, so an extra key
+    // or a long summary does not discard the Agent's whole contribution.
+    const decisionReason = typeof receipt.taskDecision?.reason === 'string'
+      ? receipt.taskDecision.reason
+      : ''
+    const summary = typeof receipt.summary === 'string' && receipt.summary.trim()
+      ? receipt.summary.trim().slice(0, V4_RECEIPT_MAX_TEXT)
+      : (decisionReason.trim().slice(0, V4_RECEIPT_MAX_TEXT) || 'Discussion contribution')
+    receipt = {
+      version: 1,
+      phase: 'discussion',
+      summary,
+      ...(receipt.taskDecision != null ? { taskDecision: receipt.taskDecision } : {}),
+    }
+  }
+  text = stripV4InternalProtocolTail(
+    text.replace(LEGACY_CONSENSUS_MARKER, '').replace(/\n{3,}/g, '\n\n'),
+  )
+  if (receipt == null && options.allowMissingV4Receipt === true && text) {
+    return { text, collaboration: null }
+  }
+  if (receipt == null && expectedPhase === 'proposal'
+      && options.allowMissingProposalReceipt === true && text) {
+    receipt = {
+      version: 1,
+      phase: 'proposal',
+      summary: 'The Agent returned a visible proposal without a structured receipt.',
+      capabilities: ['Delivered a user-facing proposal'],
+      intendedWork: ['Addressed the current user task'],
+      deliverables: ['Visible Agent response'],
+      dependencies: [],
+    }
+  }
+  if (receipt == null) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_REQUIRED')
+  if (!receipt || typeof receipt !== 'object' || Array.isArray(receipt)) {
+    throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  }
+  try {
+    if (Buffer.byteLength(canonicalJson(receipt)) > V4_RECEIPT_MAX_BYTES) {
+      throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+    }
+  } catch (error) {
+    if (error?.message === 'LOCAL_RUN_COLLABORATION_RECEIPT_INVALID') throw error
+    throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  }
+  const phase = String(expectedPhase || '')
+  const summary = boundedReceiptText(receipt.summary)
+  const allowed = new Set([
+    'version', 'phase', 'summary',
+    ...(phase === 'discussion' ? ['taskDecision'] : []),
+    ...(['challenge', 'verification'].includes(phase) ? ['verdict'] : []),
+    ...(phase === 'proposal'
+      ? ['capabilities', 'intendedWork', 'deliverables', 'dependencies'] : []),
+    ...(phase === 'challenge'
+      ? [
+          'proposedAssignments', 'finalizerKind', 'verifierKinds',
+          'supportedPlanHash', 'agreeToPlan',
+        ] : []),
+    ...(phase === 'work' ? ['workItemId', 'deliverables'] : []),
+    ...(phase === 'synthesis' ? ['resolvedIssueIds'] : []),
+  ])
+  if (!V4_AGENT_PHASES.has(phase) || !summary
+      || Object.keys(receipt).some(key => !allowed.has(key))
+      || (receipt.version != null && receipt.version !== 1)
+      || (receipt.phase != null && receipt.phase !== phase)) {
+    throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  }
+  if (['challenge', 'verification'].includes(phase)
+      && !['support', 'contradict'].includes(receipt.verdict)) {
+    throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  }
+  const resolvedIssueIds = phase === 'synthesis'
+    ? boundedReceiptList(receipt.resolvedIssueIds)
+    : null
+  if (phase === 'synthesis' && !Array.isArray(receipt.resolvedIssueIds)) {
+    throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  }
+  if (phase === 'challenge' && receipt.supportedPlanHash != null
+      && !SHA256.test(String(receipt.supportedPlanHash))) {
+    throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+  }
+  const proposalList = (field, requireItem = false) => {
+    if (!Array.isArray(receipt[field])) {
+      throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+    }
+    const list = boundedReceiptList(receipt[field])
+    if (requireItem && list.length === 0) {
+      throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+    }
+    return list
+  }
+  const boundedList = value => {
+    if (value == null) return []
+    if (!Array.isArray(value) || value.length > V4_RECEIPT_MAX_ITEMS) {
+      throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+    }
+    const result = value.map(item => boundedReceiptText(item))
+    if (result.some(item => !item)) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+    return result
+  }
+  const normalized = {
+    version: 1,
+    phase,
+    summary,
+    ...(phase === 'discussion' && receipt.taskDecision != null
+      ? leniencyTaskDecision(receipt.taskDecision, summary) : {}),
+    ...(['challenge', 'verification'].includes(phase) ? { verdict: receipt.verdict } : {}),
+    ...(phase === 'proposal' ? {
+      capabilities: proposalList('capabilities', true),
+      intendedWork: proposalList('intendedWork', true),
+      deliverables: proposalList('deliverables', true),
+      dependencies: proposalList('dependencies'),
+    } : {}),
+    ...(phase === 'challenge' ? {
+      ...(Array.isArray(receipt.proposedAssignments)
+        ? {
+            proposedAssignments: receipt.proposedAssignments.map(assignment => (
+              assignment && typeof assignment === 'object' && !Array.isArray(assignment)
+                && assignment.role === 'finalizer'
+                ? { ...assignment, role: 'integrator' }
+                : assignment
+            )),
+          }
+        : {}),
+      ...(receipt.finalizerKind ? { finalizerKind: boundedReceiptText(receipt.finalizerKind, 120) } : {}),
+      ...(Array.isArray(receipt.verifierKinds) ? { verifierKinds: boundedList(receipt.verifierKinds) } : {}),
+      ...(receipt.supportedPlanHash ? { supportedPlanHash: receipt.supportedPlanHash } : {}),
+      ...(typeof receipt.agreeToPlan === 'boolean' ? { agreeToPlan: receipt.agreeToPlan } : {}),
+    } : {}),
+    ...(phase === 'work' ? {
+      ...(receipt.workItemId ? { workItemId: boundedReceiptText(receipt.workItemId, 120) } : {}),
+      deliverables: boundedList(receipt.deliverables),
+    } : {}),
+    ...(phase === 'synthesis' ? { resolvedIssueIds } : {}),
+  }
+  return { text, collaboration: Object.freeze(JSON.parse(canonicalJson(normalized))) }
+}
+
+function parseV4CollaborationReceipt(result, required = false, expectedPhase = '', options = {}) {
+  try {
+    return parseV4CollaborationReceiptStrict(result, required, expectedPhase, options)
+  } catch (error) {
+    const code = String(error?.message || '')
+    if (expectedPhase !== 'proposal'
+        || options.allowMissingProposalReceipt !== true
+        || !['LOCAL_RUN_COLLABORATION_RECEIPT_INVALID', 'LOCAL_RUN_COLLABORATION_RECEIPT_REQUIRED']
+          .includes(code)) {
+      throw error
+    }
+    const rawText = typeof result?.text === 'string' ? result.text : ''
+    const markerIndex = rawText.search(/\[\[MELDWORK_COLLABORATION/i)
+    const visibleText = (markerIndex >= 0 ? rawText.slice(0, markerIndex) : rawText)
+      .replace(LEGACY_CONSENSUS_MARKER, '')
+      .replace(/\n{3,}/g, '\n\n')
+      .trim()
+    if (!visibleText) throw error
+    return {
+      text: visibleText,
+      collaboration: Object.freeze({
+        version: 1,
+        phase: 'proposal',
+        summary: 'The Agent returned a visible proposal without a valid structured receipt.',
+        capabilities: ['Delivered a user-facing proposal'],
+        intendedWork: ['Addressed the current user task'],
+        deliverables: ['Visible Agent response'],
+        dependencies: [],
+      }),
+    }
+  }
+}
+
+function cleanReceiptText(value, limit) {
+  return String(value || '').trim().slice(0, limit)
+}
+
+module.exports = {
+  V4_AGENT_PHASES,
+  V4_RECEIPT_MARKER_PREFIX,
+  parseV4CollaborationReceipt,
+}

desktop/src/collaboration/collaboration-records.cjs

@@ -533,6 +533,37 @@ function publicCollaborationText(value, limit = MAX_TEXT_CHARS) {
     .slice(0, limit)
 }
 
+const PRIVATE_FILESYSTEM_ROOT = /^\/(?:Users|home|private|var|tmp|etc|opt|Volumes|root|usr|System|Library|Applications|mnt|srv)(?:\/|$)/
+
+function peerPathText(candidate, workspaceRoot) {
+  const trailing = candidate.match(/[.,;:!?]+$/)?.[0] || ''
+  const clean = trailing ? candidate.slice(0, -trailing.length) : candidate
+  if (workspaceRoot && (clean === workspaceRoot || clean.startsWith(`${workspaceRoot}/`))) {
+    return `${clean === workspaceRoot ? '.' : clean.slice(workspaceRoot.length + 1)}${trailing}`
+  }
+  if (PRIVATE_FILESYSTEM_ROOT.test(clean)) return `[path]${trailing}`
+  // Routes and commands such as /api/users or /login are not local files.
+  return candidate
+}
+
+// Peer-visible text for natural discussion transcripts. Unlike
+// publicCollaborationText, files inside the shared workspace stay readable as
+// workspace-relative paths and non-filesystem tokens (API routes, slash
+// commands) are kept; other absolute local paths and secrets stay redacted.
+function peerCollaborationText(value, options = {}) {
+  const workspaceRoot = typeof options.workspaceRoot === 'string'
+    ? options.workspaceRoot.replace(/\/+$/, '')
+    : ''
+  return redactSecrets(String(value || ''))
+    .replace(/\bfile:\/\/(\/[^\s"'`<>]+)/giu, (_match, filePath) => peerPathText(filePath, workspaceRoot))
+    .replace(/(^|[\s("'`])(\/(?!\/)[^\s"'`<>)]*)/gmu, (_match, lead, candidate) => (
+      `${lead}${peerPathText(candidate, workspaceRoot)}`
+    ))
+    .replace(/\b[A-Za-z]:\\(?:[^\s"'`<>]+\\)*[^\s"'`<>]*/gu, '[path]')
+    .replace(/[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/g, '')
+    .trim()
+}
+
 function collaborationPackageText(handoff, entries) {
   const lines = [
     'MELDWORK_COLLABORATION_PACKAGE_V1',
@@ -557,6 +588,7 @@ function collaborationPackageText(handoff, entries) {
 }
 
 module.exports = {
+  peerCollaborationText,
   BLACKBOARD_ENTRY_VERSION,
   COLLABORATION_STATE_VERSION,
   HANDOFF_VERSION,

desktop/src/collaboration/orchestration-v4-records.cjs

@@ -1,4 +1,5 @@
 const crypto = require('node:crypto')
+const { parseTaskDecision } = require('./task-decision.cjs')
 
 const { normalizeContentBlobRef } = require('../attachments/content-blob-store.cjs')
 const { canonicalJson } = require('./context-pack-records.cjs')
@@ -9,7 +10,7 @@ const V4_TEMPLATES = new Set(['concurrent-batch', 'discussion'])
 const V4_WORKFLOWS = new Set(['manual', 'auto', 'concurrent-batch', 'discussion'])
 const V4_PHASES = new Set([
   'prepare', 'dispatch', 'running', 'reconcile',
-  'proposal', 'challenge', 'coordination', 'work', 'synthesis', 'verification',
+  'proposal', 'discussion', 'challenge', 'coordination', 'work', 'synthesis', 'verification',
   'commit', 'committed', 'completed', 'failed', 'stopped', 'human-gate',
 ])
 const V4_SLOT_STATUSES = new Set([
@@ -60,7 +61,7 @@ const MESSAGE_ID = /^message-[a-f0-9]{64}$/u
 const EVIDENCE_ID = /^evidence-[a-f0-9]{64}$/u
 
 const ORCHESTRATION_FIELDS = new Set([
-  'version', 'workflow', 'template', 'phase', 'batchId', 'round',
+  'version', 'workflow', 'template', 'phase', 'discussionStyle', 'discussionWriterKind', 'batchId', 'round',
   'currentKind', 'currentKinds', 'pendingKinds', 'activeKinds',
   'successfulKinds', 'agreementKinds', 'attachmentRecipients',
   'totalSuccesses', 'terminalFailureOccurred', 'collaboration', 'taskGraph',
@@ -829,7 +830,7 @@ function normalizeDeliveryState(input, targetKinds, index) {
   if (!isRecord(input) || !exactFields(input, [
     'recipientKind', 'sessionRefHash', 'sessionProvenanceHash',
     'sourceAgentKind', 'sourcePhase', 'watermark', 'snapshotHash', 'operationId',
-    'packageHash', 'deliveryId', 'status', 'updatedAt',
+    'packageHash', 'deliveryId', 'status', 'updatedAt', 'sourceMessages',
   ])) fail('ORCHESTRATION_V4_DELIVERY_STATE_INVALID')
   const recipientKind = cleanId(input.recipientKind)
   const sourceAgentKind = cleanId(input.sourceAgentKind)
@@ -841,6 +842,20 @@ function normalizeDeliveryState(input, targetKinds, index) {
   const packageHash = String(input.packageHash || '')
   const deliveryId = cleanId(input.deliveryId)
   const status = String(input.status || '')
+  let sourceMessages
+  if (hasOwn(input, 'sourceMessages')) {
+    if (!Array.isArray(input.sourceMessages) || !input.sourceMessages.length
+        || input.sourceMessages.length > 100) fail('ORCHESTRATION_V4_DELIVERY_STATE_INVALID')
+    sourceMessages = input.sourceMessages.map(message => {
+      if (!isRecord(message) || !exactFields(message, ['id', 'hash'])
+          || !cleanId(message.id) || cleanId(message.id) !== message.id
+          || typeof message.hash !== 'string' || !SHA256.test(message.hash)) fail('ORCHESTRATION_V4_DELIVERY_STATE_INVALID')
+      return { id: message.id, hash: message.hash }
+    })
+    if (new Set(sourceMessages.map(message => message.id)).size !== sourceMessages.length) {
+      fail('ORCHESTRATION_V4_DELIVERY_STATE_INVALID')
+    }
+  }
   if (!recipientKind || !sourceAgentKind || !targetKinds.includes(recipientKind)
       || !targetKinds.includes(sourceAgentKind) || !V4_PHASES.has(sourcePhase)
       || !Number.isSafeInteger(input.watermark) || input.watermark < 0
@@ -861,6 +876,7 @@ function normalizeDeliveryState(input, targetKinds, index) {
     packageHash,
     deliveryId,
     status,
+    ...(sourceMessages ? { sourceMessages } : {}),
     updatedAt: safeTimestamp(input.updatedAt, index),
   }
 }
@@ -1601,10 +1617,18 @@ function validateV4Relations(orchestration, targetKinds) {
       || orchestration.synthesisBinding?.writerKind
       || orchestration.coordinationPlan?.finalizerKind
       || orchestration.commitState.writerKind
+      || orchestration.discussionWriterKind
     if (writableSlots.length > 1 || writableSlots.some(slot => (
       slot.agentKind !== effectiveWriterKind
-        || !['synthesis', 'human-gate'].includes(orchestration.phase)
+        || !['discussion', 'synthesis', 'human-gate'].includes(orchestration.phase)
+        || (orchestration.phase === 'discussion' && !orchestration.discussionWriterKind)
     ))) fail('ORCHESTRATION_V4_PERMISSION_INVALID')
+    for (const assignment of orchestration.phase === 'discussion' ? orchestration.plan.assignments : []) {
+      const slot = orchestration.slots.find(item => item.slotId === assignment.slotId)
+      if (slot && assignment.readOnly !== (slot.permission !== 'workspace-write')) {
+        fail('ORCHESTRATION_V4_PERMISSION_INVALID')
+      }
+    }
   }
   const watermarkKeys = orchestration.deliveryWatermarks.map(item => `${item.agentKind}\u0000${item.phase}`)
   if (new Set(watermarkKeys).size !== watermarkKeys.length) {
@@ -1784,6 +1808,14 @@ function parseOrchestrationV4(input, options = {}) {
   }
   const { workflow, template } = normalizeWorkflow(input.workflow, input.template)
   const phase = phaseForTemplate(template, String(input.phase || ''))
+  const discussionStyle = hasOwn(input, 'discussionStyle')
+    ? String(input.discussionStyle || '')
+    : ''
+  if (discussionStyle && (
+    workflow !== 'auto'
+    || template !== 'discussion'
+    || !['sequential', 'agent-led'].includes(discussionStyle)
+  )) fail('ORCHESTRATION_V4_SCHEMA_INVALID')
   const batchId = cleanId(input.batchId)
   const snapshotHash = String(input.snapshotHash || '')
   if (!batchId || !SHA256.test(snapshotHash)) fail('ORCHESTRATION_V4_SCHEMA_INVALID')
@@ -1891,6 +1923,15 @@ function parseOrchestrationV4(input, options = {}) {
   if (coordinationPlan && (workflow !== 'auto' || template !== 'discussion')) {
     fail('ORCHESTRATION_V4_COORDINATION_PLAN_INVALID')
   }
+  const discussionWriterKind = hasOwn(input, 'discussionWriterKind') ? cleanId(input.discussionWriterKind) : ''
+  if (hasOwn(input, 'discussionWriterKind') && (!discussionWriterKind
+      || input.discussionWriterKind !== discussionWriterKind
+      || !effectiveKinds.includes(discussionWriterKind)
+      || workflow !== 'auto' || template !== 'discussion'
+      || !['proposal', 'discussion', 'stopped', 'failed'].includes(phase)
+      || synthesisBinding || coordinationPlan || commitState.writerKind)) {
+    fail('ORCHESTRATION_V4_PERMISSION_INVALID')
+  }
   if (coordinationPlan) {
     validateCoordinationAgreement(
       coordinationPlan, slots, challengeBindings, participantKinds, snapshotHash,
@@ -2017,6 +2058,8 @@ function parseOrchestrationV4(input, options = {}) {
     workflow,
     template,
     phase,
+    ...(discussionStyle ? { discussionStyle } : {}),
+    ...(discussionWriterKind ? { discussionWriterKind } : {}),
     batchId,
     currentKinds,
     activeKinds,
@@ -2161,7 +2204,7 @@ function parseCollaborationControlBlock(input, options = {}) {
     'claims', 'findings', 'refs', 'writer', 'writerKind', 'batchId', 'entryIds',
     'capabilities', 'intendedWork', 'deliverables', 'dependencies',
     'proposedAssignments', 'finalizerKind', 'verifierKinds', 'supportedPlanHash',
-    'agreeToPlan', 'workItemId',
+    'agreeToPlan', 'workItemId', 'taskDecision',
   ]) || input.version !== 1) {
     fail('COLLABORATION_CONTROL_BLOCK_INVALID')
   }
@@ -2171,6 +2214,8 @@ function parseCollaborationControlBlock(input, options = {}) {
     fail('COLLABORATION_CONTROL_BLOCK_INVALID')
   }
   const summary = safeText(input.summary, MAX_V4_SUMMARY_CHARS, true)
+  const taskDecision = input.taskDecision == null ? null : parseTaskDecision(input.taskDecision)
+  if (taskDecision && phase !== 'discussion') fail('COLLABORATION_CONTROL_BLOCK_INVALID')
   const boundedList = (value, limit = MAX_V4_REFS, textLimit = MAX_V4_TEXT_CHARS) => {
     if (value == null) return []
     if (!Array.isArray(value) || value.length > limit) fail('COLLABORATION_CONTROL_BLOCK_INVALID')
@@ -2301,6 +2346,7 @@ function parseCollaborationControlBlock(input, options = {}) {
     status,
     summary,
     conclusion,
+    ...(taskDecision ? { taskDecision } : {}),
     artifactIds,
     evidenceIds,
     findingIds,
@@ -2339,6 +2385,7 @@ function collaborationChars(block) {
     ...(Array.isArray(block.intendedWork) ? block.intendedWork : []),
     ...(Array.isArray(block.deliverables) ? block.deliverables : []),
     ...(Array.isArray(block.dependencies) ? block.dependencies : []),
+    ...(block.taskDecision ? [block.taskDecision.reason, ...block.taskDecision.deliverables] : []),
     ...(Array.isArray(block.proposedAssignments)
       ? block.proposedAssignments.flatMap(item => [item.objective, item.expectedOutput])
       : []),
@@ -2547,6 +2594,7 @@ function createCollaborationReceipt(input, options = {}) {
     ...(input.supportedPlanHash ? { supportedPlanHash: input.supportedPlanHash } : {}),
     ...(typeof input.agreeToPlan === 'boolean' ? { agreeToPlan: input.agreeToPlan } : {}),
     ...(input.workItemId ? { workItemId: input.workItemId } : {}),
+    ...(input.taskDecision ? { taskDecision: input.taskDecision } : {}),
     deliveryWatermark: input.deliveryWatermark || 0,
     snapshotHash: input.snapshotHash,
     ...(typeof input.writer === 'boolean' ? { writer: input.writer } : {}),

desktop/src/collaboration/task-decision.cjs

@@ -0,0 +1,88 @@
+const { redactSecrets } = require('../security/secret-redaction.cjs')
+
+function parseTaskDecision(input) {
+  const invalid = () => { throw new Error('LOCAL_RUN_TASK_DECISION_INVALID') }
+  if (!input || typeof input !== 'object' || Array.isArray(input)
+      || Object.keys(input).some(key => !['status', 'reason', 'deliverables', 'handoffTo', 'nextKinds'].includes(key))
+      || !['completed', 'continue', 'blocked', 'needs-human'].includes(input.status)) invalid()
+  const text = (value, limit) => {
+    if (typeof value !== 'string' || !value.trim() || value.length > limit
+        || /[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/.test(value)) invalid()
+    return redactSecrets(value.trim())
+  }
+  const reason = text(input.reason, 1600)
+  if (!Array.isArray(input.deliverables) || input.deliverables.length > 16) invalid()
+  const deliverables = input.deliverables.map(value => text(value, 800))
+  if (input.status === 'completed' && !deliverables.length) invalid()
+  if (input.handoffTo != null && (input.status !== 'continue'
+      || typeof input.handoffTo !== 'string'
+      || !/^[a-z0-9][a-z0-9-]{0,79}$/.test(input.handoffTo))) invalid()
+  if (input.nextKinds != null && (!Array.isArray(input.nextKinds)
+      || input.nextKinds.length > 32
+      || input.nextKinds.some(kind => typeof kind !== 'string'
+        || !/^[a-z0-9][a-z0-9-]{0,79}$/i.test(kind))
+      || (input.nextKinds.length > 0 && input.status !== 'continue'))) invalid()
+  const nextKinds = input.nextKinds?.map(kind => kind.toLowerCase())
+  if (nextKinds && new Set(nextKinds).size !== nextKinds.length) invalid()
+  return {
+    status: input.status, reason, deliverables,
+    ...(input.handoffTo != null ? { handoffTo: input.handoffTo } : {}),
+    ...(nextKinds ? { nextKinds } : {}),
+  }
+}
+
+const STATUS_SYNONYMS = new Map([
+  ['completed', 'completed'], ['complete', 'completed'], ['done', 'completed'],
+  ['finished', 'completed'], ['success', 'completed'], ['succeeded', 'completed'],
+  ['continue', 'continue'], ['continuing', 'continue'], ['in-progress', 'continue'],
+  ['in_progress', 'continue'], ['ongoing', 'continue'], ['pending', 'continue'],
+  ['blocked', 'blocked'], ['blocker', 'blocked'],
+  ['needs-human', 'needs-human'], ['needs_human', 'needs-human'],
+  ['needs human', 'needs-human'], ['needshuman', 'needs-human'], ['human', 'needs-human'],
+])
+
+function agentKindToken(value) {
+  return typeof value === 'string'
+    ? value.trim().replace(/^@+/, '').toLowerCase()
+    : ''
+}
+
+// Agents write task decisions by hand at the end of free text. Small slips
+// (capitalised status, "done", an @handle, a single string instead of a list,
+// a completed answer without a separate file) used to fail the whole turn
+// and silently end the discussion. Normalise them first; the strict parser
+// still validates the result.
+function normalizeTaskDecision(input, options = {}) {
+  if (!input || typeof input !== 'object' || Array.isArray(input)) return input
+  const status = STATUS_SYNONYMS.get(String(input.status || '').trim().toLowerCase())
+    || String(input.status || '')
+  const textList = value => (Array.isArray(value) ? value : (value == null ? [] : [value]))
+    .filter(item => typeof item === 'string' && item.trim())
+    .map(item => item.trim().slice(0, 800))
+    .slice(0, 16)
+  const reason = typeof input.reason === 'string' && input.reason.trim()
+    ? input.reason.trim().slice(0, 1600)
+    : String(options.fallbackReason || '').trim().slice(0, 1600) || 'No reason given.'
+  let deliverables = textList(input.deliverables)
+  // A completed answer is itself the deliverable when nothing else was named.
+  if (status === 'completed' && !deliverables.length) deliverables = ['The answer in this message']
+  const nextKinds = [...new Set(textList(input.nextKinds).map(agentKindToken).filter(Boolean))]
+  const handoffTo = agentKindToken(input.handoffTo)
+  return {
+    status,
+    reason,
+    deliverables,
+    ...(status === 'continue' && handoffTo ? { handoffTo } : {}),
+    // Keep an explicit list (even an empty one) so the stored decision says
+    // what the Agent wrote; peers can only be requested while continuing.
+    ...(input.nextKinds != null
+      ? { nextKinds: status === 'continue' ? nextKinds.slice(0, 32) : [] }
+      : {}),
+  }
+}
+
+function parseLenientTaskDecision(input, options = {}) {
+  return parseTaskDecision(normalizeTaskDecision(input, options))
+}
+
+module.exports = { normalizeTaskDecision, parseLenientTaskDecision, parseTaskDecision }

desktop/src/gates/human-gate-coordinator.cjs

@@ -208,9 +208,11 @@ class HumanGateCoordinator {
       status,
       optionId: option.optionId,
       actorId: input.actorId || 'local-user',
-      decidedAt: input.decidedAt || this.now(),
+      decidedAt: input.decidedAt || current.decision?.decidedAt || this.now(),
       ...(response ? { response } : {}),
     })
+    // Recovery owns recorded decisions; repeated submissions must not dispatch them again.
+    if (current.status !== 'pending') return publicHumanGate(record)
     this.clearExpiry(record.gateId)
     try {
       this.onResumed(record, record.decision)

desktop/src/main.cjs

@@ -1,6 +1,8 @@
 const electron = require('electron')
 const { app, BrowserWindow, dialog, ipcMain, Menu, Notification, protocol, shell } = electron
+const fs = require('node:fs')
 const path = require('node:path')
+const { redactSecrets } = require('./security/secret-redaction.cjs')
 const { AttachmentStore } = require('./attachments/attachment-store.cjs')
 const { ContentBlobStore } = require('./attachments/content-blob-store.cjs')
 const { ContextPackStore } = require('./collaboration/context-pack-store.cjs')
@@ -25,6 +27,7 @@ const {
   shutdownAcpSessionRuntime,
 } = require('./agents/cli/cli-adapters.cjs')
 const { AgentInstaller } = require('./agents/installer/agent-installer.cjs')
+const { networkEnvironment } = require('./agents/cli/cli-network-environment.cjs')
 const { AgentConnectorInstanceStore } = require('./agents/connectors/agent-connector-instance-store.cjs')
 const { LocalAgentConnectors } = require('./agents/connectors/agent-connector-local.cjs')
 const { AgentConnectorPackageStore } = require('./agents/connectors/agent-connector-package-store.cjs')
@@ -47,8 +50,6 @@ const {
   resolveNativeShellEnvironment,
 } = require('./agents/local-agent-readiness.cjs')
 const { LocalWorkspace } = require('./workspace/local-workspace.cjs')
-const { MediaGenerationRuntime } = require('./media/media-generation-runtime.cjs')
-const { resolveMediaProvider } = require('./media/media-provider-resolution.cjs')
 const { registerDesktopIpc } = require('./shell/main-ipc.cjs')
 const {
   isAllowedExternalUrl,
@@ -70,26 +71,37 @@ const {
   resolveKnowledgeBaseSources,
 } = require('./knowledge/local-knowledge-base.cjs')
 const { LocalKnowledgeConnectors } = require('./knowledge/local-knowledge-connectors.cjs')
-const { ProviderStore } = require('./providers/provider-store.cjs')
+const { nativeOpenClawOptions } = require('./agents/cli/openclaw-runtime.cjs')
 const { migrateLegacyProductData } = require('./product-data-migration.cjs')
-const {
-  EXTERNAL_PROVIDER_KINDS,
-  providerAgentKind,
-  providerOptionsFor,
-} = require('./providers/provider-options.cjs')
 const LOCAL_AGENT_KINDS = new Set([
   'codex', 'hermes', 'openclaw', 'workbuddy', 'pi', 'kimi', 'mimo', 'claude', 'gemini', 'opencode', 'qwen',
   'opencodereview',
 ])
 const QUIT_CLEANUP_TIMEOUT_MS = 2500
 let mainWindow = null
+
+// Faults outside a request handler are written to a small private log in
+// userData (secrets redacted, size-capped) instead of disappearing or
+// taking the app down, so instability reports have evidence.
+function recordMainProcessFault(kind, detail) {
+  const text = String(detail?.stack || detail?.message || detail || '').slice(0, 4000)
+  console.error(`[meldwork] ${kind}`, text)
+  try {
+    const file = path.join(app.getPath('userData'), 'meldwork-diagnostics.log')
+    try {
+      if (fs.statSync(file).size > 1024 * 1024) fs.renameSync(file, `${file}.1`)
+    } catch { /* first entry */ }
+    fs.appendFileSync(file, `${new Date().toISOString()} ${kind} ${redactSecrets(text)}\n`, { mode: 0o600 })
+  } catch { /* diagnostics must never throw */ }
+}
+process.on('unhandledRejection', reason => recordMainProcessFault('unhandledRejection', reason))
+process.on('uncaughtException', error => recordMainProcessFault('uncaughtException', error))
 let workspace = null
 let workspaceChangedListener = null
 let workspaceRunFinishedListener = null
 let workspaceRunEventListener = null
 let installer = null
 let localAgentRefreshQueue = Promise.resolve()
-let providerStore = null
 let customAgentStore = null
 let agentConnectors = null
 let cloudAgentRuntime = null
@@ -226,21 +238,9 @@ function openExternalUrl(value) {
   return true
 }
 
-function providerOptions(kind, context = {}) {
-  if (!EXTERNAL_PROVIDER_KINDS.has(kind)) return {}
-  const status = providerStore?.status(kind) || {}
-  if (!status.configured && !(kind === 'openclaw' && context.nativeRuntime)) return {}
-  return providerOptionsFor(
-    kind,
-    status.configured ? providerStore.envForAgent(kind) : {},
-    context,
-    status,
-  )
-}
-
-async function nativeShellEnvironment() {
+async function nativeShellEnvironment(options = {}) {
   const home = app.getPath('home')
-  const resolved = await resolveNativeShellEnvironment({ home })
+  const resolved = await resolveNativeShellEnvironment({ home, refresh: options.refresh === true })
   return {
     ...resolved,
     env: { ...process.env, ...resolved.env },
@@ -253,37 +253,6 @@ async function detectLocalAgents() {
   return detectAgents({ env: shellEnvironment.env, home: shellEnvironment.home })
 }
 
-function mediaFallbackProviders(env = process.env) {
-  const apiKey = String(
-    env.ZGCI_MEDIA_API_KEY
-      || env.ZGCI_API_KEY
-      || env.ZCGI_API_KEY
-      || env.ZGCI_LLM_API_KEY
-      || '',
-  ).trim()
-  if (!apiKey) return []
-  const baseUrl = String(env.ZGCI_MEDIA_BASE_URL || 'https://hub.zgci.org/v1')
-    .trim()
-    .replace(/\/+$/, '')
-    .replace(/\/chat\/completions$/i, '')
-  let parsed
-  try { parsed = new URL(baseUrl) } catch { return [] }
-  if (parsed.protocol !== 'https:' || parsed.username || parsed.password
-      || parsed.search || parsed.hash) {
-    return []
-  }
-  return [{
-    kind: 'zgci-media',
-    status: {
-      configured: true,
-      provider: 'ZGCI Media',
-      baseUrl,
-      model: 'glm',
-    },
-    credentials: { OPENAI_API_KEY: apiKey },
-  }]
-}
-
 function workspaceStoragePath(userData = app.getPath('userData')) {
   return path.join(userData, 'meldwork-workspace.json')
 }
@@ -373,23 +342,25 @@ async function runCoreAgent(agent, prompt, workdir, options = {}) {
   const connectorCredentialIsolation = options.connectorCredentialIsolation === true
   const runOptions = { ...options }
   delete runOptions.connectorCredentialIsolation
-  const status = providerStore.status(agent.kind)
   const shellEnvironment = connectorCredentialIsolation ? null : await nativeShellEnvironment()
-  const nativeRuntime = !connectorCredentialIsolation
-    && agent.kind === 'openclaw' && !status.configured
+  // OpenClaw runs through its isolated native runtime scope; every other Agent
+  // CLI uses its own native credentials.
+  const nativeRuntime = !connectorCredentialIsolation && agent.kind === 'openclaw'
     ? await resolveNativeOpenClawRuntime({
         executable: agent.executable,
         env: shellEnvironment.env,
         home: shellEnvironment.home,
       })
     : null
-  const injected = connectorCredentialIsolation
+  const injected = connectorCredentialIsolation || agent.kind !== 'openclaw' || !nativeRuntime
     ? {}
-    : providerOptions(agent.kind, {
-        ...runOptions,
-        workdir,
+    : nativeOpenClawOptions({
         storageRoot: app.getPath('userData'),
-        nativeRuntime,
+        workdir,
+        sessionRef: runOptions.sessionRef,
+        allowWrite: runOptions.sandbox === 'workspace-write',
+        runtime: nativeRuntime,
+        openclawVersion: agent.version,
       })
   const nativeEnv = connectorCredentialIsolation || agent.kind === 'openclaw'
     ? {}
@@ -403,7 +374,10 @@ async function runCoreAgent(agent, prompt, workdir, options = {}) {
   return runAgent(agent, prompt, workdir, {
     ...runOptions,
     ...injected,
-    env: { ...nativeEnv, ...callerEnv, ...injected.env },
+    env: {
+      ...networkEnvironment(shellEnvironment?.env),
+      ...nativeEnv, ...callerEnv, ...injected.env,
+    },
   })
 }
 
@@ -533,17 +507,6 @@ function createWorkspace() {
     }),
     connectors: [],
   })
-  const mediaGeneration = new MediaGenerationRuntime({
-    getProvider: (kind, type, excludedKinds) => resolveMediaProvider({
-      requestedKind: kind,
-      type,
-      kinds: EXTERNAL_PROVIDER_KINDS,
-      excludedKinds,
-      statusFor: candidateKind => providerStore.status(candidateKind),
-      credentialsFor: candidateKind => providerStore.envForAgent(candidateKind),
-      fallbackProviders: mediaFallbackProviders(),
-    }),
-  })
   const localWorkspace = new LocalWorkspace({
     storagePath: workspaceStoragePath(),
     contentBlobStore,
@@ -586,10 +549,8 @@ function createWorkspace() {
       })
     },
     agentLabel: kind => customAgentStore.label(kind) || agentConnectors.label(kind),
-    sharedProviderReady: kind => EXTERNAL_PROVIDER_KINDS.has(kind) && providerStore.status(kind).configured,
     connectorRuntime: agentConnectors,
     naturalAgentResponses: true,
-    generateMedia: input => mediaGeneration.generate(input),
     runAgent: async (agent, prompt, workdir, options = {}) => {
       if (customAgentStore.has(agent.kind)) {
         return customAgentStore.run(agent.kind, prompt, workdir, {
@@ -666,6 +627,7 @@ async function refreshLocalAgentState() {
   if (!target) return null
   const task = localAgentRefreshQueue.then(async () => {
     if (workspace !== target) return workspace?.snapshot() || null
+    await nativeShellEnvironment({ refresh: true })
     const snapshot = await target.refreshAgents()
     if (workspace !== target) return workspace?.snapshot() || null
     return snapshot
@@ -734,6 +696,14 @@ function createWindow() {
     showOfflinePage(errorCode)
   })
   mainWindow.webContents.on('did-finish-load', flushPendingOpenGroup)
+  // A crashed renderer used to leave a blank window; reload it so the
+  // workspace (whose state lives in the main process) comes back.
+  mainWindow.webContents.on('render-process-gone', (_event, details) => {
+    recordMainProcessFault('render-process-gone', details?.reason || 'unknown')
+    if (details?.reason !== 'clean-exit' && mainWindow && !mainWindow.isDestroyed()) {
+      mainWindow.webContents.reload()
+    }
+  })
   mainWindow.on('closed', () => {
     mainWindow = null
   })
@@ -783,8 +753,6 @@ function registerIpc() {
     openExternalUrl,
     getOutcomeStore: () => workspace?.outcomeStore,
     getSkillTrustStore: () => skillTrustStore,
-    providerStore,
-    providerAgentKind,
     refreshLocalAgentState,
     registerTrustedHandle,
     skillCatalog,
@@ -817,11 +785,12 @@ if (!hasSingleInstanceLock) {
       storagePath: path.join(app.getPath('userData'), 'meldwork-cloud-agent-bridges.json'),
       fetch: globalThis.fetch,
     })
-    providerStore = new ProviderStore({
-      storagePath: path.join(app.getPath('userData'), 'meldwork-provider.json'),
-      safeStorage: lazySafeStorage,
-      allowedKinds: [...LOCAL_AGENT_KINDS],
-    })
+    // The shared Provider store was removed; drop any legacy encrypted
+    // credentials so they do not linger on disk unread.
+    try {
+      const { rmSync } = require('node:fs')
+      rmSync(path.join(app.getPath('userData'), 'meldwork-provider.json'), { force: true })
+    } catch { /* best-effort cleanup */ }
     knowledgeBaseStore = new KnowledgeBaseStore({
       storagePath: path.join(app.getPath('userData'), 'meldwork-knowledge-base.json'),
     })

desktop/src/preload.cjs

@@ -176,13 +176,6 @@ if (isLocalDocument) Object.assign(desktopApi, {
     open: id => ipcRenderer.invoke('local-attachments:open', id),
     discard: ids => ipcRenderer.invoke('local-attachments:discard', ids),
   }),
-  localAgentProvider: Object.freeze({
-    status: kind => ipcRenderer.invoke('local-agent-provider:status', kind),
-    probe: kind => ipcRenderer.invoke('local-agent-provider:probe', kind),
-    save: (kind, input) => ipcRenderer.invoke('local-agent-provider:save', kind, input),
-    activate: (kind, preset) => ipcRenderer.invoke('local-agent-provider:activate', kind, preset),
-    delete: (kind, preset) => ipcRenderer.invoke('local-agent-provider:delete', kind, preset),
-  }),
   localKnowledgeBase: Object.freeze({
     status: kind => (kind ? ipcRenderer.invoke('local-knowledge-base:status', kind) : ipcRenderer.invoke('local-knowledge-base:status')),
     openGuide: (kind, action) => ipcRenderer.invoke('local-knowledge-base:open-guide', kind, action),

desktop/src/product-data-migration.cjs

@@ -7,7 +7,6 @@ const MIGRATIONS = Object.freeze([
   ['workspace.json', 'workspace.json'],
   ['run-ledger.json', 'run-ledger.json'],
   ['custom-agents.json', 'custom-agents.json'],
-  ['provider.json', 'provider.json'],
   ['knowledge-base.json', 'knowledge-base.json'],
   ['private', 'private'],
 ])

desktop/src/providers/provider-options.cjs

@@ -1,182 +0,0 @@
-const { managedOpenClawOptions, nativeOpenClawOptions } = require('../agents/cli/openclaw-runtime.cjs')
-
-const EXTERNAL_PROVIDER_KINDS = new Set([
-  'codex', 'hermes', 'openclaw', 'workbuddy', 'pi', 'kimi', 'mimo', 'claude', 'gemini', 'opencode', 'qwen',
-  'opencodereview',
-])
-
-const OFFICIAL_PROVIDER_BASE_URLS = Object.freeze({
-  codex: 'https://api.openai.com/v1',
-  hermes: 'https://api.openai.com/v1',
-  openclaw: 'https://api.openai.com/v1',
-  workbuddy: '',
-  pi: '',
-  kimi: 'https://api.moonshot.cn/v1',
-  mimo: '',
-  claude: 'https://api.anthropic.com',
-  gemini: 'https://generativelanguage.googleapis.com/v1beta',
-  opencode: '',
-  qwen: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
-  opencodereview: 'https://api.openai.com/v1',
-})
-
-function chatCompletionsUrl(baseUrl) {
-  const normalized = String(baseUrl || '').trim().replace(/\/+$/, '')
-  if (!normalized) return ''
-  return /\/chat\/completions$/i.test(normalized)
-    ? normalized
-    : `${normalized}/chat/completions`
-}
-
-function providerPresetFromStatus(kind, status = {}) {
-  if (['official', 'openrouter', 'custom'].includes(status.activePreset)) {
-    return status.activePreset
-  }
-  const provider = String(status.provider || '').trim().toLowerCase()
-  const baseUrl = String(status.baseUrl || '').trim().replace(/\/+$/, '')
-  if (/openrouter/i.test(provider) || /openrouter\.ai$/i.test(new URL(baseUrl || 'https://example.com').hostname)) {
-    return 'openrouter'
-  }
-  if (OFFICIAL_PROVIDER_BASE_URLS[kind]
-      && baseUrl
-      && baseUrl === OFFICIAL_PROVIDER_BASE_URLS[kind]) {
-    return 'official'
-  }
-  return 'custom'
-}
-
-function providerOptionsFor(kind, generic, context = {}, status = {}) {
-  const preset = providerPresetFromStatus(kind, status)
-  if (kind === 'codex') return { env: generic }
-  if (kind === 'hermes') {
-    return {
-      provider: { id: 'openai-api', model: generic.OPENAI_MODEL },
-      env: {
-        ...generic,
-        HERMES_INFERENCE_PROVIDER: 'openai-api',
-        HERMES_INFERENCE_MODEL: generic.OPENAI_MODEL,
-      },
-    }
-  }
-  if (kind === 'workbuddy') {
-    return {
-      env: {
-        ...generic,
-        CODEBUDDY_MODEL: generic.OPENAI_MODEL,
-        CODEBUDDY_API_KEY: generic.OPENAI_API_KEY,
-        CODEBUDDY_BASE_URL: generic.OPENAI_BASE_URL,
-      },
-    }
-  }
-  if (kind === 'pi') {
-    return {
-      env: {
-        ...generic,
-        PI_MODEL: generic.OPENAI_MODEL,
-        PI_API_KEY: generic.OPENAI_API_KEY,
-        PI_BASE_URL: generic.OPENAI_BASE_URL,
-      },
-    }
-  }
-  if (kind === 'kimi') {
-    return {
-      env: {
-        ...generic,
-        MOONSHOT_API_KEY: generic.OPENAI_API_KEY,
-        KIMI_API_KEY: generic.OPENAI_API_KEY,
-        KIMI_MODEL_API_KEY: generic.OPENAI_API_KEY,
-        KIMI_MODEL_BASE_URL: generic.OPENAI_BASE_URL,
-        KIMI_MODEL_NAME: generic.OPENAI_MODEL,
-      },
-    }
-  }
-  if (kind === 'mimo') {
-    return {
-      env: {
-        ...generic,
-        MIMO_API_KEY: generic.OPENAI_API_KEY,
-        MIMO_BASE_URL: generic.OPENAI_BASE_URL,
-        MIMO_MODEL: generic.OPENAI_MODEL,
-      },
-    }
-  }
-  if (kind === 'claude') {
-    return {
-      env: {
-        ...generic,
-        ANTHROPIC_API_KEY: generic.OPENAI_API_KEY,
-        ANTHROPIC_BASE_URL: generic.OPENAI_BASE_URL,
-        ANTHROPIC_MODEL: generic.OPENAI_MODEL,
-      },
-    }
-  }
-  if (kind === 'gemini') {
-    return {
-      env: {
-        ...generic,
-        GEMINI_API_KEY: generic.OPENAI_API_KEY,
-        GOOGLE_API_KEY: generic.OPENAI_API_KEY,
-        GEMINI_MODEL: generic.OPENAI_MODEL,
-      },
-    }
-  }
-  if (kind === 'qwen') {
-    return {
-      provider: { id: 'openai', model: generic.OPENAI_MODEL },
-      env: {
-        ...generic,
-        DASHSCOPE_API_KEY: generic.OPENAI_API_KEY,
-      },
-    }
-  }
-  if (kind === 'opencode') {
-    return {
-      env: {
-        ...generic,
-        ...(preset === 'openrouter' ? { OPENROUTER_API_KEY: generic.OPENAI_API_KEY } : {}),
-      },
-    }
-  }
-  if (kind === 'opencodereview') {
-    return {
-      env: {
-        ...generic,
-        OCR_LLM_URL: chatCompletionsUrl(generic.OPENAI_BASE_URL),
-        OCR_LLM_TOKEN: generic.OPENAI_API_KEY,
-        OCR_LLM_MODEL: generic.OPENAI_MODEL,
-        OCR_USE_ANTHROPIC: 'false',
-      },
-    }
-  }
-  if (kind === 'openclaw' && context.storageRoot && context.workdir) {
-    if (!generic?.OPENAI_API_KEY) {
-      return nativeOpenClawOptions({
-        storageRoot: context.storageRoot,
-        workdir: context.workdir,
-        sessionRef: context.sessionRef,
-        allowWrite: context.sandbox === 'workspace-write',
-        runtime: context.nativeRuntime,
-      })
-    }
-    return managedOpenClawOptions({
-      storageRoot: context.storageRoot,
-      workdir: context.workdir,
-      sessionRef: context.sessionRef,
-      allowWrite: context.sandbox === 'workspace-write',
-      provider: generic,
-    })
-  }
-  return {}
-}
-
-function providerAgentKind(value) {
-  const kind = String(value || '').trim()
-  if (!EXTERNAL_PROVIDER_KINDS.has(kind)) throw new Error('PROVIDER_AGENT_UNSUPPORTED')
-  return kind
-}
-
-module.exports = {
-  EXTERNAL_PROVIDER_KINDS,
-  providerAgentKind,
-  providerOptionsFor,
-}

desktop/src/providers/provider-store.cjs

@@ -1,378 +0,0 @@
-const fs = require('node:fs')
-
-const { atomicWritePrivateFile } = require('../security/private-file.cjs')
-
-const STORE_VERSION = 3
-const SINGLE_PROFILE_STORE_VERSION = 2
-const LEGACY_STORE_VERSION = 1
-const PROVIDER_KIND = /^[a-z0-9][a-z0-9-]{0,31}$/
-const PROVIDER_PRESETS = new Set(['official', 'openrouter', 'custom'])
-const EMPTY_METADATA = Object.freeze({ provider: '', baseUrl: '', model: '' })
-const OPENROUTER_PROVIDER = Object.freeze({
-  provider: 'OpenRouter',
-  baseUrl: 'https://openrouter.ai/api/v1',
-})
-const OFFICIAL_PROVIDERS = Object.freeze({
-  codex: { provider: 'OpenAI', baseUrl: 'https://api.openai.com/v1' },
-  hermes: { provider: 'OpenAI API', baseUrl: 'https://api.openai.com/v1' },
-  openclaw: { provider: 'OpenAI API', baseUrl: 'https://api.openai.com/v1' },
-  workbuddy: { provider: 'WorkBuddy Official', baseUrl: '' },
-  pi: { provider: 'Pi Native', baseUrl: '' },
-  kimi: { provider: 'Moonshot AI', baseUrl: 'https://api.moonshot.cn/v1' },
-  mimo: { provider: 'MiMo Native', baseUrl: '' },
-  claude: { provider: 'Anthropic', baseUrl: 'https://api.anthropic.com' },
-  gemini: { provider: 'Google AI Studio', baseUrl: 'https://generativelanguage.googleapis.com/v1beta' },
-  opencode: { provider: 'OpenCode Provider', baseUrl: '' },
-  qwen: { provider: 'DashScope', baseUrl: 'https://dashscope.aliyuncs.com/compatible-mode/v1' },
-  opencodereview: { provider: 'OpenAI API', baseUrl: 'https://api.openai.com/v1' },
-})
-function normalizeMetadata(input) {
-  const provider = String(input?.provider || '').trim()
-  let baseUrl = String(input?.baseUrl || '').trim().replace(/\/+$/, '')
-  baseUrl = baseUrl.replace(/\/chat\/completions$/i, '')
-  const model = String(input?.model || '').trim()
-  if (!provider || provider.length > 80 || !baseUrl || baseUrl.length > 300
-      || !model || model.length > 160) {
-    throw new Error('PROVIDER_INVALID_METADATA')
-  }
-  let parsed
-  try { parsed = new URL(baseUrl) } catch { throw new Error('PROVIDER_INVALID_METADATA') }
-  const loopback = ['localhost', '127.0.0.1', '[::1]'].includes(parsed.hostname)
-  if (parsed.protocol !== 'https:' && !(parsed.protocol === 'http:' && loopback)) {
-    throw new Error('PROVIDER_INSECURE_BASE_URL')
-  }
-  if (parsed.username || parsed.password || parsed.search || parsed.hash) {
-    throw new Error('PROVIDER_INVALID_METADATA')
-  }
-  return { provider, baseUrl, model }
-}
-
-function sortedKeys(value) {
-  return value && typeof value === 'object' && !Array.isArray(value)
-    ? Reflect.ownKeys(value).sort().join(',')
-    : ''
-}
-
-function inferStoredPreset(entry) {
-  const provider = String(entry?.provider || '').toLowerCase()
-  const baseUrl = String(entry?.baseUrl || '')
-  return /openrouter/i.test(provider) || /(^|\.)openrouter\.ai$/i.test(hostname(baseUrl))
-    ? 'openrouter'
-    : 'custom'
-}
-
-function hostname(value) {
-  try { return new URL(value).hostname } catch { return '' }
-}
-
-function validatePresetMetadata(kind, preset, metadata) {
-  if (preset === 'custom') return metadata
-  const expected = preset === 'openrouter' ? OPENROUTER_PROVIDER : OFFICIAL_PROVIDERS[kind]
-  if (!expected?.baseUrl
-      || metadata.provider !== expected.provider
-      || metadata.baseUrl !== expected.baseUrl) {
-    throw new Error('PROVIDER_INVALID_METADATA')
-  }
-  return metadata
-}
-
-class ProviderStore {
-  constructor({ storagePath, safeStorage, allowedKinds = [] }) {
-    if (typeof storagePath !== 'string' || !storagePath) {
-      throw new Error('PROVIDER_STORAGE_PATH_REQUIRED')
-    }
-    const normalizedKinds = [...new Set(allowedKinds.map(kind => String(kind || '').trim()))]
-    if (!normalizedKinds.length || normalizedKinds.some(kind => !PROVIDER_KIND.test(kind))) {
-      throw new Error('PROVIDER_AGENT_KINDS_REQUIRED')
-    }
-    this.storagePath = storagePath
-    this.safeStorage = safeStorage
-    this.allowedKinds = new Set(normalizedKinds)
-  }
-
-  encryptionAvailable() {
-    try {
-      return Boolean(this.safeStorage?.isEncryptionAvailable?.())
-    } catch {
-      return false
-    }
-  }
-
-  status(kind, { probeEncryption = false } = {}) {
-    const targetKind = this.normalizeKind(kind)
-    const credentialStored = fs.existsSync(this.storagePath)
-    if (!credentialStored && !probeEncryption) {
-      return {
-        provider: '',
-        baseUrl: '',
-        model: '',
-        activePreset: 'official',
-        profiles: {},
-        encryptionAvailable: null,
-        configured: false,
-      }
-    }
-
-    const encryptionAvailable = this.encryptionAvailable()
-    let metadata = EMPTY_METADATA
-    let configured = false
-    let activePreset = 'official'
-    const profiles = {}
-    if (encryptionAvailable && credentialStored) {
-      try {
-        const agent = this.readDocument().agents[targetKind]
-        if (agent) {
-          activePreset = agent.activePreset
-          for (const [preset, entry] of Object.entries(agent.profiles)) {
-            try {
-              this.decryptEntry(entry)
-              profiles[preset] = {
-                provider: entry.provider,
-                baseUrl: entry.baseUrl,
-                model: entry.model,
-                configured: true,
-              }
-            } catch { /* keep an unreadable profile unavailable */ }
-          }
-          const activeEntry = profiles[activePreset]
-          if (activeEntry) {
-            metadata = activeEntry
-            configured = true
-          }
-        }
-      } catch { /* unreadable credentials stay unavailable */ }
-    }
-    return {
-      provider: metadata.provider,
-      baseUrl: metadata.baseUrl,
-      model: metadata.model,
-      activePreset,
-      profiles,
-      encryptionAvailable,
-      configured,
-    }
-  }
-
-  save(kind, input) {
-    const targetKind = this.normalizeKind(kind)
-    if (sortedKeys(input) !== 'apiKey,baseUrl,model,preset,provider'
-        || typeof input.apiKey !== 'string' || !input.apiKey.trim()
-        || input.apiKey.trim().length > 8192) {
-      throw new Error('PROVIDER_INVALID_CREDENTIAL')
-    }
-    this.requireEncryption()
-
-    const apiKey = input.apiKey.trim()
-    const preset = this.normalizePreset(input.preset)
-    const metadata = validatePresetMetadata(targetKind, preset, normalizeMetadata(input))
-    let encrypted
-    try {
-      encrypted = this.safeStorage.encryptString(apiKey)
-      if (!Buffer.isBuffer(encrypted) || encrypted.length === 0
-          || this.safeStorage.decryptString(encrypted) !== apiKey) {
-        throw new Error('invalid encryption result')
-      }
-    } catch {
-      throw new Error('PROVIDER_ENCRYPTION_FAILED')
-    }
-
-    const document = fs.existsSync(this.storagePath)
-      ? this.readDocument()
-      : { version: STORE_VERSION, agents: {} }
-    const agent = document.agents[targetKind] || { activePreset: 'official', profiles: {} }
-    agent.profiles[preset] = {
-      ...metadata,
-      encrypted: encrypted.toString('base64'),
-    }
-    agent.activePreset = preset
-    document.agents[targetKind] = agent
-    atomicWritePrivateFile(this.storagePath, JSON.stringify(document))
-    return this.status(targetKind)
-  }
-
-  activate(kind, preset) {
-    const targetKind = this.normalizeKind(kind)
-    const targetPreset = this.normalizePreset(preset)
-    if (!fs.existsSync(this.storagePath)) {
-      if (targetPreset !== 'official') throw new Error('PROVIDER_PROFILE_UNAVAILABLE')
-      return this.status(targetKind, { probeEncryption: true })
-    }
-    this.requireEncryption()
-    const document = this.readDocument()
-    const agent = document.agents[targetKind] || { activePreset: 'official', profiles: {} }
-    if (targetPreset !== 'official' && !agent.profiles[targetPreset]) {
-      throw new Error('PROVIDER_PROFILE_UNAVAILABLE')
-    }
-    if (agent.profiles[targetPreset]) this.decryptEntry(agent.profiles[targetPreset])
-    agent.activePreset = targetPreset
-    document.agents[targetKind] = agent
-    atomicWritePrivateFile(this.storagePath, JSON.stringify(document))
-    return this.status(targetKind)
-  }
-
-  delete(kind, preset = '') {
-    const targetKind = this.normalizeKind(kind)
-    if (fs.existsSync(this.storagePath)) {
-      const document = this.readDocument()
-      if (preset) {
-        const targetPreset = this.normalizePreset(preset)
-        const agent = document.agents[targetKind]
-        if (agent) {
-          delete agent.profiles[targetPreset]
-          if (agent.activePreset === targetPreset) agent.activePreset = 'official'
-          if (Object.keys(agent.profiles).length) document.agents[targetKind] = agent
-          else delete document.agents[targetKind]
-        }
-      } else {
-        delete document.agents[targetKind]
-      }
-      if (Object.keys(document.agents).length) {
-        atomicWritePrivateFile(this.storagePath, JSON.stringify(document))
-      } else {
-        fs.unlinkSync(this.storagePath)
-      }
-    }
-    return this.status(targetKind, { probeEncryption: true })
-  }
-
-  envForAgent(kind) {
-    const targetKind = this.normalizeKind(kind)
-    this.requireEncryption()
-    const agent = this.readDocument().agents[targetKind]
-    const entry = agent?.profiles?.[agent.activePreset]
-    if (!entry) throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-    return {
-      OPENAI_API_KEY: this.decryptEntry(entry),
-      OPENAI_BASE_URL: entry.baseUrl,
-      OPENAI_MODEL: entry.model,
-    }
-  }
-
-  requireEncryption() {
-    if (!this.encryptionAvailable()) throw new Error('PROVIDER_ENCRYPTION_UNAVAILABLE')
-  }
-
-  normalizeKind(kind) {
-    const normalized = String(kind || '').trim()
-    if (!PROVIDER_KIND.test(normalized) || !this.allowedKinds.has(normalized)) {
-      throw new Error('PROVIDER_AGENT_UNSUPPORTED')
-    }
-    return normalized
-  }
-
-  normalizePreset(preset) {
-    const normalized = String(preset || '').trim().toLowerCase()
-    if (!PROVIDER_PRESETS.has(normalized)) throw new Error('PROVIDER_PRESET_UNSUPPORTED')
-    return normalized
-  }
-
-  decryptEntry(entry) {
-    this.requireEncryption()
-    try {
-      const encrypted = Buffer.from(entry.encrypted, 'base64')
-      if (!encrypted.length || encrypted.toString('base64') !== entry.encrypted) {
-        throw new Error('invalid ciphertext')
-      }
-      const apiKey = this.safeStorage.decryptString(encrypted)
-      if (typeof apiKey !== 'string' || !apiKey.trim()) throw new Error('invalid credential')
-      return apiKey
-    } catch {
-      throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-    }
-  }
-
-  readDocument() {
-    let payload
-    try {
-      payload = JSON.parse(fs.readFileSync(this.storagePath, 'utf8'))
-    } catch {
-      throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-    }
-    if (payload?.version === LEGACY_STORE_VERSION) {
-      if (sortedKeys(payload) !== 'baseUrl,encrypted,model,provider,version'
-          || typeof payload.encrypted !== 'string') {
-        throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-      }
-      const rawEntry = {
-        provider: payload.provider,
-        baseUrl: payload.baseUrl,
-        model: payload.model,
-        encrypted: payload.encrypted,
-      }
-      return {
-        version: STORE_VERSION,
-        agents: Object.fromEntries([...this.allowedKinds].map((kind) => {
-          const entry = this.normalizeEntry(rawEntry)
-          const preset = inferStoredPreset(entry)
-          return [kind, {
-            activePreset: preset,
-            profiles: { [preset]: { ...entry } },
-          }]
-        })),
-      }
-    }
-    if (payload?.version === SINGLE_PROFILE_STORE_VERSION) {
-      if (sortedKeys(payload) !== 'agents,version'
-          || !payload.agents || typeof payload.agents !== 'object'
-          || Array.isArray(payload.agents)) {
-        throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-      }
-      const agents = {}
-      for (const [kind, rawEntry] of Object.entries(payload.agents)) {
-        if (!this.allowedKinds.has(kind)) throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-        const entry = this.normalizeEntry(rawEntry)
-        const preset = inferStoredPreset(entry)
-        agents[kind] = { activePreset: preset, profiles: { [preset]: entry } }
-      }
-      return { version: STORE_VERSION, agents }
-    }
-    if (sortedKeys(payload) !== 'agents,version'
-        || payload.version !== STORE_VERSION
-        || !payload.agents || typeof payload.agents !== 'object'
-        || Array.isArray(payload.agents)) {
-      throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-    }
-    const agents = {}
-    for (const [kind, entry] of Object.entries(payload.agents)) {
-      if (!this.allowedKinds.has(kind)) throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-      agents[kind] = this.normalizeAgentEntry(entry)
-    }
-    return { version: STORE_VERSION, agents }
-  }
-
-  normalizeAgentEntry(entry) {
-    try {
-      if (sortedKeys(entry) !== 'activePreset,profiles'
-          || !entry.profiles || typeof entry.profiles !== 'object'
-          || Array.isArray(entry.profiles)) {
-        throw new Error('invalid agent entry')
-      }
-      const activePreset = this.normalizePreset(entry.activePreset)
-      const profiles = {}
-      for (const [preset, profile] of Object.entries(entry.profiles)) {
-        const normalizedPreset = this.normalizePreset(preset)
-        profiles[normalizedPreset] = this.normalizeEntry(profile)
-      }
-      if (activePreset !== 'official' && !profiles[activePreset]) {
-        throw new Error('missing active profile')
-      }
-      return { activePreset, profiles }
-    } catch {
-      throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-    }
-  }
-
-  normalizeEntry(entry) {
-    if (sortedKeys(entry) !== 'baseUrl,encrypted,model,provider'
-        || typeof entry.encrypted !== 'string') {
-      throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-    }
-    const encrypted = Buffer.from(entry.encrypted, 'base64')
-    if (!encrypted.length || encrypted.toString('base64') !== entry.encrypted) {
-      throw new Error('PROVIDER_CREDENTIAL_UNAVAILABLE')
-    }
-    return { ...normalizeMetadata(entry), encrypted: entry.encrypted }
-  }
-}
-
-module.exports = { ProviderStore }

desktop/src/runs/failure-policy.cjs

@@ -1,3 +1,11 @@
+const {
+  authFailureCertainty,
+  isNetworkFailure,
+  isProviderUnavailable,
+  isRateLimited,
+  retryAfterFromText,
+} = require('./failure-signals.cjs')
+
 const FAILURE_CATEGORIES = Object.freeze([
   'authentication',
   'compatibility',
@@ -121,23 +129,47 @@ function retryAfterMs(error, now) {
     const parsed = parseRetryAfter(headerValue(headers, 'retry-after'), now)
     if (parsed != null) return parsed
   }
-  return null
+  // Local CLIs surface provider throttling as text ("retry after 20s").
+  return retryAfterFromText(error?.diagnostic || error?.message)
+}
+
+// A local CLI process failure carries a code-derived category (usually
+// protocol). Its diagnostic often proves a transient provider condition, so
+// refine it before deciding whether to retry and how long to wait.
+function transientCategoryFromText(detail) {
+  if (!detail) return ''
+  if (isRateLimited(detail)) return 'rate_limit'
+  if (isNetworkFailure(detail)) return 'network'
+  if (isProviderUnavailable(detail)) return 'provider'
+  return ''
 }
 
 function classifiedCategory(error, code, status) {
   const explicit = String(error?.failure?.category || '')
-  if (FAILURE_CATEGORY_SET.has(explicit)) return explicit
   const detail = String(error?.diagnostic || error?.message || '')
+  if (FAILURE_CATEGORY_SET.has(explicit)) {
+    if (['protocol', 'execution'].includes(explicit)) {
+      const refined = transientCategoryFromText(error?.diagnostic || '')
+      if (refined) return refined
+    }
+    return explicit
+  }
   if (code === 'LOCAL_BUDGET_EXHAUSTED') return 'budget'
   if (/PERMISSION/.test(code)) return 'permission'
   if (code === 'LOCAL_AGENT_EXECUTION_STOPPED' || /CANCEL|ABORT/.test(code)) {
     return 'cancellation'
   }
   if (code === 'LOCAL_AGENT_SESSION_INVALID') return 'session'
-  if (status === 401 || status === 403 || code === 'LOCAL_AGENT_AUTH_REQUIRED'
-      || /unauthori[sz]ed|forbidden|invalid token/i.test(detail)) return 'authentication'
+  // Only explicit sign-in statements prove that credentials are gone. A 403,
+  // a proxy refusal, or "unauthorized" inside unrelated output is retried as
+  // a provider-side refusal instead of stopping the Agent for the run.
+  const authCertainty = authFailureCertainty(detail)
+  if (status === 401 || code === 'LOCAL_AGENT_AUTH_REQUIRED'
+      || authCertainty === 'definitive') return 'authentication'
+  if (status === 403 || code === 'LOCAL_AGENT_ACCESS_DENIED'
+      || authCertainty === 'ambiguous') return 'provider'
   if (status === 429 || /RATE_LIMIT|TOO_MANY_REQUESTS/.test(code)
-      || /rate limit|too many requests/i.test(detail)) return 'rate_limit'
+      || isRateLimited(detail)) return 'rate_limit'
   if (NETWORK_CODES.has(code) || /NETWORK|CONNECTION/.test(code)
       || /socket hang up|temporary failure|network error/i.test(detail)) return 'network'
   if (status != null && status >= 500) return 'provider'
@@ -244,18 +276,30 @@ function retryDecision(input, options = {}) {
   if (safety?.sideEffectsPossible === true && safety.idempotencyMode !== 'durable') {
     return Object.freeze({ ...base, action: 'human_gate', exhausted: false })
   }
-  const maxDelayMs = Math.max(1, boundedInteger(
+  const configuredMaxDelayMs = Math.max(1, boundedInteger(
     options.maxDelayMs, 10 * 1000, 24 * 60 * 60 * 1000,
   ))
+  // Provider throttling needs a longer and server-directed wait than a crashed
+  // process. Honor Retry-After up to the rate-limit cap instead of hammering
+  // the provider with the generic sub-second backoff.
+  const rateLimited = failure.category === 'rate_limit'
+  const maxDelayMs = rateLimited
+    ? Math.max(configuredMaxDelayMs, boundedInteger(
+      options.rateLimitMaxDelayMs, 60 * 1000, 24 * 60 * 60 * 1000,
+    ))
+    : configuredMaxDelayMs
   const backoff = boundedBackoffDelay(attempt, { ...options, maxDelayMs })
   const requestedDelay = failure.retryAfterMs == null
     ? 0
     : Math.min(maxDelayMs, boundedInteger(failure.retryAfterMs, 0, maxDelayMs))
+  const minimumDelay = rateLimited && failure.retryAfterMs == null
+    ? Math.min(maxDelayMs, boundedInteger(options.rateLimitMinDelayMs, 0, maxDelayMs))
+    : 0
   return Object.freeze({
     ...base,
     action: 'retry',
     exhausted: false,
-    delayMs: Math.max(backoff, requestedDelay),
+    delayMs: Math.max(backoff, requestedDelay, minimumDelay),
   })
 }
 

desktop/src/runs/failure-signals.cjs

@@ -0,0 +1,129 @@
+// Pure text classifiers for local Agent failure diagnostics.
+//
+// Agent CLIs report failures as free text on stderr or inside structured error
+// results. Classification decides whether Meldwork retries, asks the user to
+// sign in again, or rebuilds a native session, so every rule here errs on the
+// side of "ambiguous": only explicit login and API-key statements count as a
+// definitive authentication failure. A bare 403, a proxy refusal, or a word
+// such as "unauthorized" inside unrelated output must never mark an installed
+// Agent as signed out.
+
+const MAX_SIGNAL_TEXT_CHARS = 64 * 1024
+
+const LOG_PREFIX = /^(?:\[[^\]]{1,40}\]\s*|(?:error|fatal|err|warning)\s*[:!-]\s*|[✗✘×!]\s*)+/i
+
+const DEFINITIVE_AUTH_LINE = [
+  /^(?:invalid|expired|missing|incorrect|revoked)\s+(?:api[ _-]?key|(?:access|refresh|auth|oauth|bearer)[ _-]?token|credentials?)\b/i,
+  /^(?:authentication|authorization)\s+(?:failed|required)\b/i,
+  /^(?:please\s+)?(?:log|sign)\s+in\b/i,
+  /^(?:please\s+)?run\s+\S*\s*(?:\/)?login\b/i,
+  /^not\s+(?:logged|signed)\s+in\b/i,
+  /^you\s+are\s+not\s+(?:logged|signed)\s+in\b/i,
+  /^select an auth type\b/i,
+  /^oauth\s+token\s+(?:has\s+)?(?:expired|is\s+invalid|was\s+revoked)\b/i,
+  /^HTTP\s+401\s*:/i,
+  /^provider rejected\b/i,
+  /^(?:认证|鉴权|登录)(?:失败|过期)/,
+  /^(?:令牌|凭据|密钥)(?:无效|缺失|已过期)/,
+  /^请(?:先|重新)登录/,
+  /^未登录/,
+]
+
+const DEFINITIVE_AUTH_ANYWHERE = /"type"\s*:\s*"authentication_error"|\binvalid_api_key\b|\bincorrect api key provided\b|\binvalid x-api-key\b/i
+
+// Ambiguous refusals must lead a diagnostic line. Tool output such as
+// "Failed to download docs: HTTP 403 Forbidden" mentions a refusal of some
+// unrelated resource and stays a plain process failure.
+const AMBIGUOUS_AUTH_LINE = [
+  /^HTTP\s+403\s*:/i,
+  /^(?:403\s+)?forbidden\b/i,
+  /^unauthori[sz]ed\b/i,
+  /^access denied\b/i,
+  /^invalid token\b/i,
+]
+
+const AMBIGUOUS_AUTH_ANYWHERE = /"type"\s*:\s*"permission_error"/i
+
+const RATE_LIMIT = /\b429\b|too many requests|rate[ _-]?limit(?:ed|ing)?\b|\boverloaded(?:_error)?\b|\bquota (?:exceeded|exhausted)\b|\bresource[ _-]?exhausted\b|\binsufficient_quota\b|\b529\b/i
+
+const NETWORK = /\b(?:ECONNRESET|ECONNREFUSED|ECONNABORTED|ENOTFOUND|ETIMEDOUT|EAI_AGAIN|ENETUNREACH|ENETDOWN|EHOSTUNREACH)\b|socket hang up|fetch failed|network (?:error|is unreachable)|connection (?:reset|refused|closed|timed out)|getaddrinfo|tls handshake|ssl handshake|proxy (?:error|connection)/i
+
+const PROVIDER_UNAVAILABLE = /\b(?:HTTP\s*)?50[0234]\b\s*(?::|-)?\s*(?:internal server error|bad gateway|service unavailable|gateway timeout)|\bservice unavailable\b|\bbad gateway\b|\bgateway timeout\b|\binternal server error\b|"type"\s*:\s*"api_error"|\bupstream connect error\b/i
+
+const CONTEXT_OVERFLOW = /prompt is too long|context[_ ]length[_ ]exceeded|maximum context length|context window (?:exceeded|is full|limit)|input is too long|too many (?:input )?tokens|exceeds? the (?:maximum|model'?s?) context|context_window_exceeded|request too large for model/i
+
+function signalText(value) {
+  return String(value || '').slice(0, MAX_SIGNAL_TEXT_CHARS)
+}
+
+function diagnosticLines(value) {
+  return signalText(value)
+    .split(/\r?\n/)
+    .map(line => line.trim().replace(LOG_PREFIX, '').trim())
+    .filter(Boolean)
+}
+
+// Returns 'definitive', 'ambiguous', or ''.
+function authFailureCertainty(value) {
+  const text = signalText(value)
+  if (!text.trim()) return ''
+  const lines = diagnosticLines(text)
+  if (DEFINITIVE_AUTH_ANYWHERE.test(text)
+      || lines.some(line => DEFINITIVE_AUTH_LINE.some(pattern => pattern.test(line)))) {
+    return 'definitive'
+  }
+  if (lines.some(line => AMBIGUOUS_AUTH_LINE.some(pattern => pattern.test(line)))
+      || AMBIGUOUS_AUTH_ANYWHERE.test(text)) {
+    return 'ambiguous'
+  }
+  return ''
+}
+
+function isRateLimited(value) {
+  return RATE_LIMIT.test(signalText(value))
+}
+
+function isNetworkFailure(value) {
+  return NETWORK.test(signalText(value))
+}
+
+function isProviderUnavailable(value) {
+  return PROVIDER_UNAVAILABLE.test(signalText(value))
+}
+
+function isContextOverflow(value) {
+  return CONTEXT_OVERFLOW.test(signalText(value))
+}
+
+const RETRY_AFTER_PATTERNS = [
+  /retry[- ]after["']?\s*[:=]?\s*(\d+(?:\.\d+)?)\s*(ms|milliseconds?|s|secs?|seconds?|m|mins?|minutes?)?/i,
+  /(?:try|retry) again in\s+(\d+(?:\.\d+)?)\s*(ms|milliseconds?|s|secs?|seconds?|m|mins?|minutes?)?/i,
+  /(?:reset|resets) in\s+(\d+(?:\.\d+)?)\s*(ms|milliseconds?|s|secs?|seconds?|m|mins?|minutes?)/i,
+]
+
+const MAX_RETRY_AFTER_MS = 5 * 60 * 1000
+
+function retryAfterFromText(value) {
+  const text = signalText(value)
+  for (const pattern of RETRY_AFTER_PATTERNS) {
+    const match = text.match(pattern)
+    if (!match) continue
+    const amount = Number(match[1])
+    if (!Number.isFinite(amount) || amount < 0) continue
+    const unit = String(match[2] || 's').toLowerCase()
+    const multiplier = unit.startsWith('ms') || unit.startsWith('milli')
+      ? 1
+      : unit.startsWith('m') ? 60 * 1000 : 1000
+    return Math.min(MAX_RETRY_AFTER_MS, Math.round(amount * multiplier))
+  }
+  return null
+}
+
+module.exports = {
+  authFailureCertainty,
+  isContextOverflow,
+  isNetworkFailure,
+  isProviderUnavailable,
+  isRateLimited,
+  retryAfterFromText,
+}

desktop/src/runs/invocation-watchdog.cjs

@@ -0,0 +1,139 @@
+const REAL_CLOCK = Object.freeze({
+  now: () => Date.now(),
+  setTimeout: (callback, ms) => setTimeout(callback, ms),
+  clearTimeout: handle => clearTimeout(handle),
+})
+
+function positiveFinite(value) {
+  return Number.isFinite(value) && value > 0
+}
+
+// Stops an Agent invocation that has gone quiet. Three windows apply:
+//  - startup: strict turns (autonomous discussion) must show any sign of life
+//    within `timeoutMs`; once they do, the deadline is lifted.
+//  - inactivity: any activity refreshes `timeoutMs`, or `toolTimeoutMs` while a
+//    tool call is running.
+//  - stall: strict turns also stop after three inactivity windows without
+//    substantive progress, so a CLI that emits empty heartbeats forever still
+//    ends and the remaining participants can continue. Other turns keep
+//    running while the process is active; the user can stop them.
+// Raw process output only proves liveness (`liveness`); mapped events, answer
+// text, tool lifecycle and structured progress count as progress (`progress`).
+function createInvocationWatchdog({
+  timeoutMs,
+  toolTimeoutMs,
+  strict = false,
+  isStopped = () => false,
+  onTimeout = () => {},
+  clock = REAL_CLOCK,
+} = {}) {
+  const enabled = positiveFinite(timeoutMs)
+  const startedAt = clock.now()
+  const startupDeadlineAt = strict && enabled ? startedAt + timeoutMs : Number.POSITIVE_INFINITY
+  let lastProgressAt = startedAt
+  let lastSubstantiveAt = startedAt
+  let paused = false
+  let established = false
+  let toolActive = false
+  let closed = false
+  let timedOut = false
+  let timeoutError = null
+  let timer = null
+  let rejectTimeout = () => {}
+  const promise = new Promise((_, reject) => { rejectTimeout = reject })
+  promise.catch(() => {})
+
+  function disarm() {
+    clock.clearTimeout(timer)
+    timer = null
+  }
+
+  function arm() {
+    if (timer || paused || isStopped() || !enabled) return
+    const now = clock.now()
+    const inactivityWindowMs = toolActive ? toolTimeoutMs : timeoutMs
+    const inactivityRemainingMs = Math.max(0, inactivityWindowMs - (now - lastProgressAt))
+    const startupRemainingMs = Math.max(
+      0, (established ? Number.POSITIVE_INFINITY : startupDeadlineAt) - now,
+    )
+    const stallWindowMs = !strict
+      ? Number.POSITIVE_INFINITY
+      : (toolActive ? toolTimeoutMs : timeoutMs * 3)
+    const stallRemainingMs = Math.max(0, stallWindowMs - (now - lastSubstantiveAt))
+    const remainingMs = Math.min(inactivityRemainingMs, startupRemainingMs, stallRemainingMs)
+    timer = clock.setTimeout(() => {
+      timer = null
+      if (paused || isStopped()) return
+      timedOut = true
+      timeoutError = new Error('LOCAL_AGENT_TIMEOUT')
+      onTimeout(timeoutError)
+      rejectTimeout(timeoutError)
+    }, remainingMs)
+  }
+
+  function rearm() {
+    if (paused) return
+    disarm()
+    arm()
+  }
+
+  function inactive() {
+    return closed || isStopped()
+  }
+
+  // The process produced something meaningful: refresh every window.
+  function progress() {
+    if (inactive()) return
+    lastProgressAt = clock.now()
+    lastSubstantiveAt = lastProgressAt
+    rearm()
+  }
+
+  arm()
+
+  return {
+    promise,
+    get timedOut() { return timedOut },
+    get error() { return timeoutError },
+    progress,
+    // The process is alive but has not said anything yet (stdout/stderr bytes,
+    // ACP traffic): lift the startup deadline and refresh the inactivity
+    // window outside tool calls.
+    liveness() {
+      if (inactive()) return
+      established = true
+      if (!toolActive) lastProgressAt = clock.now()
+      rearm()
+    },
+    // First proof that the run started. Lifts the startup deadline only.
+    establish() {
+      if (established || isStopped()) return
+      established = true
+      rearm()
+    },
+    toolStarted() {
+      toolActive = true
+      progress()
+    },
+    toolFinished() {
+      toolActive = false
+      progress()
+    },
+    // Waiting for a person (Human Gate) is not the Agent being quiet.
+    pause() {
+      paused = true
+      disarm()
+    },
+    resume() {
+      paused = false
+      lastProgressAt = clock.now()
+      arm()
+    },
+    close() {
+      closed = true
+      disarm()
+    },
+  }
+}
+
+module.exports = { createInvocationWatchdog }

desktop/src/runs/run-harness-context.cjs

@@ -122,6 +122,8 @@ function normalizeSessionMeta(input) {
     meta.inheritedTaskIds = inheritedTaskIds
     meta.provenanceCompleteness = provenanceCompleteness
   }
+  const lastDeliveredMessageId = cleanId(input.lastDeliveredMessageId)
+  if (lastDeliveredMessageId) meta.lastDeliveredMessageId = lastDeliveredMessageId
   return meta
 }
 
@@ -157,6 +159,8 @@ function nextSessionMeta(meta, usage = {}) {
     if (Object.hasOwn(usage, field)) next[field] = usage[field]
     else if (Object.hasOwn(previous, field)) next[field] = previous[field]
   }
+  if (usage.lastDeliveredMessageId) next.lastDeliveredMessageId = usage.lastDeliveredMessageId
+  else if (previous.lastDeliveredMessageId) next.lastDeliveredMessageId = previous.lastDeliveredMessageId
   return normalizeSessionMeta(next)
 }
 

desktop/src/runs/run-harness-normalization.cjs

@@ -1,4 +1,5 @@
 const MAX_CAPSULE_EVENTS = 12
+const { parseTaskDecision } = require('../collaboration/task-decision.cjs')
 
 const { redactSecrets } = require('../security/secret-redaction.cjs')
 const { normalizeContentBlobRef } = require('../attachments/content-blob-store.cjs')
@@ -252,6 +253,10 @@ function normalizeContextStats(input) {
     const value = String(input[field] || '')
     if (SHA256.test(value)) context[field] = value
   }
+  const operationId = cleanId(input.operationId)
+  if (operationId) context.operationId = operationId
+  const snapshotHash = String(input.snapshotHash || '')
+  if (SHA256.test(snapshotHash)) context.snapshotHash = snapshotHash
   if (input.sessionRotated === true) context.sessionRotated = true
   const contextPackId = normalizeContextPackId(input.contextPackId)
   if (contextPackId) context.contextPackId = contextPackId
@@ -268,6 +273,10 @@ function normalizeContextStats(input) {
   if (externalRunRef) context.externalRunRef = externalRunRef
   const outcomeRefs = normalizeOutcomeRefs(input.outcomeRefs)
   if (Object.keys(outcomeRefs).length) context.outcomeRefs = outcomeRefs
+  if (input.taskDecision != null) {
+    try { context.taskDecision = parseTaskDecision(input.taskDecision) }
+    catch { /* malformed task judgments cannot authorize completion */ }
+  }
   const connectorContext = normalizeConnectorContext(input)
   if (connectorContext) Object.assign(context, connectorContext)
   return context

desktop/src/runs/run-harness.cjs

@@ -149,7 +149,7 @@ class RunHarness {
     return next
   }
 
-  beginAgent(kind, round = 0, sourceMessageIds = []) {
+  beginAgent(kind, round = 0, sourceMessageIds = [], options = {}) {
     const safeKind = cleanId(kind)
     if (!safeKind) throw new Error('RUN_HARNESS_AGENT_REQUIRED')
     if (!this.targetKinds.includes(safeKind)) {
@@ -184,6 +184,10 @@ class RunHarness {
       seenSeqs: [],
       eventCursor: 0,
       context: {},
+      // Per-topic dispatch order; the live bubble and the saved reply share it.
+      ...(Number.isSafeInteger(options?.turnSeq) && options.turnSeq > 0
+        ? { turnSeq: options.turnSeq }
+        : {}),
     }
     this.agentRuns.push(run)
     while (this.agentRuns.length > this.maxAgentRuns) {
@@ -328,6 +332,7 @@ class RunHarness {
       seenSeqs: [...run.seenSeqs],
       eventCursor: run.eventCursor,
       context: { ...run.context },
+      ...(run.turnSeq ? { turnSeq: run.turnSeq } : {}),
     }))
   }
 }

desktop/src/runs/run-journal.cjs

@@ -28,8 +28,8 @@ const CONTEXT_FIELDS = new Set([
   'includedCount', 'omittedCount', 'charCount', 'sessionRotated', 'externalRunRef',
   'contextMode', 'promptChars', 'contextPackId', 'contextPackState', 'deliveryRecordIds',
   'sessionProvenance', 'outcomeRefs', 'sourceCount', 'sourceHash', 'promptBytes',
-  'promptHash', 'wirePayloadBytes', 'wirePayloadHash',
-  'connector', 'connectorEventState',
+  'promptHash', 'wirePayloadBytes', 'wirePayloadHash', 'operationId', 'snapshotHash',
+  'connector', 'connectorEventState', 'taskDecision',
 ])
 
 function clone(value) {

desktop/src/runs/run-ledger-records.cjs

@@ -1,4 +1,5 @@
 const { redactSecrets } = require('../security/secret-redaction.cjs')
+const { parseTaskDecision } = require('../collaboration/task-decision.cjs')
 const { normalizeExternalRunRef } = require('../agents/agent-runtime-contract.cjs')
 const { parseConnectorRunSnapshot } = require('../agents/connectors/agent-connector-registry.cjs')
 const { parseRunEventState } = require('./run-event-protocol.cjs')
@@ -42,8 +43,8 @@ const CONTEXT_FIELDS = new Set([
   'includedCount', 'omittedCount', 'charCount', 'sessionRotated', 'externalRunRef',
   'contextMode', 'promptChars', 'contextPackId', 'contextPackState', 'deliveryRecordIds',
   'sessionProvenance', 'outcomeRefs', 'sourceCount', 'sourceHash', 'promptBytes',
-  'promptHash', 'wirePayloadBytes', 'wirePayloadHash',
-  'connector', 'connectorEventState',
+  'promptHash', 'wirePayloadBytes', 'wirePayloadHash', 'operationId', 'snapshotHash',
+  'connector', 'connectorEventState', 'taskDecision',
 ])
 const CONTEXT_PACK_STATES = new Set(['captured', 'legacy-unavailable'])
 const REQUIRED_BUDGET_FIELDS = new Set(['limits', 'used', 'source', 'enforcement', 'startedAt'])
@@ -62,11 +63,11 @@ const CONTINUATION_FIELDS = new Set([
 ])
 const CONTINUATION_INVOCATION_FIELDS = ['phase', 'slotId', 'operationId', 'snapshotHash']
 const CONTINUATION_V4_PHASES = new Set([
-  'proposal', 'challenge', 'work', 'synthesis', 'verification',
+  'proposal', 'discussion', 'challenge', 'work', 'synthesis', 'verification',
 ])
 const CONTINUATION_GATE_TYPES = new Set(['permission', 'budget', 'decision', 'retry', 'input'])
 const CONTINUATION_RESUME_KINDS = new Set([
-  'agent_slot', 'role_review_decision', 'v4_human_gate', 'v4_synthesis_recovery',
+  'agent_slot', 'role_review_decision', 'v4_human_gate', 'v4_synthesis_recovery', 'v4_task_decision',
 ])
 const CONTINUATION_STATES = new Set([
   'pending', 'ready', 'resuming', 'completed', 'failed', 'cancelled',
@@ -75,8 +76,9 @@ const ACTIVE_CONTINUATION_STATES = new Set(['pending', 'ready', 'resuming'])
 const ORCHESTRATION_FIELDS = new Set([
   'version', 'workflow', 'currentKind', 'pendingKinds', 'activeKinds',
   'successfulKinds', 'agreementKinds', 'attachmentRecipients',
-  'totalSuccesses', 'terminalFailureOccurred', 'collaboration', 'template', 'taskGraph',
-  'phase', 'batchId', 'round', 'currentKinds', 'snapshotHash', 'snapshot', 'plan',
+  'totalSuccesses', 'terminalFailureOccurred', 'executionSequence',
+  'collaboration', 'template', 'taskGraph',
+  'phase', 'discussionStyle', 'discussionWriterKind', 'batchId', 'round', 'currentKinds', 'snapshotHash', 'snapshot', 'plan',
   'slots', 'deliveryWatermarks', 'deliveryState', 'commitState', 'challengeBindings', 'synthesisBinding',
   'synthesisRecovery', 'convergence', 'coordinationPlan', 'workReceipts', 'candidateCommit',
 ])
@@ -239,7 +241,7 @@ function normalizeAgentRun(input, parent, fallbackTimestamp) {
     'includedCount', 'omittedCount', 'charCount', 'sessionRotated', 'externalRunRef',
     'contextMode', 'promptChars', 'contextPackId', 'deliveryRecordIds', 'sessionProvenance', 'outcomeRefs',
     'sourceCount', 'sourceHash', 'promptBytes', 'promptHash', 'wirePayloadBytes', 'wirePayloadHash',
-    'connector', 'connectorEventState',
+    'operationId', 'snapshotHash', 'connector', 'connectorEventState', 'taskDecision',
   ].some(key => hasOwn(input.context, key))
   const rawEvents = Array.isArray(input.events) ? input.events : []
   const rawSourceIds = Array.isArray(input.sourceMessageIds) ? input.sourceMessageIds : []
@@ -403,16 +405,18 @@ function normalizeContinuation(input) {
       || !CONTINUATION_STATES.has(state) || !agentRunId || !agentKind
       || (hasPublicAgentRunId && !publicAgentRunId)
       || (resumeKind === 'role_review_decision' && gateType !== 'decision')
+      || (resumeKind === 'v4_task_decision'
+        && (gateType !== 'input' || !hasInvocationBinding || phase !== 'discussion'))
       || (resumeKind === 'v4_human_gate'
         && (gateType !== 'decision' || !hasStateEpoch || stateEpoch < 1))
       || (resumeKind === 'v4_synthesis_recovery'
         && (gateType !== 'decision' || !hasStateEpoch))
       || (!['v4_human_gate', 'v4_synthesis_recovery'].includes(resumeKind) && hasStateEpoch)
       || (invocationFieldCount > 0 && !hasInvocationBinding)
-      || (hasInvocationBinding && (resumeKind !== 'agent_slot'
+      || (hasInvocationBinding && (!['agent_slot', 'v4_task_decision'].includes(resumeKind)
         || !CONTINUATION_V4_PHASES.has(phase) || !slotId || !operationId
         || !SHA256.test(snapshotHash)))
-      || (gateType === 'input' && !hasRequestBinding)
+      || (gateType === 'input' && resumeKind !== 'v4_task_decision' && !hasRequestBinding)
       || (!hasRequestBinding && [requestId, requestHash, sessionRefHash, sessionProvenanceHash]
         .some(Boolean))) return undefined
   return {
@@ -441,7 +445,8 @@ function continuationBindingMatchesOrchestration(continuation, orchestration) {
   }
   const hasBinding = CONTINUATION_INVOCATION_FIELDS.every(field => hasOwn(continuation || {}, field))
   if (!hasBinding) return true
-  if (continuation.resumeKind !== 'agent_slot' || orchestration?.version !== 4) return false
+  if (!['agent_slot', 'v4_task_decision'].includes(continuation.resumeKind)
+      || orchestration?.version !== 4) return false
   if (!['pending', 'ready', 'resuming'].includes(continuation.state)) return true
   const slots = orchestration.slots.filter(slot => slot.slotId === continuation.slotId)
   const orchestrationRound = orchestration.workflow === 'manual'
@@ -489,8 +494,9 @@ function normalizeOrchestration(input) {
   const agreementKinds = normalizeKinds(input.agreementKinds)
   const attachmentRecipients = normalizeKinds(input.attachmentRecipients)
   const totalSuccesses = boundedNumber(input.totalSuccesses, 0, 1000000)
+  const executionSequence = boundedNumber(input.executionSequence, 0, 1000000)
   const v4OnlyFields = [
-    'phase', 'batchId', 'round', 'currentKinds', 'snapshotHash', 'snapshot', 'plan',
+    'phase', 'discussionStyle', 'discussionWriterKind', 'batchId', 'round', 'currentKinds', 'snapshotHash', 'snapshot', 'plan',
     'slots', 'deliveryWatermarks', 'deliveryState', 'commitState', 'challengeBindings',
     'synthesisBinding', 'synthesisRecovery', 'convergence', 'candidateCommit',
   ]
@@ -498,6 +504,10 @@ function normalizeOrchestration(input) {
       || (version === 1 && hasOwn(input, 'collaboration'))
       || (version === 2 && workflow !== 'auto')
       || (version < 3 && (hasOwn(input, 'template') || hasOwn(input, 'taskGraph')))
+      || (hasOwn(input, 'executionSequence')
+        && (!Number.isInteger(input.executionSequence)
+          || input.executionSequence < 0
+          || input.executionSequence > 1000000))
       || (version < 4 && v4OnlyFields.some(field => hasOwn(input, field)))
       || (version === 3 && (workflow !== 'auto' || input.template !== 'task-graph'))) {
     return undefined
@@ -529,6 +539,7 @@ function normalizeOrchestration(input) {
     attachmentRecipients,
     totalSuccesses,
     terminalFailureOccurred: input.terminalFailureOccurred === true,
+    ...(hasOwn(input, 'executionSequence') ? { executionSequence } : {}),
     ...(version >= 2 ? { collaboration } : {}),
     ...(version === 3 ? { template: 'task-graph', taskGraph } : {}),
   }
@@ -694,6 +705,11 @@ function hasValidStoredRecordShape(input) {
             && (typeof agentRun.context[field] !== 'string'
               || !SHA256.test(agentRun.context[field]))) return false
       }
+      if (hasOwn(agentRun.context, 'operationId')
+          && cleanId(agentRun.context.operationId) !== agentRun.context.operationId) return false
+      if (hasOwn(agentRun.context, 'snapshotHash')
+          && (typeof agentRun.context.snapshotHash !== 'string'
+            || !SHA256.test(agentRun.context.snapshotHash))) return false
       if (hasOwn(agentRun.context, 'contextPackId')
           && normalizeContextPackId(agentRun.context.contextPackId)
             !== agentRun.context.contextPackId) return false
@@ -721,6 +737,12 @@ function hasValidStoredRecordShape(input) {
             || canonicalJson(normalizedProvenance)
               !== canonicalJson(agentRun.context.sessionProvenance)) return false
       }
+      if (hasOwn(agentRun.context, 'taskDecision')) {
+        try {
+          if (canonicalJson(parseTaskDecision(agentRun.context.taskDecision))
+              !== canonicalJson(agentRun.context.taskDecision)) return false
+        } catch { return false }
+      }
       if (hasOwn(agentRun.context, 'outcomeRefs')) {
         if (!isRecord(agentRun.context.outcomeRefs)) return false
         const normalizedOutcomeRefs = normalizeOutcomeRefs(agentRun.context.outcomeRefs)

desktop/src/runs/run-ledger.cjs

@@ -311,7 +311,7 @@ class RunLedger {
   persist(runs = this.runs) {
     atomicWritePrivateFile(
       this.storagePath,
-      `${JSON.stringify({ version: STORE_VERSION, runs }, null, 2)}\n`,
+      `${JSON.stringify({ version: STORE_VERSION, runs })}\n`,
     )
   }
 

desktop/src/runs/run-scheduler.cjs

@@ -58,6 +58,7 @@ class RunScheduler {
     this.activeGlobal = 0
     this.activeTasks = new Map()
     this.activeWorkspaces = new Map()
+    this.activeWorkspaceWriters = new Set()
     this.queue = []
     this.sequence = 0
     schedulerPauseTokens.set(this, new Map())
@@ -69,6 +70,8 @@ class RunScheduler {
 
   canGrant(entry) {
     return taskPauseAllows(this, entry)
+      && (entry.permissionMode !== 'workspace-write'
+        || !this.activeWorkspaceWriters.has(entry.workspaceKey))
       && belowLimit(this.activeGlobal, this.limits.global)
       && belowLimit(this.resourceCount(this.activeTasks, entry.taskId), this.limits.task)
       && belowLimit(this.resourceCount(this.activeWorkspaces, entry.workspaceKey), this.limits.workspace)
@@ -84,12 +87,14 @@ class RunScheduler {
     this.activeGlobal += 1
     this.changeCount(this.activeTasks, entry.taskId, 1)
     this.changeCount(this.activeWorkspaces, entry.workspaceKey, 1)
+    if (entry.permissionMode === 'workspace-write') this.activeWorkspaceWriters.add(entry.workspaceKey)
     entry.signal?.removeEventListener('abort', entry.abortHandler)
     let released = false
     const lease = Object.freeze({
       leaseId: entry.requestId,
       taskId: entry.taskId,
       workspaceKey: entry.workspaceKey,
+      permissionMode: entry.permissionMode,
       queuedAt: entry.queuedAt,
       acquiredAt: this.now(),
       release: () => {
@@ -98,6 +103,7 @@ class RunScheduler {
         this.activeGlobal -= 1
         this.changeCount(this.activeTasks, entry.taskId, -1)
         this.changeCount(this.activeWorkspaces, entry.workspaceKey, -1)
+        if (entry.permissionMode === 'workspace-write') this.activeWorkspaceWriters.delete(entry.workspaceKey)
         this.drain()
         return true
       },
@@ -105,6 +111,7 @@ class RunScheduler {
     schedulerLeaseBindings.set(lease, Object.freeze({
       taskId: entry.taskId,
       workspaceKey: entry.workspaceKey,
+      permissionMode: entry.permissionMode,
       signal: entry.signal,
     }))
     entry.resolve(lease)
@@ -129,9 +136,13 @@ class RunScheduler {
   #enqueueAcquire(input = {}, pauseToken = null) {
     let taskId
     let workspaceKey
+    const permissionMode = input.permissionMode == null ? 'read-only' : input.permissionMode
     try {
       taskId = resourceKey(input.taskId, 'RUN_SCHEDULER_TASK_REQUIRED')
       workspaceKey = resourceKey(input.workspaceKey, 'RUN_SCHEDULER_WORKSPACE_REQUIRED')
+      if (!['read-only', 'workspace-write'].includes(permissionMode)) {
+        throw schedulerError('RUN_SCHEDULER_PERMISSION_INVALID')
+      }
     } catch (error) {
       return Promise.reject(error)
     }
@@ -142,6 +153,7 @@ class RunScheduler {
         requestId: `lease-${++this.sequence}`,
         taskId,
         workspaceKey,
+        permissionMode,
         queuedAt: this.now(),
         signal,
         resolve,
@@ -176,6 +188,7 @@ class RunScheduler {
       get leaseId() { return activeLease?.leaseId || '' },
       get taskId() { return activeLease?.taskId || '' },
       get workspaceKey() { return activeLease?.workspaceKey || '' },
+      get permissionMode() { return activeLease?.permissionMode || '' },
       get queuedAt() { return activeLease?.queuedAt || 0 },
       get acquiredAt() { return activeLease?.acquiredAt || 0 },
       release: () => {

desktop/src/shell/main-ipc.cjs

@@ -185,8 +185,6 @@ function registerDesktopIpc(options) {
     loadKnowledgeBaseSources,
     localAgentCatalog,
     openExternalUrl,
-    providerStore,
-    providerAgentKind,
     refreshLocalAgentState,
     registerTrustedHandle,
     skillCatalog,
@@ -660,39 +658,6 @@ function registerDesktopIpc(options) {
   registerTrustedHandle('local-attachments:preview', id => attachments.preview(String(id || '')))
   registerTrustedHandle('local-attachments:open', id => attachments.open(String(id || '')))
   registerTrustedHandle('local-attachments:discard', ids => attachments.discardUnreferenced(ids))
-  registerTrustedHandle('local-agent-provider:status', kind => (
-    providerStore.status(providerAgentKind(kind))
-  ))
-  registerTrustedHandle('local-agent-provider:probe', kind => (
-    providerStore.status(providerAgentKind(kind), { probeEncryption: true })
-  ))
-  registerTrustedHandle('local-agent-provider:save', async (kind, input) => {
-    const selectedKind = providerAgentKind(kind)
-    const result = providerStore.save(selectedKind, {
-      apiKey: input?.apiKey,
-      provider: input?.provider,
-      baseUrl: input?.baseUrl,
-      model: input?.model,
-      preset: input?.preset,
-    })
-    getWorkspace()?.markRuntimeCredential(selectedKind, 'unknown')
-    await refreshLocalAgentState()
-    return result
-  })
-  registerTrustedHandle('local-agent-provider:activate', async (kind, preset) => {
-    const selectedKind = providerAgentKind(kind)
-    const result = providerStore.activate(selectedKind, preset)
-    getWorkspace()?.markRuntimeCredential(selectedKind, 'unknown')
-    await refreshLocalAgentState()
-    return result
-  })
-  registerTrustedHandle('local-agent-provider:delete', async (kind, preset) => {
-    const selectedKind = providerAgentKind(kind)
-    const result = providerStore.delete(selectedKind, preset)
-    getWorkspace()?.markRuntimeCredential(selectedKind, 'unknown')
-    await refreshLocalAgentState()
-    return result
-  })
   registerTrustedHandle('local-knowledge-connector:list', (...args) => {
     if (args.length) throw new Error('LOCAL_KNOWLEDGE_CONNECTOR_REQUEST_INVALID')
     const connectors = getKnowledgeConnectors?.()

desktop/src/workspace/invocation-budget.cjs

@@ -0,0 +1,37 @@
+const { outboundWirePayloadBytes } = require('../collaboration/outbound-payload.cjs')
+
+// What the Agent was sent: wire bytes are reported, tokens are an estimate.
+function recordOutboundUsage(budget, outbound) {
+  if (!budget) return
+  const bytes = outboundWirePayloadBytes(outbound)
+  // A payload that was not built by the transport has no recorded wire size;
+  // skipping it is better than failing the Agent's turn over bookkeeping.
+  if (!bytes) return
+  budget.addUsageBatch([
+    { dimension: 'outboundBytes', amount: bytes.length, source: 'reported' },
+    { dimension: 'inputTokens', amount: Math.ceil(bytes.length / 4), source: 'estimated' },
+  ])
+}
+
+// What the Agent produced: prefer the usage the CLI reported over an estimate.
+function recordReplyUsage(budget, { usage, replyText, toolCallCount }) {
+  if (!budget) return
+  const reportedOutputTokens = Number(usage?.outputTokens)
+  const hasReportedOutputTokens = Number.isSafeInteger(reportedOutputTokens)
+    && reportedOutputTokens >= 0
+  const reportedCostMicros = Number(usage?.costMicros)
+  budget.addUsageBatch([
+    {
+      dimension: 'outputTokens',
+      amount: hasReportedOutputTokens ? reportedOutputTokens : Math.ceil(replyText.length / 4),
+      source: hasReportedOutputTokens ? 'reported' : 'estimated',
+    },
+    { dimension: 'toolCalls', amount: toolCallCount, source: 'estimated' },
+    ...(Number.isSafeInteger(reportedCostMicros) && reportedCostMicros >= 0
+      ? [{ dimension: 'costMicros', amount: reportedCostMicros, source: 'reported' }]
+      : []),
+    { dimension: 'elapsedMs', value: budget.elapsedValue(), source: 'reported' },
+  ])
+}
+
+module.exports = { recordOutboundUsage, recordReplyUsage }

desktop/src/workspace/invocation-failure.cjs

@@ -0,0 +1,72 @@
+const { agentStoppedError } = require('./local-workspace-inputs.cjs')
+
+// Marks how far an invocation got before it failed, so the run can decide
+// whether the operation may be retried safely.
+function attachInvocationFailure(error, input) {
+  if (!error || (typeof error !== 'object' && typeof error !== 'function')) return error
+  const existing = error.invocationFailure && typeof error.invocationFailure === 'object'
+    ? error.invocationFailure
+    : {}
+  Object.defineProperty(error, 'invocationFailure', {
+    value: Object.freeze({
+      outcomeCertainty: existing.outcomeCertainty || input.outcomeCertainty,
+      sideEffectsPossible: existing.sideEffectsPossible === true || input.sideEffectsPossible === true,
+      operationId: existing.operationId || input.operationId,
+      idempotencyMode: existing.idempotencyMode || input.idempotencyMode,
+    }),
+    enumerable: false,
+    configurable: true,
+  })
+  return error
+}
+
+// Hidden properties on an error that downstream code reads but that must not
+// leak into serialised error messages.
+function attachErrorMetadata(error, metadata) {
+  if (!error || (typeof error !== 'object' && typeof error !== 'function')) return error
+  for (const [name, value] of Object.entries(metadata)) {
+    if (!value) continue
+    Object.defineProperty(error, name, { value, enumerable: false, configurable: true })
+  }
+  return error
+}
+
+// Why an invocation ended and how to report it:
+//  - the parent run was stopped, timed out or is shutting down;
+//  - the watchdog cut a silent Agent off;
+//  - the Agent was cancelled on its own (e.g. a superseded attempt);
+//  - or it simply failed with `caughtError`.
+// `settleOutstanding` tells the caller to wait briefly for capture and run
+// promises that may still be winding down.
+function resolveInvocationStop({
+  caughtError, signal, parentAbortObserved, activeRun, agentController, watchdog,
+}) {
+  const parentTimedOut = Boolean(signal?.aborted && activeRun?.stopReason === 'timeout')
+  const parentStopped = Boolean(signal?.aborted || parentAbortObserved)
+  const parentInterrupted = parentStopped && activeRun?.stopReason === 'shutdown'
+  const agentCancelled = agentController.signal.aborted && !parentStopped && !watchdog.timedOut
+  let error = caughtError
+  let status = 'failed'
+  if (parentTimedOut) {
+    error = new Error('LOCAL_AGENT_TIMEOUT')
+    status = 'timeout'
+  } else if (parentStopped) {
+    error = caughtError?.message === 'LOCAL_AGENT_EXECUTION_STOPPED'
+      ? caughtError
+      : agentStoppedError()
+    status = parentInterrupted ? 'interrupted' : 'stopped'
+  } else if (watchdog.timedOut) {
+    error = watchdog.error || new Error('LOCAL_AGENT_TIMEOUT')
+    status = 'timeout'
+  } else if (agentCancelled) {
+    error = agentStoppedError()
+    status = 'stopped'
+  }
+  return {
+    error,
+    status,
+    settleOutstanding: parentStopped || watchdog.timedOut || agentCancelled,
+  }
+}
+
+module.exports = { attachErrorMetadata, attachInvocationFailure, resolveInvocationStop }

desktop/src/workspace/invocation-outputs.cjs

@@ -0,0 +1,113 @@
+const {
+  MAX_MESSAGE_ATTACHMENTS,
+  abortableOperation,
+  normalizeAttachmentMetadata,
+} = require('./local-workspace-inputs.cjs')
+
+// Absolute paths an Agent mentions in its reply. Quoted, backticked or
+// Markdown-linked paths may contain spaces; bare paths end at whitespace.
+function extractReportedOutputPaths(text) {
+  if (typeof text !== 'string') return []
+  const root = '\\/(?:Users|home|tmp|private\\/tmp|Volumes)\\/'
+  const delimited = [...text.matchAll(new RegExp(`[\`"'(<](${root}[^\`"'<>\\n)]+)[\`"')>]`, 'g'))]
+    .map(match => match[1])
+  const bare = (text.match(new RegExp(`${root}[^\\s"'<>\`\\]\\[)]+`, 'g')) || [])
+    .filter(value => !delimited.some(full => full !== value && full.startsWith(value)))
+  return [...new Set([...delimited, ...bare].map(value => value.trim().replace(/[.,;:!?]+$/, '')))]
+}
+
+// Files a writable Agent turn leaves behind: a snapshot of the workspace is
+// taken before the turn and compared afterwards to import generated media and
+// to describe produced artifacts. Capturing is best effort; only a stop
+// request is allowed to interrupt it.
+class InvocationOutputs {
+  constructor({ enabled, workdir, signal, agentKind, startedAt, deps }) {
+    this.enabled = enabled === true
+    this.workdir = workdir
+    this.signal = signal
+    this.agentKind = agentKind
+    this.startedAt = startedAt
+    this.deps = deps
+    this.baseline = null
+    this.artifactBaseline = null
+    this.capturePromise = null
+    this.artifactCapturePromise = null
+    this.importPromise = null
+    this.outcomePromise = null
+  }
+
+  // Promises that may still be running when the invocation is cut short.
+  pending() {
+    return [
+      this.capturePromise,
+      this.artifactCapturePromise,
+      this.importPromise,
+      this.outcomePromise,
+    ].filter(Boolean)
+  }
+
+  async tolerate(operation) {
+    try {
+      return await abortableOperation(operation, this.signal)
+    } catch (error) {
+      if (this.signal.aborted) throw error
+      return undefined
+    }
+  }
+
+  async begin() {
+    if (!this.enabled) return
+    this.artifactCapturePromise = Promise.resolve().then(
+      () => this.deps.captureArtifactOutputs(this.workdir, { signal: this.signal }),
+    )
+    this.artifactCapturePromise.catch(() => {})
+    try {
+      this.capturePromise = Promise.resolve().then(
+        () => this.deps.captureAgentOutputs(this.workdir, { signal: this.signal }),
+      )
+      this.capturePromise.catch(() => {})
+      this.baseline = await abortableOperation(() => this.capturePromise, this.signal)
+    } catch (error) {
+      if (this.signal.aborted) throw error
+    }
+    this.artifactBaseline = (await this.tolerate(() => this.artifactCapturePromise)) ?? null
+  }
+
+  // Media the Agent produced during this turn, as message attachments.
+  async importMedia(replyText) {
+    if (!this.enabled) return []
+    this.importPromise = Promise.resolve().then(() => this.deps.importAgentOutputs({
+      workdir: this.workdir,
+      baseline: this.baseline,
+      startedAt: this.startedAt,
+      agentKind: this.agentKind,
+      reportedPaths: extractReportedOutputPaths(replyText),
+      signal: this.signal,
+    }))
+    this.importPromise.catch(() => {})
+    const imported = await this.tolerate(() => this.importPromise)
+    return (Array.isArray(imported) ? imported : [])
+      .slice(0, MAX_MESSAGE_ATTACHMENTS)
+      .map(normalizeAttachmentMetadata)
+      .filter(Boolean)
+  }
+
+  // Descriptors of the artifacts the Agent wrote, for the evidence record.
+  async describeOutcomes({ runId, agentRunId }) {
+    if (!this.enabled || !this.artifactBaseline) return []
+    this.outcomePromise = Promise.resolve().then(() => this.deps.captureAgentOutcomeDescriptors({
+      workdir: this.workdir,
+      baseline: this.artifactBaseline,
+      startedAt: this.startedAt,
+      agentKind: this.agentKind,
+      runId,
+      agentRunId,
+      signal: this.signal,
+    }))
+    this.outcomePromise.catch(() => {})
+    const captured = await this.tolerate(() => this.outcomePromise)
+    return Array.isArray(captured) ? captured : []
+  }
+}
+
+module.exports = { InvocationOutputs, extractReportedOutputPaths }

desktop/src/workspace/invocation-prompt.cjs

@@ -0,0 +1,63 @@
+const { stagedAgentInputPrompt } = require('../agents/agent-input-staging.cjs')
+const { directPromptFor } = require('./local-workspace-context.cjs')
+const { cleanText } = require('./local-workspace-inputs.cjs')
+
+const RESPONSE_SCOPE_INSTRUCTION = [
+  'Final response scope:',
+  'Answer only the current user task stated above.',
+  'Use earlier messages only as supporting evidence when they help answer that current task.',
+  'Do not separately mention, summarize, continue, or complete any previous user task.',
+  'Do not append an answer to an older task after the current-task conclusion.',
+].join('\n')
+
+// Builds the prompt one Agent receives for one turn.
+//  - frozen / isolated invocations send exactly the approved prompt.
+//  - direct chats lead with the user's request; a resumed native session
+//    already holds the conversation, a new or rebuilt one gets the history.
+//  - group turns use the V4 prompt when the orchestrator supplied one, else
+//    the workspace prompt for the current mode.
+// Staged attachments and harness recovery instructions are appended to both.
+function assembleInvocationPrompt({
+  group, kind, mode, threadRootId, context, activeRun, isolated, frozen,
+  v4Prompt, promptMode, afterKind, packed, session, stagedInputs, promptFor, agentLabel,
+}) {
+  // Frozen and isolated prompts are fixed text. Staged attachments are the
+  // only thing added to a frozen one, so the Agent can open the user's files.
+  if (isolated || frozen) {
+    return [v4Prompt, stagedAgentInputPrompt(stagedInputs)].filter(Boolean).join('\n\n')
+  }
+  const runtimeInstruction = cleanText(context.runtimeInstruction, 3000)
+  const regenerationInstruction = cleanText(context.regenerationInstruction, 3000)
+  const additions = [
+    stagedAgentInputPrompt(stagedInputs),
+    runtimeInstruction ? `Harness recovery task:\n${runtimeInstruction}` : '',
+    regenerationInstruction ? `Fresh response instruction:\n${regenerationInstruction}` : '',
+  ]
+  if (group.conversationType === 'direct') {
+    return [
+      directPromptFor({
+        group,
+        label: agentLabel(kind),
+        packed,
+        skillHints: context.skillHints || [],
+        knowledgeBaseHints: context.knowledgeBaseHints || [],
+        includeHistory: session.rotated || !session.ref,
+      }),
+      ...additions,
+    ].filter(Boolean).join('\n\n')
+  }
+  return [
+    context.v4 === true && v4Prompt
+      ? v4Prompt
+      : promptFor(
+        group, kind, context.completionPolicy === 'typed' ? 'manual' : mode,
+        threadRootId, context.skillHints || [],
+        context.knowledgeBaseHints || [], afterKind, packed, promptMode,
+        activeRun?.unlimitedRounds === true,
+      ),
+    ...additions,
+    RESPONSE_SCOPE_INSTRUCTION,
+  ].filter(Boolean).join('\n')
+}
+
+module.exports = { assembleInvocationPrompt }

desktop/src/workspace/invocation-session.cjs

@@ -0,0 +1,156 @@
+const { createHash, randomUUID } = require('node:crypto')
+const { canonicalJson } = require('../collaboration/context-pack-records.cjs')
+const { nextSessionMeta, normalizeSessionMeta } = require('../runs/run-harness.cjs')
+
+const MAX_INHERITED_TASK_IDS = 64
+
+function sha256(value) {
+  return createHash('sha256').update(String(value || '')).digest('hex')
+}
+
+function createdSessionProvenance(taskId, stateless = false) {
+  if (stateless) {
+    return {
+      scope: 'none', reuse: false, origin: 'none', originTaskId: null,
+      inheritedTaskIds: [], completeness: 'complete',
+    }
+  }
+  return {
+    scope: 'conversation',
+    reuse: false,
+    origin: 'created',
+    originTaskId: taskId || null,
+    inheritedTaskIds: [],
+    completeness: 'complete',
+  }
+}
+
+function provenanceMeta(provenance) {
+  if (provenance.scope === 'unknown-legacy') {
+    return {
+      sessionScope: 'unknown-legacy',
+      originTaskId: '',
+      inheritedTaskIds: [],
+      provenanceCompleteness: 'unknown-legacy',
+    }
+  }
+  if (provenance.scope === 'none') return {}
+  return {
+    sessionScope: provenance.scope,
+    originTaskId: provenance.originTaskId || '',
+    inheritedTaskIds: provenance.inheritedTaskIds || [],
+    provenanceCompleteness: provenance.completeness,
+  }
+}
+
+function completedSessionMeta(meta, provenance, taskId, usage) {
+  let inheritedTaskIds = [...(provenance.inheritedTaskIds || [])]
+  let completeness = provenance.completeness
+  if (provenance.scope === 'conversation' && taskId
+      && taskId !== provenance.originTaskId && !inheritedTaskIds.includes(taskId)) {
+    inheritedTaskIds.push(taskId)
+    if (inheritedTaskIds.length > MAX_INHERITED_TASK_IDS) {
+      inheritedTaskIds = inheritedTaskIds.slice(-MAX_INHERITED_TASK_IDS)
+      completeness = 'partial'
+    }
+  }
+  return nextSessionMeta(meta, {
+    ...usage,
+    ...provenanceMeta({ ...provenance, inheritedTaskIds, completeness }),
+  })
+}
+
+// Binds a pending permission or connector request to the exact native session
+// it was raised in, so a decision cannot be replayed against another session.
+function connectorSessionBinding(sessionRef, sessionProvenance) {
+  const stableProvenance = {
+    scope: String(sessionProvenance?.scope || 'none'),
+    originTaskId: String(sessionProvenance?.originTaskId || ''),
+    inheritedTaskIds: Array.isArray(sessionProvenance?.inheritedTaskIds)
+      ? [...sessionProvenance.inheritedTaskIds]
+      : [],
+    completeness: String(sessionProvenance?.completeness || 'complete'),
+  }
+  return {
+    sessionRefHash: sha256(sessionRef),
+    sessionProvenanceHash: sha256(canonicalJson(stableProvenance)),
+  }
+}
+
+// The native Agent session one invocation runs in. A conversation owns one
+// native session for its lifetime: it is resumed on every turn and rebuilt
+// only when the Agent reports that the stored session is no longer valid.
+class InvocationSession {
+  // `resolved` is what the workspace stored for this conversation; `deps`
+  // carries the state hooks so this class never touches workspace state itself.
+  constructor({ resolved, group, kind, taskId, deps }) {
+    this.group = group
+    this.kind = kind
+    this.taskId = taskId
+    this.deps = deps
+    this.key = resolved.key
+    this.storedRef = String(resolved.sessionRef || '')
+    this.ref = this.storedRef
+    this.meta = normalizeSessionMeta(resolved.sessionMeta)
+    this.provenance = resolved.provenance
+    this.rotated = resolved.sessionReset === true
+    this.transport = this.ref ? String(this.meta.transport || '') : ''
+  }
+
+  binding() {
+    return connectorSessionBinding(this.ref, this.provenance)
+  }
+
+  // Forget the native session so the next turn starts a fresh one. OpenClaw
+  // addresses its sessions by name, so it gets a new name straight away.
+  discard() {
+    this.meta = {}
+    this.provenance = createdSessionProvenance(this.taskId)
+    this.transport = ''
+    this.rotated = true
+    this.deps.commitSessionState((nextState) => {
+      delete nextState.sessions[this.key]
+      delete nextState.sessionMeta[this.key]
+      if (this.kind === 'openclaw') {
+        const generation = randomUUID().replace(/[^a-zA-Z0-9]/g, '').slice(0, 12) || 'session'
+        this.ref = this.deps.openClawSessionRef(
+          this.group, generation, this.group.conversationType === 'direct' ? '' : this.taskId,
+        )
+        nextState.sessions[this.key] = this.ref
+        nextState.sessionMeta[this.key] = normalizeSessionMeta(provenanceMeta(this.provenance))
+      } else {
+        this.ref = ''
+      }
+    })
+  }
+
+  // The Agent reported (or created) its native session id.
+  adopt(nextRef, metadata = {}) {
+    const transport = ['legacy', 'acp'].includes(metadata?.transport) ? metadata.transport : ''
+    if (transport) this.transport = transport
+    this.ref = String(nextRef || '')
+    this.deps.persistSessionState(this.key, this.ref, normalizeSessionMeta({
+      ...this.meta,
+      ...provenanceMeta(this.provenance),
+      ...(this.transport ? { transport: this.transport } : {}),
+    }))
+  }
+
+  // A turn finished: record what it cost and which message it had seen.
+  complete(resultRef, usage) {
+    this.deps.persistSessionState(
+      this.key,
+      resultRef || this.ref,
+      completedSessionMeta(this.meta, this.provenance, this.taskId, usage),
+    )
+  }
+}
+
+module.exports = {
+  InvocationSession,
+  completedSessionMeta,
+  connectorSessionBinding,
+  createdSessionProvenance,
+  provenanceMeta,
+  sha256,
+}

desktop/src/workspace/local-workspace-agent-catalog.cjs

@@ -3,6 +3,21 @@ const {
 } = require('../agents/agent-runtime-contract.cjs')
 
 const RECENT_VERIFICATION_MS = 24 * 60 * 60 * 1000
+const AUTH_FAILURE_RETRY_MS = 30 * 1000
+// A readiness probe that times out or crashes says nothing about whether the
+// Agent works: it is not a "not signed in" answer. An Agent that was usable
+// stays usable through such results for this long; a definitive answer
+// (signed out, incompatible, a recorded auth failure) applies immediately.
+const INDETERMINATE_PROBE_GRACE_MS = 10 * 60 * 1000
+const INDETERMINATE_PROBE_SOURCES = new Set(['native-runtime-unavailable', 'native-probe-failed'])
+
+function expiredAuthFailure(runtime, now) {
+  if (runtime?.credentialState !== 'missing') return false
+  const checkedAt = Date.parse(String(runtime.checkedAt || ''))
+  const current = Date.parse(String(now || ''))
+  return !Number.isFinite(checkedAt) || (Number.isFinite(current)
+    && (current < checkedAt || current - checkedAt >= AUTH_FAILURE_RETRY_MS))
+}
 
 function agentVersionIdentified(agent) {
   if (agent?.custom === true) return true
@@ -34,7 +49,6 @@ class LocalWorkspaceAgentCatalog {
     this.setDetectedAgents = options.setDetectedAgents
     this.detectAgents = options.detectAgents
     this.credentialState = options.credentialState
-    this.sharedProviderReady = options.sharedProviderReady
     this.attachmentSupport = options.attachmentSupport || (() => ({}))
     this.save = options.save
     this.emitChanged = options.emitChanged
@@ -42,6 +56,7 @@ class LocalWorkspaceAgentCatalog {
     this.now = options.now
     this.refreshTask = null
     this.refreshAgain = false
+    this.indeterminateSince = new Map()
   }
 
   async refresh() {
@@ -66,27 +81,30 @@ class LocalWorkspaceAgentCatalog {
 
   async refreshOnce() {
     const runtimeAtRefreshStart = new Map(Object.entries(this.state().agentRuntime))
+    const previousAgents = new Map(this.detectedAgents().map(agent => [agent.kind, agent]))
     const detected = await this.detectAgents()
-    const nativeStates = await Promise.all(detected.map((agent) => {
-      const runtime = runtimeAtRefreshStart.get(agent.kind)
-      const sharedProviderReady = Boolean(this.sharedProviderReady(agent.kind))
-      if (sharedProviderReady && runtime?.credentialState === 'missing') {
-        return { state: 'missing', source: 'runtime-auth-failure' }
+    const nativeStates = await Promise.all(detected.map(async (agent) => {
+      try {
+        return await this.credentialState(agent.kind, agent)
+      } catch {
+        return { state: 'unknown', source: 'native-probe-failed' }
       }
-      if (sharedProviderReady) {
-        return { state: 'ready', source: 'shared-provider' }
-      }
-      return this.credentialState(agent.kind, agent)
     }))
     const state = this.state()
     let recoveredRuntimeCredential = false
     const agents = detected.map((agent, index) => {
       const native = nativeStates[index]
       let runtime = state.agentRuntime[agent.kind]
-      const sharedProviderReady = Boolean(this.sharedProviderReady(agent.kind))
       const nativeState = ['ready', 'missing'].includes(native?.state) ? native.state : 'unknown'
       const authoritativeNativeState = native?.source === 'native-auth-status'
-      const sharedProviderRequired = native?.source === 'shared-provider-required'
+      // Give corrected native credentials a retry after the cooldown, never erase a newer failure.
+      if (expiredAuthFailure(runtime, this.now())
+          && !(authoritativeNativeState && nativeState === 'ready')
+          && runtime === runtimeAtRefreshStart.get(agent.kind)) {
+        delete state.agentRuntime[agent.kind]
+        runtime = undefined
+        recoveredRuntimeCredential = true
+      }
       // A probe that started before a newer runtime failure cannot clear that failure.
       if (authoritativeNativeState && nativeState === 'ready'
           && ['missing', 'unknown'].includes(runtime?.credentialState)
@@ -106,30 +124,36 @@ class LocalWorkspaceAgentCatalog {
       const nativeReadySource = nativeState === 'ready'
         ? (native.source || 'native-credential')
         : ''
+      const retained = this.retainedAvailability({
+        agent, previous: previousAgents.get(agent.kind), native, runtimeMissing,
+      })
       let credentialState = 'unknown'
-      if (runtimeMissing || sharedProviderRequired || nativeState === 'missing') {
+      if (runtimeMissing || nativeState === 'missing') {
         credentialState = 'missing'
       }
-      if (!runtimeMissing && sharedProviderReady) credentialState = 'ready'
-      else if (!runtimeMissing && nativeState === 'ready') credentialState = 'ready'
+      if (!runtimeMissing && nativeState === 'ready') credentialState = 'ready'
       else if (!runtimeMissing && verifiedReady && nativeState !== 'missing') credentialState = 'ready'
+      if (retained && retained.credentialState === 'ready') credentialState = 'ready'
       const installed = true
       const versionIdentified = agentVersionIdentified(agent)
       const compatible = agentCompatible(agent)
       const nativeConfigurationUnknown = !runtimeMissing
         && nativeState === 'unknown'
         && native?.source !== 'native-runtime-unavailable'
-      const configured = sharedProviderReady
+        && native?.source !== 'native-probe-failed'
+      const configured = retained !== null
         || nativeState === 'ready'
         || native?.source === 'native-auth-status'
         || native?.source === 'native-runtime-unavailable'
         || verifiedReady
         || runtimeMissing
         || nativeConfigurationUnknown
-      const authenticated = !runtimeMissing && nativeState !== 'missing' && (
-        sharedProviderReady || nativeState === 'ready' || verifiedReady || nativeConfigurationUnknown
-      )
-      const runtimePrerequisitesReady = native?.source !== 'native-runtime-unavailable'
+      const authenticated = retained !== null || (!runtimeMissing && nativeState !== 'missing' && (
+        nativeState === 'ready' || verifiedReady || nativeConfigurationUnknown
+      ))
+      const runtimePrerequisitesReady = retained !== null
+        || (native?.source !== 'native-runtime-unavailable'
+          && (native?.source !== 'native-probe-failed' || recentlyVerified(runtime, this.now())))
       const invocable = installed && versionIdentified && compatible && configured
         && authenticated && runtimePrerequisitesReady
       const verifiedRecently = (authoritativeNativeState && nativeState === 'ready')
@@ -143,11 +167,12 @@ class LocalWorkspaceAgentCatalog {
       let availabilitySource = 'unverified'
       if (!compatible) availabilitySource = 'incompatible'
       else if (runtimeMissing) availabilitySource = 'runtime-auth-failure'
-      else if (sharedProviderReady) availabilitySource = nativeReadySource || 'shared-provider'
       else if (nativeState === 'missing') availabilitySource = native.source || 'none'
+      else if (native?.source === 'native-probe-failed') availabilitySource = 'native-probe-failed'
       else if (nativeReadySource) availabilitySource = nativeReadySource
       else if (verifiedReady) availabilitySource = 'verified-run'
       else if (nativeConfigurationUnknown) availabilitySource = 'local-cli'
+      if (retained) availabilitySource = retained.availabilitySource
       return {
         ...agent,
         installed,
@@ -163,8 +188,7 @@ class LocalWorkspaceAgentCatalog {
         task: capabilities.task,
         resumable: capabilities.resumable,
         capabilities,
-        showInSidebar: available
-          && (typeof preferred === 'boolean' ? preferred : true),
+        showInSidebar: typeof preferred === 'boolean' ? preferred : true,
       }
     })
     this.setDetectedAgents(agents)
@@ -173,13 +197,33 @@ class LocalWorkspaceAgentCatalog {
     return this.snapshot()
   }
 
+  // The state an Agent keeps while a readiness probe is inconclusive, or null
+  // when the probe result must be taken at face value.
+  retainedAvailability({ agent, previous, native, runtimeMissing }) {
+    if (!INDETERMINATE_PROBE_SOURCES.has(native?.source)) {
+      this.indeterminateSince.delete(agent.kind)
+      return null
+    }
+    const nowMs = Date.parse(String(this.now()))
+    if (!this.indeterminateSince.has(agent.kind) && Number.isFinite(nowMs)) {
+      this.indeterminateSince.set(agent.kind, nowMs)
+    }
+    const since = this.indeterminateSince.get(agent.kind)
+    if (previous?.available !== true || runtimeMissing || !Number.isFinite(nowMs)
+        || !Number.isFinite(since) || nowMs - since >= INDETERMINATE_PROBE_GRACE_MS
+        || !agentVersionIdentified(agent) || !agentCompatible(agent)) return null
+    return {
+      credentialState: previous.credentialState,
+      availabilitySource: previous.availabilitySource || 'native-probe-failed',
+    }
+  }
+
   setSidebarVisibility(kind, visible) {
     const agent = this.detectedAgents().find(item => item.kind === kind)
     if (!agent) throw new Error('LOCAL_AGENT_NOT_INSTALLED')
-    if (visible && !agent.available) throw new Error('LOCAL_AGENT_UNAVAILABLE')
     const state = this.state()
     state.agentPreferences[kind] = { showInSidebar: Boolean(visible) }
-    agent.showInSidebar = agent.available && Boolean(visible)
+    agent.showInSidebar = Boolean(visible)
     this.save()
     this.emitChanged()
     return this.snapshot()
@@ -191,8 +235,9 @@ class LocalWorkspaceAgentCatalog {
       credentialState,
       checkedAt: this.now(),
     }
+    if (credentialState === 'ready') this.indeterminateSince.delete(kind)
     const agent = this.detectedAgents().find(item => item.kind === kind)
-    if (agent) {
+    if (agent && !(credentialState === 'unknown' && agent.authenticated === true)) {
       agent.credentialState = credentialState
       agent.versionIdentified = agentVersionIdentified(agent)
       agent.compatible = agentCompatible(agent)
@@ -210,8 +255,7 @@ class LocalWorkspaceAgentCatalog {
             ? 'runtime-auth-failure'
             : 'unverified'
       const preferred = state.agentPreferences[kind]?.showInSidebar
-      agent.showInSidebar = agent.available
-        && (typeof preferred === 'boolean' ? preferred : true)
+      agent.showInSidebar = typeof preferred === 'boolean' ? preferred : true
     }
     this.save()
     this.emitChanged()

desktop/src/workspace/local-workspace-auto-runner-collaboration.cjs

@@ -0,0 +1,261 @@
+// Collaboration records written while an automatic run progresses (roles, objectives, results).
+const { createHash } = require('node:crypto')
+const {
+  appendBlackboardEntry,
+  appendHandoff,
+  collaborationPackageText,
+  emptyCollaborationState,
+  publicCollaborationText,
+  roleForIndex,
+  visibleBlackboardEntries,
+} = require('../collaboration/collaboration-records.cjs')
+const { createTaskGraphCursor } = require('../collaboration/task-graph-records.cjs')
+
+const collaborationMethods = {
+  orchestrationCursor(targetKinds, taskGraph = null) {
+    if (taskGraph) {
+      return {
+        version: 3,
+        workflow: 'auto',
+        template: 'task-graph',
+        currentKind: '',
+        pendingKinds: [...targetKinds],
+        activeKinds: [...targetKinds],
+        successfulKinds: [],
+        agreementKinds: [],
+        attachmentRecipients: [],
+        totalSuccesses: 0,
+        terminalFailureOccurred: false,
+        collaboration: emptyCollaborationState(),
+        taskGraph: createTaskGraphCursor(taskGraph),
+      }
+    }
+    return {
+      version: 2,
+      workflow: 'auto',
+      currentKind: '',
+      pendingKinds: [...targetKinds],
+      activeKinds: [...targetKinds],
+      successfulKinds: [],
+      agreementKinds: [],
+      attachmentRecipients: [],
+      totalSuccesses: 0,
+      terminalFailureOccurred: false,
+      collaboration: emptyCollaborationState(),
+    }
+  },
+
+  collaborationRole(activeKinds, kind) {
+    return roleForIndex(Math.max(0, activeKinds.indexOf(kind)))
+  },
+
+  collaborationAudience(role) {
+    return {
+      roles: role === 'primary'
+        ? ['reviewer', 'arbiter']
+        : (role === 'reviewer' ? ['primary', 'arbiter'] : ['primary', 'reviewer']),
+      agentKinds: [],
+    }
+  },
+
+  collaborationAcceptance(role) {
+    if (role === 'reviewer') {
+      return [
+        'Check the selected claims and outcome references.',
+        'Identify unsupported assumptions, conflicts, or missing evidence.',
+      ]
+    }
+    if (role === 'arbiter') {
+      return [
+        'Resolve visible conflicts using the selected evidence.',
+        'Return one bounded conclusion and state whether consensus is reached.',
+      ]
+    }
+    return [
+      'Produce one concrete conclusion for the current task.',
+      'Publish only claims and durable outcome references needed downstream.',
+    ]
+  },
+
+  collaborationExpectedOutput(role) {
+    if (role === 'reviewer') return 'A concise review with corrections or explicit acceptance.'
+    if (role === 'arbiter') return 'A final adjudication of the selected claims and conflicts.'
+    return 'A concise primary conclusion with any durable Artifact or Evidence references.'
+  },
+
+  collaborationObjective(group, controller, threadRootId) {
+    const root = this.state().messages.find(message => (
+      message.groupId === group.id && message.id === threadRootId && message.role === 'user'
+    ))
+    return publicCollaborationText(root?.content || 'Complete the current user task.', 3000)
+      || 'Complete the current user task.'
+  },
+
+  prepareCollaborationPackage(
+    group, controller, activeKinds, kind, threadRootId, options = {},
+  ) {
+    const collaboration = controller.orchestration?.collaboration || emptyCollaborationState()
+    const destination = {
+      agentKind: kind,
+      role: options.role || this.collaborationRole(activeKinds, kind),
+    }
+    const selectedEntryIds = Array.isArray(options.selectedEntryIds)
+      ? new Set(options.selectedEntryIds)
+      : null
+    const selectedEntries = selectedEntryIds
+      ? collaboration.entries.filter(entry => selectedEntryIds.has(entry.entryId))
+      : visibleBlackboardEntries(collaboration, destination)
+    const objective = this.collaborationObjective(group, controller, threadRootId)
+    const expectedOutput = options.expectedOutput
+      || this.collaborationExpectedOutput(destination.role)
+    const acceptanceCriteria = options.acceptanceCriteria
+      || this.collaborationAcceptance(destination.role)
+    const selectedIds = selectedEntries.map(entry => entry.entryId)
+    const reusable = [...collaboration.handoffs].reverse().find(handoff => (
+      handoff.destination.agentKind === destination.agentKind
+      && handoff.destination.role === destination.role
+      && handoff.objective === objective
+      && handoff.expectedOutput === expectedOutput
+      && handoff.selectedEntryIds.length === selectedIds.length
+      && handoff.selectedEntryIds.every((id, index) => id === selectedIds[index])
+      && handoff.acceptanceCriteria.length === acceptanceCriteria.length
+      && handoff.acceptanceCriteria.every((value, index) => value === acceptanceCriteria[index])
+    ))
+    if (reusable) {
+      return {
+        handoff: reusable,
+        entries: selectedEntries,
+        text: collaborationPackageText(reusable, selectedEntries),
+      }
+    }
+    const previousOwner = [...collaboration.entries].reverse().find(entry => (
+      entry.owner.type === 'agent'
+    ))?.owner
+    const createdAt = Date.now()
+    const next = appendHandoff(collaboration, {
+      source: previousOwner || { type: 'harness' },
+      destination,
+      objective,
+      selectedEntryIds: selectedIds,
+      expectedOutput,
+      acceptanceCriteria,
+      provenance: {
+        runId: controller.runId,
+        taskId: controller.taskId,
+        round: controller.currentRound,
+        agentRunId: null,
+        artifactIds: [],
+        evidenceIds: [],
+      },
+      createdAt,
+    })
+    const handoff = next.handoffs.at(-1)
+    this.checkpointOrchestration(group, controller, { collaboration: next })
+    return {
+      handoff,
+      entries: selectedEntries,
+      text: collaborationPackageText(handoff, selectedEntries),
+    }
+  },
+
+  recordCollaborationResult(group, controller, activeKinds, kind, result, options = {}) {
+    const role = options.role || this.collaborationRole(activeKinds, kind)
+    const conclusion = publicCollaborationText(result?.message?.content || '')
+    const trace = result?.message?.trace
+    const provenance = {
+      runId: controller.runId,
+      taskId: controller.taskId,
+      round: controller.currentRound,
+      agentRunId: trace?.agentRunId || null,
+      artifactIds: result?.outcomeRefs?.artifactIds || [],
+      evidenceIds: result?.outcomeRefs?.evidenceIds || [],
+    }
+    let collaboration = controller.orchestration?.collaboration || emptyCollaborationState()
+    const previousCollaboration = collaboration
+    const previousEntryIds = new Set(collaboration.entries.map(entry => entry.entryId))
+    let sequence = collaboration.entries.length + 1
+    const recordedAt = Date.now()
+    const conclusionValue = conclusion
+      ? createHash('sha256').update(conclusion).digest('hex')
+      : ''
+    const entryType = options.entryType || 'claim'
+    const subject = options.subject || `task:${controller.taskId}:conclusion`
+    const entryRefs = Array.isArray(options.refs) ? options.refs : []
+    const duplicateConclusion = conclusion && collaboration.entries.some(entry => (
+      entry.entryType === entryType
+      && entry.lifecycle.state === 'active'
+      && entry.subject === subject
+      && entry.value === conclusionValue
+      && entry.owner.type === 'agent'
+      && entry.owner.agentKind === kind
+    ))
+    if (conclusion && !duplicateConclusion) {
+      collaboration = appendBlackboardEntry(collaboration, {
+        entryType,
+        subject,
+        statement: conclusion,
+        value: entryType === 'claim' ? conclusionValue : '',
+        owner: { type: 'agent', agentKind: kind, role },
+        audience: this.collaborationAudience(role),
+        lifecycle: {
+          state: 'active', sequence, recordedAt, supersedesEntryId: null,
+        },
+        provenance,
+        refs: entryRefs,
+      })
+      sequence = collaboration.entries.length + 1
+    }
+    for (const artifactId of duplicateConclusion ? [] : provenance.artifactIds.slice(0, 1)) {
+      collaboration = appendBlackboardEntry(collaboration, {
+        entryType: 'artifact-ref',
+        subject: `task:${controller.taskId}:artifact`,
+        statement: 'Durable Artifact produced by the Agent.',
+        value: '',
+        owner: { type: 'agent', agentKind: kind, role },
+        audience: this.collaborationAudience(role),
+        lifecycle: {
+          state: 'active', sequence, recordedAt, supersedesEntryId: null,
+        },
+        provenance,
+        refs: [artifactId],
+      })
+      sequence = collaboration.entries.length + 1
+    }
+    for (const evidenceId of duplicateConclusion ? [] : provenance.evidenceIds.slice(0, 1)) {
+      collaboration = appendBlackboardEntry(collaboration, {
+        entryType: 'evidence-ref',
+        subject: `task:${controller.taskId}:evidence`,
+        statement: 'Durable Evidence recorded for the Agent conclusion.',
+        value: '',
+        owner: { type: 'agent', agentKind: kind, role },
+        audience: this.collaborationAudience(role),
+        lifecycle: {
+          state: 'active', sequence, recordedAt, supersedesEntryId: null,
+        },
+        provenance,
+        refs: [evidenceId],
+      })
+      sequence = collaboration.entries.length + 1
+    }
+    if (collaboration !== previousCollaboration) {
+      this.checkpointOrchestration(group, controller, { collaboration })
+    }
+    return collaboration.entries
+      .filter(entry => !previousEntryIds.has(entry.entryId))
+      .map(entry => entry.entryId)
+  },
+
+  checkpointOrchestration(group, controller, updates = {}) {
+    const previous = controller.orchestration
+    controller.orchestration = { ...previous, ...updates }
+    // A thrown callback has an unknown write outcome; only explicit failure is safe to roll back.
+    const persisted = this.checkpointRun?.(group.id, controller)
+    if (this.hasRunLedger() && persisted !== true) {
+      controller.orchestration = previous
+      if (previous?.version === 4) controller.currentRound = previous.round
+      throw new Error('LOCAL_RUN_PERSIST_FAILED')
+    }
+  },
+}
+
+module.exports = { collaborationMethods }

desktop/src/workspace/local-workspace-auto-runner-natural-decisions.cjs

@@ -0,0 +1,229 @@
+const { parseLenientTaskDecision } = require('../collaboration/task-decision.cjs')
+
+const v4NaturalDecisionMethods = {
+  v4NaturalRouteDecision(decision, activeKinds) {
+    const participants = Array.isArray(activeKinds) ? activeKinds : []
+    if (decision == null) return { status: 'none', kinds: [], dropped: [] }
+    let parsed
+    try { parsed = parseLenientTaskDecision(decision) }
+    catch { return { status: 'none', kinds: [], dropped: [] } }
+    const requested = parsed.nextKinds || []
+    // A request for an absent or paused participant is dropped on its own;
+    // it used to stop the whole discussion with a misleading message.
+    const dropped = requested.filter(kind => !participants.includes(kind))
+    const kinds = participants.filter(kind => requested.includes(kind))
+    return { status: kinds.length ? 'valid' : 'none', kinds, dropped }
+  },
+
+  v4NaturalNotice(group, controller, threadRootId, key, text, params = {}) {
+    const runParams = { ...params, runId: controller.runId }
+    const duplicate = this.state().messages.some(message => (
+      message.threadRootId === threadRootId
+        && message.system?.key === key
+        && JSON.stringify(message.system?.params || {}) === JSON.stringify(runParams)
+    ))
+    if (!duplicate) {
+      this.addMessage(group.id, 'system', text, '', threadRootId, { key, params: runParams })
+    }
+  },
+
+  async v4NaturalRoundOutcome(group, controller, threadRootId, round, activeKinds) {
+    const messages = this.v4NaturalDiscussionRoundMessages(group, controller, threadRootId, round)
+    if (!activeKinds.length) return { status: 'partial' }
+    const nominalOwner = this.v4NaturalOwner(group, controller, threadRootId)
+    const owner = this.v4NaturalOwner(group, controller, threadRootId, activeKinds)
+    if (owner !== nominalOwner) {
+      // The owner failed or was paused: the next selected participant takes
+      // over instead of the discussion ending silently as partial.
+      this.v4NaturalNotice(group, controller, threadRootId, 'system.autoOwnerChanged',
+        `${this.agentLabel(owner)} took over as delivery owner because ${this.agentLabel(nominalOwner)} is no longer available.`,
+        { agent: this.agentLabel(owner), previous: this.agentLabel(nominalOwner) })
+    }
+    const routes = messages.map(message => this.v4NaturalRouteDecision(
+      message.trace?.context?.taskDecision, activeKinds,
+    ))
+    const dropped = [...new Set(routes.flatMap(route => route.dropped || []))]
+    if (dropped.length) {
+      this.v4NaturalNotice(group, controller, threadRootId, 'system.autoRouteIgnored',
+        `Requests for ${dropped.map(kind => this.agentLabel(kind)).join(', ')} were skipped because they are not active in this discussion.`,
+        { agents: dropped.map(kind => this.agentLabel(kind)).join(', ') })
+    }
+    const ownerMessage = messages.find(message => message.agentKind === owner)
+    // A peer's blocker is input for the delivery owner, who judges the whole
+    // task. It used to end the discussion as partial on the spot.
+    const peerBlocker = messages.find(message => {
+      if (message.agentKind === owner) return false
+      const status = message.trace?.context?.taskDecision?.status
+      return status === 'blocked' || status === 'needs-human'
+    })
+    const ownerSawBlocker = peerBlocker && ownerMessage
+      && controller.discussionStyle === 'sequential'
+      && messages.indexOf(ownerMessage) > messages.indexOf(peerBlocker)
+    if (peerBlocker && !ownerSawBlocker) {
+      const peerDecision = peerBlocker.trace?.context?.taskDecision
+      this.v4NaturalNotice(group, controller, threadRootId, 'system.autoPeerBlockerForwarded',
+        `${this.agentLabel(peerBlocker.agentKind)} reported a blocker; the delivery owner will decide how to proceed.`,
+        { agent: this.agentLabel(peerBlocker.agentKind), reason: peerDecision.reason })
+      return { nextKinds: [owner], ownerReview: true }
+    }
+    const nextKinds = this.v4NaturalNextKinds(group, controller, threadRootId, round, activeKinds)
+    if (nextKinds.length) return { nextKinds, ownerReview: false }
+    const decision = ownerMessage?.trace?.context?.taskDecision
+    if (decision?.handoffTo && !activeKinds.includes(decision.handoffTo)) {
+      this.v4NaturalNotice(group, controller, threadRootId, 'system.autoHandoffIgnored',
+        `The handoff to ${this.agentLabel(decision.handoffTo)} was ignored because that Agent is not active; ${this.agentLabel(owner)} remains the delivery owner.`,
+        { agent: this.agentLabel(decision.handoffTo), owner: this.agentLabel(owner) })
+    }
+    if (decision?.status === 'continue') return { nextKinds: [owner], ownerReview: false }
+    const sawAllContributions = messages.length === 1
+      || (controller.discussionStyle === 'sequential' && messages.at(-1)?.agentKind === owner)
+    if (!ownerMessage || !sawAllContributions) return { nextKinds: [owner], ownerReview: true }
+    if (!decision) {
+      // Ask the owner once more for a judgment before giving up: a missing
+      // or unreadable receipt is usually a formatting slip, not a verdict.
+      if (!this.v4NaturalOwnerMissedDecisionBefore(group, controller, threadRootId, round, owner)) {
+        this.v4NaturalNotice(group, controller, threadRootId, 'system.autoTaskDecisionRequested',
+          `${this.agentLabel(owner)} did not include a readable task decision; asking for one before continuing.`,
+          { agent: this.agentLabel(owner), round })
+        return { nextKinds: [owner], ownerReview: true }
+      }
+      this.addMessage(group.id, 'system', 'Discussion stopped without a valid task completion decision.',
+        '', threadRootId, { key: 'system.autoTaskDecisionMissing' })
+      return { status: 'partial' }
+    }
+    if (decision.status === 'completed') {
+      // The owner may not close the task before selected peers had one discussion
+      // round of their own; otherwise a single full round collapses into an
+      // owner-only review and peers never get to challenge the result.
+      if (activeKinds.some(kind => kind !== owner)
+          && !this.v4NaturalPriorPeerDiscussion(group, controller, threadRootId, round, owner)) {
+        this.addMessage(
+          group.id, 'system',
+          'The delivery owner signaled completion before peer discussion; running one cross-review round first.',
+          '', threadRootId, { key: 'system.autoCrossReview' },
+        )
+        return { nextKinds: [...activeKinds], ownerReview: false }
+      }
+      if (controller.failedKinds.length) {
+        this.v4NaturalNotice(group, controller, threadRootId, 'system.autoCompletedWithout',
+          `The delivery owner completed the task without ${controller.failedKinds.map(kind => this.agentLabel(kind)).join(', ')}, which stopped earlier in this discussion.`,
+          { agents: controller.failedKinds.map(kind => this.agentLabel(kind)).join(', ') })
+        return { status: 'partial' }
+      }
+      return { status: 'completed' }
+    }
+    if (decision.status === 'needs-human' && this.requestTaskDecision) {
+      const response = await this.requestTaskDecision(group, controller, ownerMessage)
+      if (response?.status === 'approved') return { nextKinds: [owner], ownerReview: true }
+      if (response?.status === 'rejected') {
+        controller.stopReason = 'human_gate_rejected'
+        return { status: 'stopped' }
+      }
+    }
+    this.addMessage(group.id, 'system', decision.reason, '', threadRootId, {
+      key: decision.status === 'needs-human' ? 'system.autoTaskNeedsHuman' : 'system.autoTaskBlocked',
+      params: { reason: decision.reason },
+    })
+    return { status: 'partial' }
+  },
+
+  v4NaturalNextKinds(group, controller, threadRootId, round, activeKinds) {
+    const current = this.v4NaturalDiscussionRoundMessages(group, controller, threadRootId, round)
+    const selected = new Set(current.flatMap(message => (
+      this.v4NaturalRouteDecision(message.trace?.context?.taskDecision, activeKinds).kinds
+        .filter(kind => kind !== message.agentKind)
+    )))
+    if (!selected.size) return []
+    return activeKinds.filter(kind => selected.has(kind))
+  },
+
+  v4NaturalPriorPeerDiscussion(group, controller, threadRootId, round, owner) {
+    return this.state().messages.some(message => (
+      message.groupId === group.id
+      && message.threadRootId === threadRootId
+      && message.role === 'agent'
+      && message.agentKind !== owner
+      && message.trace?.runId === controller.runId
+      && message.trace?.phase === 'discussion'
+      && Number(message.trace?.round) < round
+      && this.v4NaturalMessageMatchesBinding(message, controller)
+    ))
+  },
+
+  v4NaturalOwnerMissedDecisionBefore(group, controller, threadRootId, round, owner) {
+    const earlier = this.state().messages.filter(message => (
+      message.groupId === group.id
+      && message.threadRootId === threadRootId
+      && message.role === 'agent'
+      && message.agentKind === owner
+      && message.trace?.runId === controller.runId
+      && message.trace?.phase === 'discussion'
+      && Number(message.trace?.round) < round
+      && this.v4NaturalMessageMatchesBinding(message, controller)
+    ))
+    const previous = earlier.at(-1)
+    return Boolean(previous && !previous.trace?.context?.taskDecision)
+  },
+
+  // With activeKinds, an owner who failed or was paused hands ownership to
+  // the first still-active selected participant.
+  v4NaturalOwner(group, controller, threadRootId, activeKinds = null) {
+    let owner = controller.targetKinds[0] || ''
+    for (const message of this.state().messages) {
+      if (message.groupId !== group.id || message.threadRootId !== threadRootId
+          || message.role !== 'agent' || message.agentKind !== owner
+          || message.trace?.phase !== 'discussion'
+          || !this.v4NaturalMessageMatchesBinding(message, controller)) continue
+      const decision = message.trace?.context?.taskDecision
+      if (decision?.status === 'continue' && decision.handoffTo
+          && controller.targetKinds.includes(decision.handoffTo)) owner = decision.handoffTo
+    }
+    if (Array.isArray(activeKinds) && activeKinds.length && !activeKinds.includes(owner)) {
+      return controller.targetKinds.find(kind => activeKinds.includes(kind)) || owner
+    }
+    return owner
+  },
+
+  v4NaturalDiscussionIsRepeating(group, controller, threadRootId, round, activeKinds) {
+    const window = Math.max(4, activeKinds.length * 2)
+    if (round < window + 2) return false
+    const firstRound = round - window + 1
+    const previous = new Map()
+    const observedRounds = new Set()
+    const messages = this.state().messages
+    for (let index = messages.length - 1; index >= 0; index -= 1) {
+      const message = messages[index]
+      const messageRound = Number(message.trace?.round) || 0
+      if (message.groupId !== group.id || message.threadRootId !== threadRootId
+          || message.role !== 'agent' || message.trace?.phase !== 'discussion'
+          || message.trace?.runId !== controller.runId || messageRound < 2
+          || messageRound > round || !this.v4NaturalMessageMatchesBinding(message, controller)) continue
+      const contribution = this.v4NaturalMessageContent(message)
+      if (contribution.partial) return false
+      const content = contribution.text
+        .replace(/@[A-Za-z0-9][A-Za-z0-9_-]*/gu, '')
+        .replace(/\s+/gu, ' ').trim()
+      if (previous.has(message.agentKind) && previous.get(message.agentKind) !== content) return false
+      if (messageRound >= firstRound) {
+        observedRounds.add(messageRound)
+        previous.set(message.agentKind, content)
+      } else {
+        previous.delete(message.agentKind)
+        if (observedRounds.size === window && previous.size === 0) return true
+      }
+    }
+    return false
+  },
+
+  v4NaturalStopRepeatingDiscussion(group, controller, threadRootId, round, activeKinds) {
+    if (!this.v4NaturalDiscussionIsRepeating(group, controller, threadRootId, round, activeKinds)) return false
+    this.addMessage(
+      group.id, 'system',
+      'Automatic discussion stopped because repeated handoffs produced no new contribution.',
+      '', threadRootId, { key: 'system.autoDiscussionStalled' },
+    )
+    return true
+  },
+}
+
+module.exports = { v4NaturalDecisionMethods }

desktop/src/workspace/local-workspace-auto-runner-natural-discussion.cjs

@@ -0,0 +1,540 @@
+// Natural discussion, the production path for automatic group runs: transcript, prompts and the sequential and agent-led loops.
+const path = require('node:path')
+const { MAX_MESSAGE_TEXT_CHARS } = require('./local-workspace-contracts.cjs')
+const { hashValue } = require('../collaboration/orchestration-v4-records.cjs')
+const { packNaturalDiscussionTranscript, v4Prompt } = require('./local-workspace-context.cjs')
+const { naturalDiscussionPrompt } = require('./prompts/natural-discussion-prompt.cjs')
+const { terminalRunStatusForReason } = require('./local-workspace-auto-runner-support.cjs')
+const { traceCapsuleFromAgentRun } = require('../runs/run-harness-normalization.cjs')
+
+const naturalDiscussionMethods = {
+  v4NaturalMessageMatchesBinding(message, controller) {
+    const phase = String(message?.trace?.phase || '')
+    const kind = String(message?.agentKind || '')
+    const round = Number(message?.trace?.round) || 0
+    if (!['proposal', 'discussion'].includes(phase) || !kind || round < 1) return false
+    if (message.trace?.runId !== controller.runId) return false
+    const slot = controller.orchestration?.slots?.find(candidate => candidate.agentKind === kind)
+    const snapshotHash = String(controller.orchestration?.snapshotHash || '')
+    if (!slot?.slotId || !snapshotHash) return false
+    const operationPhase = phase === 'discussion' ? `discussion:${round}` : phase
+    const expectedOperationId = this.v4OperationId(
+      controller, kind, operationPhase, slot.slotId, round,
+    )
+    return message.trace?.context?.operationId === expectedOperationId
+      && message.trace?.context?.snapshotHash === snapshotHash
+  },
+
+  v4NaturalMessageContent(message) {
+    const text = String(message.content || '')
+    if (text.length < MAX_MESSAGE_TEXT_CHARS && message.trace?.truncated !== true) {
+      return { text, partial: false }
+    }
+    for (const artifactId of message.trace?.context?.outcomeRefs?.artifactIds || []) {
+      try {
+        const { artifact, content } = this.v4ArtifactIdentity(artifactId)
+        if (artifact.name !== `${message.agentKind}-conclusion.txt`
+            || artifact.producedBy?.runId !== message.trace?.runId
+            || artifact.producedBy?.agentRunId !== message.trace?.agentRunId
+            || artifact.producedBy?.agentKind !== message.agentKind
+            || !content.startsWith(text)) continue
+        return { text: content, partial: false }
+      } catch { /* Historical blobs may be unavailable; identify the excerpt below. */ }
+    }
+    return { text, partial: true }
+  },
+
+  v4NaturalThreadEntries(group, threadRootId, options = {}) {
+    const rounds = Array.isArray(options.rounds) ? new Set(options.rounds) : null
+    const entries = this.state().messages.filter(message => (
+      message.groupId === group.id
+      && message.threadRootId === threadRootId
+      && message.role === 'agent'
+      && message.agentKind
+      && (!rounds || rounds.has(Number(message.trace?.round) || 0))
+      && (!options.controller || this.v4NaturalMessageMatchesBinding(
+        message, options.controller,
+      ))
+    )).map((message, index) => ({ message, index }))
+      // Peers read turns in dispatch order, the same order the user sees,
+      // instead of the order replies happened to be saved in.
+      .sort((left, right) => {
+        const leftTurn = Number.isSafeInteger(left.message.turnSeq) ? left.message.turnSeq : 0
+        const rightTurn = Number.isSafeInteger(right.message.turnSeq) ? right.message.turnSeq : 0
+        return (leftTurn ? 1 : 0) - (rightTurn ? 1 : 0)
+          || leftTurn - rightTurn
+          || left.index - right.index
+      })
+      .map(({ message }) => message)
+      .map(message => ({
+      ...this.v4NaturalMessageContent(message),
+      id: message.id,
+      agentKind: message.agentKind,
+      round: Number(message.trace?.round) || 0,
+      phase: message.trace?.phase || 'proposal',
+      operationId: message.trace?.context?.operationId || '',
+    })).map(entry => ({
+      // Peers collaborating on the same workspace must be able to name its
+      // files; keep workspace paths readable as relative paths.
+      ...entry, text: this.v4SanitizeDeliveryText(entry.text, Number.MAX_SAFE_INTEGER, {
+        workspaceRoot: group.workdir ? path.resolve(group.workdir) : '',
+      }),
+    }))
+    return entries.map(entry => ({ ...entry, hash: hashValue(entry) }))
+  },
+
+  v4NaturalThreadTranscript(group, threadRootId, options = {}) {
+    return packNaturalDiscussionTranscript(this.v4NaturalThreadEntries(group, threadRootId, options))
+  },
+
+  v4NaturalDeliveryPrompt(group, controller, threadRootId, kind, slot, binding, promptBuilder) {
+    const entries = this.v4NaturalThreadEntries(group, threadRootId, { controller })
+    const previousEntries = binding.hasSession && !binding.sessionRotated
+      ? (controller.orchestration?.deliveryState || []).filter(entry => (
+          entry.recipientKind === kind && entry.status === 'acknowledged'
+          && entry.sessionRefHash === binding.sessionRefHash
+          && entry.sessionProvenanceHash === binding.sessionProvenanceHash
+          && entry.snapshotHash === controller.orchestration.snapshotHash
+          && entry.sourceMessages
+        ))
+      : []
+    const seen = new Set(previousEntries.flatMap(entry => entry.sourceMessages.map(message => (
+      [entry.sourceAgentKind, entry.sourcePhase, message.id, message.hash].join('\u0000')
+    ))))
+    const unseen = entries.filter(entry => !seen.has([
+      entry.agentKind, entry.phase, entry.id, entry.hash,
+    ].join('\u0000')))
+    const packed = packNaturalDiscussionTranscript(unseen, {
+      withEntries: true,
+      prefix: unseen.length < entries.length
+        ? 'Previously delivered turns are not repeated here. They were supplied to this native session; ask peers to restate missing evidence if your current context is incomplete. Their omission does not imply agreement or task completion.'
+        : '',
+    })
+    const included = new Set(packed.completeMessageIds)
+    const sources = new Map()
+    for (const entry of unseen) {
+      if (!included.has(entry.id)) continue
+      const key = `${entry.agentKind}\u0000${entry.phase}`
+      const source = sources.get(key) || {
+        sourceAgentKind: entry.agentKind, sourcePhase: entry.phase,
+        watermark: 0, operationId: entry.operationId, sourceMessages: [],
+      }
+      source.watermark = Math.max(source.watermark, entry.round)
+      source.sourceMessages.push({ id: entry.id, hash: entry.hash })
+      sources.set(key, source)
+    }
+    const packageHash = hashValue(packed.text)
+    const deliveryId = `delivery-${hashValue({
+      runId: controller.runId, kind, operationId: slot.operationId,
+      packageHash, binding, now: Date.now(),
+    })}`
+    const delivery = sources.size ? {
+      natural: true, previousEntries, packageHash, deliveryId,
+      entries: [...sources.values()].map(source => ({
+        ...source, recipientKind: kind,
+        sessionRefHash: binding.sessionRefHash,
+        sessionProvenanceHash: binding.sessionProvenanceHash,
+        snapshotHash: controller.orchestration.snapshotHash,
+        packageHash, deliveryId, status: 'prepared', updatedAt: Date.now(),
+      })),
+    } : null
+    this.v4PrepareDelivery(group, controller, delivery)
+    return { prompt: promptBuilder(packed.text), delivery }
+  },
+
+  v4NaturalRecoveredPhaseResult(group, controller, threadRootId, phase, kind, slot) {
+    if (!this.naturalAgentResponses || !slot || controller.currentRound < 1) return null
+    if (!['proposal', 'discussion'].includes(phase)) return null
+    const operationId = String(slot.operationId || '')
+    const snapshotHash = String(slot.snapshotHash || '')
+    if (!operationId || !snapshotHash) return null
+    const matchesBinding = context => (
+      context?.operationId === operationId && context?.snapshotHash === snapshotHash
+    )
+    const scopedMessages = [...this.state().messages].reverse().filter(message => (
+      message.groupId === group.id
+      && message.threadRootId === threadRootId
+      && message.role === 'agent'
+      && message.agentKind === kind
+      && message.trace?.runId === controller.runId
+      && message.trace?.phase === phase
+      && Number(message.trace?.round) === controller.currentRound
+    ))
+    if (scopedMessages.some(message => !matchesBinding(message.trace?.context))) {
+      throw new Error('LOCAL_RUN_COLLABORATION_SCOPE_INVALID')
+    }
+    const existingMessage = scopedMessages.find(message => matchesBinding(message.trace?.context))
+    let content = String(existingMessage?.content || '')
+    let trace = existingMessage?.trace || null
+    let metadata = existingMessage ? {
+      ...(existingMessage.elapsedMs != null ? { elapsedMs: existingMessage.elapsedMs } : {}),
+      ...(Array.isArray(existingMessage.toolCalls) ? { toolCalls: existingMessage.toolCalls } : {}),
+      ...(Array.isArray(existingMessage.attachments)
+        ? { attachments: existingMessage.attachments } : {}),
+      ...(existingMessage.responseVersionRootId
+        ? { responseVersionRootId: existingMessage.responseVersionRootId } : {}),
+      trace,
+    } : null
+    if (!content || !trace?.agentRunId) {
+      const scopedRuns = [...(controller.harness?.snapshot?.() || [])].reverse().filter(run => (
+        run.kind === kind
+        && run.round === controller.currentRound
+        && run.status === 'completed'
+        && String(run.output || '')
+      ))
+      if (scopedRuns.some(run => !matchesBinding(run.context))) {
+        throw new Error('LOCAL_RUN_COLLABORATION_SCOPE_INVALID')
+      }
+      const recoveredRun = scopedRuns.find(run => matchesBinding(run.context))
+      if (!recoveredRun) return null
+      content = String(recoveredRun.output)
+      trace = traceCapsuleFromAgentRun(recoveredRun, {
+        runId: controller.runId,
+        status: recoveredRun.status,
+        phase,
+      })
+      if (!trace?.agentRunId) return null
+      metadata = { trace }
+    }
+    return {
+      text: content,
+      pendingMessage: {
+        groupId: group.id,
+        role: 'agent',
+        content,
+        agentKind: kind,
+        threadRootId,
+        system: null,
+        metadata,
+      },
+      collaboration: trace.context?.taskDecision
+        ? { summary: trace.context.taskDecision.reason, taskDecision: trace.context.taskDecision }
+        : null,
+      outcomeRefs: trace.context?.outcomeRefs || {},
+      operationId: slot.operationId,
+      consensus: false,
+    }
+  },
+
+  v4NaturalRecoveredDiscussionResult(group, controller, threadRootId, kind, slot) {
+    return this.v4NaturalRecoveredPhaseResult(
+      group, controller, threadRootId, 'discussion', kind, slot,
+    )
+  },
+
+  v4NaturalPhasePrompt(group, kind, phase, snapshot, receiptRecords, activeKinds, options = {}) {
+    const humanResponses = options.runId ? this.taskDecisionResponses(options.runId) : []
+    return naturalDiscussionPrompt({
+      brief: v4Prompt({
+        group, kind, phase, snapshot, role: 'participant',
+        skillHints: options.skillHints,
+        knowledgeBaseHints: options.knowledgeBaseHints,
+        naturalResponse: true,
+      }),
+      group,
+      kind,
+      phase,
+      activeKinds,
+      transcript: String(options.transcript || ''),
+      humanResponsesText: humanResponses.length ? packNaturalDiscussionTranscript(humanResponses) : '',
+      writerKind: options.writerKind,
+      ownerKind: options.ownerKind,
+      allowRouting: options.allowRouting === true,
+    })
+  },
+
+  v4NaturalDiscussionOperationIds(controller, slots, kinds, round) {
+    return new Map(kinds.map((kind) => {
+      const slot = slots.find(candidate => candidate.agentKind === kind)
+      return [kind, this.v4OperationId(
+        controller, kind, `discussion:${round}`, slot?.slotId || '', round,
+      )]
+    }))
+  },
+
+  v4NaturalDiscussionRoundMessages(group, controller, threadRootId, round) {
+    return this.state().messages.filter(message => (
+      message.groupId === group.id
+      && message.threadRootId === threadRootId
+      && message.role === 'agent'
+      && message.trace?.runId === controller.runId
+      && message.trace?.phase === 'discussion'
+      && Number(message.trace?.round) === round
+      && this.v4NaturalMessageMatchesBinding(message, controller)
+    ))
+  },
+
+  async runV4SequentialDiscussion(group, controller, threadRootId, context, input) {
+    const {
+      targetKinds, activeKinds: initialActiveKinds, writerKind, batchId,
+      snapshot, snapshotRecord, snapshotHash, slots: initialSlots, receiptRecords,
+      checkpointPhase, removePhaseFailures, addRoundLimitNotice,
+    } = input
+    let activeKinds = [...initialActiveKinds]
+    let slots = initialSlots
+    let round = Math.max(1, Number(controller.currentRound) || 1)
+    let pendingKinds = controller.orchestration?.discussionStyle === 'sequential'
+      && controller.orchestration?.phase === 'discussion'
+      && controller.orchestration?.round === round
+      ? (controller.orchestration.pendingKinds || []).filter(kind => activeKinds.includes(kind))
+      : [...activeKinds]
+    if (!pendingKinds.length && controller.orchestration?.phase === 'discussion') {
+      const outcome = await this.v4NaturalRoundOutcome(group, controller, threadRootId, round, activeKinds)
+      if (outcome.status) return outcome.status
+      if (!controller.unlimitedRounds && round >= (controller.maxRounds || 6)) {
+        addRoundLimitNotice()
+        return 'round-limit'
+      }
+      round += 1
+      pendingKinds = outcome.ownerReview ? outcome.nextKinds : [...activeKinds]
+    }
+    while (!controller.signal.aborted) {
+      controller.currentRound = round
+      while (pendingKinds.length && !controller.signal.aborted) {
+        const kind = pendingKinds[0]
+        const remainingKinds = pendingKinds.slice(1)
+        const dispatch = await this.runV4Phase(group, controller, threadRootId, context, {
+          phase: 'discussion',
+          activeKinds: [kind],
+          targetKinds,
+          participantKinds: activeKinds,
+          writerKind,
+          batchId,
+          snapshot,
+          snapshotRecord,
+          snapshotHash,
+          slots,
+          receiptRecords,
+          challengeBindings: null,
+          coordinationPlan: null,
+          workReceipts: [],
+          operationIds: this.v4NaturalDiscussionOperationIds(
+            controller, slots, [kind], round,
+          ),
+          pendingKinds,
+          remainingKinds,
+          sessionPolicy: null,
+          resumedGate: input.resumedGate,
+          disableDeliveryPrompt: true,
+          promptBuilder: ({ kind: promptKind, transcript }) => this.v4NaturalPhasePrompt(
+            group, promptKind, 'discussion', snapshot, receiptRecords, activeKinds, {
+              runId: controller.runId,
+              transcript: transcript ?? this.v4NaturalThreadTranscript(
+                group, threadRootId, { controller },
+              ),
+              skillHints: context.rootSkillsByKind.get(promptKind) || [],
+              knowledgeBaseHints: context.rootKnowledgeBasesByKind.get(promptKind) || [],
+              ownerKind: this.v4NaturalOwner(group, controller, threadRootId, activeKinds),
+              writerKind,
+              allowRouting: false,
+            },
+          ),
+        })
+        slots = dispatch.slots || slots
+        if (!dispatch.ok) {
+          if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+          const failedKinds = removePhaseFailures(dispatch)
+          activeKinds = activeKinds.filter(agentKind => !failedKinds.includes(agentKind))
+          if (!activeKinds.length) {
+            return dispatch.phasePendingMessages?.length ? 'partial' : 'failed'
+          }
+        }
+        pendingKinds = remainingKinds.filter(agentKind => activeKinds.includes(agentKind))
+      }
+      const outcome = await this.v4NaturalRoundOutcome(group, controller, threadRootId, round, activeKinds)
+      if (outcome.status) return outcome.status
+      if (!controller.unlimitedRounds && round >= (controller.maxRounds || 6)) {
+        addRoundLimitNotice()
+        return 'round-limit'
+      }
+      round += 1
+      controller.currentRound = round
+      pendingKinds = outcome.ownerReview ? outcome.nextKinds : [...activeKinds]
+    }
+    return terminalRunStatusForReason(controller.stopReason)
+  },
+
+  async runV4AgentLedDiscussion(group, controller, threadRootId, context, input) {
+    const {
+      targetKinds, activeKinds: initialActiveKinds, writerKind, batchId,
+      snapshot, snapshotRecord, snapshotHash, slots: initialSlots, receiptRecords,
+      latestFor, runPhase, checkpointPhase, removePhaseFailures, addRoundLimitNotice,
+    } = input
+    let activeKinds = [...initialActiveKinds]
+    let slots = initialSlots
+    let round = Math.max(1, Number(controller.currentRound) || 1)
+    const proposalCompleteBeforeRun = activeKinds.every(kind => Boolean(latestFor('proposal', kind)))
+    if (!proposalCompleteBeforeRun) {
+      controller.currentRound = 1
+      const proposal = await runPhase('proposal', activeKinds)
+      slots = proposal.slots || slots
+      if (!proposal.ok) {
+        const failedKinds = removePhaseFailures(proposal)
+        activeKinds = activeKinds.filter(kind => !failedKinds.includes(kind))
+        checkpointPhase('proposal', [], slots, null)
+        if (!activeKinds.length) return proposal.phasePendingMessages?.length ? 'partial' : 'failed'
+      }
+    }
+    const proposalComplete = activeKinds.every(kind => Boolean(latestFor('proposal', kind)))
+    if (!proposalComplete) {
+      return activeKinds.some(kind => Boolean(latestFor('proposal', kind))) ? 'partial' : 'failed'
+    }
+    if (!controller.unlimitedRounds && (controller.maxRounds || 6) <= 1) {
+      addRoundLimitNotice()
+      return 'round-limit'
+    }
+
+    let nextKinds = []
+    if (controller.orchestration?.phase === 'discussion') {
+      round = Math.max(2, Number(controller.orchestration.round) || 2)
+      nextKinds = (controller.orchestration.pendingKinds || [])
+        .filter(kind => activeKinds.includes(kind))
+      if (!nextKinds.length) {
+        if (this.v4NaturalStopRepeatingDiscussion(
+          group, controller, threadRootId, round, activeKinds,
+        )) return 'partial'
+        const outcome = await this.v4NaturalRoundOutcome(group, controller, threadRootId, round, activeKinds)
+        if (outcome.status) return outcome.status
+        nextKinds = outcome.nextKinds
+        if (!controller.unlimitedRounds && round >= (controller.maxRounds || 6)) {
+          addRoundLimitNotice()
+          return 'round-limit'
+        }
+        round += 1
+      }
+    } else {
+      round = 2
+      // The first discussion round is a full participant round: every selected
+      // Agent reacts to the proposal set before the delivery owner reviews and
+      // judges completion, so a group run cannot collapse to owner-only turns.
+      nextKinds = [...activeKinds]
+    }
+
+    while (!controller.signal.aborted) {
+      controller.currentRound = round
+      const dispatch = await this.runV4Phase(group, controller, threadRootId, context, {
+        phase: 'discussion',
+        activeKinds: nextKinds,
+        targetKinds,
+        participantKinds: activeKinds,
+        writerKind,
+        batchId,
+        snapshot,
+        snapshotRecord,
+        snapshotHash,
+        slots,
+        receiptRecords,
+        challengeBindings: null,
+        coordinationPlan: null,
+        workReceipts: [],
+        operationIds: this.v4NaturalDiscussionOperationIds(
+          controller, slots, nextKinds, round,
+        ),
+        pendingKinds: nextKinds,
+        remainingKinds: [],
+        sessionPolicy: null,
+        resumedGate: input.resumedGate,
+        disableDeliveryPrompt: true,
+        promptBuilder: ({ kind, transcript }) => this.v4NaturalPhasePrompt(
+          group, kind, 'discussion', snapshot, receiptRecords, activeKinds, {
+            runId: controller.runId,
+            transcript: transcript ?? this.v4NaturalThreadTranscript(group, threadRootId, { controller }),
+            skillHints: context.rootSkillsByKind.get(kind) || [],
+            knowledgeBaseHints: context.rootKnowledgeBasesByKind.get(kind) || [],
+            allowRouting: true,
+            ownerKind: this.v4NaturalOwner(group, controller, threadRootId, activeKinds),
+            writerKind,
+          },
+        ),
+      })
+      slots = dispatch.slots || slots
+      if (!dispatch.ok) {
+        if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+        const failedKinds = removePhaseFailures(dispatch)
+        activeKinds = activeKinds.filter(kind => !failedKinds.includes(kind))
+        if (!activeKinds.length) return dispatch.phasePendingMessages?.length ? 'partial' : 'failed'
+      }
+      if (this.v4NaturalStopRepeatingDiscussion(
+        group, controller, threadRootId, round, activeKinds,
+      )) return 'partial'
+      const outcome = await this.v4NaturalRoundOutcome(group, controller, threadRootId, round, activeKinds)
+      if (outcome.status) return outcome.status
+      nextKinds = outcome.nextKinds
+      if (!controller.unlimitedRounds && round >= (controller.maxRounds || 6)) {
+        addRoundLimitNotice()
+        return 'round-limit'
+      }
+      round += 1
+      controller.currentRound = round
+    }
+    return terminalRunStatusForReason(controller.stopReason)
+  },
+
+  async runV4NaturalDiscussion(group, controller, threadRootId, context, input) {
+    if (input.resume && controller.orchestration?.phase === 'discussion') {
+      for (const slot of input.slots) {
+        if (slot.permission !== 'workspace-write' || !slot.agentRunId
+            || slot.commitStatus === 'committed'
+            || !['running', 'waiting', 'stopped', 'failed'].includes(slot.status)) continue
+        if (this.v4NaturalRecoveredDiscussionResult(group, controller, threadRootId, slot.agentKind, slot)) continue
+        if (input.resumedGate?.operationId === slot.operationId
+            && input.resumedGate.agentKind === slot.agentKind
+            && input.resumedGate.status === 'approved') continue
+        if (!this.requestHumanGate) throw new Error('LOCAL_RUN_RETRY_GATE_INVALID')
+        const binding = {
+          agentRunId: slot.agentRunId, agentKind: slot.agentKind,
+          round: controller.currentRound, phase: 'discussion', slotId: slot.slotId,
+          operationId: slot.operationId, snapshotHash: input.snapshotHash,
+        }
+        const request = {
+          phase: 'discussion-writer-recovery', batchId: input.batchId,
+          slotId: slot.slotId, operationId: slot.operationId, attempt: slot.attempt,
+          outcomeCertainty: 'unknown_outcome', sideEffectsPossible: true,
+        }
+        const gate = await this.requestHumanGate({
+          type: 'retry', runId: controller.runId, agentRunId: slot.agentRunId,
+          agentKind: slot.agentKind,
+          summary: 'The interrupted discussion writer may already have changed the workspace.',
+          options: [
+            { optionId: 'retry-once', name: 'Retry once', kind: 'allow_once' },
+            { optionId: 'cancel-retry', name: 'Do not retry', kind: 'reject_once' },
+          ],
+          request,
+        }, {
+          signal: controller.signal,
+          preserveOnAbort: () => controller.stopReason === 'shutdown',
+          continuation: { resumeKind: 'agent_slot', ...binding },
+        })
+        if (gate.status !== 'approved') {
+          if (this.completeHumanGateContinuation?.(controller.runId, gate.gateId, 'cancelled') !== true
+              && this.hasRunLedger()) throw new Error('LOCAL_RUN_PERSIST_FAILED')
+          controller.stopReason = 'human_gate_rejected'
+          return 'stopped'
+        }
+        if (gate.optionId !== 'retry-once') throw new Error('LOCAL_RUN_CONTINUATION_INVALID')
+        input.resumedGate = { ...binding, gateId: gate.gateId, type: 'retry',
+          status: gate.status, optionId: gate.optionId, request, used: false }
+      }
+    }
+    if (input.resumedGate?.type === 'retry'
+        && input.resumedGate.request?.phase === 'discussion-writer-recovery') {
+      const gate = input.resumedGate
+      const slot = input.slots.find(item => item.slotId === gate.slotId)
+      if (gate.status !== 'approved' || gate.optionId !== 'retry-once'
+          || !slot || slot.agentKind !== gate.agentKind
+          || slot.operationId !== gate.operationId
+          || slot.snapshotHash !== gate.snapshotHash
+          || slot.attempt !== gate.request.attempt) throw new Error('LOCAL_RUN_CONTINUATION_INVALID')
+      if (this.completeHumanGateContinuation?.(controller.runId, gate.gateId, 'completed') !== true
+          && this.hasRunLedger()) throw new Error('LOCAL_RUN_PERSIST_FAILED')
+    }
+    const style = controller.orchestration?.discussionStyle
+      || controller.discussionStyle
+      || 'sequential'
+    return style === 'agent-led'
+      ? this.runV4AgentLedDiscussion(group, controller, threadRootId, context, input)
+      : this.runV4SequentialDiscussion(group, controller, threadRootId, context, input)
+  },
+}
+
+module.exports = { naturalDiscussionMethods }

desktop/src/workspace/local-workspace-auto-runner-recovery.cjs

@@ -0,0 +1,443 @@
+// Failed, interrupted and unauthorized invocations: retry policy, pausing an Agent for the run, and recovery prompts.
+const { createHash } = require('node:crypto')
+const {
+  MAX_RUN_AGENT_ATTEMPTS,
+  boundedBackoffDelay,
+  normalizeAttemptHistoryEntry,
+  normalizeFailure,
+  normalizeFailureOutcome,
+  retryDecision,
+} = require('../runs/failure-policy.cjs')
+const {
+  circuitBreakerError,
+  sanitizedAuthenticationError,
+  unauthorizedFailure,
+} = require('./local-workspace-auto-runner-support.cjs')
+
+const recoveryMethods = {
+  setCurrentAgent(controller, kind) {
+    controller.currentKind = kind
+    controller.progress = []
+    this.emitChanged()
+  },
+
+  pauseFailedAgentForRun(group, controller, kind, threadRootId, error = null) {
+    // Failures are scoped to the current run. Group membership is user
+    // configuration: editing it after a provider or login hiccup made the
+    // Agent disappear from every later conversation turn, and removing the
+    // last member made the whole group (and its history) unloadable.
+    if (group.conversationType === 'direct' || !group.agentKinds.includes(kind)) return false
+    if (!controller.failedKinds.includes(kind)) controller.failedKinds.push(kind)
+    const alreadyReported = this.state().messages.some(message => (
+      message.threadRootId === threadRootId
+        && message.agentKind === kind
+        && message.system?.key === 'system.agentPausedAfterFailure'
+        && message.system?.params?.runId === controller.runId
+    ))
+    if (!alreadyReported) {
+      const agent = this.agentLabel(kind)
+      const reason = normalizeFailure(error).code || 'LOCAL_AGENT_PROCESS_FAILED'
+      this.addMessage(
+        group.id,
+        'system',
+        `${agent} is paused for the rest of this discussion (${reason}). It stays in the group and will be tried again on the next message.`,
+        kind,
+        threadRootId,
+        {
+          key: 'system.agentPausedAfterFailure',
+          params: { agent, reason, runId: controller.runId },
+        },
+      )
+    }
+    return true
+  },
+
+  shouldPauseFailedAgent(error) {
+    const category = normalizeFailure(error).category
+    // Runtime, provider, rate-limit, protocol, and timeout failures stay
+    // silent run-scoped exclusions; their own failure message explains them.
+    // Authentication and compatibility failures get an explicit notice
+    // because the user usually has to act (sign in or update the CLI).
+    return ['authentication', 'compatibility'].includes(category)
+  },
+
+  replacementInstruction(failedKind) {
+    return [
+      `You are replacing ${this.agentLabel(failedKind)} for this turn.`,
+      'Complete the interrupted Agent slot using the shared task context, then return your own conclusion.',
+      'Do not claim that the interrupted Agent completed this work.',
+    ].join('\n')
+  },
+
+  retryDelay(failedAttempt) {
+    return boundedBackoffDelay(failedAttempt, {
+      baseDelayMs: this.retryBaseDelayMs,
+      maxDelayMs: this.retryMaxDelayMs,
+    })
+  },
+
+  operationIdFor(controller, kind, mode) {
+    const digest = createHash('sha256').update(JSON.stringify({
+      runId: controller.runId,
+      kind,
+      mode,
+      round: controller.currentRound || 0,
+    })).digest('hex')
+    return `agent-operation-${digest}`
+  },
+
+  recordAttempt(group, controller, input) {
+    const history = Array.isArray(controller.attemptHistory) ? controller.attemptHistory : []
+    const previousSequence = history.at(-1)?.sequence || 0
+    const entry = normalizeAttemptHistoryEntry({
+      sequence: previousSequence + 1,
+      agentKind: input.agentKind,
+      phase: input.phase,
+      attempt: input.attempt,
+      failureCategory: input.failureCategory ?? null,
+      policyAction: input.policyAction,
+      backoffMs: input.backoffMs || 0,
+      recoveryAgentKind: input.recoveryAgentKind || '',
+      finalOutcome: input.finalOutcome,
+      timestamp: Date.now(),
+      ...(input.outcomeCertainty ? { outcomeCertainty: input.outcomeCertainty } : {}),
+      ...(Object.hasOwn(input, 'sideEffectsPossible')
+        ? { sideEffectsPossible: input.sideEffectsPossible }
+        : {}),
+      ...(input.operationId ? { operationId: input.operationId } : {}),
+      ...(input.idempotencyMode ? { idempotencyMode: input.idempotencyMode } : {}),
+    })
+    if (!entry) throw new Error('LOCAL_RUN_ATTEMPT_INVALID')
+    controller.attemptHistory = [...history, entry].slice(-256)
+    const persisted = this.checkpointRun?.(group.id, controller)
+    if (this.hasRunLedger() && persisted !== true) {
+      throw new Error('LOCAL_RUN_PERSIST_FAILED')
+    }
+    return entry
+  },
+
+  async prepareRetry(group, kind, controller, failedAttempt, options = {}) {
+    if (options.resetSession === true) {
+      this.resetAgentSession(group, kind, true, controller.taskId)
+    }
+    this.markRuntimeCredential(kind, 'unknown')
+    try { await this.refreshAgents() } catch {
+      if (controller.signal.aborted) throw new Error('LOCAL_AGENT_EXECUTION_STOPPED')
+    }
+    const delayMs = Number.isFinite(options.delayMs)
+      ? Math.max(0, Math.floor(options.delayMs))
+      : this.retryDelay(failedAttempt)
+    await this.sleep(delayMs, controller.signal)
+    return delayMs
+  },
+
+  async invokeTransiently(invokeSource, group, kind, controller, options = {}) {
+    const operationId = options.operationId
+    const idempotencyMode = options.idempotencyMode === 'durable' ? 'durable' : 'none'
+    const sideEffectsPossible = options.sideEffectsPossible === true
+    for (let attempt = 1; ; attempt += 1) {
+      if (controller.attemptHistory.length >= MAX_RUN_AGENT_ATTEMPTS) {
+        throw circuitBreakerError()
+      }
+      const phase = attempt === 1 ? (options.phase || 'initial') : 'transient_retry'
+      const phaseAttempt = attempt === 1 ? (options.attempt || 1) : attempt - 1
+      try {
+        const result = await invokeSource()
+        this.recordAttempt(group, controller, {
+          agentKind: options.agentKind || kind,
+          phase,
+          attempt: phaseAttempt,
+          failureCategory: null,
+          policyAction: options.successAction || 'complete',
+          backoffMs: options.successBackoffMs || 0,
+          recoveryAgentKind: options.recoveryAgentKind || '',
+          finalOutcome: 'succeeded',
+          sideEffectsPossible,
+          operationId,
+          idempotencyMode,
+        })
+        return result
+      } catch (error) {
+        if (controller.signal.aborted) {
+          this.recordAttempt(group, controller, {
+            agentKind: options.agentKind || kind,
+            phase,
+            attempt: phaseAttempt,
+            failureCategory: 'cancellation',
+            policyAction: 'cancel',
+            recoveryAgentKind: options.recoveryAgentKind || '',
+            finalOutcome: 'cancelled',
+          })
+          throw error
+        }
+        if (unauthorizedFailure(error)) throw error
+        const failure = normalizeFailure(error)
+        const safety = normalizeFailureOutcome(error, {
+          category: failure.category,
+          sideEffectsPossible,
+          operationId,
+          idempotencyMode,
+        })
+        if (failure.category === 'cancellation') throw error
+        if (options.deferFailurePolicy === true) {
+          this.recordAttempt(group, controller, {
+            agentKind: options.agentKind || kind,
+            phase,
+            attempt: phaseAttempt,
+            failureCategory: failure.category,
+            policyAction: 'fail',
+            backoffMs: 0,
+            recoveryAgentKind: '',
+            finalOutcome: 'failed',
+            ...safety,
+          })
+          throw error
+        }
+        // Session recovery is owned by LocalWorkspaceAgentInvocation so a stale
+        // native Session gets exactly one fresh attempt across every run mode.
+        // An ambiguous refusal (HTTP 403, proxy or region block) gets exactly
+        // one more attempt: a transient gateway refusal recovers, while a real
+        // permission problem fails fast without signing the Agent out.
+        const maxAttempts = failure.category === 'timeout' && options.retryTimeouts !== true
+          ? 1
+          : failure.code === 'LOCAL_AGENT_ACCESS_DENIED' ? 2 : 4
+        const decision = retryDecision(failure, {
+          attempt,
+          maxAttempts,
+          baseDelayMs: this.retryBaseDelayMs,
+          maxDelayMs: this.retryMaxDelayMs,
+          rateLimitMinDelayMs: this.retryRateLimitMinDelayMs,
+          safety,
+        })
+        if (decision.action === 'human_gate') {
+          this.recordAttempt(group, controller, {
+            agentKind: options.agentKind || kind,
+            phase,
+            attempt: phaseAttempt,
+            failureCategory: failure.category,
+            policyAction: 'human_gate',
+            backoffMs: decision.delayMs,
+            recoveryAgentKind: options.recoveryAgentKind || '',
+            finalOutcome: 'failed',
+            ...safety,
+          })
+          if (!this.requestHumanGate || !error?.runTrace?.agentRunId) {
+            throw new Error('LOCAL_RUN_RETRY_GATE_INVALID')
+          }
+          const gate = await this.requestHumanGate({
+            type: 'retry',
+            runId: controller.runId,
+            agentRunId: error.runTrace.agentRunId,
+            agentKind: options.agentKind || kind,
+            summary: 'The previous write-capable Agent attempt may already have changed the workspace.',
+            options: [
+              { optionId: 'retry-once', name: 'Retry once', kind: 'allow_once' },
+              { optionId: 'cancel-retry', name: 'Do not retry', kind: 'reject_once' },
+            ],
+            request: {
+              failureCategory: failure.category,
+              outcomeCertainty: safety.outcomeCertainty,
+              sideEffectsPossible: safety.sideEffectsPossible,
+              operationId: safety.operationId,
+              idempotencyMode: safety.idempotencyMode,
+            },
+          }, {
+            signal: controller.signal,
+            preserveOnAbort: () => controller.stopReason === 'shutdown',
+            continuation: {
+              resumeKind: 'agent_slot',
+              agentRunId: error.runTrace.agentRunId,
+              agentKind: options.agentKind || kind,
+              round: controller.currentRound || 0,
+              ...(options.v4AgentSlotBinding || {}),
+            },
+          })
+          if (gate.status !== 'approved') {
+            this.completeHumanGateContinuation?.(controller.runId, gate.gateId, 'cancelled')
+            controller.stopReason = 'human_gate_rejected'
+            controller.abort()
+            throw new Error('LOCAL_AGENT_EXECUTION_STOPPED')
+          }
+          if (this.completeHumanGateContinuation?.(
+            controller.runId, gate.gateId, 'completed',
+          ) !== true && this.hasRunLedger()) {
+            throw new Error('LOCAL_RUN_PERSIST_FAILED')
+          }
+          this.markRuntimeCredential(kind, 'unknown')
+          try { await this.refreshAgents() } catch {
+            if (controller.signal.aborted) throw new Error('LOCAL_AGENT_EXECUTION_STOPPED')
+          }
+          await this.sleep(decision.delayMs, controller.signal)
+          continue
+        }
+        if (decision.action !== 'retry') {
+          this.recordAttempt(group, controller, {
+            agentKind: options.agentKind || kind,
+            phase,
+            attempt: phaseAttempt,
+            failureCategory: failure.category,
+            policyAction: options.failureAction
+              || (decision.action === 'refresh_session' ? 'refresh_session' : 'fail'),
+            backoffMs: options.failureBackoffMs || 0,
+            recoveryAgentKind: options.recoveryAgentKind || '',
+            finalOutcome: 'failed',
+            ...safety,
+          })
+          throw error
+        }
+        this.recordAttempt(group, controller, {
+          agentKind: options.agentKind || kind,
+          phase,
+          attempt: phaseAttempt,
+          failureCategory: failure.category,
+          policyAction: 'retry',
+          backoffMs: decision.delayMs,
+          recoveryAgentKind: options.recoveryAgentKind || '',
+          finalOutcome: 'failed',
+          ...safety,
+        })
+        // Transient retries keep the Agent's readiness untouched. Refreshing
+        // the whole catalog here re-ran discovery and credential probes while
+        // the Agent was busy and could flip a healthy Agent to unavailable.
+        await this.sleep(decision.delayMs, controller.signal)
+      }
+    }
+  },
+
+  async controlAfterFailure(group, kind, controller, error) {
+    const request = this.consumeAgentControl(controller, kind)
+    if (!request) return null
+    const failure = normalizeFailure(error)
+    const manualAttempt = (controller.manualRetryCounts.get(kind) || 0) + 1
+    if (request.action === 'retry') {
+      controller.manualRetryCounts.set(kind, manualAttempt)
+      const delayMs = this.retryDelay(manualAttempt)
+      this.recordAttempt(group, controller, {
+        agentKind: kind,
+        phase: 'manual_retry',
+        attempt: manualAttempt,
+        failureCategory: failure.category,
+        policyAction: 'retry',
+        backoffMs: delayMs,
+        finalOutcome: 'failed',
+      })
+      await this.prepareRetry(group, kind, controller, manualAttempt, {
+        resetSession: true,
+        delayMs,
+      })
+      return { action: 'retry' }
+    }
+    this.recordAttempt(group, controller, {
+      agentKind: kind,
+      phase: 'manual_retry',
+      attempt: manualAttempt,
+      failureCategory: failure.category,
+      policyAction: request.action === 'replace' ? 'replace_agent' : 'cancel',
+      recoveryAgentKind: request.replacementKind || '',
+      finalOutcome: request.action === 'replace' ? 'replaced' : 'cancelled',
+    })
+    return request
+  },
+
+  async invokeWithUnauthorizedRecovery(input) {
+    const { controller, kind } = input
+    if (controller.attemptHistory.length >= MAX_RUN_AGENT_ATTEMPTS) {
+      throw circuitBreakerError()
+    }
+    if (controller.agentSlotKinds.has(kind)) {
+      throw new Error('LOCAL_AGENT_ATTEMPT_RUNNING')
+    }
+    controller.agentSlotKinds.add(kind)
+    try {
+      return await this.invokeWithUnauthorizedRecoveryInSlot(input)
+    } finally {
+      controller.agentSlotKinds.delete(kind)
+    }
+  },
+
+  async invokeWithUnauthorizedRecoveryInSlot({
+    group,
+    kind,
+    controller,
+    threadRootId,
+    context,
+    mode = 'auto',
+  }) {
+    const requestedOperationId = String(context?.operationId || '')
+    const operationId = /^[A-Za-z0-9._:-]{1,120}$/.test(requestedOperationId)
+      ? requestedOperationId
+      : this.operationIdFor(controller, kind, mode)
+    const contract = this.retryContract(kind) || {}
+    const idempotencyMode = contract.idempotencyMode === 'durable' ? 'durable' : 'none'
+    const sideEffectsPossible = group.allowWrite === true
+      && context.permissionMode !== 'read-only'
+    const invokeSource = async () => {
+      if (context.parallelGraph !== true) this.setCurrentAgent(controller, kind)
+      return this.invokeAgent(group, kind, mode, controller.signal, threadRootId, {
+        ...context,
+        yolo: context.yolo ?? this.defaultYolo,
+        deferCredentialFailure: true,
+        operationId,
+      })
+    }
+    let phase = 'initial'
+    let phaseAttempt = 1
+    while (true) {
+      try {
+        return {
+          result: await this.invokeTransiently(invokeSource, group, kind, controller, {
+            phase,
+            attempt: phaseAttempt,
+            operationId,
+            idempotencyMode,
+            sideEffectsPossible,
+            deferFailurePolicy: context.v4SynthesisRecovery === true
+              || (context.v4 === true && context.resumedGate?.type === 'permission'),
+            // A timed-out automatic slot is a failed participant for this
+            // round. Retrying it serially delays every healthy peer and
+            // defeats failure isolation; manual/direct invocations retain
+            // their existing timeout retry policy.
+            retryTimeouts: mode !== 'auto',
+            v4AgentSlotBinding: mode === 'auto' && context.v4 === true ? {
+              phase: context.phase,
+              slotId: context.slotId,
+              operationId,
+              snapshotHash: context.snapshotHash,
+            } : null,
+          }),
+          removed: false,
+        }
+      } catch (error) {
+        if (controller.signal.aborted) throw error
+        const control = await this.controlAfterFailure(group, kind, controller, error)
+        if (control?.action === 'retry') {
+          phase = 'manual_retry'
+          phaseAttempt = controller.manualRetryCounts.get(kind) || 1
+          continue
+        }
+        if (control) return { result: null, removed: false, control, error }
+        if (!unauthorizedFailure(error)) throw error
+        // Only explicit sign-in failures reach this point. Mark the Agent as
+        // needing a fresh login; the catalog clears it on the next refresh
+        // after the retry window or as soon as native credentials verify.
+        this.markRuntimeCredential(kind, 'missing')
+        this.recordAttempt(group, controller, {
+          agentKind: kind,
+          phase,
+          attempt: phaseAttempt,
+          failureCategory: 'authentication',
+          policyAction: 'fail',
+          finalOutcome: 'failed',
+          ...normalizeFailureOutcome(error, {
+            sideEffectsPossible: group.allowWrite === true,
+            operationId,
+            idempotencyMode,
+          }),
+        })
+        throw sanitizedAuthenticationError(error)
+      }
+    }
+  },
+}
+
+module.exports = { recoveryMethods }

desktop/src/workspace/local-workspace-auto-runner-rounds.cjs

@@ -0,0 +1,313 @@
+// V2 round-based runs. Kept so older runs can be resumed; new runs do not use it.
+const { MAX_RUN_AGENT_ATTEMPTS } = require('../runs/failure-policy.cjs')
+const {
+  circuitBreakerError,
+  circuitBreakerFailure,
+  hardBudgetFailure,
+  terminalRunStatusForReason,
+} = require('./local-workspace-auto-runner-support.cjs')
+
+const roundsMethods = {
+  async runRounds(group, controller, threadRootId, maxRounds, context, resume = false) {
+    const {
+      rootAttachments,
+      rootSkillsByKind,
+      rootKnowledgeBasesByKind,
+    } = context
+    const cursor = resume ? controller.orchestration : null
+    const attachmentRecipients = new Set(cursor?.attachmentRecipients || [])
+    let terminalFailureOccurred = cursor?.terminalFailureOccurred === true
+    let activeKinds = cursor ? [...cursor.activeKinds] : [...controller.targetKinds]
+    let executionSequence = Number(cursor?.executionSequence) || 0
+    let totalSuccesses = cursor?.totalSuccesses || 0
+    let consensusReached = false
+    const reportedFailures = new Set()
+    const firstRound = resume ? Math.max(0, controller.currentRound - 1) : 0
+
+    for (
+      let round = firstRound;
+      (controller.unlimitedRounds || round < maxRounds) && !controller.signal.aborted;
+      round += 1
+    ) {
+      const resumedRound = resume && round === firstRound
+      const successfulKinds = new Set(resumedRound ? cursor.successfulKinds : [])
+      const agreementKinds = new Set(resumedRound ? cursor.agreementKinds : [])
+      const replacementInstructions = new Map()
+      const roundQueue = resumedRound ? [...cursor.pendingKinds] : [...activeKinds]
+      controller.currentRound = round + 1
+      controller.completedKinds = resumedRound
+        ? activeKinds.filter(kind => !roundQueue.includes(kind))
+        : []
+      controller.failedKinds = resumedRound
+        ? controller.completedKinds.filter(kind => !successfulKinds.has(kind))
+        : []
+      if (!resumedRound) {
+        this.checkpointOrchestration(group, controller, {
+          currentKind: '',
+          pendingKinds: [...roundQueue],
+          activeKinds: [...activeKinds],
+          successfulKinds: [],
+          agreementKinds: [],
+          attachmentRecipients: [...attachmentRecipients],
+          totalSuccesses,
+          terminalFailureOccurred,
+          executionSequence,
+        })
+      }
+      this.emitChanged()
+
+      while (roundQueue.length) {
+        const kind = roundQueue.shift()
+        if (!activeKinds.includes(kind)) {
+          this.checkpointOrchestration(group, controller, { pendingKinds: [...roundQueue] })
+          continue
+        }
+        if (controller.signal.aborted) break
+        const executionKind = kind
+        executionSequence += 1
+        this.setCurrentAgent(controller, executionKind)
+        this.checkpointOrchestration(group, controller, {
+          currentKind: executionKind,
+          pendingKinds: [...roundQueue],
+          activeKinds: [...activeKinds],
+          successfulKinds: [...successfulKinds],
+          agreementKinds: [...agreementKinds],
+          attachmentRecipients: [...attachmentRecipients],
+          totalSuccesses,
+          terminalFailureOccurred,
+          executionSequence,
+        })
+        try {
+          if (controller.attemptHistory.length >= MAX_RUN_AGENT_ATTEMPTS) {
+            throw circuitBreakerError()
+          }
+          const collaborationPackage = this.prepareCollaborationPackage(
+            group, controller, activeKinds, executionKind, threadRootId,
+          )
+          const attachments = attachmentRecipients.has(executionKind)
+            ? []
+            : rootAttachments.map(attachment => attachment.path)
+          const invocation = await this.invokeWithUnauthorizedRecovery({
+            group,
+            kind: executionKind,
+            controller,
+            activeKinds,
+            threadRootId,
+            context: {
+              attachments,
+              attachmentSnapshots: attachments.length ? rootAttachments : [],
+              skillHints: rootSkillsByKind.get(executionKind) || [],
+              knowledgeBaseHints: rootKnowledgeBasesByKind.get(executionKind) || [],
+              runtimeInstruction: replacementInstructions.get(executionKind) || '',
+              collaborationPackage,
+              contextOptions: {
+                focusUserMessageId: threadRootId,
+                omitAgentThreadRootId: threadRootId,
+              },
+              executionSequence,
+            },
+            reportedFailures,
+          })
+          replacementInstructions.delete(executionKind)
+          if (invocation.control?.action === 'replace') {
+            this.recordAgentInterruption(
+              group.id, executionKind, invocation.error, threadRootId, 'stopped', reportedFailures,
+            )
+            if (!controller.failedKinds.includes(executionKind)) {
+              controller.failedKinds.push(executionKind)
+            }
+            if (!controller.completedKinds.includes(executionKind)) {
+              controller.completedKinds.push(executionKind)
+            }
+            successfulKinds.delete(executionKind)
+            agreementKinds.delete(executionKind)
+            const replacementKind = invocation.control.replacementKind
+            activeKinds = activeKinds.filter(activeKind => activeKind !== executionKind)
+            replacementInstructions.set(
+              replacementKind,
+              this.replacementInstruction(executionKind),
+            )
+            if (!roundQueue.includes(replacementKind)) roundQueue.unshift(replacementKind)
+            controller.currentKind = ''
+            controller.progress = []
+            this.checkpointOrchestration(group, controller, {
+              currentKind: '',
+              pendingKinds: [...roundQueue],
+              activeKinds: [...activeKinds],
+              successfulKinds: [...successfulKinds],
+              agreementKinds: [...agreementKinds],
+            })
+            this.emitChanged()
+            continue
+          }
+          if (invocation.control?.action === 'cancel') {
+            this.recordAgentInterruption(
+              group.id, executionKind, invocation.error, threadRootId, 'stopped', reportedFailures,
+            )
+            if (!controller.failedKinds.includes(executionKind)) {
+              controller.failedKinds.push(executionKind)
+            }
+            if (!controller.completedKinds.includes(executionKind)) {
+              controller.completedKinds.push(executionKind)
+            }
+            successfulKinds.delete(executionKind)
+            agreementKinds.delete(executionKind)
+            controller.currentKind = ''
+            controller.progress = []
+            this.checkpointOrchestration(group, controller, {
+              currentKind: '',
+              pendingKinds: [...roundQueue],
+              successfulKinds: [...successfulKinds],
+              agreementKinds: [...agreementKinds],
+            })
+            this.emitChanged()
+            continue
+          }
+          const result = invocation.result
+          if (attachments.length) attachmentRecipients.add(executionKind)
+          successfulKinds.add(executionKind)
+          if (result.consensus) agreementKinds.add(executionKind)
+          else agreementKinds.delete(executionKind)
+          this.recordCollaborationResult(
+            group, controller, activeKinds, executionKind, result,
+          )
+        } catch (error) {
+          if (circuitBreakerFailure(error)) {
+            controller.stopReason = 'circuit_breaker'
+            controller.abort()
+            break
+          }
+          const hardBudget = hardBudgetFailure(error)
+          if (hardBudget) {
+            this.recordAgentFailure(
+              group.id, executionKind, error, threadRootId, reportedFailures,
+            )
+            successfulKinds.delete(executionKind)
+            agreementKinds.delete(executionKind)
+            if (!controller.failedKinds.includes(executionKind)) {
+              controller.failedKinds.push(executionKind)
+            }
+            if (!controller.completedKinds.includes(executionKind)) {
+              controller.completedKinds.push(executionKind)
+            }
+            terminalFailureOccurred = true
+            controller.stopReason = 'hard_budget'
+            controller.abort()
+            break
+          }
+          if (controller.signal.aborted) {
+            const interruptedKind = controller.currentKind || executionKind
+            if (error?.runTrace) {
+              this.recordAgentInterruption(
+                group.id,
+                interruptedKind,
+                error,
+                threadRootId,
+                controller.stopReason === 'shutdown' ? 'interrupted' : 'stopped',
+                reportedFailures,
+              )
+              if (!controller.completedKinds.includes(interruptedKind)) {
+                controller.completedKinds.push(interruptedKind)
+              }
+              controller.currentKind = ''
+              controller.progress = []
+              this.emitChanged()
+            }
+            break
+          }
+          this.recordAgentFailure(
+            group.id, executionKind, error, threadRootId, reportedFailures,
+          )
+          successfulKinds.delete(executionKind)
+          agreementKinds.delete(executionKind)
+          if (!controller.failedKinds.includes(executionKind)) {
+            controller.failedKinds.push(executionKind)
+          }
+          terminalFailureOccurred = true
+          // Failed CLIs leave only this run's active queue; the configured
+          // group keeps them so the next turn can retry.
+          if (this.shouldPauseFailedAgent(error)) {
+            this.pauseFailedAgentForRun(group, controller, executionKind, threadRootId, error)
+          }
+          activeKinds = activeKinds.filter(activeKind => activeKind !== executionKind)
+        }
+        if (controller.signal.aborted) break
+        if (!controller.completedKinds.includes(executionKind)) {
+          controller.completedKinds.push(executionKind)
+        }
+        controller.currentKind = ''
+        controller.progress = []
+        this.checkpointOrchestration(group, controller, {
+          currentKind: '',
+          pendingKinds: [...roundQueue],
+          activeKinds: [...activeKinds],
+          successfulKinds: [...successfulKinds],
+          agreementKinds: [...agreementKinds],
+          attachmentRecipients: [...attachmentRecipients],
+          totalSuccesses,
+          terminalFailureOccurred,
+          executionSequence,
+        })
+        this.emitChanged()
+      }
+
+      if (controller.signal.aborted) break
+      const successes = activeKinds.filter(kind => successfulKinds.has(kind)).length
+      const agreements = activeKinds.filter(kind => agreementKinds.has(kind)).length
+      totalSuccesses += successes
+      this.checkpointOrchestration(group, controller, {
+        currentKind: '',
+        pendingKinds: [],
+        activeKinds: [...activeKinds],
+        successfulKinds: [...successfulKinds],
+        agreementKinds: [...agreementKinds],
+        attachmentRecipients: [...attachmentRecipients],
+        totalSuccesses,
+        terminalFailureOccurred,
+        executionSequence,
+      })
+      if (activeKinds.length < 2) break
+      const crossReviewComplete = !controller.unlimitedRounds || controller.currentRound >= 2
+      if (crossReviewComplete
+          && successes === activeKinds.length
+          && agreements === activeKinds.length) {
+        consensusReached = true
+        break
+      }
+    }
+
+    let runStatus
+    if (controller.signal.aborted) {
+      runStatus = terminalRunStatusForReason(controller.stopReason)
+    } else if (terminalFailureOccurred) runStatus = totalSuccesses > 0 ? 'partial' : 'failed'
+    else if (consensusReached) runStatus = 'completed'
+    else runStatus = totalSuccesses > 0 ? 'round-limit' : 'failed'
+    if (!terminalFailureOccurred && !controller.unlimitedRounds
+        && (runStatus === 'round-limit' || runStatus === 'failed')) {
+      this.addMessage(
+        group.id,
+        'system',
+        `Automatic discussion reached its ${maxRounds}-round limit. The latest replies above are the current result.`,
+        '',
+        threadRootId,
+        { key: 'system.autoRoundLimit', params: { rounds: maxRounds } },
+      )
+    }
+    if (controller.stopReason === 'circuit_breaker') {
+      this.addMessage(
+        group.id,
+        'system',
+        `Automatic discussion stopped after ${MAX_RUN_AGENT_ATTEMPTS} Agent attempts.`,
+        '',
+        threadRootId,
+        {
+          key: 'system.runCircuitBreaker',
+          params: { maxAttempts: MAX_RUN_AGENT_ATTEMPTS },
+        },
+      )
+    }
+    return runStatus
+  },
+}
+
+module.exports = { roundsMethods }

desktop/src/workspace/local-workspace-auto-runner-task-graph.cjs

@@ -0,0 +1,419 @@
+// V3 task-graph runs. Kept so older runs can be resumed; new runs do not use it.
+const { createHash } = require('node:crypto')
+const {
+  appendBlackboardEntry,
+  emptyCollaborationState,
+  publicCollaborationText,
+} = require('../collaboration/collaboration-records.cjs')
+const { unlimitedReviewContract } = require('./local-workspace-context.cjs')
+const {
+  parseTaskGraphCursor,
+  readyTaskGraphNodes,
+  terminalTaskGraphState,
+  updateTaskGraphCursor,
+} = require('../collaboration/task-graph-records.cjs')
+const {
+  circuitBreakerFailure,
+  hardBudgetFailure,
+  terminalRunStatusForReason,
+} = require('./local-workspace-auto-runner-support.cjs')
+
+const taskGraphMethods = {
+  unresolvedTaskGraphConflicts(collaborationInput = null) {
+    const collaboration = collaborationInput || emptyCollaborationState()
+    const resolved = new Set(collaboration.entries
+      .filter(entry => entry.entryType === 'decision' && entry.lifecycle.state === 'active')
+      .flatMap(entry => entry.refs))
+    return collaboration.entries.filter(entry => (
+      entry.entryType === 'conflict'
+      && entry.lifecycle.state === 'active'
+      && !resolved.has(entry.entryId)
+    ))
+  },
+
+  taskGraphAcceptanceCriteria(node) {
+    const criteria = []
+    if (node.acceptance.requireConclusion) criteria.push('Return one concrete conclusion.')
+    if (node.acceptance.minArtifactRefs) {
+      criteria.push(`Produce at least ${node.acceptance.minArtifactRefs} durable Artifact reference(s).`)
+    }
+    if (node.acceptance.minEvidenceRefs) {
+      criteria.push(`Produce at least ${node.acceptance.minEvidenceRefs} durable Evidence reference(s).`)
+    }
+    return criteria.length ? criteria : ['Complete the explicit Human decision.']
+  },
+
+  taskGraphSelectedEntryIds(controller, node) {
+    const cursor = parseTaskGraphCursor(controller.orchestration.taskGraph)
+    const states = new Map(cursor.nodeStates.map(state => [state.nodeId, state]))
+    const selected = node.inputNodeIds.flatMap(nodeId => states.get(nodeId)?.entryIds || [])
+    if (['reviewer', 'arbiter'].includes(node.role)) {
+      selected.push(...this.unresolvedTaskGraphConflicts(
+        controller.orchestration.collaboration,
+      ).map(entry => entry.entryId))
+    }
+    return [...new Set(selected)]
+  },
+
+  taskGraphPrompt(node, objective, collaborationPackage, unlimitedRounds = false) {
+    return [
+      'MELDWORK_TASK_GRAPH_V1',
+      `Node: ${node.nodeId}`,
+      `Role: ${node.role}`,
+      `Objective: ${objective}`,
+      `Expected output: ${node.expectedOutput}`,
+      `Acceptance criteria: ${this.taskGraphAcceptanceCriteria(node).join(' | ')}`,
+      unlimitedReviewContract(unlimitedRounds),
+      collaborationPackage.text,
+      'Return the requested result directly. Completion is evaluated from durable typed records; do not emit a consensus marker.',
+    ].join('\n')
+  },
+
+  checkpointTaskGraph(group, controller, cursorInput, options = {}) {
+    let taskGraph = parseTaskGraphCursor(cursorInput)
+    taskGraph = parseTaskGraphCursor({
+      ...taskGraph,
+      terminalState: terminalTaskGraphState(taskGraph),
+    })
+    const states = new Map(taskGraph.nodeStates.map(state => [state.nodeId, state]))
+    const pendingKinds = taskGraph.graph.nodes
+      .filter(node => states.get(node.nodeId)?.status === 'pending' && node.agentKind)
+      .map(node => node.agentKind)
+    const successfulKinds = taskGraph.graph.nodes
+      .filter(node => states.get(node.nodeId)?.status === 'accepted' && node.agentKind)
+      .map(node => node.agentKind)
+    const uniquePendingKinds = [...new Set(pendingKinds)]
+    const uniqueSuccessfulKinds = [...new Set(successfulKinds)]
+    this.checkpointOrchestration(group, controller, {
+      taskGraph,
+      currentKind: options.currentKind || '',
+      pendingKinds: uniquePendingKinds,
+      successfulKinds: uniqueSuccessfulKinds,
+      agreementKinds: [],
+      totalSuccesses: taskGraph.nodeStates.filter(state => state.status === 'accepted').length,
+      terminalFailureOccurred: taskGraph.terminalState === 'failed',
+    })
+    return taskGraph
+  },
+
+  taskGraphResultAccepted(node, result) {
+    const conclusion = publicCollaborationText(result?.message?.content || '')
+    const artifactIds = result?.outcomeRefs?.artifactIds || []
+    const evidenceIds = result?.outcomeRefs?.evidenceIds || []
+    return (!node.acceptance.requireConclusion || Boolean(conclusion))
+      && artifactIds.length >= node.acceptance.minArtifactRefs
+      && evidenceIds.length >= node.acceptance.minEvidenceRefs
+  },
+
+  applyTaskGraphAgentResult(group, controller, node, result) {
+    const conflicts = this.unresolvedTaskGraphConflicts(controller.orchestration.collaboration)
+    const entryIds = this.recordCollaborationResult(
+      group,
+      controller,
+      controller.orchestration.activeKinds,
+      node.agentKind,
+      result,
+      {
+        role: node.role,
+        entryType: node.role === 'arbiter' ? 'decision' : 'claim',
+        subject: node.role === 'primary'
+          ? `task:${controller.taskId}:conclusion`
+          : `task:${controller.taskId}:node:${node.nodeId}`,
+        refs: node.role === 'arbiter' ? conflicts.map(entry => entry.entryId) : [],
+      },
+    )
+    const accepted = this.taskGraphResultAccepted(node, result)
+    const conclusion = publicCollaborationText(result?.message?.content || '')
+    const artifactIds = result?.outcomeRefs?.artifactIds || []
+    const evidenceIds = result?.outcomeRefs?.evidenceIds || []
+    let cursor = updateTaskGraphCursor(
+      controller.orchestration.taskGraph,
+      node.nodeId,
+      {
+        status: accepted ? 'accepted' : 'failed',
+        entryIds,
+        artifactIds,
+        evidenceIds,
+        conclusionHash: conclusion
+          ? createHash('sha256').update(conclusion).digest('hex')
+          : '',
+      },
+    )
+    if (accepted) {
+      if (!controller.completedKinds.includes(node.agentKind)) {
+        controller.completedKinds.push(node.agentKind)
+      }
+    } else if (!controller.failedKinds.includes(node.agentKind)) {
+      controller.failedKinds.push(node.agentKind)
+    }
+    cursor = this.checkpointTaskGraph(group, controller, cursor)
+    return accepted
+  },
+
+  taskGraphDecisionAgentRunId(graphId, nodeId) {
+    const digest = createHash('sha256').update(JSON.stringify([graphId, nodeId])).digest('hex')
+    return `graph-decision-${digest}`
+  },
+
+  taskGraphDecisionAnchor(controller, node) {
+    const graph = parseTaskGraphCursor(controller.orchestration.taskGraph)
+    const byId = new Map(graph.graph.nodes.map(candidate => [candidate.nodeId, candidate]))
+    return [...node.dependsOn].reverse()
+      .map(nodeId => byId.get(nodeId)?.agentKind)
+      .find(Boolean) || controller.targetKinds[0]
+  },
+
+  applyTaskGraphHumanDecision(group, controller, node, decision) {
+    const option = node.decisionOptions.find(candidate => candidate.optionId === decision.optionId)
+    if (!option) throw new Error('LOCAL_WORKFLOW_DECISION_INVALID')
+    const accepted = decision.status === 'approved'
+    const conflicts = this.unresolvedTaskGraphConflicts(controller.orchestration.collaboration)
+    let collaboration = controller.orchestration.collaboration
+    collaboration = appendBlackboardEntry(collaboration, {
+      entryType: 'decision',
+      subject: `task:${controller.taskId}:node:${node.nodeId}`,
+      statement: `Human selected: ${option.name}`,
+      value: '',
+      owner: { type: 'harness' },
+      audience: { roles: ['primary', 'reviewer', 'arbiter'], agentKinds: [] },
+      lifecycle: {
+        state: 'active',
+        sequence: collaboration.entries.length + 1,
+        recordedAt: Date.now(),
+        supersedesEntryId: null,
+      },
+      provenance: {
+        runId: controller.runId,
+        taskId: controller.taskId,
+        round: controller.currentRound,
+        agentRunId: this.taskGraphDecisionAgentRunId(
+          controller.orchestration.taskGraph.graph.graphId,
+          node.nodeId,
+        ),
+        artifactIds: [],
+        evidenceIds: [],
+      },
+      refs: conflicts.map(entry => entry.entryId),
+    })
+    const decisionEntryId = collaboration.entries.at(-1).entryId
+    this.checkpointOrchestration(group, controller, { collaboration })
+    let cursor = updateTaskGraphCursor(
+      controller.orchestration.taskGraph,
+      node.nodeId,
+      {
+        status: accepted ? 'accepted' : 'rejected',
+        entryIds: [decisionEntryId],
+        decisionOptionId: option.optionId,
+      },
+    )
+    cursor = this.checkpointTaskGraph(group, controller, cursor)
+    return accepted
+  },
+
+  async executeTaskGraphAgentNode(
+    group, controller, threadRootId, context, node, parallelGraph = false,
+  ) {
+    const selectedEntryIds = this.taskGraphSelectedEntryIds(controller, node)
+    const collaborationPackage = this.prepareCollaborationPackage(
+      group,
+      controller,
+      controller.orchestration.activeKinds,
+      node.agentKind,
+      threadRootId,
+      {
+        role: node.role,
+        selectedEntryIds,
+        expectedOutput: node.expectedOutput,
+        acceptanceCriteria: this.taskGraphAcceptanceCriteria(node),
+      },
+    )
+    const isolated = ['reviewer', 'arbiter'].includes(node.role)
+    const attachments = node.role === 'primary'
+      && !controller.orchestration.attachmentRecipients.includes(node.agentKind)
+      ? context.rootAttachments.map(attachment => attachment.path)
+      : []
+    const objective = this.collaborationObjective(group, controller, threadRootId)
+    return this.invokeWithUnauthorizedRecovery({
+      group,
+      kind: node.agentKind,
+      controller,
+      threadRootId,
+      context: {
+        attachments,
+        attachmentSnapshots: attachments.length ? context.rootAttachments : [],
+        skillHints: context.rootSkillsByKind.get(node.agentKind) || [],
+        knowledgeBaseHints: context.rootKnowledgeBasesByKind.get(node.agentKind) || [],
+        collaborationPackage,
+        completionPolicy: 'typed',
+        parallelGraph,
+        runtimeInstruction: isolated ? '' : [
+          `Execute task-graph node ${node.nodeId} as ${node.role}.`,
+          `Expected output: ${node.expectedOutput}`,
+          'Do not emit a consensus marker.',
+        ].join('\n'),
+        ...(isolated ? {
+          sessionPolicy: 'isolated',
+          promptOverride: this.taskGraphPrompt(
+            node, objective, collaborationPackage, controller.unlimitedRounds === true,
+          ),
+          contextPackId: controller.contextPackId,
+        } : {}),
+        contextOptions: {
+          focusUserMessageId: threadRootId,
+          omitAgentThreadRootId: threadRootId,
+        },
+      },
+      mode: 'auto',
+    })
+  },
+
+  async executeTaskGraphHumanNode(group, controller, node) {
+    if (!this.requestHumanGate) throw new Error('LOCAL_WORKFLOW_DECISION_UNAVAILABLE')
+    const anchorKind = this.taskGraphDecisionAnchor(controller, node)
+    const agentRunId = this.taskGraphDecisionAgentRunId(
+      controller.orchestration.taskGraph.graph.graphId,
+      node.nodeId,
+    )
+    let cursor = updateTaskGraphCursor(
+      controller.orchestration.taskGraph,
+      node.nodeId,
+      { status: 'waiting', attention: 'decision', attempts: 1 },
+    )
+    cursor = this.checkpointTaskGraph(group, controller, cursor, { currentKind: anchorKind })
+    const gate = await this.requestHumanGate({
+      type: 'decision',
+      runId: controller.runId,
+      agentRunId,
+      agentKind: anchorKind,
+      summary: node.expectedOutput,
+      options: node.decisionOptions,
+      request: {
+        graphId: cursor.graph.graphId,
+        nodeId: node.nodeId,
+        inputNodeIds: node.inputNodeIds,
+      },
+    }, {
+      signal: controller.signal,
+      preserveOnAbort: () => controller.stopReason === 'shutdown',
+      continuation: {
+        resumeKind: 'role_review_decision',
+        agentRunId,
+        agentKind: anchorKind,
+        round: controller.currentRound,
+      },
+    })
+    const accepted = this.applyTaskGraphHumanDecision(group, controller, node, gate)
+    if (this.completeHumanGateContinuation?.(
+      controller.runId,
+      gate.gateId,
+      accepted ? 'completed' : 'cancelled',
+    ) !== true && this.hasRunLedger()) {
+      throw new Error('LOCAL_RUN_PERSIST_FAILED')
+    }
+    return accepted
+  },
+
+  taskGraphBatch(group, readyNodes) {
+    const parallel = []
+    const kinds = new Set()
+    if (group.allowWrite !== true) {
+      for (const node of readyNodes) {
+        if (!node.parallel || node.role === 'human' || kinds.has(node.agentKind)) continue
+        parallel.push(node)
+        kinds.add(node.agentKind)
+      }
+    }
+    return parallel.length > 1 ? parallel : readyNodes.slice(0, 1)
+  },
+
+  async runTaskGraph(group, controller, threadRootId, context) {
+    while (!controller.signal.aborted) {
+      let cursor = parseTaskGraphCursor(controller.orchestration.taskGraph)
+      const terminalState = terminalTaskGraphState(cursor)
+      if (terminalState !== 'running') {
+        cursor = this.checkpointTaskGraph(group, controller, cursor)
+        break
+      }
+      if (!controller.unlimitedRounds && controller.currentRound >= controller.maxRounds) break
+      const readyNodes = readyTaskGraphNodes(cursor)
+      if (!readyNodes.length) {
+        cursor = parseTaskGraphCursor({ ...cursor, terminalState: 'failed' })
+        this.checkpointTaskGraph(group, controller, cursor)
+        break
+      }
+      const batch = this.taskGraphBatch(group, readyNodes)
+      controller.currentRound += 1
+      if (batch.length === 1 && batch[0].role === 'human') {
+        await this.executeTaskGraphHumanNode(group, controller, batch[0])
+        continue
+      }
+      const conflicts = this.unresolvedTaskGraphConflicts(controller.orchestration.collaboration)
+      for (const node of batch) {
+        cursor = updateTaskGraphCursor(cursor, node.nodeId, {
+          status: 'running',
+          attention: node.role === 'reviewer' && conflicts.length
+            ? 'review'
+            : (node.role === 'arbiter' && conflicts.length ? 'decision' : 'none'),
+          attempts: cursor.nodeStates.find(state => state.nodeId === node.nodeId).attempts + 1,
+        })
+      }
+      const currentKind = batch.length === 1 ? batch[0].agentKind : ''
+      cursor = this.checkpointTaskGraph(group, controller, cursor, { currentKind })
+      this.emitChanged()
+      const settled = await Promise.allSettled(batch.map(node => (
+        this.executeTaskGraphAgentNode(
+          group, controller, threadRootId, context, node, batch.length > 1,
+        )
+      )))
+      for (let index = 0; index < batch.length; index += 1) {
+        const node = batch[index]
+        const outcome = settled[index]
+        if (outcome.status === 'fulfilled' && outcome.value?.result) {
+          if (outcome.value.result.outcomeRefs?.artifactIds?.length
+              && !controller.orchestration.attachmentRecipients.includes(node.agentKind)) {
+            this.checkpointOrchestration(group, controller, {
+              attachmentRecipients: [...controller.orchestration.attachmentRecipients, node.agentKind],
+            })
+          }
+          this.applyTaskGraphAgentResult(group, controller, node, outcome.value.result)
+          continue
+        }
+        const error = outcome.status === 'rejected'
+          ? outcome.reason
+          : (outcome.value?.error || new Error('LOCAL_AGENT_UNKNOWN_FAILURE'))
+        if (controller.signal.aborted) continue
+        this.recordAgentFailure(group.id, node.agentKind, error, threadRootId, new Set())
+        cursor = updateTaskGraphCursor(
+          controller.orchestration.taskGraph,
+          node.nodeId,
+          { status: 'failed' },
+        )
+        this.checkpointTaskGraph(group, controller, cursor)
+        if (!controller.failedKinds.includes(node.agentKind)) {
+          controller.failedKinds.push(node.agentKind)
+        }
+        if (circuitBreakerFailure(error)) {
+          controller.stopReason = 'circuit_breaker'
+          controller.abort()
+        } else if (hardBudgetFailure(error)) {
+          controller.stopReason = 'hard_budget'
+          controller.abort()
+        }
+      }
+      controller.currentKind = ''
+      controller.progress = []
+      this.emitChanged()
+    }
+
+    if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+    const cursor = parseTaskGraphCursor(controller.orchestration.taskGraph)
+    if (cursor.terminalState === 'accepted') return 'completed'
+    if (cursor.terminalState === 'rejected') return 'stopped'
+    if (cursor.terminalState === 'failed') {
+      return cursor.nodeStates.some(state => state.status === 'accepted') ? 'partial' : 'failed'
+    }
+    return cursor.nodeStates.some(state => state.status === 'accepted') ? 'round-limit' : 'failed'
+  },
+}
+
+module.exports = { taskGraphMethods }

desktop/src/workspace/local-workspace-auto-runner-v4-context.cjs

@@ -0,0 +1,403 @@
+// What a V4 run freezes and shares: the task snapshot, stored receipts and the collaboration package.
+const { createHash } = require('node:crypto')
+const { peerCollaborationText, publicCollaborationText } = require('../collaboration/collaboration-records.cjs')
+const {
+  buildCollaborationPackage,
+  hashValue,
+  parseCollaborationControlBlock,
+  resolveCoordinationConsensus,
+} = require('../collaboration/orchestration-v4-records.cjs')
+const { canonicalJson } = require('../collaboration/context-pack-records.cjs')
+const {
+  restoreV4SnapshotSkills,
+  v4SnapshotBodyHash,
+  v4SnapshotSkillHints,
+  validateV4SnapshotBody,
+} = require('./local-workspace-context.cjs')
+
+const v4ContextMethods = {
+  v4OperationId(
+    controller, kind, phase, slotId = '', round = controller.currentRound || 0,
+  ) {
+    const digest = createHash('sha256').update(JSON.stringify([
+      controller.runId,
+      controller.taskId,
+      kind,
+      slotId,
+      phase,
+      round,
+    ])).digest('hex')
+    return `operation-${digest}`
+  },
+
+  v4SnapshotRecord(controller, snapshot, targetKinds) {
+    if (!this.contentBlobStore) throw new Error('LOCAL_RUN_SNAPSHOT_STORE_UNAVAILABLE')
+    const serialized = canonicalJson(snapshot)
+    const contentRef = this.contentBlobStore.put(serialized, { mediaType: 'application/json' })
+    const sourceIds = [snapshot?.messageId, ...(snapshot?.history || []).map(item => item.id)]
+      .filter(Boolean).slice(-64)
+    const taskId = snapshot?.taskId || controller.taskId || null
+    const record = {
+      contextPackId: controller.contextPackId || null,
+      taskId,
+      messageId: snapshot?.messageId || controller.threadRootId || null,
+      groupId: snapshot?.group?.id || controller.groupId || null,
+      round: Math.max(0, Number(controller.currentRound) || 0),
+      targetKinds: [...targetKinds],
+      sourceIds,
+      capturedAt: Date.now(),
+      charCount: Buffer.byteLength(serialized),
+      contentHash: contentRef.hash,
+      bodyHash: v4SnapshotBodyHash(snapshot),
+      contentRef,
+    }
+    return { record, snapshotHash: hashValue(record) }
+  },
+
+  v4SnapshotSkillHints(snapshot, targetKinds) {
+    return v4SnapshotSkillHints(snapshot, targetKinds)
+  },
+
+  async v4RestoreSnapshotSkills(snapshot, targetKinds, persisted = null) {
+    return restoreV4SnapshotSkills({
+      snapshot,
+      targetKinds,
+      validateSkillSelections: this.validateSkillSelections,
+      persisted,
+    })
+  },
+
+  v4LoadSnapshot(orchestration, { taskId, messageId, groupId, targetKinds }) {
+    const record = orchestration?.snapshot
+    if (!record?.contentRef || !this.contentBlobStore) {
+      throw new Error('LOCAL_RUN_SNAPSHOT_INVALID')
+    }
+    let bytes
+    try { bytes = this.contentBlobStore.read(record.contentRef) } catch {
+      throw new Error('LOCAL_RUN_SNAPSHOT_INVALID')
+    }
+    const serialized = bytes.toString('utf8')
+    let body
+    try { body = JSON.parse(serialized) } catch {
+      throw new Error('LOCAL_RUN_SNAPSHOT_INVALID')
+    }
+    return validateV4SnapshotBody({
+      body,
+      serialized,
+      byteLength: bytes.length,
+      record,
+      orchestrationSnapshotHash: orchestration.snapshotHash,
+      taskId,
+      messageId,
+      groupId,
+      targetKinds,
+    })
+  },
+
+  v4RestoreReceipts(controller) {
+    if (Array.isArray(controller.v4ReceiptRecords)) {
+      return controller.v4ReceiptRecords.map(item => ({
+        receipt: item.receipt,
+        verdict: item.verdict || '',
+        resolvedIssueIds: Array.isArray(item.resolvedIssueIds)
+          ? [...item.resolvedIssueIds]
+          : [],
+      }))
+    }
+    const restored = []
+    const slots = Array.isArray(controller.orchestration?.slots)
+      ? controller.orchestration.slots
+      : []
+    for (const slot of slots) {
+      const records = slot?.resultRefs?.workflowOutcomeRefs
+      if (!Array.isArray(records)) continue
+      for (const record of records) {
+        if (!record?.receipt) continue
+        try {
+          const receipt = parseCollaborationControlBlock(record.receipt)
+          restored.push({
+            receipt,
+            verdict: typeof record.verdict === 'string' ? record.verdict : '',
+            resolvedIssueIds: Array.isArray(record.resolvedIssueIds)
+              ? record.resolvedIssueIds.filter(value => typeof value === 'string')
+              : [],
+          })
+        } catch { /* malformed historical receipts are ignored and re-run */ }
+      }
+    }
+    controller.v4ReceiptRecords = restored
+    return restored
+  },
+
+  v4Package(receiptRecords, options = {}) {
+    const recipientKind = String(options.recipientKind || '')
+    const binding = options.sessionBinding || {}
+    const deliveryState = Array.isArray(options.deliveryState) ? options.deliveryState : []
+    const latest = new Map()
+    for (const item of receiptRecords) {
+      const receipt = item?.receipt
+      if (!receipt) continue
+      const key = receipt.agentKind || `${receipt.phase}\u0000${receipt.summary}`
+      const previous = latest.get(key)
+      if (!previous || (Number(receipt.deliveryWatermark) || 0)
+          > (Number(previous.receipt.deliveryWatermark) || 0)
+          || ((Number(receipt.deliveryWatermark) || 0)
+            === (Number(previous.receipt.deliveryWatermark) || 0)
+            && receipt.phase.localeCompare(previous.receipt.phase) > 0)) {
+        latest.set(key, item)
+      }
+    }
+    let selectedItems = [...latest.values()]
+    const snapshotHash = typeof options.snapshotHash === 'string' ? options.snapshotHash : ''
+    const targetKinds = new Set(Array.isArray(options.targetKinds) ? options.targetKinds : [])
+    const resolvedIssueIds = []
+    const resolved = new Set()
+    for (const item of receiptRecords) {
+      const receipt = item?.receipt
+      if (!snapshotHash || receipt?.snapshotHash !== snapshotHash
+          || (targetKinds.size && !targetKinds.has(receipt.agentKind))) continue
+      for (const value of Array.isArray(item.resolvedIssueIds) ? item.resolvedIssueIds : []) {
+        if (typeof value !== 'string' || !value || value.length > 120 || resolved.has(value)) continue
+        resolved.add(value)
+        resolvedIssueIds.push(value)
+        if (resolvedIssueIds.length >= 256) break
+      }
+      if (resolvedIssueIds.length >= 256) break
+    }
+    if (recipientKind && binding.sessionRefHash && !options.forceFull) {
+      const acknowledged = new Map(deliveryState
+        .filter(item => item.recipientKind === recipientKind
+          && item.sessionRefHash === binding.sessionRefHash
+          && item.sessionProvenanceHash === binding.sessionProvenanceHash
+          && item.status === 'acknowledged')
+        .map(item => [`${item.sourceAgentKind}\u0000${item.sourcePhase}`, item.watermark]))
+      selectedItems = selectedItems.filter(item => {
+        const receipt = item.receipt
+        const previous = acknowledged.get(`${receipt.agentKind}\u0000${receipt.phase}`) || 0
+        return (Number(receipt.deliveryWatermark) || 0) > previous
+      })
+    }
+    const selectedReceipts = selectedItems.map(item => item.receipt)
+    const totalLimit = Number.isSafeInteger(options.totalLimit) ? options.totalLimit : 6000
+    const packageRecord = buildCollaborationPackage(selectedReceipts, {
+      targetKinds: options.targetKinds || [],
+      totalLimit,
+      resolvedIssueIds,
+    })
+    const text = this.v4SanitizeDeliveryText(packageRecord.text, packageRecord.text.length)
+    return text === packageRecord.text
+      ? packageRecord
+      : { ...packageRecord, text, totalChars: text.length, packageHash: hashValue(text) }
+  },
+
+  v4CoordinationResult({
+    receiptRecords, targetKinds, snapshotHash, snapshot = null,
+    challengeBindings = [], requireSupport = true,
+  }) {
+    const candidateReceipts = receiptRecords
+      .filter(record => record?.receipt?.phase === 'challenge')
+    const supportReceipts = requireSupport
+      ? targetKinds.map((kind) => {
+          const operationId = challengeBindings.find(binding => (
+            binding.reviewerKind === kind
+          ))?.reviewerOperationId || ''
+          return this.v4ReceiptForOperation(
+            receiptRecords, 'challenge', kind, operationId,
+          ) || null
+        })
+      : []
+    if (supportReceipts.some(receipt => !receipt)) {
+      const result = resolveCoordinationConsensus({
+        targetKinds, snapshotHash, candidateReceipts, supportReceipts: [],
+      })
+      return this.v4SemanticCoordinationResult(result, snapshot)
+    }
+    const result = resolveCoordinationConsensus({
+      targetKinds, snapshotHash, candidateReceipts, supportReceipts,
+    })
+    return this.v4SemanticCoordinationResult(result, snapshot)
+  },
+
+  v4SemanticCoordinationResult(result, snapshot) {
+    const candidates = result.candidates.filter((candidate) => {
+      try {
+        for (const assignment of candidate.assignments) {
+          this.v4DirectAssignmentReferences(assignment, snapshot)
+        }
+        return true
+      } catch (error) {
+        if (error?.message === 'LOCAL_RUN_V4_ASSIGNMENT_REFERENCE_INVALID') return false
+        throw error
+      }
+    })
+    const candidateHashes = new Set(candidates.map(candidate => candidate.planHash))
+    return {
+      ...result,
+      plan: result.plan && candidateHashes.has(result.plan.planHash) ? result.plan : null,
+      candidates,
+    }
+  },
+
+  v4SanitizeDeliveryText(value, limit = 6000, options = {}) {
+    const privateWrapper = '(?:private[ _-]?(?:reasoning|thoughts?)|analysis|chain[ _-]?of[ _-]?thought|cot)'
+    const privateHeading = '(?:private[ _-]?(?:reasoning|thoughts?)|chain[ _-]?of[ _-]?thought|cot)'
+    const stripped = this.v4StripMeldworkControls(
+      options.workspaceRoot
+        ? peerCollaborationText(value, { workspaceRoot: options.workspaceRoot })
+        : publicCollaborationText(value, Number.MAX_SAFE_INTEGER),
+    )
+    const maximum = Number.isSafeInteger(limit) && limit >= 0 ? limit : 6000
+    let privateSectionLevel = null
+    const withoutPrivateSections = stripped.split(/\r?\n/).filter((line) => {
+      const heading = line.match(/^\s*(#{1,6})\s+(.+)$/u)
+      const level = heading?.[1].length
+      if (heading && new RegExp(`^${privateHeading}\\b`, 'iu').test(heading[2].trim())) {
+        privateSectionLevel = Math.min(privateSectionLevel ?? level, level)
+        return false
+      }
+      if (privateSectionLevel !== null && heading && level <= privateSectionLevel) {
+        privateSectionLevel = null
+      }
+      return privateSectionLevel === null
+    }).join('\n')
+    return withoutPrivateSections
+      .replace(new RegExp(`<\\s*${privateWrapper}\\b[^>]*>[\\s\\S]*?<\\/\\s*${privateWrapper}\\s*>`, 'giu'), '')
+      .replace(/(?:^|\n)\s*(?:my\s+)?private[ _-]?reasoning\s*:\s*[^\n]*/giu, '\n')
+      .replace(/\b(?:session(?:[ _-]?(?:id|ref))?)\b\s*(?::|=)\s*[A-Za-z0-9][A-Za-z0-9._:-]*/giu, '[redacted]')
+      .replace(/\b(?:session(?:[ _-]?(?:id|ref))?)\b\s+[A-Za-z0-9][A-Za-z0-9._:-]*/giu, '[redacted]')
+      .replace(/\b(?:session[-_][A-Za-z0-9._:-]+|[A-Za-z0-9][A-Za-z0-9._:-]*-(?:native-)?session)\b/giu, '[redacted]')
+      .replace(/\n{3,}/g, '\n\n')
+      .trim()
+      .slice(0, maximum)
+  },
+
+  v4StripMeldworkControls(text) {
+    const marker = '[[MELDWORK_'
+    const source = String(text || '')
+    const upper = source.toUpperCase()
+    let cursor = 0
+    let output = ''
+    while (true) {
+      const start = upper.indexOf(marker, cursor)
+      if (start < 0) return output + source.slice(cursor)
+      output += source.slice(cursor, start)
+      let quote = ''
+      let escaped = false
+      let braces = 0
+      let brackets = 0
+      let end = -1
+      for (let index = start + 2; index < source.length - 1; index += 1) {
+        const character = source[index]
+        if (quote) {
+          if (escaped) escaped = false
+          else if (character === '\\') escaped = true
+          else if (character === quote) quote = ''
+          continue
+        }
+        if (character === '"' || character === "'") {
+          quote = character
+        } else if (character === '{') {
+          braces += 1
+        } else if (character === '}') {
+          braces = Math.max(0, braces - 1)
+        } else if (character === '[') {
+          brackets += 1
+        } else if (character === ']') {
+          if (braces === 0 && brackets === 0 && source[index + 1] === ']') {
+            end = index + 2
+            break
+          }
+          brackets = Math.max(0, brackets - 1)
+        }
+      }
+      if (end < 0) return output
+      const header = source.slice(start, end)
+      if (!header.includes(':')) {
+        const closing = `[[/${header.slice(2, -2)}]]`.toUpperCase()
+        const closingStart = upper.indexOf(closing, end)
+        cursor = closingStart < 0 ? end : closingStart + closing.length
+      } else {
+        cursor = end
+      }
+    }
+  },
+
+  v4CoordinationText(result, limit = 4800) {
+    const candidates = Array.isArray(result?.candidates) ? result.candidates : []
+    if (!candidates.length) return ''
+    const supportCounts = new Map()
+    for (const planHash of result.supportPlanHashes || []) {
+      if (planHash) supportCounts.set(planHash, (supportCounts.get(planHash) || 0) + 1)
+    }
+    const lines = [
+      'MELDWORK_V4_RESPONSIBILITY_CANDIDATES_V1',
+      'Fields: h=planHash;s=support;f=finalizer;v=verifiers;a=[t=taskId,o=owner,r=role,x=objective,y=expectedOutput,i=inputRefs,a=artifactRefs,d=dependencies]',
+    ]
+    const append = (line) => {
+      const next = [...lines, line].join('\n')
+      if (next.length > limit) return false
+      lines.push(line)
+      return true
+    }
+    for (const plan of candidates) {
+      let safe = true
+      const text = value => {
+        const raw = String(value || '')
+        const sanitized = this.v4SanitizeDeliveryText(raw, 2000)
+        if (/\[\[MELDWORK_/iu.test(raw) || sanitized !== raw) safe = false
+        return sanitized
+      }
+      const values = value => Array.isArray(value) ? value.map(text) : []
+      const graph = {
+        h: plan.planHash,
+        s: supportCounts.get(plan.planHash) || 0,
+        f: text(plan.finalizerKind),
+        v: values(plan.verifierKinds),
+        a: plan.assignments.map(assignment => ({
+          t: text(assignment.taskId),
+          o: text(assignment.ownerKind),
+          r: text(assignment.role),
+          x: text(assignment.objective),
+          y: text(assignment.expectedOutput),
+          i: values(assignment.inputRefs),
+          a: values(assignment.artifactIds),
+          d: values(assignment.dependsOn),
+        })),
+      }
+      if (!safe) continue
+      append(JSON.stringify(graph))
+    }
+    return lines.join('\n')
+  },
+
+  v4ChallengeProposalText(binding, reviewerKind, limit = 1600) {
+    const proposalKind = String(binding?.proposalKind || '')
+    if (!proposalKind || proposalKind === reviewerKind) {
+      throw new Error('LOCAL_RUN_V4_CHALLENGE_BINDING_INVALID')
+    }
+    for (const artifactId of binding.artifactIds || []) {
+      try {
+        const { artifact, content } = this.v4ArtifactIdentity(artifactId)
+        const body = this.v4SanitizeDeliveryText(content, limit)
+        if (artifact.producedBy?.agentKind !== proposalKind || !body) continue
+        return [
+          'MELDWORK_V4_ASSIGNED_PROPOSAL_V1',
+          `Proposal Agent: ${proposalKind}`,
+          `Proposal Artifact: ${artifactId}`,
+          'Proposal body excerpt:',
+          body,
+        ].join('\n')
+      } catch { /* skip non-text and unavailable outcome refs */ }
+    }
+    throw new Error('LOCAL_RUN_V4_CHALLENGE_BINDING_INVALID')
+  },
+
+  v4CoordinationAssignment(coordinationPlan, kind, options = {}) {
+    const owned = coordinationPlan?.assignments?.filter(item => item.ownerKind === kind) || []
+    if (options.taskId) return owned.find(item => item.taskId === options.taskId) || null
+    if (options.role) return owned.find(item => item.role === options.role) || null
+    return owned[0] || null
+  },
+}
+
+module.exports = { v4ContextMethods }

desktop/src/workspace/local-workspace-auto-runner-v4-delivery.cjs

@@ -0,0 +1,604 @@
+// How a V4 run hands context to each Agent: assignments, the blackboard owner, delivery prompts and delivery state.
+const { createHash } = require('node:crypto')
+const { collaborationPackageIndexText } = require('../collaboration/collaboration-records.cjs')
+const { hashValue } = require('../collaboration/orchestration-v4-records.cjs')
+const { v4Prompt } = require('./local-workspace-context.cjs')
+
+const v4DeliveryMethods = {
+  v4DirectAssignmentReferences(assignment, snapshot) {
+    const snapshotSources = new Map([
+      [snapshot?.messageId, snapshot?.taskText],
+      ...(Array.isArray(snapshot?.history) ? snapshot.history : []).map(item => [item?.id, item?.text]),
+    ].filter(([sourceId]) => typeof sourceId === 'string' && sourceId))
+    const safeReferenceText = (value) => {
+      const raw = String(value || '').trim()
+      const sanitized = this.v4SanitizeDeliveryText(raw, Number.MAX_SAFE_INTEGER)
+      if (!raw || sanitized !== raw || /\[\[MELDWORK_/iu.test(raw)) {
+        throw new Error('LOCAL_RUN_V4_ASSIGNMENT_REFERENCE_INVALID')
+      }
+      return sanitized
+    }
+    const inputReferences = assignment.inputRefs.map((sourceId) => {
+      if (!snapshotSources.has(sourceId)) {
+        throw new Error('LOCAL_RUN_V4_ASSIGNMENT_REFERENCE_INVALID')
+      }
+      return {
+        kind: 'input', id: sourceId, header: `Input ${sourceId}:`,
+        body: safeReferenceText(snapshotSources.get(sourceId)),
+      }
+    })
+    const directArtifacts = assignment.artifactIds.map((artifactId) => {
+      let identity
+      try { identity = this.v4ArtifactIdentity(artifactId) } catch {
+        throw new Error('LOCAL_RUN_V4_ASSIGNMENT_REFERENCE_INVALID')
+      }
+      if (identity.artifact.contentRef?.mediaType !== 'text/plain') {
+        throw new Error('LOCAL_RUN_V4_ASSIGNMENT_REFERENCE_INVALID')
+      }
+      return {
+        kind: 'artifact', id: artifactId, header: `Artifact ${artifactId}:`,
+        body: safeReferenceText(identity.content),
+      }
+    })
+    return { inputReferences, directArtifacts }
+  },
+
+  v4WorkAssignmentText(
+    coordinationPlan, kind, taskId, snapshot = null, limit = 6000, receiptRecords = [],
+  ) {
+    const assignment = this.v4CoordinationAssignment(coordinationPlan, kind, { taskId })
+    if (!assignment) throw new Error('LOCAL_RUN_V4_COORDINATION_PLAN_INVALID')
+    const { inputReferences, directArtifacts } = this.v4DirectAssignmentReferences(
+      assignment, snapshot,
+    )
+    const directReferences = [...inputReferences, ...directArtifacts]
+    const dependencies = new Set(assignment.dependsOn)
+    const latestDependencies = new Map()
+    for (const record of receiptRecords) {
+      const receipt = record?.receipt
+      if (receipt?.phase !== 'work' || !dependencies.has(receipt.workItemId)) continue
+      latestDependencies.set(receipt.workItemId, record)
+    }
+    const dependencyArtifacts = [...latestDependencies.values()]
+      .sort((left, right) => String(left.receipt.workItemId)
+        .localeCompare(String(right.receipt.workItemId)))
+      .flatMap(record => (record.receipt.artifactIds || []).map((artifactId, index) => ({
+        artifactId,
+        index,
+        receipt: record.receipt,
+      })))
+      .map(({ artifactId, index, receipt }) => {
+        let identity
+        try { identity = this.v4ArtifactIdentity(artifactId) } catch {
+          throw new Error('LOCAL_RUN_V4_REFERENCE_INVALID')
+        }
+        const mediaType = String(identity.artifact.contentRef?.mediaType || '')
+        const raw = String(identity.content || '').trim()
+        const body = mediaType === 'text/plain'
+          ? this.v4SanitizeDeliveryText(raw, Number.MAX_SAFE_INTEGER)
+          : ''
+        let omission = ''
+        if (mediaType !== 'text/plain') {
+          omission = `${artifactId} reason=non-text mediaType=${mediaType || 'unknown'}`
+        } else if (!raw) {
+          omission = `${artifactId} reason=empty`
+        } else if (body !== raw) {
+          omission = `${artifactId} reason=safety-filter`
+        }
+        return {
+          key: `${receipt.workItemId}:${index}`,
+          artifactId,
+          body,
+          omission,
+          header: `Artifact ${artifactId} from ${receipt.agentKind} work=${receipt.workItemId}`,
+        }
+      })
+    const render = (included) => {
+      const lines = [
+        'MELDWORK_V4_AGREED_WORK_PACKAGE_V1',
+        `Responsibility plan: ${coordinationPlan.planHash}`,
+        `Work item: ${assignment.taskId}`,
+        `Owner: ${assignment.ownerKind}`,
+        `Role: ${assignment.role}`,
+        `Input refs: ${assignment.inputRefs.length ? assignment.inputRefs.join(', ') : '(none)'}`,
+        `Artifact refs: ${assignment.artifactIds.length ? assignment.artifactIds.join(', ') : '(none)'}`,
+        `Dependencies: ${assignment.dependsOn.length ? assignment.dependsOn.join(', ') : '(none)'}`,
+      ]
+      if (inputReferences.length) {
+        lines.push('MELDWORK_V4_FROZEN_INPUT_REFS_V1')
+        for (const reference of inputReferences) {
+          lines.push(reference.header)
+          const body = included.get(`direct:${reference.kind}:${reference.id}`) || ''
+          if (body) lines.push(body)
+        }
+      }
+      if (directArtifacts.length) {
+        lines.push('MELDWORK_V4_DIRECT_ARTIFACT_REFS_V1')
+        for (const reference of directArtifacts) {
+          lines.push(reference.header)
+          const body = included.get(`direct:${reference.kind}:${reference.id}`) || ''
+          if (body) lines.push(body)
+        }
+      }
+      if (dependencyArtifacts.length) {
+        lines.push('MELDWORK_V4_REFERENCED_ARTIFACTS_V1')
+        for (const artifact of dependencyArtifacts) {
+          lines.push(artifact.header)
+          const body = included.get(`dependency:${artifact.key}`) || ''
+          if (body) lines.push(body)
+        }
+      }
+      const objective = included.get('objective') || ''
+      const expectedOutput = included.get('expectedOutput') || ''
+      lines.push(`Objective: ${objective}`, `Expected output: ${expectedOutput}`)
+      const freeTextOmissions = [
+        ...(objective.length < assignment.objective.length ? ['objective reason=budget'] : []),
+        ...(expectedOutput.length < assignment.expectedOutput.length
+          ? ['expectedOutput reason=budget'] : []),
+      ]
+      if (freeTextOmissions.length) {
+        lines.push('MELDWORK_V4_TRUNCATED_FREE_TEXT_V1', ...freeTextOmissions)
+      }
+      const directOmissions = directReferences.filter(reference => (
+        (included.get(`direct:${reference.kind}:${reference.id}`) || '').length
+          < reference.body.length
+      )).map(reference => `${reference.kind} ${reference.id} reason=budget`)
+      if (directOmissions.length) {
+        lines.push('MELDWORK_V4_OMITTED_DIRECT_REFERENCE_BODIES_V1', ...directOmissions)
+      }
+      const dependencyOmissions = dependencyArtifacts.flatMap((artifact) => {
+        if (artifact.omission) return [artifact.omission]
+        return (included.get(`dependency:${artifact.key}`) || '').length < artifact.body.length
+          ? [`${artifact.artifactId} reason=budget`]
+          : []
+      })
+      if (dependencyOmissions.length) {
+        lines.push('MELDWORK_V4_OMITTED_ARTIFACT_BODIES_V1', ...dependencyOmissions)
+      }
+      return lines.join('\n')
+    }
+    const maximum = Number.isSafeInteger(limit) && limit >= 0 ? limit : 6000
+    const fullIncluded = new Map([
+      ['objective', assignment.objective],
+      ['expectedOutput', assignment.expectedOutput],
+      ...directReferences.map(reference => [
+        `direct:${reference.kind}:${reference.id}`, reference.body,
+      ]),
+      ...dependencyArtifacts.filter(artifact => artifact.body).map(artifact => [
+        `dependency:${artifact.key}`, artifact.body,
+      ]),
+    ])
+    const full = render(fullIncluded)
+    if (full.length <= maximum) return full
+    const bodyFields = [...fullIncluded.entries()]
+    const bodyLineCount = directReferences.length
+      + dependencyArtifacts.filter(artifact => artifact.body).length
+    let remaining = maximum - render(new Map()).length - bodyLineCount
+    if (remaining < 0) throw new Error('LOCAL_RUN_V4_DELIVERY_BUDGET_EXCEEDED')
+    const included = new Map()
+    let pendingFields = bodyFields.length
+    for (const [key, body] of bodyFields) {
+      const share = pendingFields > 0 ? Math.floor(remaining / pendingFields) : 0
+      const value = body.slice(0, Math.max(0, share))
+      included.set(key, value)
+      remaining -= value.length
+      pendingFields -= 1
+    }
+    const result = render(included)
+    if (result.length > maximum) throw new Error('LOCAL_RUN_V4_DELIVERY_BUDGET_EXCEEDED')
+    return result
+  },
+
+  v4CheckpointRole(phase, slot, coordinationPlan, workAssignment = null) {
+    if (['proposal', 'discussion', 'challenge'].includes(phase)) return 'participant'
+    const assignment = phase === 'work'
+      ? (workAssignment || this.v4CoordinationAssignment(coordinationPlan, slot.agentKind))
+      : (phase === 'synthesis'
+        ? this.v4CoordinationAssignment(coordinationPlan, slot.agentKind, { role: 'integrator' })
+          || this.v4CoordinationAssignment(coordinationPlan, slot.agentKind)
+        : this.v4CoordinationAssignment(coordinationPlan, slot.agentKind))
+    if (phase === 'verification') {
+      if (coordinationPlan?.verifierKinds?.includes(slot.agentKind)) return 'verifier'
+      return assignment?.role || 'writer'
+    }
+    return assignment?.role || 'writer'
+  },
+
+  v4PromptRole(phase, kind, coordinationPlan, workAssignment = null) {
+    if (['proposal', 'discussion', 'challenge'].includes(phase)) return 'participant'
+    if (phase === 'verification') {
+      if (!coordinationPlan?.verifierKinds?.includes(kind)) {
+        throw new Error('LOCAL_RUN_V4_COORDINATION_PLAN_INVALID')
+      }
+      return 'verifier'
+    }
+    const assignment = phase === 'work'
+      ? workAssignment
+      : (phase === 'synthesis'
+        ? this.v4CoordinationAssignment(coordinationPlan, kind, { role: 'integrator' })
+          || this.v4CoordinationAssignment(coordinationPlan, kind)
+        : this.v4CoordinationAssignment(coordinationPlan, kind))
+    if (!assignment) throw new Error('LOCAL_RUN_V4_COORDINATION_PLAN_INVALID')
+    return assignment.role
+  },
+
+  v4BlackboardOwner(coordinationPlan, kind) {
+    const assignment = this.v4CoordinationAssignment(coordinationPlan, kind, { role: 'integrator' })
+      || this.v4CoordinationAssignment(coordinationPlan, kind)
+    if (!assignment) {
+      throw new Error('LOCAL_RUN_V4_COORDINATION_PLAN_INVALID')
+    }
+    return {
+      type: 'agent',
+      agentKind: kind,
+      role: assignment.role,
+    }
+  },
+
+  v4ArtifactContext(
+    receiptRecords, phase, coordinationPlan, kind, limit = 3600, workTaskId = '',
+  ) {
+    let selected = []
+    if (phase === 'work') {
+      const assignment = this.v4CoordinationAssignment(
+        coordinationPlan, kind, { taskId: workTaskId },
+      )
+      const dependencies = new Set(assignment?.dependsOn || [])
+      selected = receiptRecords.filter(record => (
+        record?.receipt?.phase === 'work' && dependencies.has(record.receipt.workItemId)
+      ))
+    } else if (phase === 'synthesis') {
+      const workItems = new Set(coordinationPlan?.assignments?.map(item => item.taskId) || [])
+      selected = receiptRecords.filter(record => (
+        record?.receipt?.phase === 'work' && workItems.has(record.receipt.workItemId)
+      ))
+    } else if (phase === 'verification') {
+      const latest = this.v4LatestReceipt(receiptRecords, 'synthesis')
+      selected = latest ? [latest] : []
+    }
+    const latestByKey = new Map()
+    for (const record of selected) {
+      const receipt = record.receipt
+      const key = receipt.workItemId || `${receipt.phase}:${receipt.agentKind}`
+      latestByKey.set(key, record)
+    }
+    const records = [...latestByKey.values()].sort((left, right) => (
+      String(left.receipt.workItemId || left.receipt.agentKind)
+        .localeCompare(String(right.receipt.workItemId || right.receipt.agentKind))
+    ))
+    if (!records.length) return ''
+    const artifacts = records.flatMap((record, recordIndex) => {
+      const receipt = record.receipt
+      return (receipt.artifactIds || []).map((artifactId, artifactIndex) => ({
+        key: `${recordIndex}:${artifactIndex}`,
+        artifactId,
+        receipt,
+      }))
+    }).map(({ key, artifactId, receipt }) => {
+      let identity
+      try { identity = this.v4ArtifactIdentity(artifactId) } catch {
+        throw new Error('LOCAL_RUN_V4_REFERENCE_INVALID')
+      }
+      const mediaType = String(identity.artifact.contentRef?.mediaType || '')
+      const raw = String(identity.content || '').trim()
+      const body = mediaType === 'text/plain'
+        ? this.v4SanitizeDeliveryText(raw, Number.MAX_SAFE_INTEGER)
+        : ''
+      return {
+        key,
+        artifactId,
+        receipt,
+        mediaType,
+        raw,
+        body,
+        header: `Artifact ${artifactId} from ${receipt.agentKind} work=${receipt.workItemId || receipt.phase}`,
+      }
+    })
+    if (!artifacts.length) return ''
+    const maximum = Number.isSafeInteger(limit) && limit >= 0 ? limit : 3600
+    const textArtifacts = artifacts.filter(artifact => artifact.mediaType === 'text/plain'
+      && artifact.body)
+    const fixedOmissions = artifacts.flatMap((artifact) => {
+      if (artifact.mediaType !== 'text/plain') {
+        return [`${artifact.artifactId} reason=non-text mediaType=${artifact.mediaType || 'unknown'}`]
+      }
+      if (!artifact.raw) return [`${artifact.artifactId} reason=empty`]
+      return artifact.body === artifact.raw
+        ? []
+        : [`${artifact.artifactId} reason=safety-filter`]
+    })
+    const potentialBudgetOmissions = textArtifacts.map(artifact => (
+      `${artifact.artifactId} reason=budget`
+    ))
+    const structuralLines = [
+      'MELDWORK_V4_REFERENCED_ARTIFACTS_V1',
+      ...artifacts.map(artifact => artifact.header),
+    ]
+    const reservedManifest = [...fixedOmissions, ...potentialBudgetOmissions]
+    const reservedLines = reservedManifest.length
+      ? [...structuralLines, 'MELDWORK_V4_OMITTED_ARTIFACT_BODIES_V1', ...reservedManifest]
+      : structuralLines
+    const bodySeparators = textArtifacts.length
+    const bodyBudget = maximum - reservedLines.join('\n').length - bodySeparators
+    if (bodyBudget < 0) throw new Error('LOCAL_RUN_V4_DELIVERY_BUDGET_EXCEEDED')
+    let remaining = bodyBudget
+    let remainingArtifacts = textArtifacts.length
+    const includedByKey = new Map()
+    for (const artifact of textArtifacts) {
+      const share = remainingArtifacts > 0 ? Math.floor(remaining / remainingArtifacts) : 0
+      const included = artifact.body.slice(0, Math.max(0, share))
+      includedByKey.set(artifact.key, included)
+      remaining -= included.length
+      remainingArtifacts -= 1
+    }
+    const budgetOmissions = textArtifacts.filter(artifact => (
+      includedByKey.get(artifact.key).length < artifact.body.length
+    )).map(artifact => `${artifact.artifactId} reason=budget`)
+    const lines = ['MELDWORK_V4_REFERENCED_ARTIFACTS_V1']
+    for (const artifact of artifacts) {
+      lines.push(artifact.header)
+      const included = includedByKey.get(artifact.key) || ''
+      if (included) lines.push(included)
+    }
+    const omissions = [...fixedOmissions, ...budgetOmissions]
+    if (omissions.length) {
+      lines.push('MELDWORK_V4_OMITTED_ARTIFACT_BODIES_V1', ...omissions)
+    }
+    const result = lines.join('\n')
+    if (result.length > maximum) throw new Error('LOCAL_RUN_V4_DELIVERY_BUDGET_EXCEEDED')
+    return result
+  },
+
+  v4DeliveryPrompt(group, controller, input) {
+    const {
+      kind, phase, snapshot, receiptRecords, role, targetKinds, slot, options,
+      sessionBinding, skillHints = [],
+    } = input
+    const full = sessionBinding.sessionRotated === true || sessionBinding.hasSession !== true
+    const packageFloor = phase === 'proposal' ? 0 : collaborationPackageIndexText(targetKinds).length
+    let remaining = 6000 - packageFloor
+    let assignedProposalText = phase === 'challenge'
+      ? this.v4SanitizeDeliveryText(options.assignedProposalText || '', 1600)
+      : ''
+    if (assignedProposalText) {
+      remaining = Math.max(0, remaining - 2)
+      assignedProposalText = assignedProposalText.slice(0, remaining)
+      remaining -= assignedProposalText.length
+    }
+    const rawExtraContext = String(options.extraContext || '').trim()
+    const requestedExtraContext = this.v4SanitizeDeliveryText(rawExtraContext, 6000)
+    const artifactContextPlan = options.artifactContext?.coordinationPlan || null
+    const structuredArtifactContext = ['synthesis', 'verification'].includes(phase)
+      && options.artifactContext && typeof options.artifactContext === 'object'
+      && !Array.isArray(options.artifactContext)
+    if ((phase === 'work' || structuredArtifactContext) && rawExtraContext.length > 6000) {
+      throw new Error('LOCAL_RUN_V4_DELIVERY_BUDGET_EXCEEDED')
+    }
+    if ((phase === 'work' || structuredArtifactContext)
+        && requestedExtraContext !== rawExtraContext) {
+      throw new Error('LOCAL_RUN_V4_ASSIGNMENT_REFERENCE_INVALID')
+    }
+    const extraLimit = Math.max(0, remaining - 2)
+    let extraContext = requestedExtraContext
+    if (structuredArtifactContext) {
+      const artifactLimit = extraLimit - (extraContext ? extraContext.length + 2 : 0)
+      const artifactContext = this.v4ArtifactContext(
+        receiptRecords,
+        phase,
+        artifactContextPlan,
+        kind,
+        Math.max(0, artifactLimit),
+      )
+      extraContext = [extraContext, artifactContext].filter(Boolean).join('\n\n')
+      if (extraContext.length > extraLimit) {
+        throw new Error('LOCAL_RUN_V4_DELIVERY_BUDGET_EXCEEDED')
+      }
+    } else if (extraContext.length > extraLimit) {
+      if (phase === 'work') throw new Error('LOCAL_RUN_V4_DELIVERY_BUDGET_EXCEEDED')
+      const marker = 'MELDWORK_V4_CONTEXT_TRUNCATED_V1'
+      if (extraLimit <= marker.length + 1) {
+        extraContext = ''
+      } else {
+        let prefix = extraContext.slice(0, extraLimit - marker.length - 1)
+        const boundary = prefix.lastIndexOf('\n')
+        if (boundary > 0) prefix = prefix.slice(0, boundary)
+        extraContext = `${prefix.trimEnd()}\n${marker}`
+      }
+    }
+    if (extraContext) remaining -= extraContext.length + 2
+    const packageRecord = phase === 'proposal'
+      ? null
+      : this.v4Package(receiptRecords, {
+          recipientKind: kind,
+          sessionBinding,
+          deliveryState: controller.orchestration?.deliveryState || [],
+          targetKinds,
+          snapshotHash: controller.orchestration?.snapshotHash,
+          forceFull: full,
+          totalLimit: packageFloor + remaining,
+        })
+    const deliveryId = `delivery-${hashValue({
+      runId: controller.runId, recipientKind: kind, slotId: slot.slotId,
+      operationId: slot.operationId, packageHash: packageRecord?.packageHash || '', now: Date.now(),
+    }).slice(0, 64)}`
+    const delivery = packageRecord?.receipts?.length ? {
+      deliveryId,
+      packageHash: packageRecord.packageHash,
+      entries: packageRecord.receipts.map(receipt => ({
+        recipientKind: kind,
+        sessionRefHash: sessionBinding.sessionRefHash,
+        sessionProvenanceHash: sessionBinding.sessionProvenanceHash,
+        sourceAgentKind: receipt.agentKind,
+        sourcePhase: receipt.phase,
+        watermark: receipt.deliveryWatermark,
+        snapshotHash: receipt.snapshotHash,
+        operationId: receipt.operationId,
+        packageHash: packageRecord.packageHash,
+        deliveryId,
+        status: 'prepared',
+        updatedAt: Date.now(),
+      })),
+    } : null
+    this.v4PrepareDelivery(group, controller, delivery)
+    return {
+      prompt: this.v4PhasePrompt(group, kind, phase, snapshot, receiptRecords, role, {
+        reviewTarget: options.reviewTarget,
+        assignedProposalText,
+        packageText: packageRecord?.text || '',
+        extraContext,
+        snapshotHash: controller.orchestration?.snapshotHash,
+        skillHints,
+      }),
+      delivery,
+    }
+  },
+
+  v4PrepareDelivery(group, controller, delivery) {
+    if (delivery) {
+      const replaced = new Set(delivery.entries.map(entry => [
+        entry.recipientKind, entry.sessionRefHash, entry.sessionProvenanceHash,
+        entry.sourceAgentKind, entry.sourcePhase,
+      ].join('\u0000')))
+      const deliveryState = [
+        ...(controller.orchestration?.deliveryState || []).filter(entry => (
+          (!delivery.natural || entry.recipientKind !== delivery.entries[0].recipientKind
+            || (entry.sessionRefHash === delivery.entries[0].sessionRefHash
+              && entry.sessionProvenanceHash === delivery.entries[0].sessionProvenanceHash))
+          && !replaced.has([
+          entry.recipientKind, entry.sessionRefHash, entry.sessionProvenanceHash,
+          entry.sourceAgentKind, entry.sourcePhase,
+        ].join('\u0000')))),
+        ...delivery.entries,
+      ].sort((left, right) => [
+        left.recipientKind, left.sessionRefHash, left.sessionProvenanceHash,
+        left.sourceAgentKind, left.sourcePhase,
+      ].join('\u0000').localeCompare([
+        right.recipientKind, right.sessionRefHash, right.sessionProvenanceHash,
+        right.sourceAgentKind, right.sourcePhase,
+      ].join('\u0000')))
+      this.checkpointOrchestration(group, controller, { deliveryState })
+    }
+  },
+
+  v4SetDeliveryStatus(group, controller, delivery, status, binding = null) {
+    if (!delivery?.entries?.length) return
+    const deliveryState = (controller.orchestration?.deliveryState || []).map(entry => {
+      const matched = delivery.entries.some(candidate => candidate.deliveryId === entry.deliveryId
+        && candidate.sourceAgentKind === entry.sourceAgentKind
+        && candidate.sourcePhase === entry.sourcePhase
+        && entry.status === 'prepared')
+      let sourceMessages = entry.sourceMessages
+      if (matched && sourceMessages && status === 'acknowledged'
+          && binding?.sessionRefHash === entry.sessionRefHash
+          && binding?.sessionProvenanceHash === entry.sessionProvenanceHash) {
+        const previous = delivery.previousEntries?.find(candidate => (
+          candidate.sourceAgentKind === entry.sourceAgentKind
+          && candidate.sourcePhase === entry.sourcePhase
+        ))
+        sourceMessages = [...new Map([
+          ...(previous?.sourceMessages || []), ...sourceMessages,
+        ].map(message => [message.id, message])).values()].slice(-100)
+      }
+      return matched ? {
+        ...entry,
+        ...(sourceMessages ? { sourceMessages } : {}),
+        ...(binding ? {
+          sessionRefHash: binding.sessionRefHash,
+          sessionProvenanceHash: binding.sessionProvenanceHash,
+        } : {}),
+        status,
+        updatedAt: Date.now(),
+      } : entry
+    }).sort((left, right) => [
+      left.recipientKind, left.sessionRefHash, left.sessionProvenanceHash,
+      left.sourceAgentKind, left.sourcePhase,
+    ].join('\u0000').localeCompare([
+      right.recipientKind, right.sessionRefHash, right.sessionProvenanceHash,
+      right.sourceAgentKind, right.sourcePhase,
+    ].join('\u0000')))
+    this.checkpointOrchestration(group, controller, { deliveryState })
+  },
+
+  v4ChallengeTarget(kind, targetKinds, round = 1, seed = '') {
+    const kinds = Array.isArray(targetKinds) ? targetKinds : []
+    const index = kinds.indexOf(kind)
+    if (index < 0 || kinds.length < 2) return ''
+    const digest = createHash('sha256').update(JSON.stringify([seed, round, kinds])).digest('hex')
+    const offset = 1 + (Number.parseInt(digest.slice(0, 8), 16) % (kinds.length - 1))
+    return kinds[(index + offset) % kinds.length] || ''
+  },
+
+  v4ChallengeBindings({ controller, targetKinds, round, snapshotRecord, slots, receiptRecords }) {
+    const kinds = [...new Set(Array.isArray(targetKinds) ? targetKinds : [])].sort()
+    if (kinds.length < 2 || !Number.isSafeInteger(round) || round < 2) {
+      throw new Error('LOCAL_RUN_V4_CHALLENGE_BINDING_INVALID')
+    }
+    const digest = createHash('sha256').update(JSON.stringify([
+      snapshotRecord?.bodyHash || '', kinds,
+    ])).digest('hex')
+    const initialOffset = 1 + (Number.parseInt(digest.slice(0, 8), 16) % (kinds.length - 1))
+    const offset = 1 + ((initialOffset - 1 + (round - 2)) % (kinds.length - 1))
+    const slotsByKind = new Map(slots.map(slot => [slot.agentKind, slot]))
+    const proposalsByKind = new Map()
+    for (const record of receiptRecords) {
+      const receipt = record?.receipt
+      if (receipt?.phase === 'proposal' && kinds.includes(receipt.agentKind)) {
+        proposalsByKind.set(receipt.agentKind, receipt)
+      }
+    }
+    return kinds.map((reviewerKind, index) => {
+      const proposalKind = kinds[(index + offset) % kinds.length]
+      const reviewerSlot = slotsByKind.get(reviewerKind)
+      const proposalSlot = slotsByKind.get(proposalKind)
+      const proposalReceipt = proposalsByKind.get(proposalKind)
+      if (!reviewerSlot || !proposalSlot || !proposalReceipt) {
+        throw new Error('LOCAL_RUN_V4_CHALLENGE_BINDING_INVALID')
+      }
+      return {
+        round,
+        reviewerKind,
+        reviewerSlotId: reviewerSlot.slotId,
+        reviewerOperationId: this.v4OperationId(
+          controller, reviewerKind, 'challenge', reviewerSlot.slotId,
+        ),
+        proposalKind,
+        proposalSlotId: proposalSlot.slotId,
+        proposalOperationId: proposalReceipt.operationId,
+        proposalReceiptId: proposalReceipt.receiptId,
+        artifactIds: [...(proposalReceipt.artifactIds || [])],
+        evidenceIds: [...(proposalReceipt.evidenceIds || [])],
+      }
+    })
+  },
+
+  v4PhasePrompt(group, kind, phase, snapshot, receiptRecords, role, options = {}) {
+    const collaborationPackage = phase === 'proposal'
+      ? ''
+      : String(options.packageText || this.v4Package(receiptRecords, {
+          recipientKind: kind,
+          forceFull: options.forceFullPackage === true,
+          snapshotHash: options.snapshotHash,
+        }).text)
+    const reviewTarget = phase === 'challenge' ? String(options.reviewTarget || '') : ''
+    const assignedProposalText = phase === 'challenge'
+      ? this.v4SanitizeDeliveryText(options.assignedProposalText || '', 1600) : ''
+    const extraContext = this.v4SanitizeDeliveryText(options.extraContext || '', 6000)
+    return [
+      v4Prompt({
+        group, kind, phase, snapshot, role, skillHints: options.skillHints,
+        knowledgeBaseHints: options.knowledgeBaseHints,
+        naturalResponse: this.naturalAgentResponses,
+      }),
+      reviewTarget
+        ? `Coverage responsibility: explicitly incorporate or challenge ${reviewTarget}'s proposal while negotiating with every peer. This coverage duty does not assign responsibilities or authorize you to allocate work for others.`
+        : '',
+      assignedProposalText,
+      collaborationPackage
+        ? `Selected records from completed phases:\n${collaborationPackage}`
+        : '',
+      extraContext,
+    ].filter(Boolean).join('\n\n')
+  },
+}
+
+module.exports = { v4DeliveryMethods }

desktop/src/workspace/local-workspace-auto-runner-v4-discussion.cjs

@@ -0,0 +1,964 @@
+// The V4 discussion run: snapshot, phases, resume and commit.
+const { normalizeFailure } = require('../runs/failure-policy.cjs')
+const { commitDeliveryObservable } = require('../collaboration/orchestration-v4-records.cjs')
+const { v4Snapshot } = require('./local-workspace-context.cjs')
+const { terminalRunStatusForReason } = require('./local-workspace-auto-runner-support.cjs')
+
+const v4DiscussionMethods = {
+  async runV4Discussion(group, controller, threadRootId, context, resume = false, resumedGate = null) {
+    const snapshotTargetKinds = [...controller.targetKinds]
+    const existing = controller.orchestration?.version === 4
+      ? controller.orchestration
+      : null
+    const targetKinds = [...snapshotTargetKinds]
+    controller.discussionStyle = existing?.discussionStyle
+      || controller.discussionStyle
+      || 'sequential'
+    let activeKinds = Array.isArray(existing?.activeKinds) && existing.activeKinds.length
+      ? existing.activeKinds.filter(kind => targetKinds.includes(kind))
+      : [...targetKinds]
+    const removedKinds = new Set(targetKinds.filter(kind => !activeKinds.includes(kind)))
+    if (existing?.synthesisBinding) {
+      throw new Error('LOCAL_RUN_V4_LEGACY_SYNTHESIS_BINDING_UNSUPPORTED')
+    }
+    let synthesisBinding = existing?.synthesisBinding
+      ? {
+          ...existing.synthesisBinding,
+          candidates: existing.synthesisBinding.candidates.map(candidate => ({
+            ...candidate,
+            ...(candidate.evidence ? { evidence: { ...candidate.evidence } } : {}),
+          })),
+          verificationKinds: [...existing.synthesisBinding.verificationKinds],
+        }
+      : null
+    let synthesisRecovery = existing?.synthesisRecovery
+      ? {
+          ...existing.synthesisRecovery,
+          verificationKinds: [...existing.synthesisRecovery.verificationKinds],
+          rankedKinds: [...existing.synthesisRecovery.rankedKinds],
+          triedWriters: [...existing.synthesisRecovery.triedWriters],
+          attempts: existing.synthesisRecovery.attempts.map(attempt => ({ ...attempt })),
+          ...(existing.synthesisRecovery.pendingGate
+            ? { pendingGate: { ...existing.synthesisRecovery.pendingGate } }
+            : {}),
+        }
+      : null
+    let coordinationPlan = existing?.coordinationPlan
+      ? {
+          ...existing.coordinationPlan,
+          assignments: existing.coordinationPlan.assignments.map(assignment => ({
+            ...assignment,
+            inputRefs: [...assignment.inputRefs],
+            artifactIds: [...assignment.artifactIds],
+            dependsOn: [...assignment.dependsOn],
+          })),
+          verifierKinds: [...existing.coordinationPlan.verifierKinds],
+          agreedBy: [...existing.coordinationPlan.agreedBy],
+          ...(existing.coordinationPlan.supportReceiptIds
+            ? { supportReceiptIds: [...existing.coordinationPlan.supportReceiptIds] }
+            : {}),
+        }
+      : null
+    const postSynthesisPhases = new Set([
+      'synthesis', 'verification', 'human-gate', 'commit', 'committed', 'completed',
+    ])
+    const synthesisObservable = existing && (
+      postSynthesisPhases.has(existing.phase)
+      || commitDeliveryObservable(existing.commitState)
+      || existing.slots?.some(slot => (
+        postSynthesisPhases.has(slot.phase)
+        || (slot.resultRefs?.workflowOutcomeRefs || []).some(item => (
+          ['synthesis', 'verification'].includes(item?.receipt?.phase)
+        ))
+      ))
+    )
+    if (!synthesisBinding && !coordinationPlan && synthesisObservable) {
+      throw new Error('LOCAL_RUN_V4_SYNTHESIS_BINDING_REQUIRED')
+    }
+    let writerKind = synthesisRecovery?.activeWriterKind
+      || coordinationPlan?.finalizerKind || synthesisBinding?.writerKind
+      || existing?.discussionWriterKind
+      || (!existing && this.naturalAgentResponses && group.allowWrite === true
+        ? targetKinds.find(kind => this.canWriteWorkspace(kind)) : '') || ''
+    controller.currentRound = Math.max(1, controller.currentRound || existing?.round || 0)
+    const rootMessage = existing ? null : this.state().messages.find(message => (
+      message.id === threadRootId && message.groupId === group.id && message.role === 'user'
+    ))
+    const snapshot = existing
+      ? this.v4LoadSnapshot(existing, {
+          taskId: controller.taskId,
+          messageId: threadRootId,
+          groupId: controller.groupId,
+          targetKinds: snapshotTargetKinds,
+        })
+      : v4Snapshot({
+          state: this.state(),
+          group,
+          taskId: threadRootId,
+          targetKinds: snapshotTargetKinds,
+          message: rootMessage,
+          skillHintsByKind: context.rootSkillsByKind,
+          phase: 'proposal',
+          writerKind,
+        })
+    if (existing && this.naturalAgentResponses && !coordinationPlan && !synthesisBinding
+        && (snapshot.writerKind || '') !== writerKind) {
+      throw new Error('LOCAL_RUN_COLLABORATION_SCOPE_INVALID')
+    }
+    if (existing) {
+      const persistedSkills = this.v4SnapshotSkillHints(snapshot, snapshotTargetKinds)
+      const hasPersistedSkills = snapshotTargetKinds.some(kind => (
+        (persistedSkills.get(kind) || []).length > 0
+      ))
+      context = {
+        rootAttachments: [],
+        rootSkillsByKind: hasPersistedSkills
+          ? await this.v4RestoreSnapshotSkills(snapshot, snapshotTargetKinds, persistedSkills)
+          : persistedSkills,
+        rootKnowledgeBasesByKind: new Map(snapshotTargetKinds.map(kind => [kind, []])),
+      }
+    }
+    const builtSnapshot = existing
+      ? { record: existing.snapshot, snapshotHash: existing.snapshotHash }
+      : this.v4SnapshotRecord(controller, snapshot, snapshotTargetKinds)
+    const snapshotRecord = builtSnapshot.record
+    const snapshotHash = builtSnapshot.snapshotHash
+    const batchId = existing?.batchId || this.v4BatchId(controller, 'discussion')
+    const receiptRecords = this.v4RestoreReceipts(controller)
+    let workReceipts = Array.isArray(existing?.workReceipts)
+      ? existing.workReceipts.map(receipt => ({
+          ...receipt,
+          snapshotContentRef: { ...receipt.snapshotContentRef },
+          collaborationReceipt: { ...receipt.collaborationReceipt },
+          artifacts: receipt.artifacts.map(artifact => ({
+            ...artifact,
+            contentRef: { ...artifact.contentRef },
+          })),
+        }))
+      : []
+    controller.v4WorkReceipts = workReceipts
+    let challengeBindings = Array.isArray(existing?.challengeBindings)
+      ? existing.challengeBindings.map(binding => ({
+          ...binding,
+          artifactIds: [...binding.artifactIds],
+          evidenceIds: [...binding.evidenceIds],
+        }))
+      : null
+    controller.v4Watermarks = existing?.deliveryWatermarks
+      ? existing.deliveryWatermarks.map(item => ({ ...item }))
+      : []
+    const initialSlots = existing?.slots?.length
+      ? existing.slots.map(slot => ({ ...slot }))
+      : targetKinds.map((kind, index) => ({
+          slotId: `slot-${index + 1}-${kind}`,
+          agentKind: kind,
+          phase: 'proposal',
+          status: 'planned',
+          operationId: this.v4OperationId(
+            controller, kind, 'proposal', `slot-${index + 1}-${kind}`,
+          ),
+          queuePosition: index,
+          snapshotHash,
+          deliveryWatermark: 0,
+          receiptId: '',
+          resultHash: '',
+          assignedAt: Date.now(),
+          startedAt: Date.now(),
+          finishedAt: null,
+          commitStatus: 'pending',
+          attempt: 0,
+          permission: 'read-only',
+          resultRefs: { artifactIds: [], evidenceIds: [], workflowOutcomeRefs: [] },
+        }))
+    let slots = initialSlots
+    if (!existing) {
+      this.v4CheckpointPhase(group, controller, {
+        targetKinds,
+        participantKinds: activeKinds,
+        phase: 'proposal',
+        batchId,
+        snapshotRecord,
+        snapshotHash,
+        slots,
+        writerKind,
+        currentKinds: targetKinds,
+        pendingKinds: targetKinds,
+        receipts: receiptRecords,
+      })
+      this.emitChanged()
+    }
+
+    const latestFor = (phase, kind) => this.v4LatestReceipt(receiptRecords, phase, kind)
+    const pendingKindsFor = (phase, activeKinds, workAssignments = []) => activeKinds.filter(kind => {
+      if (phase === 'work') {
+        const assignment = workAssignments.find(candidate => candidate.ownerKind === kind)
+        return !assignment || !workReceipts.some(receipt => (
+          receipt.planHash === coordinationPlan.planHash
+          && receipt.taskId === assignment.taskId
+          && receipt.ownerKind === kind
+        ))
+      }
+      const latest = latestFor(phase, kind)
+      if (!latest) return true
+      if (phase === 'proposal') return false
+      const slot = slots.find(candidate => candidate.agentKind === kind)
+      const synthesisAttempt = phase === 'synthesis'
+        ? [...(synthesisRecovery?.attempts || [])].reverse().find(attempt => (
+            attempt.writerKind === kind
+          ))
+        : null
+      const expected = synthesisAttempt?.operationId || (slot
+        ? this.v4OperationId(controller, kind, phase, slot.slotId)
+        : '')
+      return latest.receipt.operationId !== expected
+    })
+    const checkpointPhase = (phase, activeKinds, phaseSlots = slots, bindings = challengeBindings) => {
+      this.v4CheckpointPhase(group, controller, {
+        targetKinds,
+        participantKinds: activeKindsForRun(),
+        phase,
+        batchId,
+        snapshotRecord,
+        snapshotHash,
+        slots: phaseSlots,
+        writerKind,
+        currentKinds: [...activeKinds],
+        pendingKinds: [...activeKinds],
+        receipts: receiptRecords,
+        challengeBindings: bindings,
+        synthesisBinding,
+        synthesisRecovery,
+        coordinationPlan,
+        workReceipts,
+      })
+    }
+    const activeKindsForRun = () => [...activeKinds]
+    const runPhase = async (phase, requestedKinds, bindings = challengeBindings, options = {}) => {
+      const workAssignments = Array.isArray(options.workAssignments)
+        ? options.workAssignments
+        : []
+      const pending = pendingKindsFor(phase, requestedKinds, workAssignments)
+      if (!pending.length) {
+        checkpointPhase(phase, [], slots)
+        return { ok: true, phaseReceipts: [], synthesisPendingMessage: null }
+      }
+      const runPending = async (phaseKinds) => {
+        const result = await this.runV4Phase(group, controller, threadRootId, context, {
+          phase,
+          activeKinds: phaseKinds,
+          targetKinds,
+          participantKinds: activeKindsForRun(),
+          writerKind,
+          batchId,
+          snapshot,
+          snapshotRecord,
+          snapshotHash,
+          slots,
+          receiptRecords,
+          challengeBindings: bindings,
+          synthesisBinding,
+          synthesisRecovery,
+          coordinationPlan,
+          workReceipts,
+          deliveryReceiptRecords: options.deliveryReceiptRecords,
+          coordinationText: options.coordinationText || '',
+          workAssignments,
+          resumedGate,
+        })
+        slots = result.slots
+        if (result.synthesisRecovery) synthesisRecovery = result.synthesisRecovery
+        if (result.workReceipts) {
+          workReceipts = result.workReceipts
+          controller.v4WorkReceipts = workReceipts
+        }
+        return result
+      }
+      if (!['challenge', 'verification'].includes(phase) || pending.length <= 1) {
+        return runPending(pending)
+      }
+      const aggregate = {
+        slots,
+        receiptRecords,
+        phaseReceipts: [],
+        phasePendingMessages: [],
+        failures: [],
+        synthesisPendingMessage: null,
+        ok: true,
+      }
+      for (const kind of pending) {
+        const result = await runPending([kind])
+        aggregate.slots = result.slots
+        aggregate.phaseReceipts.push(...(result.phaseReceipts || []))
+        aggregate.phasePendingMessages.push(...(result.phasePendingMessages || []))
+        aggregate.failures.push(...(result.failures || []))
+        if (!result.ok) aggregate.ok = false
+      }
+      return aggregate
+    }
+    const commitPhaseMessages = (phase, result) => {
+      // Proposal/work runners commit each pending Agent message at the same
+      // time they persist its receipt and slot. Replaying those messages here
+      // creates a second commit attempt with live metadata and can reject a
+      // valid first-round result as LOCAL_RUN_COMMIT_INVALID.
+      return (result?.phasePendingMessages || []).filter(item => (
+        item?.pendingMessage?.content && item.kind
+      )).length
+    }
+    const removePhaseFailures = (result) => {
+      const failedKinds = [...new Set((result?.failures || [])
+        .filter((item) => {
+          const failure = normalizeFailure(item.error)
+          return !['budget', 'cancellation', 'permission'].includes(failure.category)
+            && !failure.code.startsWith('COLLABORATION_CONTROL_')
+        })
+        .map(item => item.kind)
+        .filter(Boolean))]
+      const removed = []
+      for (const kind of failedKinds) {
+        const failure = (result?.failures || []).find(item => item.kind === kind)?.error
+        // Every failure is a run-scoped exclusion. Group membership is never
+        // edited, so a later turn can use the Agent again.
+        const didRemove = this.shouldPauseFailedAgent(failure)
+          ? this.pauseFailedAgentForRun(group, controller, kind, threadRootId, failure)
+          : false
+        // The failed slot is still isolated from this run's active queue even
+        // when its configured group membership is retained.
+        removed.push(kind)
+        if (didRemove) {
+          removedKinds.add(kind)
+        }
+      }
+      return removed
+    }
+    const finalStatus = (status) => {
+      if (status !== 'completed') return status
+      if (!removedKinds.size && !controller.failedKinds.length) return status
+      controller.failedKinds = [...new Set([...controller.failedKinds, ...removedKinds])]
+      return 'partial'
+    }
+    const addRoundLimitNotice = () => {
+      if (this.state().messages.some(message => (
+        message.threadRootId === threadRootId
+          && message.system?.key === 'system.autoRoundLimit'
+      ))) return
+      this.addMessage(
+        group.id,
+        'system',
+        `Automatic discussion reached its ${controller.maxRounds || 6}-round limit. The latest replies above are the current result.`,
+        '',
+        threadRootId,
+        { key: 'system.autoRoundLimit', params: { rounds: controller.maxRounds || 6 } },
+      )
+    }
+    const requestStableGate = async (convergence) => {
+      if (!this.requestHumanGate) return 'round-limit'
+      const gateAgentRunId = this.v4OperationId(controller, writerKind, 'human-gate',
+        slots.find(slot => slot.agentKind === writerKind)?.slotId || '')
+      const publicAgentRunId = slots.find(slot => slot.agentKind === writerKind)?.agentRunId || ''
+      checkpointPhase('human-gate', [], slots)
+      const gate = await this.requestHumanGate({
+        type: 'decision',
+        runId: controller.runId,
+        agentRunId: gateAgentRunId,
+        agentKind: writerKind,
+        summary: 'The candidate and unresolved issues have remained unchanged for two rounds.',
+        options: [
+          { optionId: 'continue-discussion', name: 'Continue', kind: 'accept' },
+          { optionId: 'stop-discussion', name: 'Stop', kind: 'reject' },
+        ],
+        request: {
+          phase: 'discussion',
+          round: convergence.lastCompletedRound,
+          candidateHash: convergence.candidateContentHash,
+          unresolvedIssueIds: [...convergence.openIssueIds],
+          stateEpoch: convergence.stateEpoch,
+        },
+      }, {
+        signal: controller.signal,
+        preserveOnAbort: () => controller.stopReason === 'shutdown',
+        continuation: {
+          resumeKind: 'v4_human_gate',
+          agentRunId: gateAgentRunId,
+          ...(publicAgentRunId ? { publicAgentRunId } : {}),
+          agentKind: writerKind,
+          round: convergence.lastCompletedRound,
+          stateEpoch: convergence.stateEpoch,
+        },
+      })
+      if (gate.status !== 'approved') {
+        if (gate.gateId) this.completeHumanGateContinuation?.(
+          controller.runId, gate.gateId, 'cancelled',
+        )
+        return 'stopped'
+      }
+      const current = controller.orchestration?.convergence
+      if (!current || current.stateEpoch !== convergence.stateEpoch
+          || current.stateKey !== convergence.stateKey) {
+        throw new Error('LOCAL_RUN_CONTINUATION_INVALID')
+      }
+      controller.orchestration = {
+        ...controller.orchestration,
+        convergence: { ...current, acknowledgedGateEpoch: current.stateEpoch },
+      }
+      if (gate.gateId && this.completeHumanGateContinuation?.(
+        controller.runId, gate.gateId, 'completed',
+      ) !== true && this.hasRunLedger()) throw new Error('LOCAL_RUN_PERSIST_FAILED')
+      return 'continue'
+    }
+    const requestSynthesisRecoveryGate = async () => {
+      const binding = synthesisRecovery?.pendingGate
+      if (!binding || !this.requestHumanGate) return 'failed'
+      const publicAgentRunId = slots.find(slot => (
+        slot.slotId === binding.slotId && slot.agentKind === binding.writerKind
+      ))?.agentRunId || ''
+      const permissionContinuation = controller.continuation
+      if (permissionContinuation?.gateType === 'permission'
+          && permissionContinuation.resumeKind === 'agent_slot'
+          && permissionContinuation.phase === 'synthesis'
+          && permissionContinuation.agentKind === binding.writerKind
+          && permissionContinuation.slotId === binding.slotId
+          && permissionContinuation.operationId === binding.operationId) {
+        if (this.completeHumanGateContinuation?.(
+          controller.runId, permissionContinuation.gateId, 'completed',
+        ) !== true && this.hasRunLedger()) throw new Error('LOCAL_RUN_PERSIST_FAILED')
+      }
+      slots = slots.map(slot => slot.status === 'running'
+        ? { ...slot, status: 'waiting' }
+        : slot)
+      checkpointPhase('human-gate', [], slots)
+      const gate = await this.requestHumanGate({
+        type: 'decision',
+        runId: controller.runId,
+        agentRunId: binding.operationId,
+        agentKind: binding.writerKind,
+        summary: 'The workspace-write synthesis attempt may have produced side effects, but its result is unknown.',
+        options: [
+          { optionId: 'retry-original-writer', name: 'Retry original writer', kind: 'accept' },
+          ...(binding.proposedReplacementKind ? [{
+            optionId: 'replace-next-writer', name: 'Replace with next writer', kind: 'accept',
+          }] : []),
+          { optionId: 'stop-discussion', name: 'Stop', kind: 'reject' },
+        ],
+        request: {
+          phase: 'synthesis-recovery',
+          bindingHash: binding.bindingHash,
+          writerKind: binding.writerKind,
+          slotId: binding.slotId,
+          operationId: binding.operationId,
+          attempt: binding.attempt,
+          proposedReplacementKind: binding.proposedReplacementKind,
+          round: binding.round,
+          stateEpoch: binding.stateEpoch,
+          outcomeCertainty: 'unknown_outcome',
+        },
+      }, {
+        signal: controller.signal,
+        preserveOnAbort: () => controller.stopReason === 'shutdown',
+        continuation: {
+          resumeKind: 'v4_synthesis_recovery',
+          agentRunId: binding.operationId,
+          ...(publicAgentRunId ? { publicAgentRunId } : {}),
+          agentKind: binding.writerKind,
+          round: binding.round,
+          stateEpoch: binding.stateEpoch,
+        },
+      })
+      if (gate.status !== 'approved' || gate.optionId === 'stop-discussion') {
+        synthesisRecovery = this.v4CancelSynthesisRecovery(synthesisRecovery)
+        checkpointPhase('human-gate', [], slots)
+        if (gate.gateId) this.completeHumanGateContinuation?.(
+          controller.runId, gate.gateId, 'cancelled',
+        )
+        return 'stopped'
+      }
+      if (!['retry-original-writer', 'replace-next-writer'].includes(gate.optionId)) {
+        throw new Error('LOCAL_RUN_CONTINUATION_INVALID')
+      }
+      const nextWriterKind = gate.optionId === 'replace-next-writer'
+        ? binding.proposedReplacementKind
+        : binding.writerKind
+      if (!nextWriterKind) throw new Error('LOCAL_RUN_CONTINUATION_INVALID')
+      if (gate.gateId && this.completeHumanGateContinuation?.(
+        controller.runId, gate.gateId, 'completed',
+      ) !== true && this.hasRunLedger()) throw new Error('LOCAL_RUN_PERSIST_FAILED')
+      synthesisRecovery = this.v4AppendSynthesisAttempt(
+        controller, coordinationPlan || synthesisBinding, synthesisRecovery, slots, nextWriterKind,
+      )
+      writerKind = synthesisRecovery.activeWriterKind
+      const nextAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+      slots = slots.map((slot) => {
+        if (slot.agentKind !== writerKind) return { ...slot, permission: 'read-only' }
+        const { agentRunId: _agentRunId, ...retainedSlot } = slot
+        return {
+          ...retainedSlot,
+          phase: 'synthesis',
+          status: 'planned',
+          operationId: nextAttempt.operationId,
+          receiptId: '',
+          resultHash: '',
+          finishedAt: null,
+          commitStatus: 'pending',
+          permission: group.allowWrite === true ? 'workspace-write' : 'read-only',
+        }
+      })
+      checkpointPhase('synthesis', [writerKind], slots)
+      return 'continue'
+    }
+
+    let round = Math.max(1, Number(existing?.round) || controller.currentRound || 1)
+    const acknowledgedGateResume = resume === true
+      && existing?.phase === 'challenge'
+      && existing.convergence?.acknowledgedGateEpoch === existing.convergence?.stateEpoch
+      && existing.convergence?.lastCompletedRound === round
+      && controller.continuation?.resumeKind === 'v4_human_gate'
+      && controller.continuation?.state === 'completed'
+    if (acknowledgedGateResume) {
+      this.assertV4PhaseCanAdvance(controller)
+      round += 1
+      controller.currentRound = round
+      challengeBindings = null
+    }
+    let proposalComplete = activeKinds.every(kind => Boolean(latestFor('proposal', kind)))
+    const legacyBoundRun = Boolean(synthesisBinding && !coordinationPlan)
+    const postChallengePhases = new Set([
+      'coordination', 'work', 'synthesis', 'verification',
+      'human-gate', 'commit', 'committed', 'completed',
+    ])
+    const challengeObservable = existing && (
+      postChallengePhases.has(existing.phase)
+      || existing.slots?.some(slot => (
+        (slot.phase === 'challenge' && slot.status !== 'planned')
+        || postChallengePhases.has(slot.phase)
+        || (slot.resultRefs?.workflowOutcomeRefs || []).some(item => (
+          item?.receipt?.phase === 'challenge'
+        ))
+      ))
+    )
+    if (!coordinationPlan && !legacyBoundRun
+        && !challengeBindings && challengeObservable && !acknowledgedGateResume) {
+      throw new Error('LOCAL_RUN_V4_CHALLENGE_BINDING_REQUIRED')
+    }
+    if (!coordinationPlan && !legacyBoundRun
+        && challengeBindings?.some(binding => binding.round !== round)) {
+      throw new Error('LOCAL_RUN_V4_CHALLENGE_BINDING_INVALID')
+    }
+    if (existing?.candidateCommit) {
+      const verificationKinds = synthesisRecovery?.verificationKinds
+        || coordinationPlan?.verifierKinds
+        || synthesisBinding?.verificationKinds
+        || []
+      const validationSlots = slots.map(slot => (
+        slot.agentKind === writerKind && slot.phase === 'completed'
+          ? { ...slot, phase: 'synthesis' }
+          : slot
+      ))
+      const acceptance = this.v4Acceptance(
+        receiptRecords, activeKinds, writerKind, verificationKinds,
+        {
+          runId: controller.runId,
+          round,
+          slots: validationSlots,
+          challengeBindings,
+          previousConvergence: null,
+          coordinationPlan,
+          workReceipts,
+        },
+      )
+      if (!acceptance.accepted
+          || acceptance.candidateArtifactId !== existing.candidateCommit.candidateArtifactId
+          || acceptance.candidateHash !== existing.candidateCommit.candidateContentHash) {
+        throw new Error('LOCAL_RUN_V4_CANDIDATE_INVALID')
+      }
+      return this.v4CommitAcceptedCandidate({
+        group, controller, threadRootId, targetKinds, batchId, snapshotRecord, snapshotHash,
+        slots, receiptRecords, challengeBindings, synthesisBinding, synthesisRecovery,
+        coordinationPlan, workReceipts, convergence: existing.convergence,
+        writerKind, verificationKinds, acceptance, participantKinds: activeKinds,
+      })
+    }
+    if (this.naturalAgentResponses && !coordinationPlan && !synthesisBinding && !synthesisRecovery) {
+      return this.runV4NaturalDiscussion(group, controller, threadRootId, context, {
+        resume,
+        resumedGate,
+        targetKinds,
+        activeKinds,
+        writerKind,
+        batchId,
+        snapshot,
+        snapshotRecord,
+        snapshotHash,
+        slots,
+        receiptRecords,
+        latestFor,
+        runPhase,
+        checkpointPhase,
+        removePhaseFailures,
+        addRoundLimitNotice,
+      })
+    }
+    while (!controller.signal.aborted) {
+      controller.currentRound = round
+      if (!proposalComplete) {
+        const proposal = await runPhase('proposal', activeKinds)
+        commitPhaseMessages('proposal', proposal)
+        if (!proposal.ok) {
+          const failedKinds = removePhaseFailures(proposal)
+          activeKinds = activeKinds.filter(kind => !failedKinds.includes(kind))
+          checkpointPhase('proposal', [], slots)
+          if (!failedKinds.length || activeKinds.length < 2) {
+            return proposal.phasePendingMessages?.length ? 'partial' : 'failed'
+          }
+        }
+        proposalComplete = activeKinds.every(kind => Boolean(latestFor('proposal', kind)))
+        if (!proposalComplete) {
+          return activeKinds.some(kind => Boolean(latestFor('proposal', kind))) ? 'partial' : 'failed'
+        }
+      }
+
+      if (!coordinationPlan && !legacyBoundRun) {
+        if (round === 1) {
+          if (!controller.unlimitedRounds && round >= (controller.maxRounds || 6)) {
+            addRoundLimitNotice()
+            return 'round-limit'
+          }
+          this.assertV4PhaseCanAdvance(controller)
+          round = 2
+          controller.currentRound = round
+        }
+
+        if (!challengeBindings || challengeBindings[0]?.round !== round) {
+          const activeSet = new Set(activeKinds)
+          slots = slots.map((slot) => {
+            if (!activeSet.has(slot.agentKind)) return slot
+            const { agentRunId: _agentRunId, ...retainedSlot } = slot
+            return {
+              ...retainedSlot,
+              phase: 'challenge',
+              status: 'planned',
+              operationId: this.v4OperationId(
+                controller, slot.agentKind, 'challenge', slot.slotId,
+              ),
+              receiptId: '',
+              resultHash: '',
+              startedAt: Date.now(),
+              finishedAt: null,
+              commitStatus: 'pending',
+              permission: 'read-only',
+            }
+          })
+          challengeBindings = this.v4ChallengeBindings({
+            controller,
+            targetKinds: activeKinds,
+            round,
+            snapshotRecord,
+            slots,
+            receiptRecords,
+          })
+          checkpointPhase('challenge', activeKinds, slots, challengeBindings)
+          this.emitChanged()
+        }
+
+        const knownCoordination = this.v4CoordinationResult({
+          receiptRecords,
+          targetKinds: activeKinds,
+          snapshotHash,
+          snapshot,
+          requireSupport: false,
+        })
+        const challenge = await runPhase('challenge', activeKinds, challengeBindings, {
+          coordinationText: this.v4CoordinationText(knownCoordination),
+        })
+        commitPhaseMessages('challenge', challenge)
+        if (!challenge.ok) {
+          if (controller.signal.aborted) return 'stopped'
+          const failedKinds = removePhaseFailures(challenge)
+          activeKinds = activeKinds.filter(kind => !failedKinds.includes(kind))
+          if (!failedKinds.length || activeKinds.length < 2
+              || (!controller.unlimitedRounds && round >= (controller.maxRounds || 6))) {
+            return challenge.phasePendingMessages?.length ? 'partial' : 'failed'
+          }
+          this.assertV4PhaseCanAdvance(controller)
+          round += 1
+          controller.currentRound = round
+          challengeBindings = null
+          this.emitChanged()
+          continue
+        }
+        if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+
+        const coordination = this.v4CoordinationResult({
+          receiptRecords,
+          targetKinds: activeKinds,
+          snapshotHash,
+          snapshot,
+          challengeBindings,
+        })
+        if (!coordination.plan) {
+          if (!controller.unlimitedRounds && round >= (controller.maxRounds || 6)) {
+            addRoundLimitNotice()
+            return 'round-limit'
+          }
+          this.assertV4PhaseCanAdvance(controller)
+          round += 1
+          controller.currentRound = round
+          challengeBindings = null
+          this.emitChanged()
+          continue
+        }
+        coordinationPlan = coordination.plan
+        writerKind = coordinationPlan.finalizerKind
+        this.v4CheckpointPhase(group, controller, {
+          targetKinds,
+          participantKinds: activeKinds,
+          phase: 'coordination',
+          batchId,
+          snapshotRecord,
+          snapshotHash,
+          slots,
+          writerKind,
+          currentKinds: [],
+          pendingKinds: [],
+          receipts: receiptRecords,
+          challengeBindings,
+          coordinationPlan,
+        })
+        this.emitChanged()
+      }
+
+      if (coordinationPlan) {
+        while (!controller.signal.aborted) {
+          const workState = this.v4WorkState(workReceipts, coordinationPlan)
+          if (!workState.pending.length) break
+          if (!workState.ready.length) return 'failed'
+          const readyAssignments = []
+          const readyOwners = new Set()
+          for (const assignment of workState.ready) {
+            if (readyOwners.has(assignment.ownerKind)) continue
+            readyOwners.add(assignment.ownerKind)
+            readyAssignments.push(assignment)
+          }
+          const readyKinds = activeKinds.filter(kind => readyOwners.has(kind))
+          const completedBefore = workState.completed.size
+          const work = await runPhase('work', readyKinds, challengeBindings, {
+            deliveryReceiptRecords: [...receiptRecords],
+            workAssignments: readyAssignments,
+          })
+          if (!work.ok) return controller.signal.aborted ? 'stopped' : 'failed'
+          const completedAfter = this.v4WorkState(workReceipts, coordinationPlan).completed.size
+          if (completedAfter <= completedBefore) return 'failed'
+        }
+        if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+        writerKind = coordinationPlan.finalizerKind
+      } else {
+        writerKind = synthesisRecovery?.activeWriterKind || synthesisBinding.writerKind
+      }
+
+      const roleBinding = coordinationPlan || synthesisBinding
+      if (!roleBinding) return 'failed'
+      if (!synthesisRecovery) {
+        synthesisRecovery = this.v4CreateSynthesisRecovery(
+          controller,
+          roleBinding,
+          slots,
+          { stateEpoch: controller.orchestration?.convergence?.stateEpoch || 0 },
+        )
+        writerKind = synthesisRecovery.activeWriterKind
+      } else if (!this.v4ActiveSynthesisAttempt(synthesisRecovery)) {
+        const latestAttempt = synthesisRecovery.attempts.at(-1)
+        const currentRoundOperationId = latestAttempt
+          ? this.v4SynthesisAttemptOperationId(
+              controller,
+              latestAttempt.writerKind,
+              latestAttempt.slotId,
+              latestAttempt.attempt,
+            )
+          : ''
+        if (latestAttempt?.status === 'completed'
+            && latestAttempt.operationId !== currentRoundOperationId) {
+          synthesisRecovery = this.v4AppendSynthesisAttempt(
+            controller,
+            roleBinding,
+            synthesisRecovery,
+            slots,
+            synthesisRecovery.activeWriterKind,
+          )
+        }
+      }
+      let activeAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+      if (activeAttempt?.status === 'leased'
+          && activeAttempt.permission === 'workspace-write'
+          && activeAttempt.leaseAcquired === true) {
+        const acceptedReceipt = this.v4TrustedSynthesisResult({
+          controller,
+          slots,
+          activeAttempt,
+          synthesisRecovery,
+          coordinationPlan,
+          snapshotHash,
+        })
+        if (!acceptedReceipt) {
+          synthesisRecovery = this.v4UpdateSynthesisAttempt(
+            synthesisRecovery,
+            activeAttempt.operationId,
+            {
+              status: 'unknown_outcome',
+              sideEffectsPossible: true,
+              outcomeCertainty: 'unknown_outcome',
+            },
+          )
+          synthesisRecovery = {
+            ...synthesisRecovery,
+            pendingGate: this.v4SynthesisRecoveryGate(controller, synthesisRecovery),
+          }
+          activeAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+        }
+      }
+      if (activeAttempt?.status === 'unknown_outcome') {
+        if (!synthesisRecovery.pendingGate) {
+          synthesisRecovery = {
+            ...synthesisRecovery,
+            pendingGate: this.v4SynthesisRecoveryGate(controller, synthesisRecovery),
+          }
+        }
+        const recoveryResult = await requestSynthesisRecoveryGate()
+        if (recoveryResult !== 'continue') return recoveryResult
+      }
+      let verificationKinds = [...synthesisRecovery.verificationKinds]
+      let synthesis = null
+      while (!controller.signal.aborted) {
+        writerKind = synthesisRecovery.activeWriterKind
+        verificationKinds = [...synthesisRecovery.verificationKinds]
+        synthesis = await runPhase('synthesis', [writerKind])
+        if (synthesis.ok) break
+        if (controller.signal.aborted) return 'stopped'
+        const failedAttempt = [...synthesisRecovery.attempts].reverse()[0]
+        if (failedAttempt?.status === 'unknown_outcome') {
+          const recoveryResult = await requestSynthesisRecoveryGate()
+          if (recoveryResult !== 'continue') return recoveryResult
+          continue
+        }
+        const replacement = coordinationPlan
+          ? null
+          : this.v4ReplaceSynthesisWriter(
+              controller, synthesisBinding, synthesisRecovery, slots,
+            )
+        if (!replacement) return 'failed'
+        synthesisRecovery = replacement
+        writerKind = replacement.activeWriterKind
+        verificationKinds = [...replacement.verificationKinds]
+      }
+      if (!synthesis?.ok) return controller.signal.aborted ? 'stopped' : 'failed'
+      let synthesisPendingMessage = synthesis.synthesisPendingMessage
+      const synthesisReceipt = latestFor('synthesis', writerKind)
+      if (!synthesisPendingMessage && synthesisReceipt) {
+        const candidateArtifactId = synthesisReceipt.receipt.artifactIds?.[0]
+        let candidateContent = ''
+        if (candidateArtifactId) {
+          try { candidateContent = this.v4ArtifactIdentity(candidateArtifactId).content } catch {}
+        }
+        synthesisPendingMessage = {
+          groupId: group.id,
+          role: 'agent',
+          content: candidateContent
+            || synthesisReceipt.receipt.conclusion
+            || synthesisReceipt.receipt.summary,
+          agentKind: writerKind,
+          threadRootId,
+          system: null,
+          metadata: {},
+        }
+      }
+      if (!synthesisPendingMessage) return 'failed'
+      const previousConvergence = controller.orchestration?.convergence || null
+      this.v4RoundState(
+        receiptRecords, activeKinds, writerKind, verificationKinds,
+        {
+          runId: controller.runId,
+          round,
+          slots,
+          challengeBindings,
+          previousConvergence,
+          coordinationPlan,
+          includeVerification: false,
+        },
+      )
+      const verification = await runPhase('verification', verificationKinds)
+      if (!verification.ok) return controller.signal.aborted ? 'stopped' : 'partial'
+
+      const acceptance = this.v4Acceptance(
+        receiptRecords, activeKinds, writerKind, verificationKinds,
+        {
+          runId: controller.runId,
+          round,
+          slots,
+          challengeBindings,
+          previousConvergence,
+          coordinationPlan,
+          workReceipts,
+        },
+      )
+      const convergence = this.v4NextConvergence(
+        previousConvergence,
+        {
+          artifactId: acceptance.candidateArtifactId,
+          contentHash: acceptance.candidateHash,
+        },
+        acceptance.openIssueIds,
+        round,
+        writerKind,
+      )
+      this.v4CheckpointPhase(group, controller, {
+        targetKinds,
+        participantKinds: activeKinds,
+        phase: 'verification',
+        batchId,
+        snapshotRecord,
+        snapshotHash,
+        slots,
+        writerKind,
+        currentKinds: [],
+        pendingKinds: [],
+        receipts: receiptRecords,
+        challengeBindings,
+        synthesisBinding,
+        coordinationPlan,
+        convergence,
+      })
+      const acceptedStable = controller.unlimitedRounds
+        ? convergence.consecutiveStableRounds >= 2
+        : round >= (controller.maxRounds || 6)
+      if (acceptance.accepted && acceptedStable) {
+        return finalStatus(await this.v4CommitAcceptedCandidate({
+          group, controller, threadRootId, targetKinds, batchId, snapshotRecord, snapshotHash,
+          slots, receiptRecords, challengeBindings, synthesisBinding, synthesisRecovery,
+          coordinationPlan, workReceipts, convergence,
+          writerKind, verificationKinds, acceptance, participantKinds: activeKinds,
+        }))
+      }
+
+      if (this.v4ShouldOpenStableGate(convergence, controller.unlimitedRounds)) {
+        const gateResult = await requestStableGate(convergence)
+        if (gateResult !== 'continue') return gateResult
+      }
+      if (!controller.unlimitedRounds && round >= (controller.maxRounds || 6)) {
+        addRoundLimitNotice()
+        return 'round-limit'
+      }
+      if (controller.signal.aborted) break
+      this.assertV4PhaseCanAdvance(controller)
+      round += 1
+      controller.currentRound = round
+      if (coordinationPlan) {
+        challengeBindings = null
+        if (Object.hasOwn(controller.orchestration || {}, 'challengeBindings')) {
+          const { challengeBindings: _challengeBindings, ...orchestration } = controller.orchestration
+          controller.orchestration = orchestration
+        }
+      }
+      this.emitChanged()
+    }
+    return terminalRunStatusForReason(controller.stopReason)
+  },
+}
+
+module.exports = { v4DiscussionMethods }

desktop/src/workspace/local-workspace-auto-runner-v4-phase.cjs

@@ -0,0 +1,872 @@
+// Running one V4 phase for a set of Agents.
+const { collaborationPackageIndexText, emptyCollaborationState } = require('../collaboration/collaboration-records.cjs')
+const { hashValue } = require('../collaboration/orchestration-v4-records.cjs')
+const { stalePermissionResumeFailure } = require('./local-workspace-auto-runner-support.cjs')
+
+const v4PhaseMethods = {
+  v4PhaseAdvanceBlocked(controller) {
+    const orchestration = controller.orchestration
+    const continuation = controller.continuation
+    return orchestration?.version === 4
+      && continuation?.resumeKind === 'agent_slot'
+      && continuation.phase === orchestration.phase
+      && continuation.round === orchestration.round
+      // `resuming` means the gate decision has already been accepted and the
+      // continuation is actively replaying the gated slot. Blocking that
+      // state makes the resume path fail with its own phase-gate error.
+      && ['pending', 'ready'].includes(continuation.state)
+  },
+
+  assertV4PhaseCanAdvance(controller) {
+    if (this.v4PhaseAdvanceBlocked(controller)) {
+      throw new Error('LOCAL_RUN_V4_PHASE_GATE_PENDING')
+    }
+  },
+
+  v4PhaseSlots({ controller, targetKinds, phase, activeKinds, slots, snapshotHash,
+    writerKind, operationIds = null, allowWrite = false, now = Date.now() }) {
+    const active = new Set(activeKinds)
+    return slots.map((slot, index) => {
+      const running = active.has(slot.agentKind)
+      const previousHistory = Array.isArray(slot.resultRefs?.workflowOutcomeRefs)
+        ? slot.resultRefs.workflowOutcomeRefs
+        : []
+      const operationId = running
+        ? (operationIds?.get(slot.agentKind)
+          || this.v4OperationId(controller, slot.agentKind, phase, slot.slotId))
+        : slot.operationId
+      const { agentRunId: previousAgentRunId, ...retainedSlot } = slot
+      return {
+        ...retainedSlot,
+        ...(!running && operationId === slot.operationId && previousAgentRunId
+          ? { agentRunId: previousAgentRunId }
+          : {}),
+        slotId: slot.slotId || `slot-${index + 1}-${slot.agentKind}`,
+        phase: running ? phase : (slot.phase || phase),
+        status: running ? 'planned' : (slot.status || 'completed'),
+        operationId,
+        queuePosition: index,
+        snapshotHash,
+        deliveryWatermark: Number(slot.deliveryWatermark) || 0,
+        receiptId: running ? '' : (slot.receiptId || ''),
+        resultHash: running ? '' : (slot.resultHash || ''),
+        assignedAt: slot.assignedAt || now,
+        startedAt: running ? now : (slot.startedAt || now),
+        finishedAt: running ? null : (slot.finishedAt || now),
+        commitStatus: running ? 'pending' : (slot.commitStatus || 'committed'),
+        attempt: (Number.isSafeInteger(slot.attempt) ? slot.attempt : 0) + (running ? 1 : 0),
+        permission: allowWrite && slot.agentKind === writerKind
+          && (phase === 'synthesis' || (this.naturalAgentResponses && phase === 'discussion'))
+          ? 'workspace-write'
+          : 'read-only',
+        resultRefs: {
+          artifactIds: Array.isArray(slot.resultRefs?.artifactIds)
+            ? [...slot.resultRefs.artifactIds]
+            : [],
+          evidenceIds: Array.isArray(slot.resultRefs?.evidenceIds)
+            ? [...slot.resultRefs.evidenceIds]
+            : [],
+          workflowOutcomeRefs: [...previousHistory],
+        },
+      }
+    })
+  },
+
+  v4CheckpointPhase(group, controller, {
+    targetKinds, phase, batchId, snapshotRecord, snapshotHash, slots,
+    writerKind, currentKinds = [], pendingKinds = [], receipts = [], commitState = null,
+    participantKinds = targetKinds,
+    challengeBindings = controller.orchestration?.round === controller.currentRound
+      ? controller.orchestration?.challengeBindings
+      : null,
+    synthesisBinding = controller.orchestration?.synthesisBinding || null,
+    synthesisRecovery = controller.orchestration?.synthesisRecovery || null,
+    convergence = controller.orchestration?.convergence || null,
+    coordinationPlan = controller.orchestration?.coordinationPlan || null,
+    workReceipts = controller.orchestration?.workReceipts || [],
+    candidateCommit = controller.orchestration?.candidateCommit || null,
+    workAssignments = [],
+  }) {
+    const current = controller.orchestration
+    const naturalWriterKind = this.naturalAgentResponses && ['proposal', 'discussion'].includes(phase)
+      && !coordinationPlan && !synthesisBinding ? writerKind : ''
+    if (current?.version === 4
+        && (current.phase !== phase || current.round !== controller.currentRound)) {
+      this.assertV4PhaseCanAdvance(controller)
+    }
+    const workAssignmentsByKind = new Map(workAssignments.map(assignment => (
+      [assignment.ownerKind, assignment]
+    )))
+    const assignments = slots.map((slot, index) => ({
+      agentKind: slot.agentKind,
+      slotId: slot.slotId,
+      role: this.v4CheckpointRole(
+        phase, slot, coordinationPlan, workAssignmentsByKind.get(slot.agentKind) || null,
+      ),
+      operationId: slot.operationId,
+      objective: phase === 'proposal'
+        ? 'Produce an independent bounded proposal.'
+        : (phase === 'work' && workAssignmentsByKind.has(slot.agentKind)
+          ? workAssignmentsByKind.get(slot.agentKind).objective
+          : `Complete the ${phase} collaboration phase using the peer-agreed responsibility plan.`),
+      expectedOutput: phase === 'work' && workAssignmentsByKind.has(slot.agentKind)
+        ? workAssignmentsByKind.get(slot.agentKind).expectedOutput
+        : 'Return a concise structured collaboration result.',
+      inputRefs: receipts.map(item => item.receipt?.receiptId).filter(Boolean).slice(-64),
+      readOnly: slot.permission !== 'workspace-write',
+      index,
+    }))
+    const plan = {
+      version: 1,
+      snapshotHash,
+      assignments: assignments.map(({ index: _index, ...assignment }) => assignment),
+      createdAt: snapshotRecord.capturedAt,
+      barrier: 'phase',
+    }
+    const watermarks = Array.isArray(controller.v4Watermarks)
+      ? controller.v4Watermarks
+      : []
+    const commit = commitState || {
+      status: 'pending',
+      writerKind: naturalWriterKind ? null : writerKind || null,
+      committedKinds: [],
+      pendingKinds: [...participantKinds],
+      operationId: '',
+      attempt: 0,
+      updatedAt: Date.now(),
+      committedSlotIds: [],
+      messageIds: [],
+      blackboardEntryIds: [],
+    }
+    const record = {
+      version: 4,
+      workflow: 'auto',
+      template: 'discussion',
+      phase,
+      discussionStyle: controller.discussionStyle
+        || controller.orchestration?.discussionStyle
+        || 'sequential',
+      ...(naturalWriterKind ? { discussionWriterKind: naturalWriterKind } : {}),
+      batchId,
+      round: Math.max(0, Number(controller.currentRound) || 0),
+      currentKind: '',
+      currentKinds: [...currentKinds],
+      pendingKinds: [...pendingKinds],
+      activeKinds: [...participantKinds],
+      successfulKinds: slots
+        .filter(slot => ['completed', 'partial'].includes(slot.status))
+        .map(slot => slot.agentKind),
+      agreementKinds: [],
+      attachmentRecipients: [],
+      totalSuccesses: slots.filter(slot => slot.status === 'completed').length,
+      terminalFailureOccurred: slots.some(slot => slot.status === 'failed'),
+      snapshotHash,
+      snapshot: snapshotRecord,
+      plan,
+      slots,
+      deliveryWatermarks: watermarks,
+      ...(Array.isArray(controller.orchestration?.deliveryState)
+        ? { deliveryState: controller.orchestration.deliveryState.map(entry => ({ ...entry })) }
+        : {}),
+      commitState: commit,
+      ...(candidateCommit ? { candidateCommit: { ...candidateCommit } } : {}),
+      collaboration: controller.orchestration?.collaboration || emptyCollaborationState(),
+      ...(Array.isArray(challengeBindings)
+        ? { challengeBindings: challengeBindings.map(binding => ({
+            ...binding,
+            artifactIds: [...binding.artifactIds],
+            evidenceIds: [...binding.evidenceIds],
+          })) }
+        : {}),
+      ...(synthesisBinding ? {
+        synthesisBinding: {
+          ...synthesisBinding,
+          candidates: synthesisBinding.candidates.map(candidate => ({
+            ...candidate,
+            ...(candidate.evidence ? { evidence: { ...candidate.evidence } } : {}),
+          })),
+          verificationKinds: [...synthesisBinding.verificationKinds],
+        },
+      } : {}),
+      ...(synthesisRecovery ? {
+        synthesisRecovery: {
+          ...synthesisRecovery,
+          verificationKinds: [...synthesisRecovery.verificationKinds],
+          rankedKinds: [...synthesisRecovery.rankedKinds],
+          triedWriters: [...synthesisRecovery.triedWriters],
+          attempts: synthesisRecovery.attempts.map(attempt => ({ ...attempt })),
+          ...(synthesisRecovery.pendingGate
+            ? { pendingGate: { ...synthesisRecovery.pendingGate } }
+            : {}),
+        },
+      } : {}),
+      ...(convergence ? { convergence: { ...convergence, openIssueIds: [...convergence.openIssueIds] } } : {}),
+      ...(coordinationPlan ? {
+        coordinationPlan: {
+          ...coordinationPlan,
+          assignments: coordinationPlan.assignments.map(assignment => ({ ...assignment })),
+          verifierKinds: [...coordinationPlan.verifierKinds],
+          agreedBy: [...coordinationPlan.agreedBy],
+          ...(coordinationPlan.supportReceiptIds
+            ? { supportReceiptIds: [...coordinationPlan.supportReceiptIds] }
+            : {}),
+        },
+      } : {}),
+      ...(coordinationPlan ? {
+        workReceipts: workReceipts.map(receipt => ({
+          ...receipt,
+          snapshotContentRef: { ...receipt.snapshotContentRef },
+          collaborationReceipt: { ...receipt.collaborationReceipt },
+          artifacts: receipt.artifacts.map(artifact => ({
+            ...artifact,
+            contentRef: { ...artifact.contentRef },
+          })),
+        })),
+      } : {}),
+    }
+    this.checkpointOrchestration(group, controller, record)
+    return record
+  },
+
+  async runV4Phase(group, controller, threadRootId, context, phaseInput) {
+    const {
+      phase, activeKinds, targetKinds, writerKind, batchId,
+      snapshot, snapshotRecord, snapshotHash, slots, receiptRecords, challengeBindings,
+    } = phaseInput
+    const participantKinds = Array.isArray(phaseInput.participantKinds)
+      ? phaseInput.participantKinds
+      : targetKinds
+    const coordinationPlan = phaseInput.coordinationPlan || null
+    const workAssignments = phase === 'work' && Array.isArray(phaseInput.workAssignments)
+      ? phaseInput.workAssignments
+      : []
+    const workAssignmentsByKind = new Map(workAssignments.map(assignment => (
+      [assignment.ownerKind, assignment]
+    )))
+    if (phase === 'work' && (workAssignments.length !== activeKinds.length
+        || workAssignmentsByKind.size !== activeKinds.length
+        || activeKinds.some(kind => {
+          const assignment = workAssignmentsByKind.get(kind)
+          return !assignment || !coordinationPlan?.assignments?.some(candidate => (
+            candidate.taskId === assignment.taskId && candidate.ownerKind === kind
+          ))
+        }))) {
+      throw new Error('LOCAL_RUN_V4_COORDINATION_PLAN_INVALID')
+    }
+    let workReceipts = Array.isArray(phaseInput.workReceipts)
+      ? [...phaseInput.workReceipts]
+      : []
+    const deliveryReceiptRecords = Array.isArray(phaseInput.deliveryReceiptRecords)
+      ? phaseInput.deliveryReceiptRecords
+      : [...receiptRecords]
+    const workContextLimit = Math.max(
+      0, 6000 - collaborationPackageIndexText(participantKinds).length - 2,
+    )
+    const extraContextFor = kind => phase === 'work'
+      ? this.v4WorkAssignmentText(
+          coordinationPlan,
+          kind,
+          workAssignmentsByKind.get(kind)?.taskId,
+          snapshot,
+          workContextLimit,
+          deliveryReceiptRecords,
+        )
+      : (phase === 'challenge' ? phaseInput.coordinationText || '' : '')
+    let synthesisRecovery = phaseInput.synthesisRecovery || null
+    const synthesisAttempt = phase === 'synthesis'
+      ? this.v4ActiveSynthesisAttempt(synthesisRecovery)
+      : null
+    const operationIds = phaseInput.operationIds instanceof Map
+      ? phaseInput.operationIds
+      : synthesisAttempt
+      ? new Map([[synthesisAttempt.writerKind, synthesisAttempt.operationId]])
+      : (phase === 'work'
+        ? new Map(activeKinds.map((kind) => {
+            const slot = slots.find(candidate => candidate.agentKind === kind)
+            const assignment = workAssignmentsByKind.get(kind)
+            return [kind, this.v4OperationId(
+              controller, kind, `work:${assignment.taskId}`, slot?.slotId || '',
+            )]
+          }))
+        : null)
+    const phaseSlots = this.v4PhaseSlots({
+      controller,
+      targetKinds,
+      phase,
+      activeKinds,
+      slots,
+      snapshotHash,
+      writerKind,
+      operationIds,
+      allowWrite: group.allowWrite === true,
+    })
+    const resumedGate = phaseInput.resumedGate
+    const resumedGateFor = slot => (
+      resumedGate?.phase === phase
+        && resumedGate.agentKind === slot.agentKind
+        && resumedGate.slotId === slot.slotId
+        && resumedGate.operationId === slot.operationId
+        && resumedGate.snapshotHash === snapshotHash
+        && resumedGate.round === controller.currentRound
+        ? resumedGate
+        : null
+    )
+    const activeSet = new Set(activeKinds)
+    const phaseReceipts = []
+    const snapshotSourceMessageIds = [snapshot.messageId, ...snapshot.history.map(item => item.id)]
+    for (const slot of phaseSlots) {
+      if (activeSet.has(slot.agentKind)) {
+        slot.status = phase === 'synthesis' ? 'queued' : 'running'
+      }
+    }
+    const cursorPendingKinds = Array.isArray(phaseInput.pendingKinds)
+      ? phaseInput.pendingKinds
+      : activeKinds
+    this.v4CheckpointPhase(group, controller, {
+      targetKinds,
+      phase,
+      batchId,
+      snapshotRecord,
+      snapshotHash,
+      slots: phaseSlots,
+      writerKind,
+      currentKinds: activeKinds,
+      pendingKinds: cursorPendingKinds,
+      participantKinds,
+      receipts: receiptRecords,
+      challengeBindings,
+      synthesisRecovery,
+      coordinationPlan,
+      workReceipts,
+      workAssignments,
+    })
+    this.emitChanged()
+
+    if (phase === 'proposal') {
+      const proposalTurns = this.reserveTurnSeqs(group.id, threadRootId, activeKinds.length)
+      const dispatches = activeKinds.map((kind, kindIndex) => {
+        const slot = phaseSlots.find(candidate => candidate.agentKind === kind)
+        const prompt = this.v4PhasePrompt(
+          group, kind, phase, snapshot, receiptRecords, 'participant', {
+            reviewTarget: '',
+            packageText: '',
+            skillHints: context.rootSkillsByKind.get(kind) || [],
+            knowledgeBaseHints: context.rootKnowledgeBasesByKind.get(kind) || [],
+          },
+        )
+        const sessionPolicy = this.naturalAgentResponses ? null : 'frozen'
+        return {
+          kind,
+          slot,
+          context: {
+            v4: true,
+            phase,
+            batchId,
+            slotId: slot.slotId,
+            snapshotHash,
+            ...(proposalTurns[kindIndex] ? { turnSeq: proposalTurns[kindIndex] } : {}),
+            ...(sessionPolicy ? { sessionPolicy } : {}),
+            promptOverride: prompt,
+            contextPackId: controller.contextPackId,
+            snapshotSourceMessageIds,
+            snapshotSourceEntries: snapshot.history,
+            permissionMode: 'read-only',
+            singleWriterKind: writerKind,
+            parallelGraph: true,
+            deferMessage: true,
+            allowMissingProposalReceipt: true,
+            allowMissingV4Receipt: this.naturalAgentResponses,
+            operationId: slot.operationId,
+            ...(resumedGateFor(slot) ? { resumedGate: resumedGateFor(slot) } : {}),
+            skillHints: context.rootSkillsByKind.get(kind) || [],
+            knowledgeBaseHints: context.rootKnowledgeBasesByKind.get(kind) || [],
+            ...(this.naturalAgentResponses && context.rootAttachments?.length
+              ? { attachmentSnapshots: context.rootAttachments }
+              : {}),
+          },
+        }
+      })
+      return this.runV4ProposalBatch(group, controller, threadRootId, {
+        targetKinds,
+        participantKinds,
+        writerKind,
+        batchId,
+        snapshotRecord,
+        snapshotHash,
+        phaseSlots,
+        receiptRecords,
+        dispatches,
+        resumedGate,
+      })
+    }
+
+    const failures = []
+    const phasePendingMessages = []
+    let synthesisPendingMessage = null
+    let workAdmissionClosed = false
+    let workAdmissionTail = Promise.resolve()
+    const admitWork = (item) => {
+      const admission = workAdmissionTail.then(() => {
+        if (workAdmissionClosed) return { ...item, workSettled: true }
+        const admitted = this.v4AdmitWorkSettlement(
+          group, controller, threadRootId, item, {
+            targetKinds,
+            participantKinds,
+            writerKind,
+            batchId,
+            snapshotRecord,
+            snapshotHash,
+            phaseSlots,
+            receiptRecords,
+            challengeBindings,
+            coordinationPlan,
+            workReceipts,
+            phaseReceipts,
+            workAssignments,
+            resumedGate: resumedGateFor(phaseSlots.find(candidate => (
+              candidate.agentKind === item.kind
+            ))),
+          },
+        )
+        workReceipts = admitted.workReceipts
+        if (admitted.failure) failures.push(admitted.failure)
+        return { ...item, workSettled: true }
+      })
+      workAdmissionTail = admission.catch(() => { workAdmissionClosed = true })
+      return admission
+    }
+    // Batched peers keep dispatch order in the timeline whatever order they
+    // finish in.
+    const phaseTurns = this.reserveTurnSeqs(group.id, threadRootId, activeKinds.length)
+    const turnSeqByKind = new Map(activeKinds.map((kind, index) => [kind, phaseTurns[index]]))
+    const settled = await Promise.all(activeKinds.map(async kind => {
+      const slot = phaseSlots.find(candidate => candidate.agentKind === kind)
+      const challengeBinding = phase === 'challenge'
+        ? challengeBindings?.find(binding => binding.reviewerKind === kind) || null
+        : null
+      const assignedProposalText = challengeBinding
+        ? this.v4ChallengeProposalText(challengeBinding, kind)
+        : ''
+      const role = this.v4PromptRole(
+        phase, kind, coordinationPlan, workAssignmentsByKind.get(kind) || null,
+      )
+      const extraContext = extraContextFor(kind)
+      const promptOptions = {
+        reviewTarget: phase === 'challenge'
+          ? challengeBinding?.proposalKind || ''
+          : '',
+        assignedProposalText,
+        packageText: '',
+        extraContext,
+        snapshotHash,
+        skillHints: context.rootSkillsByKind.get(kind) || [],
+        knowledgeBaseHints: context.rootKnowledgeBasesByKind.get(kind) || [],
+      }
+      const prompt = typeof phaseInput.promptBuilder === 'function'
+        ? phaseInput.promptBuilder({
+            kind, slot, role, options: promptOptions,
+            ...(phaseInput.disableDeliveryPrompt === true ? { transcript: '' } : {}),
+          })
+        : this.v4PhasePrompt(
+          group,
+          kind,
+          phase,
+          snapshot,
+          deliveryReceiptRecords,
+          role,
+          promptOptions,
+        )
+      const sessionPolicy = phaseInput.sessionPolicy === undefined
+        ? 'frozen'
+        : phaseInput.sessionPolicy
+      let preparedDelivery = null
+      let item
+      const resumedSlotGate = resumedGateFor(slot)
+      const recoveredResult = phase === 'discussion'
+        ? this.v4NaturalRecoveredDiscussionResult(
+            group, controller, threadRootId, kind, slot,
+          )
+        : null
+      if (recoveredResult) return { kind, invocation: { result: recoveredResult } }
+      try {
+        const invocation = await this.invokeWithUnauthorizedRecovery({
+          group,
+          kind,
+          controller,
+          activeKinds: participantKinds,
+          threadRootId,
+          context: {
+            v4: true,
+            phase,
+            batchId,
+            slotId: slot.slotId,
+            snapshotHash,
+            ...(turnSeqByKind.get(kind) ? { turnSeq: turnSeqByKind.get(kind) } : {}),
+            ...(sessionPolicy ? { sessionPolicy } : {}),
+            promptOverride: prompt,
+            contextPackId: controller.contextPackId,
+            snapshotSourceMessageIds,
+            snapshotSourceEntries: snapshot.history,
+            permissionMode: group.allowWrite === true
+              && (phase === 'synthesis' || (this.naturalAgentResponses && phase === 'discussion'))
+              && kind === writerKind
+              ? 'workspace-write'
+              : 'read-only',
+            singleWriterKind: writerKind,
+            parallelGraph: true,
+            deferMessage: true,
+            allowMissingV4Receipt: this.naturalAgentResponses,
+            operationId: slot.operationId,
+            ...(resumedSlotGate ? { resumedGate: resumedSlotGate } : {}),
+            v4SynthesisRecovery: phase === 'synthesis',
+            onLeaseAcquired: (leaseState) => {
+              if (!leaseState?.agentRunId) throw new Error('LOCAL_RUN_PERSIST_FAILED')
+              slot.agentRunId = leaseState.agentRunId
+              if (phase === 'synthesis') {
+                const activeAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+                if (!activeAttempt || activeAttempt.operationId !== slot.operationId) {
+                  throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+                }
+                synthesisRecovery = this.v4UpdateSynthesisAttempt(
+                  synthesisRecovery,
+                  activeAttempt.operationId,
+                  {
+                    status: 'leased',
+                    permission: leaseState?.permissionMode === 'workspace-write'
+                      ? 'workspace-write' : 'read-only',
+                    leaseAcquired: true,
+                    sideEffectsPossible: leaseState?.sideEffectsPossible === true,
+                    outcomeCertainty: leaseState?.sideEffectsPossible === true
+                      ? 'unknown_outcome' : 'not_started',
+                  },
+                )
+              }
+              slot.status = 'running'
+              this.v4CheckpointPhase(group, controller, {
+                targetKinds,
+                phase,
+                batchId,
+                snapshotRecord,
+                snapshotHash,
+                slots: phaseSlots,
+                writerKind,
+                currentKinds: activeKinds,
+                pendingKinds: cursorPendingKinds,
+                participantKinds,
+                receipts: receiptRecords,
+                challengeBindings,
+                synthesisRecovery,
+                coordinationPlan,
+                workReceipts,
+                workAssignments,
+              })
+            },
+            ...(phaseInput.disableDeliveryPrompt === true
+              ? {
+                  v4PromptBuilder: sessionBinding => {
+                    const built = this.v4NaturalDeliveryPrompt(
+                      group, controller, threadRootId, kind, slot, sessionBinding,
+                      transcript => phaseInput.promptBuilder({ kind, slot, role, options: promptOptions, transcript }),
+                    )
+                    preparedDelivery = built.delivery
+                    return built
+                  },
+                }
+              : {
+                  v4PromptBuilder: (sessionBinding) => {
+                    const built = this.v4DeliveryPrompt(group, controller, {
+                      kind, phase, snapshot, receiptRecords: deliveryReceiptRecords,
+                      role, targetKinds: participantKinds, slot,
+                      options: {
+                        reviewTarget: phase === 'challenge'
+                          ? challengeBinding?.proposalKind || ''
+                          : '',
+                        assignedProposalText,
+                        extraContext,
+                        ...(['synthesis', 'verification'].includes(phase)
+                          ? { artifactContext: { coordinationPlan } }
+                          : {}),
+                      },
+                      sessionBinding,
+                      skillHints: context.rootSkillsByKind.get(kind) || [],
+                    })
+                    preparedDelivery = built.delivery
+                    return built
+                  },
+                }),
+            skillHints: context.rootSkillsByKind.get(kind) || [],
+            knowledgeBaseHints: context.rootKnowledgeBasesByKind.get(kind) || [],
+            ...(this.naturalAgentResponses && context.rootAttachments?.length
+              ? { attachmentSnapshots: context.rootAttachments }
+              : {}),
+          },
+        })
+        if (invocation.result?.v4Delivery) {
+          this.v4SetDeliveryStatus(
+            group, controller, invocation.result.v4Delivery, 'acknowledged',
+            invocation.result.v4SessionBinding,
+          )
+        } else if (preparedDelivery) {
+          this.v4SetDeliveryStatus(group, controller, preparedDelivery, 'uncertain')
+        }
+        item = { kind, invocation }
+      } catch (error) {
+        if (preparedDelivery) this.v4SetDeliveryStatus(group, controller, preparedDelivery, 'uncertain')
+        if (stalePermissionResumeFailure(error, resumedSlotGate)) {
+          this.resetAgentSession(group, kind, true, controller.taskId)
+          item = { kind, error, permissionRestartRecovery: true }
+        } else {
+          item = { kind, error }
+        }
+      }
+      return phase === 'work' ? admitWork(item) : item
+    }))
+
+    for (const item of phase === 'work' ? [] : settled) {
+      const slot = phaseSlots.find(candidate => candidate.agentKind === item.kind)
+      const result = item.invocation?.result
+      if (controller.signal.aborted) {
+        slot.status = 'stopped'
+        slot.commitStatus = 'partial'
+        slot.finishedAt = Date.now()
+        continue
+      }
+      if (!result) {
+        const error = item.error || item.invocation?.error
+          || new Error('LOCAL_AGENT_UNKNOWN_FAILURE')
+        slot.status = 'failed'
+        slot.commitStatus = 'failed'
+        slot.finishedAt = Date.now()
+        failures.push({ ...item, error })
+        if (!controller.signal.aborted) {
+          this.recordAgentFailure(group.id, item.kind, error, threadRootId, new Set())
+        }
+        if (phase === 'synthesis' && synthesisRecovery) {
+          const activeAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+          if (activeAttempt?.operationId === slot.operationId) {
+            const failure = error?.invocationFailure || {}
+            const unknownWrite = activeAttempt.permission === 'workspace-write'
+              && activeAttempt.leaseAcquired
+              && (item.permissionRestartRecovery === true
+                || (failure.outcomeCertainty !== 'not_started'
+                  && failure.outcomeCertainty !== 'known_failed'))
+            synthesisRecovery = this.v4UpdateSynthesisAttempt(
+              synthesisRecovery,
+              activeAttempt.operationId,
+              unknownWrite ? {
+                status: 'unknown_outcome',
+                sideEffectsPossible: true,
+                outcomeCertainty: 'unknown_outcome',
+              } : {
+                status: 'failed',
+                sideEffectsPossible: false,
+                outcomeCertainty: failure.outcomeCertainty === 'not_started'
+                  ? 'not_started' : 'known_failed',
+              },
+            )
+          }
+        }
+        continue
+      }
+      try {
+        const reportedOperationId = result?.message?.trace?.context?.operationId
+          || result?.pendingMessage?.metadata?.trace?.context?.operationId
+        if (reportedOperationId && reportedOperationId !== slot.operationId) {
+          throw new Error('LOCAL_RUN_COLLABORATION_SCOPE_INVALID')
+        }
+        const receiptRecord = this.v4ReceiptForResult(
+          result,
+          phase,
+          item.kind,
+          slot,
+          snapshotHash,
+          {
+            controller,
+            ...(phase === 'work'
+              ? { workItemId: workAssignmentsByKind.get(item.kind)?.taskId || '' }
+              : {}),
+            reviewedArtifactId: phase === 'challenge'
+              ? challengeBindings?.find(binding => (
+                binding.reviewerKind === item.kind
+              ))?.artifactIds?.[0] || ''
+              : (phase === 'verification'
+                ? [...receiptRecords].reverse().find(record => (
+                  record.receipt?.phase === 'synthesis'
+                ))?.receipt?.artifactIds?.[0] || ''
+                : ''),
+          },
+        )
+        if (phase === 'work') {
+          const assignment = workAssignmentsByKind.get(item.kind)
+          if (!assignment || receiptRecord.receipt.workItemId !== assignment.taskId) {
+            throw new Error('LOCAL_RUN_V4_WORK_RECEIPT_INVALID')
+          }
+          if (!this.v4ReceiptReferencesComplete(receiptRecord)) {
+            throw new Error('LOCAL_RUN_V4_WORK_RECEIPT_INVALID')
+          }
+        }
+        const duplicate = receiptRecords.find(existing => (
+          existing?.receipt?.receiptId === receiptRecord.receipt.receiptId
+        ))
+        if (!duplicate) {
+          phaseReceipts.push(receiptRecord)
+          receiptRecords.push(receiptRecord)
+        }
+        const effectiveReceiptRecord = duplicate || receiptRecord
+        controller.v4ReceiptRecords = receiptRecords
+        slot.status = 'completed'
+        slot.commitStatus = 'pending'
+        slot.finishedAt = Date.now()
+        slot.receiptId = effectiveReceiptRecord.receipt.receiptId
+        slot.deliveryWatermark = effectiveReceiptRecord.receipt.deliveryWatermark
+        slot.resultHash = hashValue(effectiveReceiptRecord.receipt)
+        slot.resultRefs = {
+          artifactIds: result.outcomeRefs?.artifactIds || [],
+          evidenceIds: result.outcomeRefs?.evidenceIds || [],
+          workflowOutcomeRefs: [
+            ...(slot.resultRefs?.workflowOutcomeRefs || []),
+            {
+              receipt: effectiveReceiptRecord.receipt,
+              ...(effectiveReceiptRecord.verdict ? { verdict: effectiveReceiptRecord.verdict } : {}),
+              ...(effectiveReceiptRecord.resolvedIssueIds.length
+                ? { resolvedIssueIds: effectiveReceiptRecord.resolvedIssueIds }
+                : {}),
+            },
+          ],
+        }
+        if (phase === 'synthesis' && !duplicate) {
+          synthesisPendingMessage = result.pendingMessage || result.message || null
+        } else if (result.pendingMessage?.content) {
+          phasePendingMessages.push({ kind: item.kind, pendingMessage: result.pendingMessage })
+          if (['discussion', 'challenge'].includes(phase)) {
+            this.commitV4PhaseMessage({
+              group, controller, threadRootId, phase, kind: item.kind,
+              pendingMessage: result.pendingMessage,
+            })
+            slot.commitStatus = 'committed'
+          }
+        }
+        if (phase === 'synthesis' && synthesisRecovery) {
+          const activeAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+          if (!activeAttempt || activeAttempt.operationId !== slot.operationId) {
+            throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+          }
+          synthesisRecovery = this.v4UpdateSynthesisAttempt(
+            synthesisRecovery,
+            activeAttempt.operationId,
+            { status: 'completed', outcomeCertainty: 'succeeded' },
+          )
+        }
+      } catch (error) {
+        slot.status = 'failed'
+        slot.commitStatus = 'failed'
+        slot.finishedAt = Date.now()
+        failures.push({ ...item, error })
+        if (phase === 'synthesis' && synthesisRecovery) {
+          const activeAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+          if (activeAttempt?.operationId === slot.operationId) {
+            const failure = error?.invocationFailure || {}
+            const unknownWrite = activeAttempt.permission === 'workspace-write'
+              && activeAttempt.leaseAcquired
+              && failure.outcomeCertainty !== 'not_started'
+              && failure.outcomeCertainty !== 'known_failed'
+            synthesisRecovery = this.v4UpdateSynthesisAttempt(
+              synthesisRecovery,
+              activeAttempt.operationId,
+              unknownWrite ? {
+                status: 'unknown_outcome',
+                sideEffectsPossible: true,
+                outcomeCertainty: 'unknown_outcome',
+              } : {
+                status: 'failed',
+                sideEffectsPossible: false,
+                outcomeCertainty: failure.outcomeCertainty === 'not_started'
+                  ? 'not_started' : 'known_failed',
+              },
+            )
+          }
+        }
+        if (!controller.signal.aborted) {
+          this.recordAgentFailure(group.id, item.kind, error, threadRootId, new Set())
+        }
+      }
+    }
+    const newWatermarks = phaseReceipts.map(item => ({
+      agentKind: item.receipt.agentKind,
+      phase: item.receipt.phase,
+      watermark: item.receipt.deliveryWatermark,
+      operationId: item.receipt.operationId,
+      snapshotHash: item.receipt.snapshotHash,
+      updatedAt: Date.now(),
+    }))
+    controller.v4Watermarks = [
+      ...(controller.v4Watermarks || []).filter(existing => (
+        !newWatermarks.some(next => (
+          next.agentKind === existing.agentKind && next.phase === existing.phase
+        ))
+      )),
+      ...newWatermarks,
+    ]
+    if (phase === 'synthesis') {
+      const activeAttempt = this.v4ActiveSynthesisAttempt(synthesisRecovery)
+      if (activeAttempt?.status === 'unknown_outcome' && !synthesisRecovery.pendingGate) {
+        synthesisRecovery = {
+          ...synthesisRecovery,
+          pendingGate: this.v4SynthesisRecoveryGate(controller, synthesisRecovery),
+        }
+      }
+    }
+    const resumedSlot = phaseSlots.find(slot => (
+      slot.status === 'completed' && resumedGateFor(slot)
+    ))
+    if (resumedSlot
+        && !this.completeV4ResumedAgentSlotContinuation(controller, resumedGate)) {
+      throw new Error('LOCAL_RUN_PERSIST_FAILED')
+    }
+    const remainingKinds = Array.isArray(phaseInput.remainingKinds)
+      ? phaseInput.remainingKinds
+      : []
+    this.v4CheckpointPhase(group, controller, {
+      targetKinds,
+      phase,
+      batchId,
+      snapshotRecord,
+      snapshotHash,
+      slots: phaseSlots,
+      writerKind,
+      currentKinds: [],
+      pendingKinds: remainingKinds,
+      participantKinds,
+      receipts: receiptRecords,
+      challengeBindings,
+      synthesisRecovery,
+      coordinationPlan,
+      workReceipts,
+      workAssignments,
+    })
+    controller.currentKind = ''
+    controller.progress = []
+    controller.completedKinds = [...new Set([
+      ...(controller.completedKinds || []),
+      ...activeKinds.filter(kind => !failures.some(item => item.kind === kind)),
+    ])]
+    controller.failedKinds = [...new Set([
+      ...(controller.failedKinds || []),
+      ...failures.map(item => item.kind),
+    ])]
+    this.emitChanged()
+    return {
+      slots: phaseSlots,
+      receiptRecords,
+      phaseReceipts,
+      phasePendingMessages,
+      failures,
+      synthesisPendingMessage,
+      synthesisRecovery,
+      workReceipts,
+      ok: failures.length === 0,
+    }
+  },
+}
+
+module.exports = { v4PhaseMethods }

desktop/src/workspace/local-workspace-auto-runner-v4-results.cjs

@@ -0,0 +1,835 @@
+// Turning an Agent result into a V4 receipt and committing it; proposal batches and work settlement.
+const {
+  appendBlackboardEntry,
+  createBlackboardEntryRecord,
+  emptyCollaborationState,
+  publicCollaborationText,
+} = require('../collaboration/collaboration-records.cjs')
+const {
+  appendWorkReceipt,
+  candidateCommitBlackboardStatement,
+  candidateCommitSinkId,
+  createCollaborationReceipt,
+  hashValue,
+} = require('../collaboration/orchestration-v4-records.cjs')
+const { createReviewerFindingRecord } = require('../collaboration/outcome-records.cjs')
+const { canonicalJson } = require('../collaboration/context-pack-records.cjs')
+const { parseTaskDecision } = require('../collaboration/task-decision.cjs')
+const { terminalRunStatusForReason } = require('./local-workspace-auto-runner-support.cjs')
+
+const v4ResultMethods = {
+  recordV4ReviewerFindings({ controller, phase, kind, result, verdict, summary,
+    reviewedArtifactId = '' }) {
+    if (!this.outcomeStore || !['challenge', 'verification'].includes(phase)
+        || !['support', 'contradict'].includes(verdict) || !reviewedArtifactId) return []
+    const trace = result?.pendingMessage?.metadata?.trace || result?.message?.trace || {}
+    const runId = String(controller?.runId || '')
+    const agentRunId = String(trace.agentRunId || '')
+    if (!runId || !agentRunId) return []
+    const sourceEvidenceIds = Array.isArray(result?.outcomeRefs?.evidenceIds)
+      ? result.outcomeRefs.evidenceIds.filter(value => typeof value === 'string').slice(0, 62)
+      : []
+    try {
+      const { artifact, contentHash } = this.v4ArtifactIdentity(reviewedArtifactId)
+      const reviewEvidence = this.outcomeStore.putEvidence({
+        kind: 'review',
+        level: 'observed',
+        subject: { type: 'artifact', artifactId: reviewedArtifactId },
+        summary: publicCollaborationText(summary || '', 4000) || 'Agent review recorded.',
+        recordedBy: { kind: 'agent', runId, agentRunId, agentKind: kind },
+        refs: [
+          { type: 'artifact', artifactId: reviewedArtifactId },
+          { type: 'blob', contentRef: artifact.contentRef, contentHash },
+          ...sourceEvidenceIds.map(evidenceId => ({ type: 'evidence', evidenceId })),
+        ],
+      })
+      const input = {
+        artifactId: reviewedArtifactId,
+        relation: verdict,
+        summary: publicCollaborationText(summary || '', 4000) || 'Agent review recorded.',
+        reviewer: { kind: 'agent', runId, agentRunId, agentKind: kind },
+        evidenceIds: [reviewEvidence.evidenceId],
+      }
+      const candidate = createReviewerFindingRecord(input)
+      try { return [this.outcomeStore.getReviewerFinding(candidate.reviewerFindingId)] }
+      catch { return [this.outcomeStore.putReviewerFinding(candidate)] }
+    } catch {
+      return []
+    }
+  },
+
+  v4ReceiptForResult(result, phase, kind, slot, snapshotHash, options = {}) {
+    let raw = result?.collaboration
+    if ((!raw || typeof raw !== 'object' || Array.isArray(raw))
+        && this.naturalAgentResponses && ['proposal', 'discussion'].includes(phase)) {
+      const visibleSummary = (publicCollaborationText(
+        result?.text || result?.message?.content || '', 800,
+      ) || `Completed the ${phase} phase.`).replace(/\s+/gu, ' ').trim()
+      raw = {
+        summary: visibleSummary,
+        capabilities: ['Delivered a natural-language response'],
+        intendedWork: ['Addressed the current user task'],
+        deliverables: ['Natural-language Markdown response'],
+        dependencies: [],
+      }
+    }
+    if (!raw || typeof raw !== 'object' || Array.isArray(raw)) {
+      throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_REQUIRED')
+    }
+    const summary = publicCollaborationText(raw.summary || '', 800)
+    if (!summary) throw new Error('LOCAL_RUN_COLLABORATION_RECEIPT_INVALID')
+    const parsedStatus = publicCollaborationText(raw.status || '', 40)
+    const status = ['completed', 'accepted', 'continue', 'needs-review', 'rejected']
+      .includes(parsedStatus)
+      ? parsedStatus
+      : 'completed'
+    const verdict = publicCollaborationText(raw.verdict || '', 80)
+    const deliveryWatermark = (Number(slot.deliveryWatermark) || 0) + 1
+    const outcomeRefs = phase === 'work'
+      ? (result?.producedOutcomeRefs || {})
+      : (result?.outcomeRefs || {})
+    const artifactIds = Array.isArray(outcomeRefs.artifactIds)
+      ? outcomeRefs.artifactIds.filter(value => typeof value === 'string').slice(0, 64)
+      : []
+    const evidenceIds = Array.isArray(outcomeRefs.evidenceIds)
+      ? outcomeRefs.evidenceIds.filter(value => typeof value === 'string')
+        .slice(0, 64)
+      : []
+    const findingIds = Array.isArray(raw.findingIds)
+      ? raw.findingIds.filter(value => typeof value === 'string').slice(0, 64)
+      : []
+    let unresolved = Array.isArray(raw.unresolved)
+      ? raw.unresolved.filter(value => value && typeof value === 'object')
+        .slice(0, 32)
+        .map(value => ({
+          id: publicCollaborationText(value.id || '', 120),
+          summary: publicCollaborationText(value.summary || '', 800),
+          refs: Array.isArray(value.refs)
+            ? value.refs.filter(ref => typeof ref === 'string').slice(0, 64)
+            : [],
+        }))
+        .filter(value => value.id && value.summary)
+      : []
+    const findings = Array.isArray(raw.findings)
+      ? raw.findings.filter(value => value && typeof value === 'object')
+        .slice(0, 16)
+        .map(value => ({
+          kind: publicCollaborationText(value.kind || 'observation', 80) || 'observation',
+          summary: publicCollaborationText(value.summary || '', 800),
+          ...(value.severity ? { severity: publicCollaborationText(value.severity, 40) } : {}),
+          refs: Array.isArray(value.refs)
+            ? value.refs.filter(ref => typeof ref === 'string').slice(0, 64)
+            : [],
+        }))
+        .filter(value => value.summary)
+      : []
+    const listFieldFrom = (value, limit = 32) => Array.isArray(value)
+      ? value.filter(item => typeof item === 'string').slice(0, limit)
+        .map(item => publicCollaborationText(item, 120))
+        .filter(Boolean)
+      : []
+    const listField = (field, limit = 32) => Array.isArray(raw[field])
+      ? raw[field].filter(value => typeof value === 'string').slice(0, limit)
+      : []
+    const proposedAssignments = Array.isArray(raw.proposedAssignments)
+      ? raw.proposedAssignments.filter(value => value && typeof value === 'object').slice(0, 32)
+        .map(value => ({
+          taskId: publicCollaborationText(value.taskId || '', 120),
+          ownerKind: publicCollaborationText(value.ownerKind || '', 120),
+          role: publicCollaborationText(value.role || '', 40),
+          objective: publicCollaborationText(value.objective || '', 2000),
+          expectedOutput: publicCollaborationText(value.expectedOutput || '', 2000),
+          inputRefs: listFieldFrom(value.inputRefs, 64),
+          artifactIds: listFieldFrom(value.artifactIds, 64),
+          dependsOn: listFieldFrom(value.dependsOn, 32),
+        }))
+        .filter(value => value.taskId && value.ownerKind && value.role
+          && value.objective && value.expectedOutput)
+      : []
+    const reviewerFindings = this.recordV4ReviewerFindings({
+      controller: options.controller,
+      phase,
+      kind,
+      result,
+      verdict,
+      summary,
+      reviewedArtifactId: options.reviewedArtifactId || '',
+    })
+    for (const finding of reviewerFindings) {
+      if (finding?.reviewerFindingId && !findingIds.includes(finding.reviewerFindingId)) {
+        findingIds.push(finding.reviewerFindingId)
+      }
+      for (const evidenceId of finding?.evidenceIds || []) {
+        if (!evidenceIds.includes(evidenceId) && evidenceIds.length < 64) {
+          evidenceIds.push(evidenceId)
+        }
+      }
+    }
+    if (['challenge', 'verification'].includes(phase)
+        && verdict === 'contradict' && unresolved.length === 0) {
+      if (reviewerFindings.length === 0) {
+        throw new Error('LOCAL_RUN_V4_REVIEW_FINDING_INVALID')
+      }
+      unresolved = reviewerFindings.slice(0, 32).map(finding => ({
+        id: this.v4IssueId({ receipt: { phase } }, finding.reviewerFindingId),
+        summary: publicCollaborationText(finding.summary || summary, 800) || summary,
+        refs: [finding.reviewerFindingId],
+      }))
+    }
+    const taskDecision = raw.taskDecision ? parseTaskDecision(raw.taskDecision) : null
+    // Public receipts exclude paths; the Agent run retains its original decision.
+    if (taskDecision) {
+      taskDecision.reason = publicCollaborationText(taskDecision.reason, 1600)
+      taskDecision.deliverables = taskDecision.deliverables
+        .map(value => publicCollaborationText(value, 800))
+    }
+    const receipt = createCollaborationReceipt({
+      phase,
+      agentKind: kind,
+      slotId: slot.slotId,
+      operationId: slot.operationId,
+      status,
+      summary,
+      conclusion: '',
+      ...(taskDecision ? { taskDecision } : {}),
+      artifactIds,
+      evidenceIds,
+      findingIds,
+      unresolved,
+      findings,
+      claims: Array.isArray(raw.claims)
+        ? raw.claims.filter(value => typeof value === 'string').slice(0, 16)
+        : [],
+      refs: Array.isArray(raw.refs)
+        ? raw.refs.filter(value => typeof value === 'string').slice(0, 64)
+        : [],
+      capabilities: listField('capabilities', 32),
+      intendedWork: listField('intendedWork', 32),
+      deliverables: listField('deliverables', 32),
+      dependencies: listField('dependencies', 32),
+      proposedAssignments,
+      ...(typeof raw.finalizerKind === 'string' && raw.finalizerKind
+        ? { finalizerKind: publicCollaborationText(raw.finalizerKind, 120) } : {}),
+      ...(Array.isArray(raw.verifierKinds)
+        ? {
+            verifierKinds: listField(
+              'verifierKinds',
+              Math.max(1, (options.controller?.targetKinds?.length || 2) - 1),
+            ),
+          }
+        : {}),
+      ...(typeof raw.supportedPlanHash === 'string' && raw.supportedPlanHash
+        ? { supportedPlanHash: raw.supportedPlanHash } : {}),
+      ...(typeof raw.agreeToPlan === 'boolean' ? { agreeToPlan: raw.agreeToPlan } : {}),
+      ...(typeof raw.workItemId === 'string' && raw.workItemId
+        ? { workItemId: publicCollaborationText(raw.workItemId, 120) } : {}),
+      deliveryWatermark,
+      snapshotHash,
+    })
+    const resolvedIssueIds = Array.isArray(raw.resolvedIssueIds)
+      ? raw.resolvedIssueIds
+        .filter(value => typeof value === 'string')
+        .map(value => publicCollaborationText(value, 120))
+        .filter(Boolean)
+        .slice(0, 32)
+      : []
+    return {
+      receipt,
+      verdict,
+      resolvedIssueIds,
+    }
+  },
+
+  completeV4ResumedAgentSlotContinuation(controller, resumedGate) {
+    if (!resumedGate?.gateId) return true
+    const current = controller.continuation
+    const exactBinding = current?.resumeKind === 'agent_slot'
+      && current.agentKind === resumedGate.agentKind
+      && current.round === resumedGate.round
+      && current.phase === resumedGate.phase
+      && current.slotId === resumedGate.slotId
+      && current.operationId === resumedGate.operationId
+      && current.snapshotHash === resumedGate.snapshotHash
+    if (current?.gateId === resumedGate.gateId) {
+      if (!exactBinding || current.gateType !== resumedGate.type) return false
+      if (current.state === 'completed') return true
+      if (!['pending', 'ready', 'resuming'].includes(current.state)) return false
+      controller.waitingGateIds.delete(resumedGate.gateId)
+      controller.continuation = {
+        ...current,
+        state: 'completed',
+        updatedAt: Math.max(Number(current.updatedAt) || 0, Date.now()),
+      }
+      return true
+    }
+    return Boolean(
+      resumedGate.type === 'permission'
+      && current?.gateType === 'permission'
+      && current.state === 'completed'
+      && exactBinding,
+    ) || !this.hasRunLedger()
+  },
+
+  commitV4PhaseMessage({ group, controller, threadRootId, phase, kind, pendingMessage }) {
+    if (!pendingMessage?.content || !kind) return null
+    const agentRunId = pendingMessage.metadata?.trace?.agentRunId || ''
+    return this.commitV4AgentMessage({
+      messageId: `message-${hashValue({
+        agentKind: kind,
+        agentRunId,
+        phase,
+        runId: controller.runId,
+        taskId: controller.taskId,
+      })}`,
+      groupId: group.id,
+      agentKind: kind,
+      threadRootId,
+      content: pendingMessage.content,
+      metadata: pendingMessage.metadata || {},
+    })
+  },
+
+  async runV4ProposalBatch(group, controller, threadRootId, phaseInput) {
+    const {
+      targetKinds, writerKind, batchId, snapshotRecord, snapshotHash,
+      phaseSlots, receiptRecords, dispatches, resumedGate = null,
+      participantKinds = targetKinds,
+    } = phaseInput
+    const phaseReceipts = []
+    const phasePendingMessages = []
+    const failures = []
+    const pendingKinds = () => phaseSlots
+      .filter(slot => ['planned', 'queued', 'running', 'waiting'].includes(slot.status))
+      .map(slot => slot.agentKind)
+    const checkpoint = () => this.v4CheckpointPhase(group, controller, {
+      targetKinds,
+      phase: 'proposal',
+      batchId,
+      snapshotRecord,
+      snapshotHash,
+      slots: phaseSlots,
+      writerKind,
+      currentKinds: pendingKinds(),
+      pendingKinds: pendingKinds(),
+      participantKinds,
+      receipts: receiptRecords,
+    })
+    const failSlot = (slot, kind, error) => {
+      slot.status = 'failed'
+      slot.commitStatus = 'failed'
+      slot.finishedAt = Date.now()
+      failures.push({ kind, error })
+      if (!controller.signal.aborted) {
+        this.recordAgentFailure(group.id, kind, error, threadRootId, new Set())
+      }
+    }
+
+    for (const slot of phaseSlots) {
+      if (dispatches.some(dispatch => dispatch.kind === slot.agentKind)) slot.status = 'running'
+    }
+    const settled = await Promise.all(dispatches.map(async (dispatch) => {
+      const { kind, slot, context } = dispatch
+      try {
+        const recoveredResult = this.v4NaturalRecoveredPhaseResult(
+          group, controller, threadRootId, 'proposal', kind, slot,
+        )
+        const invocation = recoveredResult
+          ? { result: recoveredResult }
+          : await this.invokeWithUnauthorizedRecovery({
+              group,
+              kind,
+              controller,
+              activeKinds: participantKinds,
+              threadRootId,
+              context: {
+                ...context,
+                onLeaseAcquired: (leaseState) => {
+                  if (!leaseState?.agentRunId) throw new Error('LOCAL_RUN_PERSIST_FAILED')
+                  slot.agentRunId = leaseState.agentRunId
+                  slot.status = 'running'
+                  checkpoint()
+                },
+              },
+            })
+        if (controller.signal.aborted) {
+          slot.status = 'stopped'
+          slot.commitStatus = 'partial'
+          slot.finishedAt = Date.now()
+          checkpoint()
+          return
+        }
+        const result = invocation.result
+        if (!result) {
+          failSlot(slot, kind, invocation.error || new Error('LOCAL_AGENT_UNKNOWN_FAILURE'))
+          checkpoint()
+          return
+        }
+        const reportedOperationId = result?.message?.trace?.context?.operationId
+          || result?.pendingMessage?.metadata?.trace?.context?.operationId
+        if (reportedOperationId && reportedOperationId !== slot.operationId) {
+          throw new Error('LOCAL_RUN_COLLABORATION_SCOPE_INVALID')
+        }
+        const receiptRecord = this.v4ReceiptForResult(
+          result, 'proposal', kind, slot, snapshotHash, { controller },
+        )
+        if (!['completed', 'accepted'].includes(receiptRecord.receipt.status)) {
+          throw new Error('LOCAL_RUN_COLLABORATION_PROPOSAL_UNSUCCESSFUL')
+        }
+        const duplicate = receiptRecords.find(existing => (
+          existing?.receipt?.receiptId === receiptRecord.receipt.receiptId
+        ))
+        if (!duplicate) {
+          phaseReceipts.push(receiptRecord)
+          receiptRecords.push(receiptRecord)
+        }
+        const effectiveReceiptRecord = duplicate || receiptRecord
+        controller.v4ReceiptRecords = receiptRecords
+        slot.status = 'completed'
+        slot.commitStatus = 'pending'
+        slot.finishedAt = Date.now()
+        slot.receiptId = effectiveReceiptRecord.receipt.receiptId
+        slot.deliveryWatermark = effectiveReceiptRecord.receipt.deliveryWatermark
+        slot.resultHash = hashValue(effectiveReceiptRecord.receipt)
+        slot.resultRefs = {
+          artifactIds: result.outcomeRefs?.artifactIds || [],
+          evidenceIds: result.outcomeRefs?.evidenceIds || [],
+          workflowOutcomeRefs: [
+            ...(slot.resultRefs?.workflowOutcomeRefs || []),
+            {
+              receipt: effectiveReceiptRecord.receipt,
+              ...(effectiveReceiptRecord.verdict
+                ? { verdict: effectiveReceiptRecord.verdict }
+                : {}),
+              ...(effectiveReceiptRecord.resolvedIssueIds.length
+                ? { resolvedIssueIds: effectiveReceiptRecord.resolvedIssueIds }
+                : {}),
+            },
+          ],
+        }
+        const watermark = {
+          agentKind: effectiveReceiptRecord.receipt.agentKind,
+          phase: effectiveReceiptRecord.receipt.phase,
+          watermark: effectiveReceiptRecord.receipt.deliveryWatermark,
+          operationId: effectiveReceiptRecord.receipt.operationId,
+          snapshotHash: effectiveReceiptRecord.receipt.snapshotHash,
+          updatedAt: Date.now(),
+        }
+        controller.v4Watermarks = [
+          ...(controller.v4Watermarks || []).filter(existing => (
+            existing.agentKind !== watermark.agentKind || existing.phase !== watermark.phase
+          )),
+          watermark,
+        ]
+        if (result.pendingMessage) {
+          phasePendingMessages.push({ kind, pendingMessage: result.pendingMessage })
+          this.commitV4PhaseMessage({
+            group, controller, threadRootId, phase: 'proposal', kind,
+            pendingMessage: result.pendingMessage,
+          })
+          slot.commitStatus = 'committed'
+        }
+        if (resumedGate?.agentKind === slot.agentKind
+            && !this.completeV4ResumedAgentSlotContinuation(controller, resumedGate)) {
+          throw new Error('LOCAL_RUN_PERSIST_FAILED')
+        }
+        checkpoint()
+      } catch (error) {
+        failSlot(slot, kind, error)
+        checkpoint()
+      }
+    }))
+    void settled
+    checkpoint()
+    controller.currentKind = ''
+    controller.progress = []
+    controller.completedKinds = [...new Set([
+      ...(controller.completedKinds || []),
+      ...phaseSlots.filter(slot => slot.status === 'completed').map(slot => slot.agentKind),
+    ])]
+    controller.failedKinds = [...new Set([
+      ...(controller.failedKinds || []),
+      ...failures.map(item => item.kind),
+    ])]
+    this.emitChanged()
+    return {
+      slots: phaseSlots,
+      receiptRecords,
+      phaseReceipts,
+      failures,
+      phasePendingMessages,
+      synthesisPendingMessage: null,
+      ok: failures.length === 0,
+    }
+  },
+
+  v4AdmitWorkSettlement(group, controller, threadRootId, item, input) {
+    const {
+      targetKinds, writerKind, batchId, snapshotRecord, snapshotHash,
+      phaseSlots, receiptRecords, challengeBindings, coordinationPlan, workAssignments,
+      participantKinds = targetKinds,
+    } = input
+    const slot = phaseSlots.find(candidate => candidate.agentKind === item.kind)
+    const result = item.invocation?.result
+    const liveSlot = controller.orchestration?.slots?.find(candidate => (
+      candidate.slotId === slot?.slotId
+    ))
+    if (!slot || controller.signal.aborted
+        || ['completed', 'failed', 'stopped', 'committed'].includes(
+          controller.orchestration?.phase,
+        )
+        || !liveSlot || liveSlot.operationId !== slot.operationId
+        || !['planned', 'queued', 'running', 'waiting'].includes(liveSlot.status)) {
+      if (slot) {
+        slot.status = 'stopped'
+        slot.commitStatus = 'partial'
+        slot.finishedAt = Date.now()
+      }
+      return { workReceipts: input.workReceipts, failure: null }
+    }
+    if (!result) {
+      const error = item.error || item.invocation?.error
+        || new Error('LOCAL_AGENT_UNKNOWN_FAILURE')
+      slot.status = 'failed'
+      slot.commitStatus = 'failed'
+      slot.finishedAt = Date.now()
+      this.recordAgentFailure(group.id, item.kind, error, threadRootId, new Set())
+      return { workReceipts: input.workReceipts, failure: { ...item, error } }
+    }
+    try {
+      const resultTrace = result?.message?.trace || result?.pendingMessage?.metadata?.trace
+      const reportedOperationId = result?.operationId || resultTrace?.context?.operationId
+      if (reportedOperationId !== slot.operationId
+          || resultTrace?.runId !== controller.runId) {
+        throw new Error('LOCAL_RUN_COLLABORATION_SCOPE_INVALID')
+      }
+      const assignment = workAssignments?.find(candidate => candidate.ownerKind === item.kind)
+      const receiptRecord = this.v4ReceiptForResult(
+        result, 'work', item.kind, slot, snapshotHash, {
+          controller,
+          workItemId: assignment?.taskId || '',
+        },
+      )
+      if (!assignment || receiptRecord.receipt.workItemId !== assignment.taskId
+          || !this.v4ReceiptReferencesComplete(receiptRecord, '', {
+            runId: controller.runId,
+            agentRunId: resultTrace?.agentRunId,
+            agentKind: item.kind,
+            operationId: slot.operationId,
+          })) {
+        throw new Error('LOCAL_RUN_V4_WORK_RECEIPT_INVALID')
+      }
+      const exactReceipt = this.v4ExactWorkReceipt(
+        receiptRecord, assignment, slot, snapshotRecord, snapshotHash, coordinationPlan,
+      )
+      const durableReceipts = Array.isArray(controller.orchestration?.workReceipts)
+        ? controller.orchestration.workReceipts
+        : input.workReceipts
+      const appended = appendWorkReceipt(durableReceipts, exactReceipt)
+      const assignmentOrder = new Map(coordinationPlan.assignments.map((candidate, index) => (
+        [candidate.taskId, index]
+      )))
+      const workReceipts = [...appended].sort((left, right) => (
+        assignmentOrder.get(left.taskId) - assignmentOrder.get(right.taskId)
+      ))
+      const duplicate = receiptRecords.find(existing => (
+        existing?.receipt?.receiptId === receiptRecord.receipt.receiptId
+      ))
+      if (!duplicate) receiptRecords.push(receiptRecord)
+      const effectiveReceiptRecord = duplicate || receiptRecord
+      if (!duplicate) input.phaseReceipts.push(receiptRecord)
+      controller.v4ReceiptRecords = receiptRecords
+      controller.v4WorkReceipts = workReceipts
+      slot.status = 'completed'
+      slot.commitStatus = 'pending'
+      slot.finishedAt = Date.now()
+      slot.receiptId = effectiveReceiptRecord.receipt.receiptId
+      slot.deliveryWatermark = effectiveReceiptRecord.receipt.deliveryWatermark
+      slot.resultHash = hashValue(effectiveReceiptRecord.receipt)
+      slot.resultRefs = {
+        artifactIds: result.producedOutcomeRefs?.artifactIds || [],
+        evidenceIds: result.producedOutcomeRefs?.evidenceIds || [],
+        workflowOutcomeRefs: [
+          ...(slot.resultRefs?.workflowOutcomeRefs || []),
+          ...(!duplicate ? [{ receipt: effectiveReceiptRecord.receipt }] : []),
+        ],
+      }
+      const watermark = {
+        agentKind: effectiveReceiptRecord.receipt.agentKind,
+        phase: effectiveReceiptRecord.receipt.phase,
+        watermark: effectiveReceiptRecord.receipt.deliveryWatermark,
+        operationId: effectiveReceiptRecord.receipt.operationId,
+        snapshotHash: effectiveReceiptRecord.receipt.snapshotHash,
+        updatedAt: Date.now(),
+      }
+      controller.v4Watermarks = [
+        ...(controller.v4Watermarks || []).filter(existing => (
+          existing.agentKind !== watermark.agentKind || existing.phase !== watermark.phase
+        )),
+        watermark,
+      ]
+      const pendingKinds = phaseSlots.filter(candidate => (
+        ['planned', 'queued', 'running', 'waiting'].includes(candidate.status)
+      )).map(candidate => candidate.agentKind)
+      if (input.resumedGate
+          && !this.completeV4ResumedAgentSlotContinuation(controller, input.resumedGate)) {
+        throw new Error('LOCAL_RUN_PERSIST_FAILED')
+      }
+      this.v4CheckpointPhase(group, controller, {
+        targetKinds,
+        phase: 'work',
+        batchId,
+        snapshotRecord,
+        snapshotHash,
+        slots: phaseSlots,
+        writerKind,
+        currentKinds: pendingKinds,
+        pendingKinds,
+        participantKinds,
+        receipts: receiptRecords,
+        challengeBindings,
+        coordinationPlan,
+        workReceipts,
+        workAssignments,
+      })
+      return { workReceipts, failure: null }
+    } catch (error) {
+      if (error?.message === 'LOCAL_RUN_PERSIST_FAILED'
+          || String(error?.message || '').startsWith('TEST_CRASH:')) throw error
+      slot.status = 'failed'
+      slot.commitStatus = 'failed'
+      slot.finishedAt = Date.now()
+      this.recordAgentFailure(group.id, item.kind, error, threadRootId, new Set())
+      return { workReceipts: input.workReceipts, failure: { ...item, error } }
+    }
+  },
+
+  async v4CommitAcceptedCandidate({
+    group, controller, threadRootId, targetKinds, batchId, snapshotRecord, snapshotHash,
+    slots, receiptRecords, challengeBindings, synthesisBinding, synthesisRecovery,
+    coordinationPlan, workReceipts, convergence, writerKind, verificationKinds, acceptance,
+    participantKinds = targetKinds,
+  }) {
+    if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+    const synthesisRecord = this.v4LatestReceipt(receiptRecords, 'synthesis', writerKind)
+    const evidenceIds = [...(synthesisRecord?.receipt?.evidenceIds || [])]
+    const { artifact, contentHash, content } = this.v4ArtifactIdentity(
+      acceptance.candidateArtifactId,
+    )
+    if (!content || contentHash !== acceptance.candidateHash
+        || !evidenceIds.length
+        || !evidenceIds.some(evidenceId => this.v4EvidenceSupportsArtifact(
+          evidenceId, acceptance.candidateArtifactId, contentHash,
+        ))) {
+      throw new Error('LOCAL_RUN_V4_CANDIDATE_INVALID')
+    }
+    const identity = {
+      runId: controller.runId,
+      taskId: controller.taskId,
+      candidateContentHash: contentHash,
+    }
+    const previous = controller.orchestration?.candidateCommit || null
+    const collaborationAtIntent = controller.orchestration?.collaboration || emptyCollaborationState()
+    const writerOwner = this.v4BlackboardOwner(coordinationPlan, writerKind)
+    const blackboardSequence = previous?.blackboardSequence
+      || (collaborationAtIntent.entries.at(-1)?.lifecycle?.sequence || 0) + 1
+    const blackboardRecordedAt = previous?.blackboardRecordedAt ?? 0
+    const commitId = candidateCommitSinkId(identity, 'commit')
+    const blackboardEntryAtIntent = createBlackboardEntryRecord({
+      entryType: 'artifact-ref',
+      subject: `candidate-commit:${commitId}`,
+      statement: candidateCommitBlackboardStatement({
+        candidateArtifactId: acceptance.candidateArtifactId,
+        candidateContentHash: contentHash,
+      }),
+      value: contentHash,
+      owner: writerOwner,
+      audience: { roles: [], agentKinds: [...targetKinds].sort() },
+      lifecycle: {
+        state: 'active', sequence: blackboardSequence,
+        recordedAt: blackboardRecordedAt, supersedesEntryId: null,
+      },
+      provenance: {
+        runId: controller.runId, taskId: controller.taskId,
+        round: controller.currentRound, agentRunId: null,
+        artifactIds: [acceptance.candidateArtifactId], evidenceIds,
+      },
+      refs: [acceptance.candidateArtifactId, ...evidenceIds],
+    })
+    const base = {
+      runId: controller.runId,
+      taskId: controller.taskId,
+      groupId: group.id,
+      threadRootId,
+      candidateArtifactId: acceptance.candidateArtifactId,
+      candidateContentHash: contentHash,
+      candidateContentRef: { ...artifact.contentRef },
+      evidenceIds,
+      writerKind,
+      writerRole: writerOwner.role,
+      commitId,
+      messageId: candidateCommitSinkId(identity, 'message'),
+      blackboardEntryId: blackboardEntryAtIntent.entryId,
+      blackboardSequence,
+      blackboardRecordedAt,
+    }
+    if (previous) {
+      const bound = Object.fromEntries(Object.keys(base).map(key => [key, previous[key]]))
+      if (canonicalJson(bound) !== canonicalJson(base)) {
+        throw new Error('LOCAL_RUN_COMMIT_INVALID')
+      }
+    }
+    let candidateCommit = previous || {
+      ...base,
+      status: 'intent',
+      messageStatus: 'pending',
+      blackboardStatus: 'pending',
+      attempt: 1,
+      updatedAt: Date.now(),
+    }
+    const checkpoint = (phase, nextCommit, collaboration, nextSlots = slots) => {
+      candidateCommit = nextCommit
+      if (collaboration) {
+        controller.orchestration = { ...controller.orchestration, collaboration }
+      }
+      this.v4CheckpointPhase(group, controller, {
+        targetKinds,
+        phase,
+        batchId,
+        snapshotRecord,
+        snapshotHash,
+        slots: nextSlots,
+        writerKind,
+        currentKinds: [],
+        pendingKinds: [],
+        participantKinds,
+        receipts: receiptRecords,
+        challengeBindings,
+        synthesisBinding,
+        synthesisRecovery,
+        coordinationPlan,
+        workReceipts,
+        convergence,
+        candidateCommit,
+      })
+    }
+    if (!previous) checkpoint('commit', candidateCommit, controller.orchestration?.collaboration)
+
+    const messageInput = {
+      messageId: candidateCommit.messageId,
+      groupId: group.id,
+      agentKind: writerKind,
+      threadRootId,
+      content,
+      metadata: {},
+    }
+    const existingMessage = this.state().messages.find(message => (
+      message.id === candidateCommit.messageId
+    ))
+    if (candidateCommit.messageStatus === 'committed' && !existingMessage) {
+      throw new Error('LOCAL_RUN_COMMIT_INVALID')
+    }
+    if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+    if (!this.commitV4AgentMessage) throw new Error('LOCAL_RUN_COMMIT_INVALID')
+    const committedMessage = this.commitV4AgentMessage(messageInput)
+    if (committedMessage.id !== candidateCommit.messageId) {
+      throw new Error('LOCAL_RUN_COMMIT_INVALID')
+    }
+    if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+    if (candidateCommit.status === 'intent') {
+      candidateCommit = {
+        ...candidateCommit,
+        status: 'message-committed',
+        messageStatus: 'committed',
+        updatedAt: Date.now(),
+      }
+      checkpoint('commit', candidateCommit, controller.orchestration.collaboration)
+    }
+
+    let collaboration = controller.orchestration.collaboration || emptyCollaborationState()
+    const existingEntry = collaboration.entries.find(entry => (
+      entry.entryId === candidateCommit.blackboardEntryId
+    ))
+    const blackboardEntry = createBlackboardEntryRecord({
+      entryType: 'artifact-ref',
+      subject: `candidate-commit:${candidateCommit.commitId}`,
+      statement: candidateCommitBlackboardStatement(candidateCommit),
+      value: contentHash,
+      owner: { type: 'agent', agentKind: writerKind, role: candidateCommit.writerRole },
+      audience: { roles: [], agentKinds: [...targetKinds].sort() },
+      lifecycle: {
+        state: 'active',
+        sequence: candidateCommit.blackboardSequence,
+        recordedAt: candidateCommit.blackboardRecordedAt,
+        supersedesEntryId: null,
+      },
+      provenance: {
+        runId: controller.runId,
+        taskId: controller.taskId,
+        round: controller.currentRound,
+        agentRunId: null,
+        artifactIds: [acceptance.candidateArtifactId],
+        evidenceIds,
+      },
+      refs: [acceptance.candidateArtifactId, ...evidenceIds],
+    })
+    if (blackboardEntry.entryId !== candidateCommit.blackboardEntryId) {
+      throw new Error('LOCAL_RUN_COMMIT_INVALID')
+    }
+    if (existingEntry && canonicalJson(existingEntry) !== canonicalJson(blackboardEntry)) {
+      throw new Error('LOCAL_RUN_COMMIT_INVALID')
+    }
+    if (candidateCommit.blackboardStatus === 'committed' && !existingEntry) {
+      throw new Error('LOCAL_RUN_COMMIT_INVALID')
+    }
+    if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+    if (!existingEntry) {
+      const {
+        entryId: _entryId,
+        version: _version,
+        recordType: _recordType,
+        ...blackboardInput
+      } = blackboardEntry
+      collaboration = appendBlackboardEntry(collaboration, blackboardInput)
+      const appended = collaboration.entries.find(entry => entry.entryId === blackboardEntry.entryId)
+      if (!appended || canonicalJson(appended) !== canonicalJson(blackboardEntry)) {
+        throw new Error('LOCAL_RUN_COMMIT_INVALID')
+      }
+    }
+    if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+    if (candidateCommit.status === 'message-committed') {
+      candidateCommit = {
+        ...candidateCommit,
+        status: 'sinks-committed',
+        blackboardStatus: 'committed',
+        updatedAt: Date.now(),
+      }
+      checkpoint('commit', candidateCommit, collaboration)
+    }
+
+    if (controller.signal.aborted) return terminalRunStatusForReason(controller.stopReason)
+    const finalSlots = slots.map(slot => ({
+      ...slot,
+      phase: slot.agentKind === writerKind ? 'completed' : slot.phase,
+      status: ['failed', 'stopped', 'timeout', 'interrupted'].includes(slot.status)
+        ? slot.status : 'completed',
+      finishedAt: slot.finishedAt || Date.now(),
+      commitStatus: 'committed',
+    }))
+    candidateCommit = {
+      ...candidateCommit,
+      status: 'completed',
+      messageStatus: 'committed',
+      blackboardStatus: 'committed',
+      updatedAt: Date.now(),
+    }
+    checkpoint('completed', candidateCommit, collaboration, finalSlots)
+    controller.completedKinds = [...participantKinds]
+    controller.failedKinds = targetKinds.filter(kind => !participantKinds.includes(kind))
+    controller.currentKind = ''
+    controller.progress = []
+    this.emitChanged()
+    return 'completed'
+  },
+}
+
+module.exports = { v4ResultMethods }

desktop/src/workspace/local-workspace-auto-runner-v4-synthesis.cjs

@@ -0,0 +1,220 @@
+// V4 synthesis recovery: choosing, replacing and cancelling the Agent that writes the final result.
+const { createHash } = require('node:crypto')
+const {
+  coordinationRecoveryKinds,
+  createSynthesisBinding,
+  hashValue,
+  rankedSynthesisKinds,
+} = require('../collaboration/orchestration-v4-records.cjs')
+
+const v4SynthesisMethods = {
+  v4SynthesisBinding({ targetKinds, snapshotRecord, routingDecision }) {
+    const selected = new Set(targetKinds)
+    const candidates = (Array.isArray(routingDecision?.candidates)
+      ? routingDecision.candidates : [])
+      .filter(candidate => selected.has(candidate?.kind))
+      .map(candidate => ({
+        kind: candidate.kind,
+        score: candidate.score,
+        ...(candidate.evidence ? { evidence: { ...candidate.evidence } } : {}),
+      }))
+    return createSynthesisBinding({
+      snapshotContentHash: snapshotRecord?.bodyHash,
+      targetKinds,
+      candidates,
+    })
+  },
+
+  v4SynthesisRankedKinds(roleBinding) {
+    return roleBinding?.planHash
+      ? coordinationRecoveryKinds(roleBinding)
+      : rankedSynthesisKinds(roleBinding)
+  },
+
+  v4SynthesisVerificationKinds(roleBinding, writerKind) {
+    if (roleBinding?.planHash) {
+      const rankedKinds = coordinationRecoveryKinds(roleBinding)
+      const verificationKinds = roleBinding.verifierKinds.filter(kind => kind !== writerKind)
+      if (!rankedKinds.includes(writerKind) || verificationKinds.length < 1) {
+        throw new Error('LOCAL_RUN_V4_COORDINATION_PLAN_INVALID')
+      }
+      return verificationKinds
+    }
+    return rankedSynthesisKinds(roleBinding)
+      .filter(kind => kind !== writerKind)
+      .slice(0, Math.min(2, roleBinding.candidates.length - 1))
+  },
+
+  v4SynthesisAttemptOperationId(controller, kind, slotId, attempt) {
+    const digest = createHash('sha256').update(JSON.stringify([
+      controller.runId,
+      controller.taskId,
+      kind,
+      slotId,
+      'synthesis',
+      controller.currentRound || 0,
+      attempt,
+    ])).digest('hex')
+    return `operation-${digest}`
+  },
+
+  v4CreateSynthesisRecovery(controller, roleBinding, slots, options = {}) {
+    const rankedKinds = this.v4SynthesisRankedKinds(roleBinding)
+    const originalWriterKind = roleBinding?.planHash
+      ? roleBinding.finalizerKind : roleBinding.writerKind
+    const writerKind = String(options.writerKind || originalWriterKind)
+    const writerSlot = slots.find(slot => slot.agentKind === writerKind)
+    if (!writerSlot || !rankedKinds.includes(writerKind)) {
+      throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+    }
+    const attempt = Math.max(1, Number(options.attempt) || 1)
+    const operationId = this.v4SynthesisAttemptOperationId(
+      controller, writerKind, writerSlot.slotId, attempt,
+    )
+    return {
+      revision: Math.max(1, Number(options.revision) || 1),
+      originalWriterKind,
+      activeWriterKind: writerKind,
+      verificationKinds: this.v4SynthesisVerificationKinds(roleBinding, writerKind),
+      rankedKinds,
+      rankingFingerprint: roleBinding?.planHash
+        ? hashValue({ planHash: roleBinding.planHash, rankedKinds })
+        : hashValue({ selectionInputHash: roleBinding.selectionInputHash, rankedKinds }),
+      stateEpoch: Math.max(0, Number(options.stateEpoch) || 0),
+      triedWriters: rankedKinds.slice(0, rankedKinds.indexOf(writerKind) + 1),
+      attempts: [{
+        attemptId: `synthesis-attempt-${hashValue({ operationId, attempt }).slice(0, 64)}`,
+        writerKind,
+        slotId: writerSlot.slotId,
+        operationId,
+        attempt,
+        status: 'intent',
+        permission: 'read-only',
+        leaseAcquired: false,
+        sideEffectsPossible: false,
+        outcomeCertainty: 'not_started',
+        updatedAt: Date.now(),
+      }],
+    }
+  },
+
+  v4ActiveSynthesisAttempt(recovery) {
+    return [...(recovery?.attempts || [])].reverse().find(attempt => (
+      ['intent', 'leased', 'unknown_outcome'].includes(attempt.status)
+    )) || null
+  },
+
+  v4UpdateSynthesisAttempt(recovery, operationId, updates) {
+    let matched = false
+    const attempts = recovery.attempts.map(attempt => {
+      if (attempt.operationId !== operationId) return attempt
+      matched = true
+      return { ...attempt, ...updates, updatedAt: Date.now() }
+    })
+    if (!matched) throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+    return { ...recovery, attempts }
+  },
+
+  v4SynthesisRecoveryGate(controller, recovery) {
+    const attempt = this.v4ActiveSynthesisAttempt(recovery)
+    if (!attempt || attempt.status !== 'unknown_outcome') {
+      throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+    }
+    const currentIndex = recovery.rankedKinds.indexOf(recovery.activeWriterKind)
+    const fields = {
+      writerKind: attempt.writerKind,
+      slotId: attempt.slotId,
+      operationId: attempt.operationId,
+      attempt: attempt.attempt,
+      proposedReplacementKind: recovery.rankedKinds[currentIndex + 1] || '',
+      round: controller.currentRound,
+      stateEpoch: recovery.stateEpoch,
+      rankingFingerprint: recovery.rankingFingerprint,
+    }
+    return { bindingHash: hashValue(fields), ...fields }
+  },
+
+  v4CancelSynthesisRecovery(recovery) {
+    const binding = recovery?.pendingGate
+    const boundAttempt = recovery?.attempts?.find(attempt => (
+      attempt.operationId === binding?.operationId
+    ))
+    if (binding && boundAttempt?.status === 'cancelled'
+        && boundAttempt.sideEffectsPossible === true
+        && boundAttempt.outcomeCertainty === 'unknown_outcome') {
+      return recovery
+    }
+    const attempt = this.v4ActiveSynthesisAttempt(recovery)
+    if (!binding || !attempt || attempt.status !== 'unknown_outcome'
+        || attempt.writerKind !== binding.writerKind
+        || attempt.slotId !== binding.slotId
+        || attempt.operationId !== binding.operationId
+        || attempt.attempt !== binding.attempt) {
+      throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+    }
+    return this.v4UpdateSynthesisAttempt(recovery, attempt.operationId, {
+      status: 'cancelled',
+      sideEffectsPossible: true,
+      outcomeCertainty: 'unknown_outcome',
+    })
+  },
+
+  v4AppendSynthesisAttempt(controller, roleBinding, recovery, slots, writerKind) {
+    let nextRecovery = recovery
+    const activeAttempt = this.v4ActiveSynthesisAttempt(nextRecovery)
+    if (activeAttempt?.status === 'unknown_outcome') {
+      nextRecovery = this.v4UpdateSynthesisAttempt(
+        nextRecovery,
+        activeAttempt.operationId,
+        {
+          status: 'superseded',
+          sideEffectsPossible: true,
+          outcomeCertainty: 'unknown_outcome',
+        },
+      )
+    } else if (activeAttempt) {
+      throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+    }
+    const { pendingGate: _pendingGate, ...withoutGate } = nextRecovery
+    const writerSlot = slots.find(slot => slot.agentKind === writerKind)
+    const writerIndex = withoutGate.rankedKinds.indexOf(writerKind)
+    if (!writerSlot || writerIndex < 0) {
+      throw new Error('LOCAL_RUN_V4_SYNTHESIS_RECOVERY_INVALID')
+    }
+    const attempt = withoutGate.attempts.filter(item => item.writerKind === writerKind).length + 1
+    const operationId = this.v4SynthesisAttemptOperationId(
+      controller, writerKind, writerSlot.slotId, attempt,
+    )
+    return {
+      ...withoutGate,
+      revision: withoutGate.revision + 1,
+      activeWriterKind: writerKind,
+      verificationKinds: this.v4SynthesisVerificationKinds(roleBinding, writerKind),
+      triedWriters: withoutGate.rankedKinds.slice(0, writerIndex + 1),
+      attempts: [...withoutGate.attempts, {
+        attemptId: `synthesis-attempt-${hashValue({ operationId, attempt }).slice(0, 64)}`,
+        writerKind,
+        slotId: writerSlot.slotId,
+        operationId,
+        attempt,
+        status: 'intent',
+        permission: 'read-only',
+        leaseAcquired: false,
+        sideEffectsPossible: false,
+        outcomeCertainty: 'not_started',
+        updatedAt: Date.now(),
+      }],
+    }
+  },
+
+  v4ReplaceSynthesisWriter(controller, roleBinding, recovery, slots) {
+    const currentIndex = recovery.rankedKinds.indexOf(recovery.activeWriterKind)
+    const writerKind = recovery.rankedKinds[currentIndex + 1] || ''
+    if (!writerKind) return null
+    return this.v4AppendSynthesisAttempt(
+      controller, roleBinding, recovery, slots, writerKind,
+    )
+  },
+}
+
+module.exports = { v4SynthesisMethods }

desktop/src/workspace/local-workspace-context.cjs

@@ -21,11 +21,25 @@ const {
   normalizeSkillHint,
   skillHintsPrompt,
 } = require('./local-workspace-inputs.cjs')
+const { MAX_MESSAGE_TEXT_CHARS } = require('./local-workspace-contracts.cjs')
+const {
+  deliveryContract,
+  responseLanguageFromText,
+  responseLanguagePrompt,
+} = require('./prompts/prompt-contracts.cjs')
+const { naturalBrief } = require('./prompts/natural-discussion-prompt.cjs')
 
 const CONTINUATION_STABLE_CONTEXT_TEXT_LIMIT = 1400
 const CONTINUATION_RECENT_CONTEXT_TEXT_LIMIT = 4600
-const CURRENT_TASK_TEXT_LIMIT = 6000
+// User messages are already capped at MAX_MESSAGE_TEXT_CHARS when they are
+// submitted. The current task must reach the Agent in full: truncating it, or
+// failing the turn when it is longer than a fixed budget, made long pasted
+// documents unusable. The extra room covers the attached-file note.
+const CURRENT_TASK_TEXT_LIMIT = MAX_MESSAGE_TEXT_CHARS + 2000
 const V4_SNAPSHOT_HISTORY_LIMIT = 16
+const NATURAL_TRANSCRIPT_TEXT_LIMIT = 48000
+const NATURAL_TRANSCRIPT_ENTRY_LIMIT = 20000
+const NATURAL_TRANSCRIPT_OMISSION_NOTICE = 'This transcript is partial: some turns or parts of turns were omitted to fit the context budget. Omission is not evidence of agreement or completion. Ask peers to restate any missing evidence needed to judge the original task.'
 const V4_SNAPSHOT_FIELDS = [
   'group', 'history', 'messageId', 'phase', 'snapshotHash',
   'skillHintsByKind', 'targetKinds', 'taskId', 'taskText', 'version', 'writerKind',
@@ -53,6 +67,50 @@ function unlimitedReviewContract(enabled = false) {
   return enabled ? UNLIMITED_REVIEW_CONTRACT : ''
 }
 
+function packNaturalDiscussionTranscript(entries, options = {}) {
+  const prefix = String(options.prefix || '')
+  const latestByKind = new Map(entries.map((entry, index) => [entry.agentKind, index]))
+  let excerpted = entries.some(entry => entry.partial === true)
+  const completeBodies = new Map()
+  const packed = packContextEntries(entries.map((entry, index) => {
+    const heading = `Round ${Number(entry.round) || 0} - @${entry.agentKind}\n`
+    let text = String(entry.text || '').trim()
+    const limit = NATURAL_TRANSCRIPT_ENTRY_LIMIT - heading.length
+    if (text.length > limit) {
+      const marker = '\n[Middle of this turn omitted]\n'
+      const edge = Math.floor((limit - marker.length) / 2)
+      // Preserve closing requests as well as opening findings, without inventing a summary.
+      const head = text.slice(0, edge).replace(/[\uD800-\uDBFF]$/u, '')
+      const tail = text.slice(-edge).replace(/^[\uDC00-\uDFFF]/u, '')
+      text = `${head}${marker}${tail}`
+      excerpted = true
+    } else if (!entry.partial) {
+      completeBodies.set(entry.id, `${heading}${text}`)
+    }
+    return {
+      id: entry.id,
+      text: `${heading}${text}`,
+      priority: index === entries.length - 1 ? 3
+        : latestByKind.get(entry.agentKind) === index ? 2 : 1,
+    }
+  }), {
+    budget: NATURAL_TRANSCRIPT_TEXT_LIMIT - NATURAL_TRANSCRIPT_OMISSION_NOTICE.length - prefix.length - 2,
+    entryLimit: NATURAL_TRANSCRIPT_ENTRY_LIMIT,
+    maxEntries: 100,
+  })
+  const text = [
+    prefix,
+    excerpted || packed.omittedCount ? NATURAL_TRANSCRIPT_OMISSION_NOTICE : '',
+    packed.text,
+  ].filter(Boolean).join('\n')
+  return options.withEntries ? {
+    text,
+    completeMessageIds: packed.sourceEntries.filter(entry => (
+      completeBodies.get(entry.id) === entry.text
+    )).map(entry => entry.id),
+  } : text
+}
+
 function v4Snapshot({
   state, group, taskId, targetKinds, message, skillHintsByKind,
   phase = 'proposal', writerKind = '',
@@ -219,7 +277,7 @@ function validateV4SnapshotBody({
 
 function v4Prompt({
   group, kind, phase, snapshot, role = 'proposal', receipt = true, skillHints = null,
-  naturalResponse = false,
+  knowledgeBaseHints = [], naturalResponse = false,
 }) {
   const history = Array.isArray(snapshot?.history) && snapshot.history.length
     ? snapshot.history.map(item => [
@@ -229,15 +287,7 @@ function v4Prompt({
     : '(none)'
   const selectedAgentCount = Array.isArray(snapshot?.targetKinds) ? snapshot.targetKinds.length : 0
   const requiredVerifierCount = Math.max(0, selectedAgentCount - 1)
-  const phaseInstruction = naturalResponse
-    ? {
-        proposal: 'Work independently on the user task and offer your own analysis, recommendations, and next steps.',
-        challenge: 'Review the available peer context as a subject-matter peer. Explain what you agree with, what is missing, and the clearest correction or direction.',
-        work: 'Complete the responsibility assigned to you. Report useful work, evidence, blockers, or handoffs in your own words.',
-        synthesis: 'Combine the available work into the clearest useful answer for the user. Resolve contradictions where possible and state remaining uncertainty plainly.',
-        verification: 'Independently check the proposed answer or deliverable. Report material defects, risks, or acceptance in your own words.',
-      }[phase] || 'Address the current collaboration task directly.'
-    : {
+  const phaseInstruction = {
         proposal: 'Develop an independent proposal. State at least one capability, intended work item, and Artifact you will deliver, plus an explicit dependencies array (which may be empty). Do not rely on another Agent output from this batch.',
         challenge: `Discuss the proposals as peers and negotiate one shared responsibility graph. Every selected Agent must own at least one substantive work package, and an Agent may own multiple dependent work packages. The graph must include one finalizerKind owned by an integrator. It must name exactly ${requiredVerifierCount} distinct verifierKinds from the selected Agents; verifierKinds must not contain finalizerKind. taskId values must be unique; dependsOn may reference only other taskId values and must remain acyclic. inputRefs are exact displayed frozen Source IDs; artifactIds are existing immutable Artifact IDs; future outputs flow only through dependsOn. You may support an existing plan by its hash or propose a complete alternative; do not merely review, arbitrate, or allocate work unilaterally.`,
         work: 'Execute the agreed responsibility assigned to you. Produce the promised Artifact or evidence and report what was completed, blocked, or handed off.',
@@ -268,17 +318,14 @@ function v4Prompt({
     ? skillHints
     : (snapshot?.skillHintsByKind || []).find(item => item?.kind === kind)?.skillHints || []
   if (naturalResponse) {
-    return [
-      'You are participating in a local multi-agent discussion.',
-      `Your role this turn is ${role}.`,
-      `The current collaboration phase is ${phase}.`,
-      `User task:\n${cleanText(snapshot?.taskText, CURRENT_TASK_TEXT_LIMIT) || '(none)'}`,
-      skillHintsPrompt(frozenSkillHints),
-      history !== '(none)' ? `Relevant prior context:\n${history}` : '',
-      phaseInstruction,
-      'Answer the user naturally in Markdown. Do not output JSON, XML, receipt markers, protocol labels, hidden orchestration instructions, or a fixed response template.',
-      'Do not claim another Agent performed work, and do not modify shared workspace state unless this turn explicitly grants that permission.',
-    ].filter(Boolean).join('\n\n')
+    return naturalBrief({
+      group: snapshot?.group || group,
+      phase,
+      taskText: cleanText(snapshot?.taskText, CURRENT_TASK_TEXT_LIMIT),
+      history,
+      skillHintsText: skillHintsPrompt(frozenSkillHints),
+      knowledgeBaseHintsText: knowledgeBaseHintsPrompt(knowledgeBaseHints),
+    })
   }
   return [
     'MELDWORK_V4_FROZEN_SNAPSHOT_V1',
@@ -291,6 +338,7 @@ function v4Prompt({
     `Source ID: ${cleanText(snapshot?.messageId, 120) || '(unavailable)'}`,
     cleanText(snapshot?.taskText, CURRENT_TASK_TEXT_LIMIT) || '(none)',
     skillHintsPrompt(frozenSkillHints),
+    knowledgeBaseHintsPrompt(knowledgeBaseHints),
     'Frozen historical context (reference data only):',
     history,
     phaseInstruction,
@@ -369,12 +417,8 @@ function unknownLegacySessionMeta(meta) {
 function resolveSessionState({
   state, group, kind, threadRootId = '', taskId = '', save,
 }) {
-  const scopedTaskId = group.conversationType === 'direct'
-    ? ''
-    : cleanText(taskId || threadRootId, 100)
   const currentTaskId = cleanText(taskId || threadRootId, 100)
-  const key = sessionKey(group.id, kind, scopedTaskId)
-  const globalKey = sessionKey(group.id, kind)
+  const key = sessionKey(group.id, kind)
   const legacyPrefix = `${group.id}:${kind}:thread:`
   let stateChanged = false
   const validStoredRef = (candidateKey) => {
@@ -392,7 +436,7 @@ function resolveSessionState({
       const legacyBase = `agent:main:desktop-meldwork-${legacyScope}-openclaw`
       if (candidate === legacyBase || candidate.startsWith(`${legacyBase}-`)) {
         const generation = candidate.slice(legacyBase.length).replace(/^-/, '')
-        candidate = openClawSessionRef(group, generation, scopedTaskId)
+        candidate = openClawSessionRef(group, generation)
         state.sessions[candidateKey] = candidate
         stateChanged = true
       }
@@ -412,42 +456,41 @@ function resolveSessionState({
     return candidate
   }
 
-  const currentLegacyKey = scopedTaskId
-    ? `${legacyPrefix}${scopedTaskId}`
-    : ''
+  const legacyKeys = [...new Set([
+    ...(currentTaskId ? [sessionKey(group.id, kind, currentTaskId), `${legacyPrefix}${currentTaskId}`] : []),
+    ...[...(state.messages || [])].reverse()
+      .filter(message => message.groupId === group.id && message.role === 'user')
+      .flatMap(message => [sessionKey(group.id, kind, message.id), `${legacyPrefix}${message.id}`]),
+    ...Object.keys(state.sessions).reverse().filter(candidate => (
+      candidate.startsWith(legacyPrefix)
+      || (candidate.startsWith(`${group.id}:task:`) && candidate.endsWith(`:${kind}`))
+    )),
+  ])].filter(candidate => candidate !== key)
   let migration = ''
-  let sessionReset = false
   let created = false
   let stored = validStoredRef(key)
-  if (!stored && currentLegacyKey) {
-    const legacyRef = validStoredRef(currentLegacyKey)
-    if (legacyRef) {
+  for (const legacyKey of legacyKeys) {
+    const legacyRef = validStoredRef(legacyKey)
+    if (!stored && legacyRef) {
       state.sessions[key] = legacyRef
-      state.sessionMeta[key] = completeSessionMeta(
-        state.sessionMeta[currentLegacyKey], 'task', scopedTaskId,
-      )
-      delete state.sessions[currentLegacyKey]
-      delete state.sessionMeta[currentLegacyKey]
+      const legacyMeta = normalizeSessionMeta(state.sessionMeta[legacyKey])
+      state.sessionMeta[key] = ['task', 'conversation'].includes(legacyMeta.sessionScope)
+        ? normalizeSessionMeta({ ...legacyMeta, sessionScope: 'conversation' })
+        : unknownLegacySessionMeta(legacyMeta)
       stored = legacyRef
       migration = 'migrated'
       stateChanged = true
     }
-  }
-  if (!stored && scopedTaskId && globalKey !== key) {
-    // A conversation-scoped legacy Session has no trustworthy task provenance.
-    // Reusing it for a new root can make the Agent continue an unrelated topic.
-    validStoredRef(globalKey)
-    if (Object.hasOwn(state.sessions, globalKey) || Object.hasOwn(state.sessionMeta, globalKey)) {
-      delete state.sessions[globalKey]
-      delete state.sessionMeta[globalKey]
-      sessionReset = true
+    if (Object.hasOwn(state.sessions, legacyKey) || Object.hasOwn(state.sessionMeta, legacyKey)) {
+      delete state.sessions[legacyKey]
+      delete state.sessionMeta[legacyKey]
       stateChanged = true
     }
   }
   if (!stored && kind === 'openclaw') {
-    state.sessions[key] = openClawSessionRef(group, '', scopedTaskId)
+    state.sessions[key] = openClawSessionRef(group)
     state.sessionMeta[key] = completeSessionMeta(
-      state.sessionMeta[key], scopedTaskId ? 'task' : 'conversation', currentTaskId,
+      state.sessionMeta[key], 'conversation', currentTaskId,
     )
     stateChanged = true
     stored = state.sessions[key]
@@ -462,7 +505,7 @@ function resolveSessionState({
   if (stateChanged) save()
   const provenance = created || !stored
     ? {
-        scope: scopedTaskId ? 'task' : 'conversation',
+        scope: 'conversation',
         reuse: false,
         origin: 'created',
         originTaskId: currentTaskId || null,
@@ -486,7 +529,7 @@ function resolveSessionState({
           inheritedTaskIds: [...(meta.inheritedTaskIds || [])],
           completeness: meta.provenanceCompleteness || 'partial',
         }
-  return { key, sessionRef: stored, sessionMeta: meta, provenance, sessionReset }
+  return { key, sessionRef: stored, sessionMeta: meta, provenance, sessionReset: false }
 }
 
 function resolveSessionRef(options) {
@@ -531,38 +574,6 @@ function scopedConversationMessages(
   })
 }
 
-function responseLanguageFromText(text) {
-  const source = String(text || '').normalize('NFKC')
-  const han = (source.match(/[\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff]/g) || []).length
-  const kana = (source.match(/[\u3040-\u30ff]/g) || []).length
-  const hangul = (source.match(/[\uac00-\ud7af]/g) || []).length
-  const cyrillic = (source.match(/[\u0400-\u052f]/g) || []).length
-  const arabic = (source.match(/[\u0600-\u06ff]/g) || []).length
-  const latin = (source.match(/[A-Za-z]/g) || []).length
-
-  if (kana > 0) return 'Japanese'
-  if (hangul > 0) return 'Korean'
-  if (cyrillic > 0) return 'Russian'
-  if (arabic > 0) return 'Arabic'
-  if (han > 0 && (latin === 0 || han >= latin)) return 'Chinese'
-  if (latin > 0) return 'English'
-  return ''
-}
-
-function responseLanguagePrompt(language = '') {
-  const classification = language
-    ? `The latest user message is written primarily in ${language}.`
-    : 'The latest user message language was not confidently classified; infer it from the message body only.'
-  return [
-    'Response language contract:',
-    'Always answer entirely in the language of the latest user message unless that message explicitly requests another language.',
-    classification,
-    'If the latest user message is English, answer entirely in English. If it is Chinese, answer entirely in Chinese.',
-    'Do not infer response language from the app UI locale, operating-system locale, Agent name, attachment filenames, or earlier conversation turns.',
-    'Do not translate or switch languages unless the latest user message explicitly asks for it.',
-  ].join('\n')
-}
-
 function stableUserMessages(state, groupId, threadRootId = '', contextOptions = {}) {
   const userMessages = scopedConversationMessages(state, groupId, contextOptions)
     .filter(message => message.role === 'user')
@@ -593,12 +604,14 @@ function stableUserInstructions(state, groupId, threadRootId = '', contextOption
 function recentTranscriptEntries(state, groupId, afterAgentKind = '', contextOptions = {}) {
   const messages = scopedConversationMessages(state, groupId, contextOptions)
   let afterIndex = -1
+  const afterMessageId = cleanText(contextOptions.afterMessageId, 100)
+  if (afterMessageId) afterIndex = messages.findIndex(message => message.id === afterMessageId)
   if (afterAgentKind) {
     for (let index = messages.length - 1; index >= 0; index -= 1) {
       const message = messages[index]
       if (message.agentKind === afterAgentKind
           && (message.role === 'agent' || isTracedAgentTerminalMessage(message))) {
-        afterIndex = index
+        afterIndex = Math.max(afterIndex, index)
         break
       }
     }
@@ -645,11 +658,13 @@ function packedPromptContext({
   excludeResponseVersionRootId = '',
   focusUserMessageId = '',
   omitAgentThreadRootId = '',
+  afterMessageId = '',
 }) {
   const contextOptions = {
     beforeMessageId,
     excludeResponseVersionRootId,
     focusUserMessageId,
+    afterMessageId: cleanText(afterMessageId, 100),
   }
   const scopedMessages = scopedConversationMessages(state, groupId, contextOptions)
   const focusId = cleanText(focusUserMessageId, 100)
@@ -741,6 +756,7 @@ function packedPromptContext({
     currentTaskText,
     latestUserLanguage: responseLanguageFromText(latestUserMessage?.content),
     latestUserMessageId: latestUserMessage?.id || '',
+    latestMessageId: scopedMessages.at(-1)?.id || '',
     ...bootstrapSelection,
     continuationSourceMessageIds: continuationSelection.sourceMessageIds,
     continuationSourceEntries: continuationSelection.sourceEntries,
@@ -755,6 +771,57 @@ function packedPromptContext({
   }
 }
 
+// The write-access and deliverable contract applies to every turn, including
+// direct chats that rely on the native Agent session for history. Without it
+// Agents do not know where to put files, so generated outputs are never
+// captured and read-only chats claim media that does not exist.
+// Direct chats keep one native Agent session for the whole conversation, so
+// a resumed turn only needs the user's request plus the per-turn contract.
+// When the session is new or was rebuilt, the earlier conversation is packed
+// before the request so the Agent never loses either of them.
+function directPromptFor({
+  group,
+  label = '',
+  packed,
+  skillHints = [],
+  knowledgeBaseHints = [],
+  includeHistory = false,
+}) {
+  const currentTaskText = packed?.currentTaskText || '(none)'
+  const latestId = String(packed?.latestUserMessageId || '')
+  const hasPriorContext = includeHistory
+    && (packed?.continuationSourceMessageIds || []).some(id => id && id !== latestId)
+  // The full contract is delivered once per native session; resumed turns
+  // carry a one-line reminder so long chats do not re-send the same block.
+  const notes = includeHistory
+    ? [
+        label ? `You are working in this Meldwork chat as ${label}.` : '',
+        deliveryContract(group),
+        responseLanguagePrompt(packed?.latestUserLanguage),
+      ]
+    : [
+        group?.allowWrite
+          ? 'Save deliverables into .meldwork-output/ in the working directory; do not claim a file exists until it was written.'
+          : 'This chat is read-only; do not claim that a file or media was generated or delivered.',
+        'Answer in the language of the current request unless it asks for another language.',
+      ]
+  notes.push(skillHintsPrompt(skillHints), knowledgeBaseHintsPrompt(knowledgeBaseHints))
+  return [
+    ...(hasPriorContext
+      ? [
+          'Earlier conversation in this chat (reference only; verify it before relying on it):',
+          packed.continuationText,
+          '',
+          'Current request:',
+        ]
+      : []),
+    currentTaskText,
+    '',
+    'Meldwork conversation notes:',
+    ...notes.filter(Boolean),
+  ].join('\n')
+}
+
 function promptFor({
   group,
   kind,
@@ -784,18 +851,7 @@ function promptFor({
     packed?.currentTaskText || '(none)',
     'Treat older group messages and conclusions as reference only. Do not continue an older task unless the current user task explicitly asks you to do so.',
   ].join('\n')
-  const mediaDelivery = group.allowWrite
-      ? [
-          'Workspace access: You may create and edit files in the conversation working directory. When the user requests a deliverable, execute the work instead of returning only a plan.',
-          'Deliverable capture contract: Save or copy each final document, code file, diff, structured-data file, media file, or bundle into .meldwork-output/ in the conversation working directory.',
-          'Only files created or changed during this run are captured as durable outputs. Do not claim a deliverable exists until the real file has been written there.',
-          'When the user asks for an image, audio, or video, do not claim it was generated unless a real file exists.',
-          'Supported chat attachments include common images, media, documents, source code, configuration text, and archives. Other recognized outputs remain durable Artifacts.',
-        ].join('\n')
-    : [
-        'This conversation is read-only. Do not claim that a media file was generated or delivered because no writable output can be attached.',
-        'If the user requests generated media, explain that workspace write access must be enabled before a real file can be delivered.',
-      ].join('\n')
+  const mediaDelivery = deliveryContract(group)
   if (promptMode === 'continuation') {
     const continuationMediaDelivery = group.allowWrite
       ? 'Workspace write access remains enabled for this task; execute requested deliverables and capture them as required.'
@@ -845,6 +901,9 @@ function promptFor({
 }
 
 module.exports = {
+  deliveryContract,
+  directPromptFor,
+  packNaturalDiscussionTranscript,
   clearSessionState,
   openClawSessionRef,
   packedPromptContext,

desktop/src/workspace/local-workspace-contracts.cjs

@@ -34,11 +34,6 @@ function defaultAgentLabel(kind) {
   return AGENT_LABELS[kind] || String(kind || 'Agent')
 }
 
-function credentialFailure(error) {
-  return /api[ _-]?key|credential|auth(?:entication|orization)?|login|log in|unauthorized|forbidden|401|403|令牌|凭据|登录|认证/i
-    .test(String(error?.message || error || ''))
-}
-
 function cleanText(value, limit = MAX_MESSAGE_TEXT_CHARS) {
   return String(value || '').trim().slice(0, limit)
 }
@@ -61,7 +56,6 @@ module.exports = {
   STABLE_USER_TURNS_PER_EDGE,
   cleanInline,
   cleanText,
-  credentialFailure,
   defaultAgentLabel,
   isSupportedAgentKind,
 }

desktop/src/workspace/local-workspace-conversations.cjs

@@ -16,6 +16,7 @@ const {
 
 class LocalWorkspaceConversations {
   constructor(options) {
+    this.reserveTurnSeq = typeof options.reserveTurnSeq === 'function' ? options.reserveTurnSeq : null
     this.state = options.state
     this.detectedAgents = options.detectedAgents
     this.save = options.save
@@ -139,6 +140,9 @@ class LocalWorkspaceConversations {
   removeAgent(groupId, kind) {
     const group = this.getGroup(groupId)
     if (group.conversationType === 'direct' || !group.agentKinds.includes(kind)) return false
+    // A group without members cannot be loaded again, which would discard
+    // its whole history. Membership edits must always keep one Agent.
+    if (group.agentKinds.length <= 1) return false
     return this.commit(() => {
       group.agentKinds = group.agentKinds.filter(agentKind => agentKind !== kind)
       group.updatedAt = this.now()
@@ -236,6 +240,16 @@ class LocalWorkspaceConversations {
       createdAt: this.now(),
     }
     if (threadRootId) message.threadRootId = threadRootId
+    // Group topics order by dispatch turn. Replies and failure notices carry
+    // the number reserved at dispatch; any other topic message (system
+    // notices, follow-up input) takes the next number when it is written.
+    const requestedTurnSeq = Number(metadata.turnSeq)
+    const turnSeq = Number.isSafeInteger(requestedTurnSeq) && requestedTurnSeq > 0
+      ? requestedTurnSeq
+      : (threadRootId && group.conversationType !== 'direct' && this.reserveTurnSeq
+        ? this.reserveTurnSeq(groupId, threadRootId)
+        : null)
+    if (turnSeq) message.turnSeq = turnSeq
     if (systemKey) {
       message.system = {
         key: systemKey,

desktop/src/workspace/local-workspace-inputs.cjs

@@ -40,7 +40,6 @@ module.exports = {
   cleanProgressSteps: runtimeContracts.cleanProgressSteps,
   cleanRunMaxRounds: runtimeContracts.cleanRunMaxRounds,
   cleanText: contracts.cleanText,
-  credentialFailure: contracts.credentialFailure,
   defaultAgentLabel: contracts.defaultAgentLabel,
   emptyState: messageRecords.emptyState,
   isTracedAgentTerminalMessage: runtimeContracts.isTracedAgentTerminalMessage,

desktop/src/workspace/local-workspace-message-records.cjs

@@ -171,12 +171,17 @@ function knowledgeBaseHintsPrompt(hints) {
   ].join('\n')
 }
 
-function normalizeLoadedGroup(input) {
+function normalizeLoadedGroup(input, options = {}) {
   if (!input || typeof input !== 'object' || Array.isArray(input)) return null
   const id = cleanText(input.id, 100)
-  const agentKinds = [...new Set((Array.isArray(input.agentKinds) ? input.agentKinds : [])
+  const normalizeKinds = values => [...new Set((Array.isArray(values) ? values : [])
     .map(kind => cleanInline(kind, 40))
     .filter(isSupportedAgentKind))]
+  let agentKinds = normalizeKinds(input.agentKinds)
+  // Earlier builds could persistently remove every member of a group after
+  // repeated failures. Dropping such a group would also drop its messages,
+  // so rebuild the membership from the Agents that replied in it instead.
+  if (!agentKinds.length) agentKinds = normalizeKinds(options.recoveredAgentKinds)
   if (!id || !agentKinds.length) return null
 
   const group = {
@@ -210,8 +215,10 @@ function normalizeLoadedMessage(input) {
   const agentKind = isSupportedAgentKind(requestedAgentKind) ? requestedAgentKind : ''
   if (!id || !groupId || !role || (role === 'agent' && !agentKind)) return null
   const systemKey = role === 'system' ? cleanInline(input.system?.key, 100) : ''
+  const turnSeq = Number(input.turnSeq)
 
   const message = {
+    ...(Number.isSafeInteger(turnSeq) && turnSeq > 0 ? { turnSeq } : {}),
     id,
     groupId,
     role,

desktop/src/workspace/local-workspace-message-submission.cjs

@@ -11,15 +11,16 @@ const {
   attachmentType,
   cleanInline,
   cleanText,
+  knowledgeBaseHintsPrompt,
   normalizeAttachmentMetadata,
   normalizeAutoRounds,
   normalizeKnowledgeBaseHint,
   normalizeSkillHint,
   normalizeTargetKinds,
+  skillHintsPrompt,
   terminalRunStatusForReason,
 } = require('./local-workspace-inputs.cjs')
 const { MAX_RUN_AGENT_ATTEMPTS } = require('../runs/failure-policy.cjs')
-const { mediaGenerationRequest } = require('../media/media-generation-request.cjs')
 const { createTaskGraph } = require('../collaboration/task-graph-records.cjs')
 const { assertLocalSkillExecution } = require('../skills/local-skill-contract.cjs')
 const {
@@ -38,12 +39,14 @@ const {
 } = require('../collaboration/orchestration-v4-records.cjs')
 const {
   restoreV4SnapshotSkills,
-  v4Prompt,
   v4Snapshot,
   v4SnapshotBodyHash,
   v4SnapshotSkillHints,
   validateV4SnapshotBody,
 } = require('./local-workspace-context.cjs')
+const { manualReplyPrompt } = require('./prompts/manual-reply-prompt.cjs')
+
+const DISCUSSION_STYLES = new Set(['sequential', 'agent-led'])
 
 function hardBudgetFailure(error) {
   return error?.code === 'LOCAL_BUDGET_EXHAUSTED'
@@ -107,6 +110,9 @@ class LocalWorkspaceMessageSubmission {
     this.bindRunTask = options.bindRunTask
     this.releasePreparation = options.releasePreparation
     this.addMessage = options.addMessage
+    this.reserveTurnSeqs = typeof options.reserveTurnSeqs === 'function'
+      ? options.reserveTurnSeqs
+      : () => []
     this.rollbackAddedMessage = options.rollbackAddedMessage
     this.startAutoRunner = options.startAutoRunner
     this.beginRun = options.beginRun
@@ -254,7 +260,6 @@ class LocalWorkspaceMessageSubmission {
     ]))
     return {
       text,
-      mediaRequest: mediaGenerationRequest(text),
       attachments,
       skillHintsByKind,
       skillHints: targetKinds.flatMap(kind => publicSkillHintsByKind.get(kind) || []),
@@ -280,7 +285,7 @@ class LocalWorkspaceMessageSubmission {
       minContextChars: Math.max(1, cleanText(input.text).length),
     })
     const targetKinds = routingDecision.selectedKinds
-    if (mode === 'auto' && targetKinds.length < 2) throw new Error('LOCAL_AUTO_AGENT_COUNT')
+    if (mode === 'auto' && !targetKinds.length) throw new Error('LOCAL_AUTO_AGENT_COUNT')
     if (input.mentionedAgentKinds != null && !Array.isArray(input.mentionedAgentKinds)) {
       throw new Error('LOCAL_MESSAGE_TARGET_REQUIRED')
     }
@@ -337,15 +342,28 @@ class LocalWorkspaceMessageSubmission {
     if (regenerateMessageId && (mode !== 'manual' || targetKinds.length !== 1)) {
       throw new Error('LOCAL_MESSAGE_REGENERATION_INVALID')
     }
+    const v4 = (mode === 'manual' || mode === 'auto')
+      && group.conversationType !== 'direct'
+      && (targetKinds.length > 1 || (mode === 'auto' && targetKinds.length === 1))
+      && !regenerateMessageId
+      && input.protocol !== 'legacy'
+      && this.v4Requested(input)
+    const requestedDiscussionStyle = input.discussionStyle == null
+      ? ''
+      : cleanInline(input.discussionStyle, 40)
+    if (requestedDiscussionStyle && !DISCUSSION_STYLES.has(requestedDiscussionStyle)) {
+      throw new Error('LOCAL_DISCUSSION_STYLE_INVALID')
+    }
+    if (requestedDiscussionStyle && (mode !== 'auto' || !v4)) {
+      throw new Error('LOCAL_DISCUSSION_STYLE_UNSUPPORTED')
+    }
     return {
       mode,
       targetKinds,
-      v4: (mode === 'manual' || mode === 'auto')
-        && group.conversationType !== 'direct'
-        && targetKinds.length > 1
-        && !regenerateMessageId
-        && input.protocol !== 'legacy'
-        && this.v4Requested(input),
+      v4,
+      discussionStyle: mode === 'auto' && v4
+        ? (requestedDiscussionStyle || 'sequential')
+        : '',
       unlimitedRounds,
       maxRounds,
       requestedThreadRootId,
@@ -532,6 +550,7 @@ class LocalWorkspaceMessageSubmission {
         ...(pendingMessage.metadata?.responseVersionRootId
           ? { responseVersionRootId: pendingMessage.metadata.responseVersionRootId } : {}),
         ...(trace ? { trace } : {}),
+        ...(pendingMessage.metadata?.turnSeq ? { turnSeq: pendingMessage.metadata.turnSeq } : {}),
       },
     }))
     const contentRef = this.contentBlobStore.put(canonicalJson(body), {
@@ -866,14 +885,22 @@ class LocalWorkspaceMessageSubmission {
     skillHintsByKind = new Map(), resumedGate = null }) {
     const snapshotSourceMessageIds = [snapshot.messageId, ...snapshot.history.map(item => item.id)]
     const frozenSnapshotHash = controller.orchestration.snapshotHash
-    const prompts = new Map(activeKinds.map(kind => [kind, v4Prompt({
+    const prompts = new Map(activeKinds.map(kind => [kind, manualReplyPrompt({
       group: { ...group, ...snapshot.group },
       kind,
-      phase: 'proposal',
-      role: kind === writerKind ? 'writer-proposal' : 'proposal',
-      snapshot,
-      skillHints: skillHintsByKind.get(kind) || [],
+      targetKinds,
+      taskText: snapshot.taskText,
+      history: snapshot.history,
+      skillHintsText: skillHintsPrompt(skillHintsByKind.get(kind) || []),
+      knowledgeBaseHintsText: knowledgeBaseHintsPrompt(
+        prepared?.knowledgeBaseHintsByKind?.get(kind) || [],
+      ),
+      writer: kind === writerKind,
     })]))
+    // Concurrent replies keep slot order in the timeline whatever order they
+    // finish in; failure notices take their slot's number too.
+    const reservedTurns = this.reserveTurnSeqs(group.id, threadRootId, activeKinds.length)
+    const turnSeqByKind = new Map(activeKinds.map((kind, index) => [kind, reservedTurns[index]]))
     const pending = await Promise.all(activeKinds.map(async (kind) => {
       const index = targetKinds.indexOf(kind)
       const slot = slots[index]
@@ -910,6 +937,7 @@ class LocalWorkspaceMessageSubmission {
             slotId: slot.slotId,
             snapshotHash: frozenSnapshotHash,
             sessionPolicy: 'frozen',
+            ...(turnSeqByKind.get(kind) ? { turnSeq: turnSeqByKind.get(kind) } : {}),
             promptOverride: prompts.get(kind),
             contextPackId: controller.contextPackId,
             snapshotSourceMessageIds,
@@ -952,6 +980,7 @@ class LocalWorkspaceMessageSubmission {
             ...(resumedGate?.agentKind === kind ? { resumedGate } : {}),
             skillHints: skillHintsByKind.get(kind) || [],
             knowledgeBaseHints: prepared?.knowledgeBaseHintsByKind?.get(kind) || [],
+            attachmentSnapshots: prepared?.attachments || [],
           },
         })
         if (controller.signal.aborted) return { kind, index, stopped: true }
@@ -1428,6 +1457,7 @@ class LocalWorkspaceMessageSubmission {
       regenerateMessageId,
       routingDecision,
       taskGraph,
+      discussionStyle,
     } = this.validateInput(group, input)
     const regeneration = this.resolveRegeneration(group, regenerateMessageId, targetKinds)
     const reservation = this.reserveRun(
@@ -1439,6 +1469,7 @@ class LocalWorkspaceMessageSubmission {
       unlimitedRounds,
     )
     reservation.v4 = v4
+    reservation.discussionStyle = discussionStyle
     try {
       if (regeneration) reservation.responseVersionRootId = regeneration.responseVersionRootId
       this.configureRunBudget(reservation, input.budget || {})
@@ -1587,7 +1618,6 @@ class LocalWorkspaceMessageSubmission {
         let activeKinds = [...targetKinds]
         const pendingKinds = [...targetKinds]
         const replacementInstructions = new Map()
-        const mediaOwnerKind = prepared.mediaRequest ? targetKinds[0] : ''
         while (pendingKinds.length) {
           const kind = pendingKinds.shift()
           if (!activeKinds.includes(kind)) continue
@@ -1619,7 +1649,6 @@ class LocalWorkspaceMessageSubmission {
                 attachmentSnapshots: prepared.attachments,
                 sessionThreadRootId: requestedThreadRootId || threadRootId,
                 runtimeInstruction: replacementInstructions.get(kind) || '',
-                mediaRequest: kind === mediaOwnerKind ? prepared.mediaRequest : null,
                 responseVersionRootId: regeneration?.responseVersionRootId || '',
                 regenerationInstruction: regeneration ? this.regenerationInstruction() : '',
                 contextOptions: regeneration
@@ -1764,7 +1793,7 @@ class LocalWorkspaceMessageSubmission {
       ? normalizeTargetKinds(input.targetKinds)
       : group.agentKinds
     const targetKinds = [...new Set(requested.filter(kind => group.agentKinds.includes(kind)))]
-    if (targetKinds.length < 2) throw new Error('LOCAL_AUTO_AGENT_COUNT')
+    if (!targetKinds.length) throw new Error('LOCAL_AUTO_AGENT_COUNT')
     const unlimitedRounds = input.unlimitedRounds === true
     const maxRounds = unlimitedRounds
       ? 0
@@ -1797,6 +1826,18 @@ class LocalWorkspaceMessageSubmission {
       group.id, 'auto', targetKinds, threadRootId, maxRounds, unlimitedRounds,
     )
     reservation.v4 = input.protocol !== 'legacy' && this.v4Requested(input)
+    const requestedDiscussionStyle = input.discussionStyle == null
+      ? ''
+      : cleanInline(input.discussionStyle, 40)
+    if (requestedDiscussionStyle && !DISCUSSION_STYLES.has(requestedDiscussionStyle)) {
+      this.releasePreparation(group.id, reservation)
+      throw new Error('LOCAL_DISCUSSION_STYLE_INVALID')
+    }
+    if (requestedDiscussionStyle && !reservation.v4) {
+      this.releasePreparation(group.id, reservation)
+      throw new Error('LOCAL_DISCUSSION_STYLE_UNSUPPORTED')
+    }
+    reservation.discussionStyle = requestedDiscussionStyle || 'sequential'
     try {
       this.configureRunBudget(reservation, input.budget || {})
       const contextPack = this.createContextPack({

desktop/src/workspace/local-workspace-run-messages.cjs

@@ -9,6 +9,11 @@ const {
   terminalStatusPrefix,
 } = require('./local-workspace-inputs.cjs')
 
+function turnSeqMetadata(error) {
+  const turnSeq = Number(error?.turnSeq)
+  return Number.isSafeInteger(turnSeq) && turnSeq > 0 ? { turnSeq } : {}
+}
+
 class LocalWorkspaceRunMessages {
   constructor(options) {
     this.state = options.state
@@ -73,7 +78,7 @@ class LocalWorkspaceRunMessages {
         kind,
         threadRootId,
         system,
-        { trace, responseVersionRootId },
+        { trace, responseVersionRootId, ...turnSeqMetadata(error) },
       )
     }
     return { label, reason }
@@ -114,7 +119,7 @@ class LocalWorkspaceRunMessages {
         key: interrupted ? 'system.agentInterrupted' : 'system.agentStopped',
         params: { agent: label },
       },
-      { trace, responseVersionRootId },
+      { trace, responseVersionRootId, ...turnSeqMetadata(error) },
     )
     return trace
   }

desktop/src/workspace/local-workspace-runtime-contracts.cjs

@@ -10,8 +10,12 @@ const {
 const AUTO_CONSENSUS_MARKER = /\[\[MELDWORK_CONSENSUS:(agree|continue)\]\]/gi
 const AUTO_FINAL_CONSENSUS_MARKER = /(?:^|\r?\n)[ \t]*\[\[MELDWORK_CONSENSUS:(agree|continue)\]\][ \t]*$/i
 const DEFAULT_RUN_SILENCE_WARNING_MS = 20 * 1000
-const DEFAULT_RUN_AGENT_TIMEOUT_MS = 3 * 60 * 1000
-const DEFAULT_RUN_AGENT_TOOL_TIMEOUT_MS = 5 * 60 * 1000
+// Inactivity windows for one Agent turn. Real CLIs regularly think, run builds
+// or wait on slow providers for several minutes; the former 3/5 minute limits
+// killed Agents that were still working. A separate no-progress cap (three
+// times the inactivity window) still stops CLIs that only emit heartbeats.
+const DEFAULT_RUN_AGENT_TIMEOUT_MS = 10 * 60 * 1000
+const DEFAULT_RUN_AGENT_TOOL_TIMEOUT_MS = 20 * 60 * 1000
 const DEFAULT_RUN_ABORT_GRACE_MS = 2500
 const DEFAULT_AUTO_ROUNDS = 6
 const MAX_AUTO_ROUNDS = 10

desktop/src/workspace/local-workspace-state.cjs

@@ -1,5 +1,7 @@
 const fs = require('node:fs')
+const { atomicWritePrivateFile } = require('../security/private-file.cjs')
 const path = require('node:path')
+const { RUN_STATUSES } = require('../runs/run-ledger-records.cjs')
 const { normalizeSessionMeta } = require('../runs/run-harness.cjs')
 const {
   SESSION_KEY,
@@ -19,7 +21,7 @@ const FAILED_AGENT_STATUSES = new Set(['failed', 'stopped', 'timeout'])
 const TERMINAL_PERSISTENCE_STATES = new Set(['pending', 'retrying', 'failed'])
 
 const V4_PHASES = new Set([
-  'prepare', 'dispatch', 'running', 'reconcile', 'proposal', 'challenge',
+  'prepare', 'dispatch', 'running', 'reconcile', 'proposal', 'discussion', 'challenge',
   'coordination', 'work', 'synthesis', 'verification', 'commit', 'committed', 'completed',
   'human-gate', 'stopped', 'failed',
 ])
@@ -183,8 +185,17 @@ function loadWorkspaceState(storagePath) {
   }
 
   try {
+    const repliedAgentKinds = new Map()
+    for (const message of Array.isArray(parsed.messages) ? parsed.messages : []) {
+      if (message?.role !== 'agent' || typeof message.groupId !== 'string') continue
+      const kinds = repliedAgentKinds.get(message.groupId) || []
+      if (!kinds.includes(message.agentKind)) kinds.push(message.agentKind)
+      repliedAgentKinds.set(message.groupId, kinds)
+    }
     const groups = parsed.groups
-      .map(normalizeLoadedGroup)
+      .map(group => normalizeLoadedGroup(group, {
+        recoveredAgentKinds: repliedAgentKinds.get(group?.id) || [],
+      }))
       .filter(Boolean)
     const groupIds = new Set(groups.map(group => group.id))
     const messages = parsed.messages
@@ -237,16 +248,15 @@ function loadWorkspaceState(storagePath) {
   }
 }
 
+// The whole workspace is rewritten on every save, synchronously on the main
+// process. Compact JSON cuts the bytes written (and the time the UI is
+// blocked) for long histories, and the private atomic writer fsyncs before
+// the rename so a crash cannot leave a truncated file behind.
 function saveWorkspaceState(storagePath, state) {
-  fs.mkdirSync(path.dirname(storagePath), { recursive: true })
-  const tempPath = `${storagePath}.tmp`
-  fs.writeFileSync(tempPath, `${JSON.stringify(state, null, 2)}\n`, { mode: 0o600 })
-  fs.renameSync(tempPath, storagePath)
+  atomicWritePrivateFile(storagePath, `${JSON.stringify(state)}\n`)
 }
 
-function durableWaitingRunSnapshots({ state, runLedger, pendingGates, liveRunIds, liveGroupIds }) {
-  let records = []
-  try { records = runLedger?.list?.() || [] } catch { return [] }
+function durableWaitingRunSnapshots({ state, records, pendingGates, liveRunIds, liveGroupIds }) {
   const groupIds = new Set(state.groups.map(group => group.id))
   const pendingById = new Map(pendingGates.map(gate => [gate.gateId, gate]))
   const projectedGroupIds = new Set()
@@ -319,6 +329,23 @@ function workspaceSnapshot({
   detectedAgents, state, preparingRuns, activeRuns, humanGateCoordinator, runLedger,
   workspaceRecovery,
 }) {
+  let records = []
+  try { records = runLedger?.list?.() || [] } catch { /* retain the visible workspace on ledger read failure */ }
+  const groupIds = new Set(state.groups.map(group => group.id))
+  const roots = new Map(state.messages.filter(message => message.role === 'user' && !message.threadRootId)
+    .map(message => [message.id, message.groupId]))
+  const latestOutcomes = new Map()
+  for (const record of records) {
+    if (!groupIds.has(record.groupId) || roots.get(record.threadRootId) !== record.groupId
+        || !RUN_STATUSES.has(record.status)) continue
+    const previous = latestOutcomes.get(record.threadRootId)
+    if (previous && previous.startedAt >= record.startedAt) continue
+    latestOutcomes.set(record.threadRootId, {
+      groupId: record.groupId, runId: record.runId, threadRootId: record.threadRootId,
+      status: record.status, targetKinds: [...record.targetKinds],
+      startedAt: record.startedAt, finishedAt: record.finishedAt || 0,
+    })
+  }
   const busyEntries = [
     ...[...preparingRuns.entries()].map(entry => [...entry, 'preparing']),
     ...[...activeRuns.entries()].map(entry => [...entry, 'running']),
@@ -330,7 +357,7 @@ function workspaceSnapshot({
   const liveRunIds = new Set(runEntries.map(([, run]) => run.runId).filter(Boolean))
   const liveGroupIds = new Set(busyEntries.map(([groupId]) => groupId))
   const durableRunEntries = durableWaitingRunSnapshots({
-    state, runLedger, pendingGates, liveRunIds, liveGroupIds,
+    state, records, pendingGates, liveRunIds, liveGroupIds,
   })
   const liveRunEntries = runEntries.map(([groupId, run, phase]) => {
     const mode = run.mode === 'auto' ? 'auto' : 'manual'
@@ -404,6 +431,7 @@ function workspaceSnapshot({
     ])],
     humanGates: publicGates,
     runs: publicRuns,
+    runOutcomes: [...latestOutcomes.values()],
   }
 }
 

desktop/src/workspace/local-workspace.cjs

@@ -2,6 +2,7 @@ const { EventEmitter } = require('node:events')
 const { createHash, randomUUID } = require('node:crypto')
 const path = require('node:path')
 const { ContentBlobStore } = require('../attachments/content-blob-store.cjs')
+const { agentRuntimeCapabilities } = require('../agents/agent-runtime-contract.cjs')
 const { ContextPackStore } = require('../collaboration/context-pack-store.cjs')
 const { HumanGateCoordinator } = require('../gates/human-gate-coordinator.cjs')
 const { HumanGateStore } = require('../gates/human-gate-store.cjs')
@@ -66,7 +67,29 @@ const TERMINAL_RUN_STATUSES = new Set([
   'completed', 'partial', 'failed', 'stopped', 'timeout', 'round-limit', 'interrupted',
   'budget-exhausted', 'circuit-breaker',
 ])
-const V4_VISIBLE_MESSAGE_PHASES = ['proposal', 'challenge', 'work', 'verification']
+const V4_VISIBLE_MESSAGE_PHASES = ['proposal', 'discussion', 'challenge', 'work', 'verification']
+
+function naturalV4MessageBindingMatches(message, record, phase) {
+  if (!record.orchestration?.discussionStyle || !['proposal', 'discussion'].includes(phase)) {
+    return true
+  }
+  const round = Number(message?.trace?.round) || 0
+  const slot = record.orchestration.slots?.find(candidate => (
+    candidate.agentKind === message.agentKind
+  ))
+  if (round < 1 || !slot?.slotId) return false
+  const operationPhase = phase === 'discussion' ? `discussion:${round}` : phase
+  const operationId = `operation-${createHash('sha256').update(JSON.stringify([
+    record.runId,
+    record.taskId,
+    message.agentKind,
+    slot.slotId,
+    operationPhase,
+    round,
+  ])).digest('hex')}`
+  return message.trace?.context?.operationId === operationId
+    && message.trace?.context?.snapshotHash === record.orchestration.snapshotHash
+}
 
 function isCommittedV4PhaseMessage(message, record) {
   const agentRunId = String(message?.trace?.agentRunId || '')
@@ -74,15 +97,16 @@ function isCommittedV4PhaseMessage(message, record) {
       || message?.trace?.runId !== record.runId || !message.agentKind || !agentRunId) {
     return false
   }
-  return V4_VISIBLE_MESSAGE_PHASES.some(phase => (
-    message.id === `message-${hashValue({
+  return V4_VISIBLE_MESSAGE_PHASES.some((phase) => {
+    if (message.id !== `message-${hashValue({
       agentKind: message.agentKind,
       agentRunId,
       phase,
       runId: record.runId,
       taskId: record.taskId,
-    })}`
-  ))
+    })}`) return false
+    return naturalV4MessageBindingMatches(message, record, phase)
+  })
 }
 
 class LocalWorkspace extends EventEmitter {
@@ -91,7 +115,6 @@ class LocalWorkspace extends EventEmitter {
     this.storagePath = options.storagePath
     this.detectAgentsFn = options.detectAgents
     this.runAgentFn = options.runAgent
-    this.generateMediaFn = options.generateMedia || null
     this.resolveAttachmentsFn = options.resolveAttachments || (async (attachments) => {
       if (attachments?.length) throw new Error('LOCAL_ATTACHMENT_STORAGE_UNAVAILABLE')
       return []
@@ -108,7 +131,6 @@ class LocalWorkspace extends EventEmitter {
       image: this.imageAttachmentLimitFn(kind),
     }))
     this.credentialStateFn = options.credentialState || (async () => ({ state: 'unknown', source: 'unverified' }))
-    this.sharedProviderReadyFn = options.sharedProviderReady || (() => false)
     this.agentLabelFn = options.agentLabel || defaultAgentLabel
     this.runSilenceWarningMs = Number.isFinite(options.runSilenceWarningMs)
       && options.runSilenceWarningMs > 0
@@ -170,6 +192,10 @@ class LocalWorkspace extends EventEmitter {
     this.detectedAgents = []
     this.preparingRuns = new Map()
     this.activeRuns = new Map()
+    // Highest turn number handed out per group topic in this process. Saved
+    // messages are the durable source; reservations only cover turns that are
+    // dispatched but not written yet.
+    this.turnSeqReservations = new Map()
     this.pendingV4ManualRecoveries = new Map()
     this.pendingV4AutoRecoveries = new Map()
     this.runCheckpointTimers = new Map()
@@ -181,7 +207,6 @@ class LocalWorkspace extends EventEmitter {
       setDetectedAgents: agents => { this.detectedAgents = agents },
       detectAgents: () => this.detectAgentsFn(),
       credentialState: (...args) => this.credentialStateFn(...args),
-      sharedProviderReady: kind => this.sharedProviderReadyFn(kind),
       attachmentSupport: kind => this.attachmentSupportFn(kind),
       save: () => this.save(),
       emitChanged: () => this.emitChanged(),
@@ -226,6 +251,7 @@ class LocalWorkspace extends EventEmitter {
       terminalRetrySleep: options.terminalRetrySleep,
     })
     this.conversations = new LocalWorkspaceConversations({
+      reserveTurnSeq: (groupId, threadRootId) => this.reserveTurnSeqs(groupId, threadRootId, 1)[0],
       state: () => this.state,
       detectedAgents: () => this.detectedAgents,
       save: () => this.save(),
@@ -245,7 +271,9 @@ class LocalWorkspace extends EventEmitter {
       addMessage: (...args) => this.addMessage(...args),
     })
     this.agentInvocation = new LocalWorkspaceAgentInvocation({
+      reserveTurnSeq: (groupId, threadRootId) => this.reserveTurnSeqs(groupId, threadRootId, 1)[0],
       state: () => this.state,
+      agentLabel: kind => this.agentLabel(kind),
       detectedAgents: () => this.detectedAgents,
       activeRuns: this.activeRuns,
       runAgentTimeoutMs: this.runAgentTimeoutMs,
@@ -287,10 +315,13 @@ class LocalWorkspace extends EventEmitter {
       completeHumanGateContinuation: (...args) => this.completeHumanGateContinuation(...args),
       connectorRuntime: options.connectorRuntime,
       attachmentSupport: (...args) => this.attachmentSupportFn(...args),
-      generateMedia: (...args) => this.generateMediaFn?.(...args),
     })
     this.autoRunner = new LocalWorkspaceAutoRunner({
+      reserveTurnSeqs: (...args) => this.reserveTurnSeqs(...args),
       state: () => this.state,
+      canWriteWorkspace: kind => agentRuntimeCapabilities(kind, {
+        agent: this.detectedAgents.find(agent => agent.kind === kind),
+      }).permissionModes.includes('workspace-write'),
       beginRun: (...args) => this.beginRun(...args),
       resolveAttachments: (...args) => this.resolveAttachments(...args),
       validateSkillSelections: (...args) => this.validateSkillSelectionsFn(...args),
@@ -306,7 +337,6 @@ class LocalWorkspace extends EventEmitter {
       recordAgentFailure: (...args) => this.recordAgentFailure(...args),
       recordAgentInterruption: (...args) => this.recordAgentInterruption(...args),
       addMessage: (...args) => this.addMessage(...args),
-      removeAgent: (...args) => this.conversations.removeAgent(...args),
       commitV4AgentMessage: input => this.commitV4AgentMessage(input),
       emitChanged: () => this.emitChanged(),
       finishRun: (...args) => this.finishRun(...args),
@@ -316,6 +346,8 @@ class LocalWorkspace extends EventEmitter {
       hasRunLedger: () => Boolean(this.runLedger),
       requestHumanGate: (...args) => this.requestHumanGate(...args),
       completeHumanGateContinuation: (...args) => this.completeHumanGateContinuation(...args),
+      requestTaskDecision: (...args) => this.requestV4TaskDecision(...args),
+      taskDecisionResponses: runId => this.v4TaskDecisionResponses(runId),
       retryContract: kind => ({
         idempotencyMode: this.detectedAgents.find(agent => agent.kind === kind)?.idempotencyMode
           === 'durable'
@@ -327,6 +359,7 @@ class LocalWorkspace extends EventEmitter {
       retrySleep: options.retrySleep,
     })
     this.messageSubmission = new LocalWorkspaceMessageSubmission({
+      reserveTurnSeqs: (...args) => this.reserveTurnSeqs(...args),
       state: () => this.state,
       detectedAgents: () => this.detectedAgents,
       isShuttingDown: () => this.shuttingDown,
@@ -476,7 +509,7 @@ class LocalWorkspace extends EventEmitter {
       const commitState = orchestration?.commitState
       const candidateCommit = orchestration?.candidateCommit
       const resumablePhase = [
-        'proposal', 'challenge', 'coordination', 'work', 'synthesis', 'verification',
+        'proposal', 'discussion', 'challenge', 'coordination', 'work', 'synthesis', 'verification',
       ].includes(orchestration?.phase)
       const resumableCommit = ['commit', 'committed', 'completed'].includes(orchestration?.phase)
         && ['committing', 'committed'].includes(commitState?.status)
@@ -636,6 +669,10 @@ class LocalWorkspace extends EventEmitter {
       const finalStatus = await this.autoRunner.resume(group, durable, controller)
       await this.finishRun(group.id, controller, finalStatus)
     } catch (error) {
+      if (controller.signal.aborted) {
+        await this.finishRun(group.id, controller, terminalRunStatusForReason(controller.stopReason))
+        return
+      }
       controller.stopReason = String(error?.message || 'LOCAL_RUN_CONTINUATION_INVALID')
       await this.finishRun(group.id, controller, 'failed')
     }
@@ -775,7 +812,7 @@ class LocalWorkspace extends EventEmitter {
     const invocationFields = ['phase', 'slotId', 'operationId', 'snapshotHash']
     const invocationFieldCount = invocationFields.filter(field => Object.hasOwn(input, field)).length
     if (invocationFieldCount > 0
-        && (input.resumeKind !== 'agent_slot'
+        && (!['agent_slot', 'v4_task_decision'].includes(input.resumeKind)
           || invocationFieldCount !== invocationFields.length)) {
       throw new Error('LOCAL_RUN_CONTINUATION_INVALID')
     }
@@ -1110,6 +1147,105 @@ class LocalWorkspace extends EventEmitter {
     return { ...match, state: continuation.state }
   }
 
+  v4TaskDecisionSource(durable, gate, request) {
+    if (durable?.mode !== 'auto' || durable.orchestration?.version !== 4
+        || gate.type !== 'input' || gate.runId !== durable.runId
+        || !request || Object.keys(request).length !== 7
+        || request.source !== 'task-decision' || request.phase !== 'discussion'
+        || request.snapshotHash !== durable.orchestration.snapshotHash) return null
+    const sources = durable.agentRuns.filter(attempt => (
+      attempt.agentRunId === gate.agentRunId && attempt.kind === gate.agentKind
+      && attempt.status === 'completed' && attempt.round === request.round
+      && attempt.context?.operationId === request.operationId
+      && attempt.context?.snapshotHash === request.snapshotHash
+      && attempt.context?.taskDecision?.status === 'needs-human'
+      && hashValue(attempt.context.taskDecision) === request.decisionHash
+    ))
+    if (sources.length !== 1 || request.operationId !== this.autoRunner.v4OperationId(
+      durable, gate.agentKind, `discussion:${request.round}`, request.slotId, request.round,
+    )) return null
+    return sources[0]
+  }
+
+  canResumeV4TaskDecision(durable, gate, request) {
+    const source = this.v4TaskDecisionSource(durable, gate, request)
+    const continuation = durable.continuation
+    const cursor = durable.orchestration
+    const group = this.state.groups.find(item => item.id === durable.groupId)
+    const slot = cursor?.slots.find(item => item.slotId === request?.slotId)
+    return Boolean(source && group && gate.options.length === 2
+      && gate.options[0].optionId === 'respond' && gate.options[0].kind === 'respond'
+      && gate.options[1].optionId === 'cancel' && gate.options[1].kind === 'reject'
+      && continuation?.resumeKind === 'v4_task_decision'
+      && cursor.phase === 'discussion' && cursor.round === request.round
+      && durable.currentRound === request.round && continuation.round === request.round
+      && continuation.slotId === request.slotId
+      && continuation.operationId === request.operationId
+      && continuation.snapshotHash === request.snapshotHash
+      && slot?.agentRunId === source.agentRunId && slot.agentKind === source.kind
+      && slot.operationId === request.operationId && slot.commitStatus === 'committed'
+      && this.autoRunner.v4NaturalOwner(
+        group, durable, durable.threadRootId, durable.orchestration?.activeKinds,
+      ) === source.kind)
+  }
+
+  v4TaskDecisionResponses(runId) {
+    const durable = this.runLedger?.get?.(runId)
+    if (!durable) return []
+    return this.humanGateStore.list({ runId }).flatMap(item => {
+      if (item.type !== 'input' || item.status !== 'approved') return []
+      try {
+        const gate = this.humanGateStore.get(item.gateId)
+        const request = this.humanGateStore.request(item.gateId)
+        const source = this.v4TaskDecisionSource(durable, gate, request)
+        return source && gate.decision?.response ? [{
+          id: gate.gateId, round: source.round, agentKind: 'human',
+          text: gate.decision.response,
+        }] : []
+      } catch { return [] }
+    })
+  }
+
+  async requestV4TaskDecision(group, controller, message) {
+    const durable = this.runLedger?.get?.(controller.runId)
+    if (!durable) return null
+    const source = durable.agentRuns.find(attempt => (
+      attempt.agentRunId === message.trace?.agentRunId && attempt.kind === message.agentKind
+    ))
+    const slot = durable.orchestration?.slots.find(item => item.agentRunId === source?.agentRunId)
+    if (!source || !slot || source.context?.taskDecision?.status !== 'needs-human'
+        || slot.commitStatus !== 'committed') throw new Error('LOCAL_RUN_CONTINUATION_INVALID')
+    const request = {
+      source: 'task-decision', phase: 'discussion', round: source.round,
+      slotId: slot.slotId, operationId: slot.operationId,
+      snapshotHash: durable.orchestration.snapshotHash,
+      decisionHash: hashValue(source.context.taskDecision),
+    }
+    const decision = await this.requestHumanGate({
+      type: 'input', runId: durable.runId, agentRunId: source.agentRunId, agentKind: source.kind,
+      createdAt: new Date(slot.finishedAt).toISOString(),
+      summary: source.context.taskDecision.reason,
+      options: [
+        { optionId: 'respond', name: 'Respond', kind: 'respond' },
+        { optionId: 'cancel', name: 'Cancel', kind: 'reject' },
+      ],
+      request,
+    }, {
+      signal: controller.signal,
+      preserveOnAbort: () => controller.stopReason === 'shutdown',
+      continuation: {
+        resumeKind: 'v4_task_decision', agentRunId: source.agentRunId, agentKind: source.kind,
+        round: source.round, phase: 'discussion', slotId: slot.slotId,
+        operationId: slot.operationId, snapshotHash: durable.orchestration.snapshotHash,
+      },
+    })
+    if (this.completeHumanGateContinuation(durable.runId, decision.gateId,
+      decision.status === 'approved' ? 'completed' : 'cancelled') !== true) {
+      throw new Error('LOCAL_RUN_PERSIST_FAILED')
+    }
+    return decision
+  }
+
   async requestHumanGate(input, options = {}) {
     const runId = String(input?.runId || '')
     const previous = this.humanGateWaitTails.get(runId) || Promise.resolve()
@@ -1217,6 +1353,11 @@ class LocalWorkspace extends EventEmitter {
     if (continuation.state !== 'pending' && gate.status === 'pending') return false
     try {
       const request = this.humanGateStore.request(gate.gateId)
+      if (continuation.resumeKind === 'v4_task_decision') {
+        const pack = this.contextPackStore.get(runRecord.contextPackId)
+        return pack.taskId === runRecord.taskId
+          && this.canResumeV4TaskDecision(runRecord, gate, request)
+      }
       const connectorBound = gate.type === 'input' || request?.source === 'connector'
       if (connectorBound && (
         continuation.requestId !== request.requestId
@@ -1437,7 +1578,7 @@ class LocalWorkspace extends EventEmitter {
     if (durable?.mode !== 'auto' || continuation?.resumeKind !== 'agent_slot'
         || cursor?.version !== 4 || cursor.workflow !== 'auto'
         || cursor.template !== 'discussion'
-        || !['proposal', 'challenge', 'work', 'synthesis', 'verification'].includes(phase)
+        || !['proposal', 'discussion', 'challenge', 'work', 'synthesis', 'verification'].includes(phase)
         || !['budget', 'permission', 'retry', 'input'].includes(gate?.type)
         || gate.runId !== durable.runId || gate.agentRunId !== continuation.agentRunId
         || gate.agentKind !== continuation.agentKind
@@ -1472,7 +1613,7 @@ class LocalWorkspace extends EventEmitter {
         || (slot.resultRefs?.workflowOutcomeRefs || []).some(item => (
           item?.receipt?.phase === phase && item.receipt.operationId === slot.operationId
         ))) return false
-    let operationPhase = phase
+    let operationPhase = phase === 'discussion' ? `discussion:${cursor.round}` : phase
     if (phase === 'work') {
       const matchingAssignments = cursor.coordinationPlan?.assignments?.filter(item => (
         item.ownerKind === slot.agentKind && item.taskId
@@ -1510,6 +1651,7 @@ class LocalWorkspace extends EventEmitter {
     }
     return slot.operationId === this.autoRunner.v4OperationId(
       controller, slot.agentKind, operationPhase, slot.slotId,
+      ...(phase === 'discussion' ? [cursor.round] : []),
     )
   }
 
@@ -1728,6 +1870,11 @@ class LocalWorkspace extends EventEmitter {
     try {
       controller = this.runCoordinator.resume(durable)
       const group = this.getGroup(durable.groupId)
+      if (durable.continuation.resumeKind === 'v4_task_decision') {
+        const finalStatus = await this.autoRunner.resume(group, durable, controller)
+        await this.finishRun(durable.groupId, controller, finalStatus)
+        return
+      }
       const resumableManual = this.canResumeManualOrchestration(durable)
       const resumableAuto = this.canResumeAutoOrchestration(durable)
       const resumableV4ManualSlot = this.canResumeV4ManualAgentSlot(durable, gate)
@@ -2293,6 +2440,22 @@ class LocalWorkspace extends EventEmitter {
     )
   }
 
+  // Reserves `count` consecutive dispatch-turn numbers for one group topic.
+  reserveTurnSeqs(groupId, threadRootId, count = 1) {
+    const amount = Math.max(0, Math.min(64, Math.floor(Number(count) || 0)))
+    if (!groupId || !threadRootId || !amount) return []
+    const key = `${groupId}\u0000${threadRootId}`
+    let cursor = this.turnSeqReservations.get(key) || 0
+    for (const message of this.state.messages) {
+      if (message.groupId === groupId && message.threadRootId === threadRootId
+          && Number.isSafeInteger(message.turnSeq) && message.turnSeq > cursor) {
+        cursor = message.turnSeq
+      }
+    }
+    this.turnSeqReservations.set(key, cursor + amount)
+    return Array.from({ length: amount }, (_, index) => cursor + index + 1)
+  }
+
   commitV4AgentMessage(input = {}) {
     const messageId = String(input.messageId || '')
     const existing = this.state.messages.find(message => message.id === messageId)
@@ -2322,6 +2485,11 @@ class LocalWorkspace extends EventEmitter {
       ...(input.metadata?.responseVersionRootId
         ? { responseVersionRootId: input.metadata.responseVersionRootId } : {}),
       ...(input.metadata?.trace ? { trace: input.metadata.trace } : {}),
+      ...(input.metadata?.turnSeq
+        ? { turnSeq: input.metadata.turnSeq }
+        : (input.threadRootId && group.conversationType !== 'direct'
+          ? { turnSeq: this.reserveTurnSeqs(group.id, input.threadRootId, 1)[0] }
+          : {})),
     })
     if (!message || message.id !== messageId) throw new Error('LOCAL_RUN_COMMIT_INVALID')
     const previousMessages = [...this.state.messages]
@@ -2375,26 +2543,47 @@ class LocalWorkspace extends EventEmitter {
   }
 
   persistSessionState(key, sessionRef, meta) {
+    const canonicalKey = this.canonicalConversationSessionKey(key)
     const nextRef = normalizeSessionRef(sessionRef)
-    if (!SESSION_KEY.test(String(key || '')) || !nextRef) return false
-    const hadSession = Object.hasOwn(this.state.sessions, key)
-    const hadMeta = Object.hasOwn(this.state.sessionMeta, key)
-    const previousSession = this.state.sessions[key]
-    const previousMeta = this.state.sessionMeta[key]
-    this.state.sessions[key] = nextRef
-    this.state.sessionMeta[key] = normalizeSessionMeta(meta)
+    if (!SESSION_KEY.test(String(canonicalKey || '')) || !nextRef) return false
+    const hadSession = Object.hasOwn(this.state.sessions, canonicalKey)
+    const hadMeta = Object.hasOwn(this.state.sessionMeta, canonicalKey)
+    const hadLegacySession = Object.hasOwn(this.state.sessions, key)
+    const hadLegacyMeta = Object.hasOwn(this.state.sessionMeta, key)
+    const previousLegacySession = this.state.sessions[key]
+    const previousLegacyMeta = this.state.sessionMeta[key]
+    const previousSession = this.state.sessions[canonicalKey]
+    const previousMeta = this.state.sessionMeta[canonicalKey]
+    this.state.sessions[canonicalKey] = nextRef
+    this.state.sessionMeta[canonicalKey] = normalizeSessionMeta(meta)
+    if (canonicalKey !== key) {
+      delete this.state.sessions[key]
+      delete this.state.sessionMeta[key]
+    }
     try {
       this.save()
     } catch (error) {
-      if (hadSession) this.state.sessions[key] = previousSession
-      else delete this.state.sessions[key]
-      if (hadMeta) this.state.sessionMeta[key] = previousMeta
-      else delete this.state.sessionMeta[key]
+      if (hadSession) this.state.sessions[canonicalKey] = previousSession
+      else delete this.state.sessions[canonicalKey]
+      if (hadMeta) this.state.sessionMeta[canonicalKey] = previousMeta
+      else delete this.state.sessionMeta[canonicalKey]
+      if (canonicalKey !== key) {
+        if (hadLegacySession) this.state.sessions[key] = previousLegacySession
+        else delete this.state.sessions[key]
+        if (hadLegacyMeta) this.state.sessionMeta[key] = previousLegacyMeta
+        else delete this.state.sessionMeta[key]
+      }
       throw error
     }
     return true
   }
 
+  canonicalConversationSessionKey(key) {
+    const value = String(key || '')
+    const match = value.match(/^([^:]+):task:[^:]+:([^:]+)$/)
+    return match ? `${match[1]}:${match[2]}` : value
+  }
+
   persistSessionRef(key, sessionRef) {
     const next = normalizeSessionRef(sessionRef)
     if (!SESSION_KEY.test(String(key || '')) || !next || next === this.state.sessions[key]) return false

desktop/src/workspace/prompts/manual-reply-prompt.cjs

@@ -0,0 +1,44 @@
+const { deliveryContract } = require('./prompt-contracts.cjs')
+const { languageNote } = require('./natural-discussion-prompt.cjs')
+
+// What each Agent is told when a user message goes to several Agents at once
+// (manual group chat). Everyone answers on their own, at the same time, and
+// nobody sees the other replies until they are all in. The user message is
+// the task; the prompt does not ask for plans, work items or receipts.
+function historyText(history) {
+  return (Array.isArray(history) ? history : [])
+    .map(item => `[${item.role}${item.agentKind ? `:${item.agentKind}` : ''}] ${item.text}`)
+    .join('\n')
+}
+
+function manualReplyPrompt({
+  group, kind, targetKinds = [], taskText, history = [], skillHintsText = '',
+  knowledgeBaseHintsText = '', writer = false,
+}) {
+  const earlier = historyText(history)
+  const peers = targetKinds.filter(peer => peer !== kind)
+  const groupName = String(group?.name || '').trim().slice(0, 120)
+  const topic = String(group?.topic || '').trim().slice(0, 240)
+  return [
+    peers.length
+      ? `You are one of several Agents answering the user's latest message in a Meldwork group chat. Everyone replies on their own at the same time; the others are ${peers.map(peer => `@${peer}`).join(', ')}. You cannot see their replies, so give your own complete answer and do not speak for them.`
+      : "You are answering the user's latest message in a Meldwork group chat.",
+    groupName ? `Group: ${groupName}${topic ? `\nTopic: ${topic}` : ''}` : '',
+    `User message:\n${taskText || '(none)'}`,
+    skillHintsText,
+    knowledgeBaseHintsText,
+    earlier
+      ? `Earlier conversation in this group (reference only; verify it before relying on it):\n${earlier}`
+      : '',
+    writer
+      ? [
+          'You are the Agent that may change files for this message. Make only the changes the user asked for within the authorized workspace, and verify the result.',
+          group?.allowWrite ? deliveryContract(group) : '',
+        ].filter(Boolean).join('\n')
+      : 'This reply is read-only: do not modify files or other shared state. If a change is needed, describe it instead.',
+    'Answer naturally in Markdown. Do not output JSON, XML, protocol labels or hidden orchestration text.',
+    languageNote(taskText),
+  ].filter(Boolean).join('\n\n')
+}
+
+module.exports = { manualReplyPrompt }

desktop/src/workspace/prompts/natural-discussion-prompt.cjs

@@ -0,0 +1,145 @@
+const { deliveryContract, responseLanguageFromText } = require('./prompt-contracts.cjs')
+
+// The prompts of a natural discussion: Agents answer in their own words, and
+// the only machine-readable part is one control receipt at the end of each
+// discussion turn. The text is split into small named sections so each can be
+// read, tested and changed on its own.
+
+const PHASE_INSTRUCTIONS = Object.freeze({
+  proposal: 'Work independently on the user task and offer your own analysis, recommendations, and next steps.',
+  discussion: 'Continue from the available peer responses. Add useful evidence, challenge weak assumptions, resolve disagreements where possible, and advance the shared result.',
+  challenge: 'Review the available peer context as a subject-matter peer. Explain what you agree with, what is missing, and the clearest correction or direction.',
+  work: 'Complete the responsibility assigned to you. Report useful work, evidence, blockers, or handoffs in your own words.',
+  synthesis: 'Combine the available work into the clearest useful answer for the user. Resolve contradictions where possible and state remaining uncertainty plainly.',
+  verification: 'Independently check the proposed answer or deliverable. Report material defects, risks, or acceptance in your own words.',
+})
+
+function phaseInstruction(phase) {
+  return PHASE_INSTRUCTIONS[phase] || 'Address the current collaboration task directly.'
+}
+
+function groupLine(group) {
+  const name = String(group?.name || '').trim().slice(0, 120)
+  const topic = String(group?.topic || '').trim().slice(0, 240)
+  if (!name && !topic) return ''
+  return `Group: ${name || 'Meldwork group'}${topic ? `\nTopic: ${topic}` : ''}`
+}
+
+function languageNote(taskText) {
+  const language = responseLanguageFromText(taskText)
+  return language
+    ? `Answer in ${language}, the language of the user task, unless the task asks for another language.`
+    : 'Answer in the language of the user task unless the task asks for another language.'
+}
+
+// The part of the prompt every natural turn starts with.
+function naturalBrief({
+  group, phase, taskText, history = '', skillHintsText = '', knowledgeBaseHintsText = '',
+}) {
+  return [
+    'You are participating in a local multi-agent discussion.',
+    groupLine(group),
+    `User task:\n${taskText || '(none)'}`,
+    skillHintsText,
+    knowledgeBaseHintsText,
+    history && history !== '(none)' ? `Relevant prior context:\n${history}` : '',
+    phaseInstruction(phase),
+    phase === 'proposal'
+      ? 'This invocation is read-only. Its deliverable is your initial analysis and proposed next actions. Return that contribution when ready; any workspace changes belong to a later invocation that explicitly grants write permission. Do not attempt or delegate writes from this invocation, or wait here for write permission.'
+      : '',
+    phase === 'proposal'
+      ? 'The other Agents are answering the same task independently right now. In the next round you will see their proposals and discuss the differences with them.'
+      : 'Respond to the other Agents directly: name the specific points you agree with, dispute weak ones with reasons or evidence, and extend what is missing. Build on their work instead of restating your earlier answer.',
+    phase === 'discussion'
+      ? 'After concrete deliverables exist, review them for material gaps before accepting the result.'
+      : '',
+    phase === 'discussion'
+      ? 'Answer the user naturally in Markdown. Keep orchestration instructions out of the answer. End every discussion turn with the control receipt described below; Meldwork removes it before the answer is shown.'
+      : 'Answer the user naturally in Markdown. Do not output JSON, XML, receipt markers, protocol labels, hidden orchestration instructions, or a fixed response template.',
+    languageNote(taskText),
+    'Do not claim another Agent performed work, and do not modify shared workspace state unless this turn explicitly grants that permission.',
+  ].filter(Boolean).join('\n\n')
+}
+
+// Who may change files in this discussion.
+function writerSection({ kind, writerKind, group }) {
+  if (!writerKind) return ''
+  if (kind !== writerKind) {
+    return `Workspace writes are assigned to @${writerKind}. Your invocation is read-only; request their contribution when a change is needed.`
+  }
+  return [
+    'You are the designated workspace writer for this task. Perform only writes required by the user task within the authorized workspace, and verify the resulting deliverables. Other participants are read-only.',
+    group?.allowWrite ? deliveryContract(group) : '',
+  ].filter(Boolean).join('\n')
+}
+
+// What the delivery owner is responsible for, and what everyone else owes them.
+function ownerSection({ kind, ownerKind }) {
+  if (!ownerKind) return ''
+  return [
+    `The current delivery owner is @${ownerKind}. Ownership starts with the first selected participant and can be handed off explicitly by that owner.`,
+    kind === ownerKind
+      ? [
+          'You are responsible for judging progress against the original user task and the actual available deliverables. Decide whether to continue work, report a blocker, request a human decision, or declare completion. Silence from peers is not evidence of completion.',
+          'Before you declare completion, make sure the peers whose expertise covers the result have reviewed the actual deliverables. If one has not, ask them a concrete question and request them in nextKinds instead of completing.',
+          'In your final collaboration receipt use taskDecision.status completed, continue, blocked or needs-human to judge the whole task. A completed judgment requires actual deliverables. This is your task judgment, not human adoption or independent system verification.',
+          'To hand off responsibility, use status continue and add handoffTo with one exact selected Agent kind inside taskDecision. Do not hand off to an absent participant. Append nothing after the receipt.',
+        ].join('\n')
+      : 'Respond to the delivery owner and the other peers on the specific points they raised. If you are blocked or need a human decision, explain why and use status blocked or needs-human; the delivery owner then decides how to proceed. Only the current delivery owner can decide task completion or hand off that responsibility.',
+  ].join('\n')
+}
+
+// How the next speaker is chosen: by the scheduler (sequential) or by the
+// Agent's own nextKinds request (agent-led).
+function routingSection({ activeKinds, allowRouting }) {
+  if (allowRouting !== true) {
+    return 'Every selected Agent speaks in order each round, so you do not need to request peers. Finish this invocation instead of waiting for another participant; Meldwork runs the next one after you.'
+  }
+  const agentList = activeKinds.map(agentKind => `@${agentKind}`).join(', ')
+  return [
+    'Read the completed peer turns across recent rounds. Decide who should contribute next from their earlier evidence, unresolved questions, and work still needed; do not automatically reply only to whoever last mentioned you.',
+    `Address peers naturally inside your explanation: ${agentList}. Put each request beside its concrete question or requested contribution. Prose mentions are discussion only; they do not dispatch peers.`,
+    'Selected peers are external Meldwork participants; their @handles are not addresses for your native agent messaging or wait tools. Meldwork dispatches requested peers only after this invocation completes. Put the request in your final response and finish this invocation so they can run; do not wait inside this invocation for their reply.',
+    `In taskDecision.nextKinds list the selected Agent identifiers whose contribution you actually request next, chosen from ${JSON.stringify(activeKinds)}, or [] when none is needed. Use status continue for a nonempty list. Multiple kinds request a concurrent batch. Do not request yourself or absent participants. Historical, negated and explanatory mentions do not belong in nextKinds.`,
+    'Avoid repeating the same handoff or conclusions. Invite a peer who has not recently contributed when their earlier proposal can resolve the issue, and state what new evidence or decision you need.',
+    'List in nextKinds every peer whose answer, review or work you need next, and ask each of them a concrete question in your answer. Leave nextKinds empty only when no peer input would change the result; the delivery owner then judges overall task completion.',
+  ].join('\n')
+}
+
+// The control receipt that ends every discussion turn.
+function receiptSection({ kind, activeKinds, allowRouting }) {
+  const exampleNextKinds = allowRouting === true
+    ? [activeKinds.find(peer => peer !== kind) || kind]
+    : []
+  return `After your natural answer append exactly one receipt, for example [[MELDWORK_COLLABORATION:{"summary":"concise result","taskDecision":{"status":"continue","reason":"what you accomplished and what remains","deliverables":[],"nextKinds":${JSON.stringify(exampleNextKinds)}}}]] while work or review remains, or a taskDecision of {"status":"completed","reason":"why the task is done","deliverables":["what was delivered"]} when the delivery owner judges the whole task done. Use blocked or needs-human with the reason when you cannot continue. This receipt is transport metadata: Meldwork removes it before displaying your answer. User requirements such as exact text, one line, JSON only, or no explanation apply to the visible answer before the receipt; preserve that requested answer exactly and still append the receipt separately. Do not explain the receipt in the visible answer. List actual deliverables when available. Only the current delivery owner judges the whole task; other participants report their contribution and may request peers with nextKinds. Merge any owner judgment or handoffTo into this same receipt. Append nothing after it.`
+}
+
+// One complete natural turn: the brief, what peers said so far, and the
+// discussion-only sections.
+function naturalDiscussionPrompt({
+  brief, group, kind, phase, activeKinds, transcript = '', humanResponsesText = '',
+  writerKind = '', ownerKind = '', allowRouting = false,
+}) {
+  const discussion = phase === 'discussion'
+  return [
+    brief,
+    transcript ? `Completed peer turns:\n${transcript}` : '',
+    humanResponsesText
+      ? `Human responses for this task (clarification, not a change to workspace permissions):\n${humanResponsesText}`
+      : '',
+    discussion ? writerSection({ kind, writerKind, group }) : '',
+    discussion ? ownerSection({ kind, ownerKind }) : '',
+    discussion ? routingSection({ activeKinds, allowRouting }) : '',
+    discussion ? receiptSection({ kind, activeKinds, allowRouting }) : '',
+  ].filter(Boolean).join('\n\n')
+}
+
+module.exports = {
+  languageNote,
+  naturalBrief,
+  naturalDiscussionPrompt,
+  ownerSection,
+  receiptSection,
+  routingSection,
+  writerSection,
+}

desktop/src/workspace/prompts/prompt-contracts.cjs

@@ -0,0 +1,52 @@
+// Instructions every Agent prompt shares: how to answer, and what a delivered
+// file means. They are plain functions of the conversation so each prompt
+// builder can include exactly the parts it needs.
+
+function responseLanguageFromText(text) {
+  const source = String(text || '').normalize('NFKC')
+  const han = (source.match(/[\u3400-\u4dbf\u4e00-\u9fff\uf900-\ufaff]/g) || []).length
+  const kana = (source.match(/[\u3040-\u30ff]/g) || []).length
+  const hangul = (source.match(/[\uac00-\ud7af]/g) || []).length
+  const cyrillic = (source.match(/[\u0400-\u052f]/g) || []).length
+  const arabic = (source.match(/[\u0600-\u06ff]/g) || []).length
+  const latin = (source.match(/[A-Za-z]/g) || []).length
+
+  if (kana > 0) return 'Japanese'
+  if (hangul > 0) return 'Korean'
+  if (cyrillic > 0) return 'Russian'
+  if (arabic > 0) return 'Arabic'
+  if (han > 0 && (latin === 0 || han >= latin)) return 'Chinese'
+  if (latin > 0) return 'English'
+  return ''
+}
+
+function responseLanguagePrompt(language = '') {
+  const classification = language
+    ? `The latest user message is written primarily in ${language}.`
+    : 'The latest user message language was not confidently classified; infer it from the message body only.'
+  return [
+    'Response language contract:',
+    'Always answer entirely in the language of the latest user message unless that message explicitly requests another language.',
+    classification,
+    'If the latest user message is English, answer entirely in English. If it is Chinese, answer entirely in Chinese.',
+    'Do not infer response language from the app UI locale, operating-system locale, Agent name, attachment filenames, or earlier conversation turns.',
+    'Do not translate or switch languages unless the latest user message explicitly asks for it.',
+  ].join('\n')
+}
+
+function deliveryContract(group) {
+  return group?.allowWrite
+    ? [
+        'Workspace access: You may create and edit files in the conversation working directory. When the user requests a deliverable, execute the work instead of returning only a plan.',
+        'Deliverable capture contract: Save or copy each final document, code file, diff, structured-data file, media file, or bundle into .meldwork-output/ in the conversation working directory.',
+        'Only files created or changed during this run are captured as durable outputs. Do not claim a deliverable exists until the real file has been written there.',
+        'When the user asks for an image, audio, or video, do not claim it was generated unless a real file exists.',
+        'Supported chat attachments include common images, media, documents, source code, configuration text, and archives. Other recognized outputs remain durable Artifacts.',
+      ].join('\n')
+    : [
+        'This conversation is read-only. Do not claim that a media file was generated or delivered because no writable output can be attached.',
+        'If the user requests generated media, explain that workspace write access must be enabled before a real file can be delivered.',
+      ].join('\n')
+}
+
+module.exports = { deliveryContract, responseLanguageFromText, responseLanguagePrompt }

desktop/test/agents/agent-output-importer.test.cjs

@@ -326,3 +326,26 @@ test('preserves the existing four-media attachment import limit', (t) => {
   const imported = importAgentOutputs({ workdir, baseline, startedAt }, store)
   assert.deepEqual(imported.map(item => item.name), ['e.png', 'd.png', 'c.png', 'b.png'])
 })
+
+test('reported paths import only files written during the run, outside input staging', (t) => {
+  const { workdir, store } = fixture(t)
+  const existing = path.join(workdir, 'README.png')
+  fs.writeFileSync(existing, PNG)
+  const old = new Date(Date.now() - 60 * 60 * 1000)
+  fs.utimesSync(existing, old, old)
+  const staged = path.join(workdir, '.meldwork-input', '.run-1', 'input.png')
+  fs.mkdirSync(path.dirname(staged), { recursive: true })
+  fs.writeFileSync(staged, PNG)
+  const baseline = captureAgentOutputState(workdir)
+  const startedAt = Date.now()
+  const written = path.join(workdir, 'charts', 'result chart.png')
+  fs.mkdirSync(path.dirname(written), { recursive: true })
+  fs.writeFileSync(written, PNG)
+
+  const imported = importAgentOutputs({
+    workdir, baseline, startedAt, reportedPaths: [existing, staged, written],
+  }, store)
+
+  // A file the Agent only mentioned (or the user's staged input) is not an output.
+  assert.deepEqual(imported.map(item => item.name), ['result chart.png'])
+})