Add connector-owned event profiles, chunk-safe redaction, terminal answer reconciliation, and renderer lifecycle handling so supported local Agents can stream text and expose sanitized tool progress consistently.
Keep Hermes sessions on persistent ACP runtimes and route OpenClaw through an authenticated isolated Gateway with bounded legacy fallback. Preserve idempotent Harness and renderer updates when authoritative final output replaces provisional deltas.
Verified with frontend tests and both frontend builds, desktop tests (1064/1064), desktop packaging, and git diff --check.
Keep retained system trace controls consistent with the message action hover behavior introduced for user and Agent messages.
Verification: covered by prior full frontend, desktop, build, and pack checks in this working tree.
Allow ready OpenCodeReview installations to appear in the sidebar and be opened as direct conversations for user review instructions.
Route explicit OpenCodeReview sends through the read-only code-review runtime so writable direct conversation defaults do not create a capability mismatch, while continuing to avoid native session persistence.
Verification: npm --prefix frontend test; npm --prefix desktop test; npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack.
Raise local attachment handling to the shared safe limit, render document/media attachments with clearer icon treatment, and keep message action controls grouped on hover with footer actions aligned to the right.
Verification: npm --prefix frontend test; npm --prefix desktop test; npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack.
Bump the desktop and frontend version to 0.1.1, refresh the macOS download links, and document the reliability and review improvements users will receive in this release.
Verified: frontend tests (263 passing); frontend production and desktop builds passing; desktop tests 991 passing with 2 pre-existing macOS CLI cancellation helper failures.
Switch the watchdog from wall-clock expiry to inactivity-based stall detection so long-running agents with observable progress do not fail spuriously.
Verified with: npm --prefix desktop test
Remove visible scrollbar styling across the app shell and conversation surfaces while preserving scroll behavior.
Verification: npm --prefix frontend test -- --run; npm --prefix frontend run build.
Inject an evidence-first adversarial review contract into every unlimited group Agent, including task-graph reviewers and arbiters. Require one complete cross-review pass before ordinary unlimited discussions can converge, and explain the stricter behavior in the bilingual confirmation, round selector, and composer placeholder.
Checks: npm --prefix frontend test (262 passed); npm --prefix desktop test (992 passed); npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack; git diff --cached --check.
Scope direct and group run state by conversation so one active chat no longer blocks another, keep completion indicators and notifications tied to the finished conversation, and stack multimodal attachments to avoid row layout breakage.
Improve local Agent readiness detection for native credentials, OpenClaw model status, shared Provider handling, and unsupported-version recovery while preserving credential secrecy.
Reorganize desktop main-process modules by domain and update README installation guidance for unsigned macOS preview builds.
Verification:
- npm --prefix frontend test
- npm --prefix desktop test
- git diff --check
Detect local Agent provider settings from the user's login shell and known config locations, prefer capability probes over stale version ranges, and reuse the same safe environment for execution. Isolate send state per conversation, add outcome-specific notifications, and document the macOS Open Anyway flow.
Checks: frontend 257 tests; desktop 454 tests; web and desktop builds; desktop pack; isolated packaged-app detection and Codex conversation smoke.
Replace the Smart Team off-state glyph with a manual routing control mark and make the tooltip explain the off/on behavior more concretely.
Checks run: npm --prefix frontend test; npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack; git diff --check.
Replace the icon-only automatic routing toggle with an explicit bilingual segmented control for selected Agents and smart team formation. Keep the existing routing behavior, Agent avatar selection, and automatic discussion controls independent.
Add focused routing-state assertions and verify responsive desktop rendering in English and Chinese.
Checks: npm --prefix frontend test (253/253); npm --prefix desktop test (962/962); npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack; packaged Electron 39.8.10 runtime at 1440x920 and 760x820 with no target-row overflow.
Add a versioned meldwork.skill.json contract covering identity, Agent compatibility, inputs, tool classes, credentials, permission mode, HTTPS destinations, and side-effect class.
Require native first-use approval for unsigned local Skills, bind trust to Agent coordinates plus manifest and snapshot hashes, persist a private hash-chained audit, and enforce compatibility and privilege checks before execution.
Checks: npm --prefix frontend test; npm --prefix desktop test; npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack; packaged Electron 39.8.10 runtime readyState=complete at 1440x920 with localSkillTrust list/revoke preload API available.
Replace core-only Connector manifest approval with a local content-addressed package lifecycle backed by a private hash-chained audit log. Add native approval, strict IPC/preload operations, bounded local echo and HTTPS JSON SDK providers, and fail-closed startup behavior without exposing paths, credentials, or shell execution to the renderer.
Add an installable non-delegate sample, SDK documentation, and a connector:conformance suite covering parsing, trust transitions, upgrades, capabilities, events, cancellation, resume, credentials, outbound enforcement, tamper handling, and failure semantics.
Checks: npm --prefix desktop run connector:conformance (25/25); npm --prefix desktop test (962/962); npm --prefix frontend test (253/253); npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack; packaged Electron 39.8.10 runtime readyState=complete at 1440x920 with preload API available.
Add strict content-addressed task graph definitions and a v3 restart-safe cursor for opt-in automatic workflows. Execute independent read-only branches in parallel, isolate Reviewer and Arbiter sessions, evaluate completion from durable Artifact and Evidence refs, and route conflicts through typed review, arbitration, or Human decision nodes while preserving bounded round-robin compatibility.
Checks: npm --prefix desktop test; npm --prefix frontend test; npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack; npm --prefix desktop run eval:deterministic; packaged Electron CDP runtime smoke.
Add strict content-addressed Eval Case, Result, and Fit Matrix records with deterministic scoring, identified reviews, local storage, and provider-opt-in execution.
Gate automatic routing on frozen evidence with at least three samples and 0.6 confidence so the synthetic CI corpus cannot change production selection.
Checks: desktop 947/947; frontend 253/253; deterministic eval 6 cases/18 results; web and desktop builds; desktop pack; packaged Electron 39.8.10 runtime.
Persist content-addressed Handoff and selective Blackboard records in automatic Run orchestration v2 while retaining exact v1 recovery compatibility. Deliver only role-authorized collaboration entries to Primary, Reviewer, and Arbiter Agents and preserve conflicting claims as append-only records.
Checks: npm --prefix desktop test; npm --prefix frontend test; npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop run pack; packaged Electron 39.8.10 CDP smoke at 1440x920.
Identify CLI versions by product, expose layered readiness, and require guarded OpenClaw runtime prerequisites before routing. Verified with npm --prefix desktop test (921/921), npm --prefix desktop run pack, and isolated packaged Electron CDP acceptance.
Persist Connector input and permission waits as restart-safe Human Gates, release scheduler capacity while suspended, and reacquire it before continuing the exact Agent attempt. Validate Run, Agent, Connector, operation, request, session, and Session provenance bindings before resume; quarantine malformed gate records and keep private text responses out of renderer-visible decisions.
Checks: desktop 912/912; focused desktop 128/128; frontend 251/251; web and desktop builds; desktop pack; packaged Electron CDP smoke; git diff --check.
Add a non-disableable 128-attempt circuit breaker, make hard budget exhaustion terminal and durable, and expose distinct restart-stable UI states. Preserve the triggering action's attempted usage and stop every later Agent call, including transient retry paths.\n\nChecks: npm --prefix desktop test (902/902); npm --prefix frontend test (249/249); frontend builds; desktop pack; packaged Electron CDP smoke.
Gate run-finished acknowledgement on a durable terminal ledger commit, retain retryable typed persistence state, and preserve every nonterminal Run under retention pressure.
Checks: desktop 894/894; frontend 249/249; web and desktop builds; desktop pack; packaged Electron CDP smoke.
Persist versioned manual and automatic orchestration cursors so approved Human Gates resume the exact Agent slot and continue remaining work without rerunning completed slots. Hydrate prior Harness traces and durable event cursors to keep post-restart events monotonic.\n\nChecks: npm test (desktop, 890 tests); focused Human Gate, auto-runner, Harness, Run Ledger, journal, and coordinator suites (208 tests).
Distinguish missing workspace files from corrupt, unreadable, and unsupported state. Untrusted state now enters read-only recovery and skips every startup reconciliation path that can mutate Run Ledger or Human Gate data.\n\nChecks: node --test test/local-workspace-state.test.cjs test/local-workspace-harness.test.cjs test/local-workspace.test.cjs test/main-security.test.cjs
Preserve the main-process sandbox decision when dispatching legacy Custom Agents and reject every invocation that is not explicitly workspace-write before process spawn.\n\nAdds direct, group, main-process, and store-level regression coverage.\n\nChecks: node --test test/custom-agent-store.test.cjs test/main-security.test.cjs test/local-workspace.test.cjs
Refresh the bilingual README around current local Agent discovery, direct multimodal output, and targeted group collaboration, with three isolated product screenshots arranged as a horizontal capability gallery.
Bound quit cleanup before terminating Electron and block activation or notifications from recreating a window during shutdown, so Cmd+Q exits cleanly and the app can reopen normally.
Checks: npm --prefix frontend test; npm --prefix frontend run build; npm --prefix frontend run build:desktop; npm --prefix desktop test; npm --prefix desktop run dist; hdiutil verify; unzip -t; git diff --check.
Ryder Sun