Prevent group completion from masking peer blockers, avoid replaying delivered group context, validate Pi protocol capability, and restrict child environments while preserving supported runtime configuration. Update the technical iteration plan and related discoverability documentation.
Recover reused ACP sessions that refuse before producing output so generic workspace recovery can establish a fresh session. Keep Pi processes alive through internal tool turns until agent_end, and extract V4 natural discussion decision methods into a dedicated module without changing the runner API or persisted structures. Verified with focused CLI, natural discussion, workspace recovery, syntax, diff, and full desktop test suites.
Use the Qwen-generated bitmap with rasterized English install guidance, keep Finder root resources hidden by repointing the background into the signed app bundle, and update packaging coverage. Verified with the full desktop suite, focused packaging tests, codesign, and hdiutil.
Use a compact 640x420 branded background and reposition the Finder icons so the installer opens without excess canvas or scrollbars. Keep key content clear when hidden DMG resources are visible.
Checks: node --test test/packaging/package-security.test.cjs; npm run dist; npx electron-builder --mac dmg; hdiutil verify; codesign --verify --deep --strict; git diff --check
Bump desktop/frontend package metadata to 0.1.5 and document CLI detection, OpenClaw migration, generic group routing and recovery improvements with their evidence boundaries.
Validation: frontend 343/343; desktop 1562/1562; deterministic eval 6 cases/18 results; web and desktop builds; pack/dist; DMG, ZIP, checksum and ad-hoc signature integrity; packaged Codex/OpenClaw group, cancellation and restart acceptance. No push, merge or installed-app replacement.
Clarify that completion receipts are stripped transport metadata, preserving user-visible exact output. Include validated dynamic Vertex model region keys in shell collection, cache identity and native Claude environment without changing packaged Electron fuses.
Validation: natural discussion/receipt 56 passed; final readiness/main 91 passed; real Codex and OpenClaw exact-answer runs completed; real Electron shell extraction passed. Full release checks remain pending.
Remove keyword-triggered shared media calls, reply-driven retries and implicit ZGCI fallback from chat orchestration. Preserve native Agent output import and visible responses.
Validation: desktop 1554/1554; five workflow routing regressions; real Electron completed with two Codex calls and zero shared media calls. Document the separate missing completion-receipt sample.
Bind clarification gates to completed task decisions and preserve them across restart. Make repeated gate responses idempotent to prevent invocation replay.
Validation: desktop 1549/1549; real Electron pending restart, input response, completion and duplicate-response checks passed with exactly one continuation call.
Use persisted taskDecision.nextKinds instead of prose mentions so negated and historical references cannot dispatch peers. Normalize identifier case generically, validate active recipients, and retain owner judgment and legacy ledger compatibility.
Validation: desktop suite 1542/1542 with native Claude integration; final protocol and discussion tests 37/37 after aligning the recipient limit to 32. Real Electron Codex/OpenClaw write, readback, completion and history reload passed. Frontend builds and release packaging remain pending.
Bind cached settings to allowed environment values and reload the login shell before serialized Agent refreshes so changed credentials and proxies do not reuse stale settings.
Validation: readiness and main security 87/87; Agent suite 459 passed, 1 optional integration skipped; isolated Electron refresh observed updated synthetic configuration with all 12 Agent entries preserved. Full desktop/build/package checks remain pending for this checkpoint.
Project minimal latest-run outcomes for visible topics through the workspace snapshot. Prefer those outcomes over historical stop messages and Agent call statuses in the topic rail and retained run card, preserving historical errors and private ledger boundaries.
Validate 1536 desktop tests, 343 frontend tests, desktop build, and real Electron reloads of recovered, completed, and blocked histories without additional Agent calls. Keep remaining V1.0.5 work explicit in the development record.
Track acknowledged source messages per native session and resend uncertain or changed input. Sanitize task decision text before hashing public receipts so absolute deliverable paths cannot strand completed group runs. Keep original Agent decisions for recovery.
Validate 1535 desktop tests including real Claude provider routing, 84 focused discussion checks, real Electron Codex/OpenClaw delivery, and recovery of the prior failed run without additional Agent calls. Synchronize the ACP stream test on received events. Record remaining recovery UI and release work.
Forward Agent-scoped cloud configuration and explicit empty overrides. Keep configured Provider credentials isolated and use Claude's native key helper to prevent settings from overriding the selected endpoint or exposing secrets in argv. Treat cloud/API-key status as configuration evidence so refresh does not erase recent auth failures.
Validation: desktop suite 1525/1525 before final credential-evidence classification; final readiness/catalog regressions 53/53; native Claude helper and endpoint tests 3/3; isolated Electron Provider save/refresh/chat passed; git diff --check passed. See progress document for platform and cloud-account limits.
Separate paid problem validation, software contribution, repeat sales, and resilience to upstream changes so service revenue does not imply software demand.
Validation: git diff --check -- docs/product-iteration-plan.md passed; documentation-only change.
Recover long replies from hash-verified, run-bound conclusion artifacts so closing peer requests survive message truncation. Bound shared history with explicit omissions and prevent incomplete excerpts from proving repetition.
Clarify read-only proposal deliverables and require external peer requests to return control to Meldwork. Update the native-session fixture to explicitly continue its requested rounds.
Verified: 1515 desktop tests, focused context and recovery tests, real Electron Codex/OpenClaw write/readback/completion, and git diff --check. Document failed validation samples and remaining release work.
Prevent stale readiness in the installer catalog and preserve a successful scan when catalog loading fails. Invalidate skills before publishing readiness so reactive loaders retain their new requests.
Verified: 339 frontend tests, desktop frontend build, real Electron unavailable-agent/settings/restart checks, and git diff --check. Record remaining release work in docs.
Separate sidebar preference from invocation readiness. Show localized failure reasons without enabling new runs, retain history access and preserve explicit hiding across recovery and restart.
Validation: frontend 334/334; related desktop 146/146; desktop frontend build; git diff --check. Real Electron verified unavailable Pi visibility, settings navigation, no chat creation and hidden preference after restart.
Share the network environment contract across login shell collection, discovery, credential probes and child execution, including isolated OpenClaw. Preserve explicit empty overrides and redact proxy credentials in diagnostics and streaming output.
Validation: 158 focused checks; Agent suite 447/447; final network/stream checks 34/34; local shell-to-curl proxy request; two real Electron scans retained 12 installations; git diff --check. Full release checks remain pending.
Freeze the capability-selected writer across snapshot, slot, plan and invocation. Require durable one-shot approval before replaying an unknown write and preserve pending approvals across shutdown. Validate capability and permission bindings, completed-result reuse and repeated crashes.
Validation: 8 natural writer tests, 139 existing ledger/schema/gate checks, 159 automatic/manual/receipt checks; git diff --check. Real Electron Codex/OpenClaw automatic task wrote and independently read back exact file bytes. Full release verification remains pending.
Identify the collector hash input and distinguish retained extracted text from the unarchived HTTP response. Verified 30 cited sources, five document link sets, report structure and git diff --check. No runtime changes.
Replace natural-discussion silence heuristics with delivery-owner decisions and explicit handoff. Preserve bound decisions through harness, ledger and journal recovery; reject fabricated responsibility receipts. Allow single-agent auto submission and avoid showing task success after a blocking judgment.
Validation: 52 natural/receipt/parser tests, 136 ledger/harness/record tests, 91 automatic workflow tests, 332 frontend tests; desktop build and diff check. Real Electron with native Codex verified completed and blocked tasks plus reopened history. Automatic writable recovery and remaining readiness work are still pending.
Tie defensibility claims to measurable customer value as native agents improve. Retain the existing research snapshot and distinguish proposed experiments from market evidence.
Validation: git diff --check -- docs/product-strategy.md. Documentation-only change; runtime tests not run.
Keep unreadable Provider credentials distinct from native login failure and prevent late runtime success from clearing that state. Preserve installed peers and display localized diagnostics. Checks: 71 main/catalog tests, 328 frontend tests, desktop build, Electron failure/recovery and diff check.
Clarify evidence limits, durable value hypotheses, scoped pilot offer and investment gates. Validate local document links, source references and git diff --check.
Refresh guarded config identity only after equivalent startup migration, including failed startup retries. Retain registered agents and per-agent probe errors. Checks: 89 focused desktop tests; 327 frontend tests; web/desktop builds and pack; real OpenClaw fixture and Electron catalog refresh. Full desktop suite still has two restart timeouts, isolated rerun passes.
Distinguish customer-owned work records from company advantages and compare against native agents with ordinary documents. Define a bounded noncoding review offer before investing in subscriptions. Checks: git diff --check; research JSON and local Markdown links validated.
Archive market sources, user feedback and corrected positioning; separate evidence from pricing and defensibility hypotheses.
Checks: staged diff whitespace, research JSON parsing and core document links. Source body hashes could not be verified and are documented as a limitation. No runtime code changed.
Distinguish discovery timeouts from missing installations, preserve native and encrypted provider readiness, correct Pi stream handling, and reuse group Agent sessions across rounds and restarts. Support inline peer routing with fairness and repetition guards.
Validation: desktop 1454 tests, frontend 326 tests, web and desktop builds, desktop pack, real Electron Pi/Claude multi-round and restart session checks; git diff --check.
Run sequential mode strictly in configured CLI order, let auto-beta route later turns from canonical mentions, and persist one native Agent session per task across rounds and restarts. Remove rigid review framing from natural prompts and add recovery, routing, and continuity coverage.
Ryder Sun