You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Adds durable long_run_start/status/list/update/review/complete/cancel tools for multi-phase work.
Persists objective, plan steps, acceptance criteria, evidence, checkpoints, review rounds, rework, background task ids and completion state under .ai-bridge/long-runs/<run_id>.json.
A step cannot be marked done without evidence. Any meaningful post-review work increments workRevision and invalidates an older PASS until the current revision is reviewed again.
FAIL reviews must identify failed criteria and actionable rework, reopen affected steps, and remain bounded.
Background bash tasks can be attached to a long run. PASS/completion is fail-closed while tasks are running/cancelling; after a process restart an unknown task must be explicitly resolved with terminal evidence.
Long shell work remains execution-unbounded while MCP polling stays short (wait_task <=30s) with adaptive 5/15/30 second poll hints, avoiding multi-hour blocking tools/call requests in ChatGPT/Codex/browser hosts.
Local loop-handoff now defaults to a 2-hour executor timeout, 1-hour reviewer/test timeouts and five evaluator/rework iterations, with durable phase/timestamp/exit/terminal-reason state.
Native MCP io.modelcontextprotocol/tasks remains an optional capability-negotiated future mapping; the ordinary-tool durable fallback works with current hosts that do not advertise that extension.
Existing v0.8.3 guarantees retained
All READY project roots are equal active roots; no entry/bootstrap project ownership.
Local and public modes both route through the shared Gateway; Gateway port/Bearer are Hub-scoped, while projects only keep internal CodexPro/Windows-MCP ports.
Windows custom install directory and Linux/SteamOS user-level build/install/live-upgrade support remain intact.
Verification
Python: 342 passed / 2 skipped.
Ruff: pass.
Pyright: 0 errors / 0 warnings.
CodexPro full smoke: pass, including durable long-run smoke, main stdio long-run flow, root scan (569 entries / 3 permission warnings), nested Git, async task, handoff and release guard.
CodexPro stress: pass.
npm production audit: 0 vulnerabilities.
Windows local full build: pass.
GitHub Actions from exact v0.8.4 tag / commit c96234dfe0a7eb271fb2f7b5f1bb2ea86109ae0c: Linux/SteamOS 3m32s pass; Windows 5m24s pass.
