- Merge finalization keeps no durable progress; cleanup runs before the task is closed
`internal/agents/mergein.go:303–324` runs push → notes → fast-forward → `RemoveTask` → `MergeTask`. A crash or error after the push leaves an active task with no worktree; a retried Merge fails in `Snapshot`, or squashes to no change and closes with an empty `merge_commit`. Have the `finalizing` marker record the squash commit; on startup, a recorded commit that is an ancestor of origin/<default> finishes notes, fast-forward, removal and MergeTask idempotently, otherwise the marker is cleared and an error is filed.
- A failed auto-commit still advances the workflow from the old HEAD
A `CommitAll` error is only logged (`internal/agents/manager.go:838–841`) and `FinishStage` still runs (`:885–889`), so `RecordStage` stores the old `git.Head` (`internal/agents/workflows.go:349–356`), the next stage branches without the turn's work, and `discardStages` later deletes the worktree that held it. On commit failure, treat the turn as unfinished: skip FinishStage, the handoff and the summary, and file the error in the task's inbox.
- Codex mid-turn delivery deletes input before the steer is accepted
Queued inputs are deleted or marked delivered and recorded as `user.message` (`internal/agents/queue.go:119`, `:158`; `internal/agents/inbox.go:404`) before `Steer`, whose error is only logged (`internal/agents/codex.go:167–174`). Return the consumed ids from `deliverMidTurn`; when Steer fails, re-insert the inputs in order and clear `delivered_at` so the turn-end Deliver sends them.
- Inbox answer validation isn't atomic with resolving the question
`CheckResponse` reads open state outside the write transaction and `Respond` updates by id alone (`internal/agents/respond.go:26–34`, `:93`; `internal/store/inbox.go:140`), so two concurrent answers both land, or a withdrawal is overwritten. Condition the update on `state='open'`, check RowsAffected, roll back the batch on conflict and return 409.
- A Codex RPC can deadlock the reader it waits on
`drive` calls `Steer` synchronously while `readLoop` pushes notifications into a 256-entry channel (`internal/codex/session.go:186`, `:440–476`, `:598`); if it fills before the response is read, each waits on the other. Route RPC responses separately from notification consumption and give RPCs bounded deadlines, without replacing backpressure with an unbounded queue.
- Agent startup holds the manager's global lock
`ensureLive` holds `m.mu` through process start and protocol init (`internal/agents/manager.go:538–629`), and `deliverMu` serializes every session (`internal/agents/queue.go:15–16`), so one slow start stalls health, interrupt and event handling everywhere. Coordinate starts per session, release m.mu during process I/O and publish the result under the lock, and narrow delivery serialization to the session.
- Git network operations have no deadline and run under a cross-project lock
`exec.Command` has no context (`internal/git/git.go:348–364`) and Merge holds the shared `mergeMu` through fetch (`internal/agents/mergein.go:68`, `:101`), so one stalled remote blocks merges in every project. Add contextual deadlines and non-interactive git/SSH, and serialize the network phase per project while keeping same-project ordering.
- A malformed file image URL breaks transcript rendering
`decodeURI` throws on e.g. `file:///tmp/100%.png` (`web/src/markdown.ts:347`), and `renderAll` appends events before mounting the composer (`web/src/task.ts:608–609`), so the rest of the transcript and the composer don't render. Catch the decode error and render the plain fallback link.
- Stale reload and focus responses overwrite newer state
`loadTasks`, `loadSessions`, `loadInbox` and `setFocus` apply every response, even a late one (`web/src/state.ts:158–179`, `:203–215`), so clicking through tabs quickly can snap the pane back. Keep a request sequence per resource key and for focus, and apply only the latest response.
- operations.md describes two incompatible event-stream designs
`docs/operations.md:320–324` says selecting tasks opens dedicated SSE streams, which is stale against `web/src/state.ts:309–340`, `:401–414`, and says only `agent.delta` is transient though `admin.health` is too. Rewrite the Events section to match the current two-stream feed design, and qualify the transaction/event guarantee at `docs/operations.md:7` for internal bookkeeping writes.
