# Task: Diagnose recurring Coding session stall

## Identity

- Task ID: 20260903-diagnose-recurring-stall-a83f5c21
- Mode: Feature
- Branch: codex/20260903-diagnose-recurring-stall-a83f5c21-diagnose-recurring-stall
- Worktree: /Users/inmanx/Documents/makelore-diagnose-recurring-stall-a83f5c21
- Base commit: 301c1496b0a3a1af3b9443a68a67de8dfea45749
- Owner: codex
- Status: Completed

## Scope

- Diagnose the recurring visible `处理中` state in the installed macOS app for
  project `肉鸽2.0`, including the active Pi logical thread that began around
  09:51 on 2026-09-03.
- Verify whether the installed artifact actually contains the previously integrated
  terminal-settlement and Bash tool-bridge fixes.
- Correlate the screenshot with the installed Main log, persisted Conversation/Pi
  session, target process state, and any active tool child before deciding whether
  this is useful work, a tool wait, a Provider failure, or another settlement gap.

## Intent And Constraints

- Keep diagnosis-time live inspection read-only: do not abort the target run,
  replay the accepted prompt, or mutate user project/session data. For the user's
  later explicit launch request, replace the old installed process only by a
  graceful quit after confirming the target had no remaining runtime/tool child;
  do not force-kill it.
- Preserve ADR-006 target-only failure and no-replay semantics.
- Do not touch the occupied `main` worktree or its uncommitted packaging changes;
  perform any source change only in this isolated worktree after evidence identifies
  a concrete defect.

## Outcome

- Confirmed that this recurrence was not an active model or Bash wait. The target
  Pi JSONL persisted an assistant terminal entry with `stopReason: "aborted"`
  and `errorMessage: "Request was aborted"` at 09:52:24 local time. Electron
  Main then completed the same Conversation with the intentional
  `project_deactivated` reason at 09:52:33. The remaining Agent Server was idle
  at 0% CPU with no tool child or network socket while Renderer still showed
  `处理中`.
- Confirmed the installed app already contains the earlier Bash write-lease and
  bounded terminal-settlement corrections. This incident instead repeated the
  foreground recovery gap documented by task
  `20260902-check-session-after-bash-fix-a6d12f39`: Renderer retained a cached
  `live` Snapshot after Main had stopped the target, and selecting an already
  selected Conversation skipped authoritative Snapshot hydration.
- Changed Conversation selection to always reconcile with Main. Existing live
  state refreshes silently, while cold and invalidated targets keep their prior
  loading/recovery behavior. Accepted or uncertain mutations are not replayed.
- Coding Chat now closes its old event stream on Main's `lifecycle:sleep`,
  refreshes the selected Conversation when the window becomes visible or focused,
  and refreshes a still-selected Conversation once when the programming view is
  mounted or its project context changes. A project/Conversation selection key
  prevents duplicate hydration caused by the selection render itself.
- Added store, Renderer, and Electron regressions that begin with a stale running
  Snapshot and verify that foreground recovery displays the persisted aborted
  terminal state without posting a prompt.

## Verification

- Correlated screenshot time, installed Main log, project registry, Conversation
  metadata, target Pi JSONL, and live PID/process/socket state. The durable turn
  ended before the visible timer continued; no running tool process remained.
- Installed `app.asar` contains `state_probe`, the bounded terminal-settlement
  failure string, and `PROMPT_FAILED_AFTER_ACCEPTANCE`, confirming the prior fix
  was packaged in the inspected app.
- Focused Vitest: `45 passed` across
  `coding-conversations-store.test.tsx` and `coding-chat-panel.test.tsx`.
- Full Vitest: `1736 passed`, `3 skipped` across 213 test files, including the
  single-worker pressure suite.
- Focused Electron E2E:
  `foreground focus rehydrates a terminal Snapshot after lifecycle sleep` passed.
- TypeScript `tsc --noEmit`, scoped ESLint, production `build:vite`, and
  `git diff --check` passed. Full `lint:check` passed with five pre-existing
  warnings and zero errors.
- `check_project_docs.py` and task-aware `check_doc_drift.py` passed.
- The old installed app was gracefully quit after confirming no target Agent
  Server or tool child remained. `pnpm run dev` then started the repaired
  worktree successfully, with the Renderer on port 5173 and Electron using the
  existing Makelore user-data directory.

## Follow-ups

- Integrate this feature branch after the occupied `main` packaging worktree is
  available, then rebuild/reinstall the macOS artifact. The currently installed
  `/Applications/Makelore.app` remains the older build; the repaired behavior is
  available from this task worktree for development verification.

## Promotion Candidates

- Target: `README.md`, `20-architecture/data-flow.md`, and
  `30-worklog/current-state.md` during Integration Mode.
  Proposal: record the Renderer foreground rehydration contract: lifecycle sleep
  closes the old Coding event stream, and view mount, project reselection,
  visibility, or focus reconciles the selected Conversation from a Main-owned
  Snapshot without replaying accepted mutations.
  Evidence: focused store/Renderer tests, real IPC Electron E2E, and the two
  installed-app stale-running incidents documented by this task and
  `20260902-check-session-after-bash-fix-a6d12f39`.
  Future impact: any new Coding navigation or lifecycle surface must preserve the
  same terminal-state reconciliation and no-replay behavior.
  Human confirmation: not required; this promotes implemented behavior without
  changing ADR-006 ownership or product direction.