Auto-review local signal is lost across restart

[cbusillo]

Summary

Auto-review did not provide a reliable visible signal for PR #348 (Fix resume replay after history snapshots) at merge-decision time.

PR: #348
Head commit: 4db78b88dc012a27c66b95d055221107c87277d9
Merge commit: f034774cbc2ab39b61ac71a29094a62e080869c1

Updated diagnosis

This is not currently a GitHub PR-commenting/review-posting failure. The codebase appears to implement auto-review as a local TUI background agent flow:

• run_background_review creates an Every Code agent with AgentSourceKind::AutoReview.
• TUI observes agent status and writes ~/.code/state/review/.../auto-review/runs.json.
• Results surface through TUI footer/notices and the local review ledger.
• I did not find a code path that posts auto-review output as a GitHub PR comment or GitHub review.

The concrete issue is that the local/TUI auto-review signal can disappear or be stale around restarts, so the user cannot tell whether the current PR/change was reviewed.

Evidence from PR #348 session

GitHub state after #348:

• gh pr view 348 --json comments,reviews,statusCheckRollup,reviewDecision showed no PR comments and no reviews.
• GitHub Blob size policy completed successfully.
• PR was mergeable and later merged.

Local auto-review store for this repo (~/.code/state/review/repo-a093bd7f/auto-review/runs.json) did contain chatwidget.rs runs from the same session, but the latest relevant run was lost after restart:

{
  "run_id": "1e5138bb-ee8a-4e1c-80c0-eb360d7b631c",
  "status": "lost",
  "freshness": "lost",
  "owner_session_id": "8e18be4f-ff91-4264-a402-d89aa6741cbc",
  "agent_id": "7bfd3c0a-41c7-451a-83ee-460c312947f1",
  "branch": "auto-review",
  "snapshot_commit": "12b9231d1c8fb41bc5a84a9018450119d69c9677",
  "base_commit": "45e92129fb94d3bdf24d014a8f7a18ce4fbb3691",
  "changed_path_count": 1,
  "listed_paths": ["code-rs/tui/src/chatwidget.rs"],
  "cancel_reason": "agent_missing_after_restart"
}

There were also older chatwidget.rs auto-review runs in the store, including completed/cancelled runs, but none provided a clear current review signal for the final PR state after restart.

Expected

After restart, Every Code should preserve or clearly restore the review state enough that the user can answer:

• Did auto-review run for the current change/snapshot?
• Is the latest review current, stale, lost, or superseded?
• If lost, should the TUI surface a retry/notice instead of silently leaving the user to infer from the local store?

Candidate fix direction

Investigate the restart reconciliation path:

• ChatWidget::reconcile_auto_review_store_after_restart() currently calls reconcile_orphaned_in_flight(std::iter::empty::<&str>(), now), which marks all in-flight review agents as lost because no live agent IDs are supplied.
• Determine whether restart can recover live agents from AGENT_MANAGER, or whether the TUI should surface a visible “auto review lost after restart; rerun review” notice/status for the latest relevant run.
• Consider whether merge-decision workflows should consult the local auto-review ledger instead of expecting GitHub comments/reviews.

Validation already done for PR #348 code fix

• cargo test --manifest-path code-rs/Cargo.toml -p code-tui replay_snapshot --features test-helpers
• ./build-fast.sh
• local read-only agent reviews against the replay patch