Status ยท 9 November 2026
State reliability fixes
Failed work remains recoverable, and late activity no longer disappears after an idle scan.
What changes
- Seventeen inventory entries receive corrections; the other thirty-five retain their existing protocols.
- Delivery recovery, turn accounting, scheduled retries, and cached evidence now have regression coverage.
- Existing INSERT/SELECT database roles keep delivering. They queue turn corrections until the optional column grant is available.
- External deployments and source records remain outside this checkout. The limits section states what local verification cannot prove.
Resulting behavior
- Webhook-only collection writes unsupported rows to a private SQLite queue before acknowledging them. Status shows pending rows.
- Replay deletes queued rows only after database commit. Stale snapshots cannot replace fuller turn evidence.
- Cursor and Pi keep one aggregate per turn. Published timestamps and human prompt identities stay stable across resumed reads.
- The scheduler serializes runs, retries incomplete daily work, and closes only its own SSH child.
- Evaluation scores follow changed inputs. Missing clones stay undecided; negative PR answers remain eligible for a later lookup.
All 52 inventory entries
These were stateful mechanisms, not 52 independent defects. Retained means no concrete failure required changing that protocol.
| Inventory entry | Disposition | Evidence or reason |
|---|---|---|
| 1.1 Batch acknowledgement | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 1.2 Collector drain loop | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 1.3 Transport delivery protocol | Corrected | Durable delivery queue; explicit retry and visible pending counts. Regression tests |
| 1.4 Scan health and heartbeat | Corrected | Incomplete labels report a failed heartbeat and nonzero exit. Regression tests |
| 1.5 State recovery and ownership | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 1.6 Repository identity initialization | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 2.1 File tail lifecycle | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 2.2 Collector state migration | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 2.3 Session, model, and turn attribution | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 2.4 Copied and imported history boundary | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 2.5 Legacy prompt recognition | Corrected | Legacy loop state survives chunk boundaries. Regression tests |
| 3.1 Claude turn and link lifecycle | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 3.2 Interrupt and queued-prompt carryover | Corrected | Queued prompt hashes survive chunk boundaries and are consumed. Regression tests |
| 3.3 Pending tool-call correlation | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 3.4 Tool deduplication and repeated-call detection | Corrected | Current open-turn tool IDs survive long turns without a truncation cap. Regression tests |
| 3.5 Turn link evidence | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 3.6 Pi turn lifecycle | Corrected | Idle Pi snapshots preserve pending calls and remain correctable. Regression tests |
| 3.7 Cursor root-list turn reconstruction | Corrected | Late Cursor work revises covered aggregates with stable IDs and timestamps. Regression tests |
| 4.1 Automated Cursor registration | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 4.2 SDK terminal-run acknowledgement | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 4.3 Cursor API catch-up and pagination | Corrected | Capped windows split; indivisible saturation refuses acknowledgement. Regression tests |
| 5.1 ActivityWatch settled-range cursor | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 5.2 ActivityWatch retention backlog | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 5.3 macOS lookup circuit breaker | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 5.4 Active-time block builder | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 6.1 Label task window and acknowledgement | Corrected | Missing episode rebuilds and budget cuts remain retryable. Regression tests |
| 6.2 Cached label unit lifecycle | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 6.3 Prompt and reply interval attribution | Corrected | Ambiguous Cursor prompt context is omitted rather than assigned incorrectly. Regression tests |
| 6.4 Evidence snapshot cache | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 6.5 Current-label selection | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 7.1 Frozen historical replay | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 7.2 Recent/mature outcome coverage | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 7.3 Local label generation clock | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 8.1 Local ensemble batch lifecycle | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 8.2 Lazy model resource lifecycle | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 8.3 llama-server process lifecycle | Corrected | Dead servers restart on the next request; spawn failures close resources. Regression tests |
| 8.4 Teacher budget and failure breaker | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 8.5 Teacher and evaluation progress | Corrected | Scores require the same complete request hash, including its revision. Regression tests |
| 9.1 Gold annotation queue and undo | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 9.2 Local HTTP save and shutdown | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 10.1 Commit outcome lifecycle (named states) | Corrected | A missing clone produces undecided, preserving prior known outcomes. Regression tests |
| 10.2 Outcome uncertainty and cache retention | Corrected | Negative PR answers cannot permanently hide a later merge. Regression tests |
| 10.3 Git source coverage lifecycle (named statuses) | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 10.4 Daily Git ledger acknowledgement | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 10.5 Git-days due and retry scheduler | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 10.6 Remote Git output parser | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 10.7 Commit shell-directory parser | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 11.1 Scheduled scan retry marker | Corrected | Kernel locking prevents overlap; tunnel failures support systemd retry. Regression tests |
| 11.2 Once-per-day outcomes and labels | Corrected | Daily markers record success; incomplete work retries after the next scan. Regression tests |
| 11.3 SSH tunnel ownership | Corrected | Cleanup targets only the owned tunnel process on normal exit and signals. Regression tests |
| 11.4 Backup attempt and last-good status | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
| 11.5 Backup retention and weekly prune | Retained | Retained protocol or source constraint. Existing regression coverage remains in the full suite. |
Verification
946 tests passed in 572.76 seconds on macOS with Python 3.13, including PostgreSQL 16 integration tests. No tests were skipped in this run.
Ruff 0.15.11 passed. Mypy 2.4.1 passed across 58 source files with optional installed packages excluded, matching the minimal CI type-checking scope. Bash syntax, whitespace, documentation links, and HTML validation passed. Two independent reviews have no unresolved findings in the corrected paths.
The final report was inspected at desktop and phone widths. The phone viewport has no page overflow, and all 52 disposition rows are present.
Regression cases failed before their corresponding fixes. PostgreSQL 16 scratch tests also cover legacy timestamps, restricted INSERT/SELECT roles, deferred updates, and replay after a column grant.
Independent correctness and contract reviews found timestamp, machine-copy, Linux retry, heartbeat, and documentation gaps. Each receives a regression or a direct documentation correction.
Limits and rollout
Update the installed package before copying the new scheduled wrapper; the wrapper calls aut scan-job.
Apply the optional UPDATE(attrs) grant as the database administrator to enable covered corrections. Until then, normal delivery continues and corrections remain queued. Run aut retry-delivery afterward.
This snapshot records verification before publication. It does not report deployment status. Database checks used a disposable PostgreSQL instance; live grant and collector checks belong to the release verification.
Deleted source blobs cannot be reconstructed. Indivisible API page saturation stops safely for operator recovery. SIGKILL cannot execute cleanup. Source-coverage hashes grow with a turn.
Crash recovery still uses at-least-once delivery and stable sink identities. That is a required safety property, rather than a defect to remove.