fix: retry failed analyses (no silent drops) + bump Bun to 1.4.2 #2

Merged

nick merged 1 commit from bugfix/retry-failed-analyses into main 2026-09-27 17:33:46 +02:00

Conversation 0 Commits 1 Files changed 17 +1203 -66

nick commented 2026-09-27 17:28:52 +02:00

Owner

Copy link

A file whose analysis FAILED (LLM timeout, download error, …) was marked processed with NO record row and was never re-analyzed — the 2026-09-23 WhatsApp appointment screenshot was dropped this way (26 files total, root cause: transient LLM TimeoutError).

What this does

• Marker state machine on processed_files (NULL=done, a:<n>=retry-pending, ae:<n>=exhausted): a failed file is NOT marked done and is re-analyzed on every poll tick until it succeeds or FILE_RETRY_MAX_ATTEMPTS (default 3) is used up.
• Visible error audit row per failed file (rec_<id>, status error, "analysis failed (attempt n): …") — new Errors filter in the dashboard.
• Retry queue in onedrive-poll: marker-driven, re-fetches each file's current Graph state by id (delta never re-delivers unchanged items); 404 → done + row retired.
• Operator actions: "Retry analysis" (POST /records/repending, re-arms the marker; double-clicks don't consume budget) and "Give up" (POST /records/giveup).
• One-time startup backfill (CAS-gated in kv): processed-but-rowless analyzable files are re-queued — this recovers the 26 silently dropped files, including the WhatsApp screenshot, on deploy.
• Bun 1.4 → 1.4.2 (Dockerfile + CI + .mise.toml): the 2026-09-23 container segfaults were on musl/alpine; 1.4.2 fixes a rare JIT crash and a GC crash on musl.
• bun test --parallel in the test script (per-file worker isolation; the process-global mock.module leaked across files).

Tests

• test/file-retry.test.ts (10 new): marker state machine, repending/giveup, auto-reject sweep skip, backfill, fail → retry-fail → retry-success and exhaustion end-to-end.
• Full suite: 224 pass / 1 skip / 0 fail; tsc --noEmit clean.

Spec updated: specs/appointment_detection.md ("Failed-analysis retry (no silent drops)").

A file whose analysis FAILED (LLM timeout, download error, …) was marked processed with NO record row and was never re-analyzed — the 2026-09-23 WhatsApp appointment screenshot was dropped this way (26 files total, root cause: transient LLM TimeoutError). ## What this does - **Marker state machine** on `processed_files` (`NULL`=done, `a:<n>`=retry-pending, `ae:<n>`=exhausted): a failed file is NOT marked done and is re-analyzed on every poll tick until it succeeds or `FILE_RETRY_MAX_ATTEMPTS` (default 3) is used up. - **Visible error audit row** per failed file (`rec_<id>`, status `error`, "analysis failed (attempt n): …") — new **Errors** filter in the dashboard. - **Retry queue** in `onedrive-poll`: marker-driven, re-fetches each file's current Graph state by id (delta never re-delivers unchanged items); 404 → done + row retired. - **Operator actions**: "Retry analysis" (`POST /records/repending`, re-arms the marker; double-clicks don't consume budget) and "Give up" (`POST /records/giveup`). - **One-time startup backfill** (CAS-gated in `kv`): processed-but-rowless analyzable files are re-queued — this recovers the 26 silently dropped files, including the WhatsApp screenshot, on deploy. - **Bun 1.4 → 1.4.2** (Dockerfile + CI + `.mise.toml`): the 2026-09-23 container segfaults were on musl/alpine; 1.4.2 fixes a rare JIT crash and a GC crash on musl. - `bun test --parallel` in the test script (per-file worker isolation; the process-global `mock.module` leaked across files). ## Tests - `test/file-retry.test.ts` (10 new): marker state machine, repending/giveup, auto-reject sweep skip, backfill, fail → retry-fail → retry-success and exhaustion end-to-end. - Full suite: 224 pass / 1 skip / 0 fail; `tsc --noEmit` clean. Spec updated: `specs/appointment_detection.md` ("Failed-analysis retry (no silent drops)").

nick added 1 commit 2026-09-27 17:28:52 +02:00

fix: retry failed analyses (no silent drops) + bump Bun to 1.4.2 ... ed4c394a84

A file whose analysis FAILED (LLM timeout, download error, ...) was marked
processed with NO record row and was never re-analyzed — the 2026-09-23
WhatsApp appointment screenshot was dropped this way (26 files total).

Fix:
- processed_files gains a marker column: NULL = done (legacy), 'a:<n>' =
  retry-pending after n failures, 'ae:<n>' = failed + budget exhausted.
- file-analysis process step is state-driven: on failure the file is NOT
  marked done; a single audit row (stable id rec_<id>, status 'error',
  'analysis failed (attempt n): ...') is written; on success the same row
  is upserted with the real classification and persist FORCES the
  pipeline status (so a retry clears the stale 'error').
- onedrive-poll collects the retry queue (marker 'a:<n>') each tick,
  re-fetches each file's current Graph state by id (delta never
  re-delivers unchanged items) and merges it into the tick; 404 → done +
  row retired.
- Budget: FILE_RETRY_MAX_ATTEMPTS (env, default 3) — a permanently broken
  file stops burning an LLM call every 5 min and is left to the operator.
- Operator UI: Errors filter; 'Retry analysis' (POST /records/repending on
  an error row: row → pending + marker re-armed; double-clicks don't
  consume budget) and 'Give up' (POST /records/giveup: row → rejected).
- auto-reject startup sweep skips retry-queued rows (pending rows whose
  raw summary is 'analysis failed ...').
- One-time startup backfill (backfill-dropped.ts, CAS-gated in kv):
  processed entries that are 'done' but have no rec_<id> row (and an
  analyzable name) are re-queued for re-analysis — recovers the 26 files
  silently dropped before this fix.
- bun test now runs with --parallel (per-file worker isolation; the
  process-global mock.module used to leak across files).

Bun 1.4 → 1.4.2 (Dockerfile + CI + .mise.toml): the 2026-09-23 container
segfaults were on musl/alpine; v1.4.2 fixes 'a rare JIT crash' and 'a GC
crash on musl' — the same failure class.

Spec: specs/appointment_detection.md gains the 'Failed-analysis retry
(no silent drops)' section; the poll-loop pseudocode now shows the retry
queue.

Tests: test/file-retry.test.ts (10) covers the marker state machine,
repending/giveup, the sweep skip, the backfill, and fail→retry-fail→
retry-success / exhaustion end-to-end. Full suite green (225 tests,
bun test --parallel) + tsc --noEmit.

nick merged commit 9d111cc871 into main 2026-09-27 17:33:46 +02:00

nick deleted branch bugfix/retry-failed-analyses 2026-09-27 17:33:46 +02:00

nick referenced this pull request from a commit 2026-09-27 17:33:47 +02:00

Merge pull request 'fix: retry failed analyses (no silent drops) + bump Bun to 1.4.2' (#2) from bugfix/retry-failed-analyses into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

home/agent!2

No description provided.