Daily — 2026-07-25 (19 workflow-fix tasks merged; ~117 problems swept; 23 filed)

What happened

• 19 workflow-fix tasks merged to main in one day — #1667–#1683, #1687, #1688 — the wave filed by the 2026-07-24 /daily. 684 commits total. 6/6 and 9/9 sessions merged in the two biggest slices; nothing archived, nothing blocked, no human intervened in any of them.
• #1689 (the 21-condition speaker × framing lattice) burned the day. 8 implementer + code-review rounds, ~4 crashed launches (GCP dispatch-arg mismatch → GCP boot-loop → RunPod ImportError → interpreter-shutdown crash → vLLM empty-prompt), and a spawn-to-first-real-work gap of ~3 h from a turn-1 API 529. It is running now on pod-1689 (4×H100), relaunched 23:14 PT.
• No clean-results promoted today. Queue: 111 awaiting_promotion, 14 proposed (10 of them daily-held needs-human), 2 plan_pending, 1 running.
• Compute is clean: one live RunPod pod (pod-1689, the active run), zero live GCE instances (eps-issue-1112/1345/1689 all TERMINATED) — no leaked spend.
• Two inline analysis rounds landed: the #825 selector-audit settle battery (eval_results/issue_825/trackm_settle_battery/) and the #1345 mapping-similarity ladder rungs 2–8 (epm:free-analysis-followup-run on #1345).

Applied workflow improvements

No route-1 self-applied fixes tonight — and that is the honest outcome, not an omission. Route 1 is the narrow no-behavior-change floor (pure prose/string/format/ comment edits, typos, stale-reference renames). Every fix-derivable problem this sweep surfaced changes what an agent or a script does, so all of them routed to independent review. The one candidate that looked route-1 shaped — a wrong git status --porcelain --cached flag — turned out to be improvised in-session, not present anywhere in the workflow surface (grep → 0 hits), so there was nothing to fix.

19 route-2 tasks filed for independent review (each auto-dispatches to /issue --auto; same-target siblings file --no-dispatch under the #1678 hold and the watcher's proposed_infra_sweep picks them up):

1. filed for review #1690 — verified-at-filing: gains 3 probe clauses (marker-existence, call-hop target tracing, suppression-predicate). /daily's own filing quality: 3 of yesterday's filings carried premises their clarifiers refuted — #1667 ("no marker posted" — it existed), #1669 (named the caller, not the constructing site; the shipped diff touched 13 files, none of them the named target), #1680 ("a park was lost" — it was correctly suppressed).
2. filed for review #1691 — watcher pass escalating a multi-day Codex quota outage. 17 days dead fleet-wide, ~12 to go; ≥25 twin dispatches skipped on 2026-07-25 alone; zero reconciler invocations all day. Four independent miners surfaced it unprompted.
3. filed for review #1692 — smoke gate must execute real paths, not mock branches, and enumerate real-vs-fallback per planned arm. 3 consecutive #1689 rounds false-passed a fatal ImportError sitting behind --mock-response; a separate round shipped literal stubs for rungs 7–9 with a green 8-phase smoke.
4. filed for review #1693 — phase idempotency + inter-phase schema assertion. The money item: every #1689 relaunch re-ran the paid Sonnet haiku_u2 phase from scratch (~4 cycles); a downstream vLLM phase then died on 33% empty prompts after GPU spin-up; and a completed phase became a whole-pipeline abort under set -e.
5. filed for review #1694 — Step 10d content-apply verification + unconditional pre-sync. A squash claiming "applied as a 3-way patch (clean)" landed a test but not the extractor it tested — main was red for 20 days.
6. filed for review #1695 — refusal/transport-aware recovery (tick_triage STALE-REDRIVE on a fresh api-error row; watcher splits boot-refusal vs boot-transport). A 529 was labelled a refusal and cost #1689 ~3 h.
7. filed for review #1696 — guard-surface reviewers post the verdict marker BEFORE the final report. 4 refusal kills in one wave; 3 of them at the reviewer's final turn after 15–24 tool calls, destroying completed verdicts.
8. filed for review #1697 — --map-files timeout dispersion margin + sibling-family surcharge. Two green legs timeout-killed at a 780 s bound (measured 728.19 s and 751.70 s — 1.04× and 0.93×), ~69 min of pure re-run.
9. filed for review #1698 — RunPod launch-path contract (branch assertion, idempotent teardown, 25–50 min launch belongs in orchestrator bg-Bash, fence read from the instance). A 60 s experimenter exit stranded a half-provisioned 4×H100.
10. filed for review #1699 — implementer pin-sweep emitted mechanically + repo-wide invariants in the local union + ruff-policy parity. Under-enumerated hit lists in 4 of 6 sessions; a canonical-resolver violation reached the gate 27 min late.
11. filed for review #1700 — verify_plan exempt-kind satisfaction + commit-SHA resolution check. A hand-dispositioned WARN in 5 of 6 sessions, one re-firing across three plan versions.
12. filed for review #1701 — inline-round scientific duties (n<d guard-rail, estimator-vs-reference diff before launch, refuted-promoted-body must file same-turn). Two headline runs launched at n=1,877 vs d=3,584.
13. filed for review #1702 — thread --env-pin into the Step 6b composition rule.
14. filed for review #1703 — sweep emits an unmatched_record_fps advisory.
15. filed for review #1704 — reconcile HF crash bundles against committed eval_results/. #1345's base-model story leg was run and lost into a bundle.
16. filed for review #1709 — SPECS family syncs paired tests/test_guard_*.py. Hit 4 times today (#1679/#1680/#1681/#1682, ~81 min), all red on the same node, two sessions landing different ad-hoc remedies.
17. filed for review #1710 — repo-root branch guard false-positives on ssh remote payloads and quoted --note text. 4 blocks carrying no repo-root mutation at all — including one this /daily run tripped on its own fingerprint helper.
18. filed for review #1711 — closed-sibling probe in the /daily filing driver.
19. filed for review #1712 — audit for landed-but-unfolded rounds.

Three of these (#1691, #1704, #1712) target autonomous_session_watch.py as distinct bugs — the dedup grain is (target_file, fingerprint), so each gets its own review.

Compose-time verification killed three filings' premises before they shipped — the verified-at-filing discipline earning its keep:

• "code-reviewer returned PASS with 7 open BLOCKERs" → refuted: code-reviewer.md rule 11 prescribes exactly that (the Step 5c-ter gate reads concerns.jsonl, not verdict prose). Dropped.
• "--map-files applies no workflow-lint surcharge" → refuted by running the function (2580 s with the lint file present). The real cause — the surcharge key is exact-path and the map is 25 sibling files — is what #1697 was filed on.
• "the RunPod bootstrap hard-sets BRANCH=main" → refuted: the threading exists (runpod.py:733, bootstrap_pod.sh:52/223); main is the fallback, so the defect is an empty repo_branch upstream. #1698 says so.

Other problems & notes

• Route 3 (needs your call) — 4 filed, 1 already tracked:
  • #1705 — #825's promoted body asserts a claim its own round refuted. With the fit module's registered dof-cap guard engaged, all four M user cells flip from −1.43…−1.65 to +0.19…+0.25; the committed #825 default is the unguarded arm (issue825_selector_audit.py:103). The body still bolds "the user's next turn is linearly unpredictable", the 12 banked ROLE-GAP deltas are all unguarded, and a comparison table you were given two days ago is superseded.
  • #1706 — #1639 is missing the #1310 assistant-test fold. The round landed on main 2026-07-23 (9e65fe09ad) and is named in #1639's own scope row, but the body carries no assistant-test result and its Repro list cites rounds 1–2 only. It sits at awaiting_promotion.
  • #1707 — #1345's base-model story leg was run and lost into a GCP crash bundle; recover-vs-re-run is a compute-spend call.
  • #1708 — deferred sweep backlog: ~60 lower-severity problems recorded but not filed tonight, so one night's sweep does not flood the 5-session infra cap.
  • #1472 (already tracked) — the Codex outage account decision. The route-3 dedup correctly suppressed a duplicate. Worth re-reading: its own 2026-07-17 analysis concludes this is a lapsed ChatGPT Pro subscription (chatgpt_subscription_active_until = 2026-05-24), not exhausted quota. It is at status blocked and carries daily-held but not needs-human, so the PM's Needs you block may not be surfacing it — which would explain 9 days of silence.
• Step C (parked workflow-fix candidates): 6 enumerated, 1 deduped, 5 routed. C0 (watcher outcome-mapping truthfulness) was deduped — its fix landed in-round as #1668 plan item D10 (verified at autonomous_session_watch.py:10574/11379), exactly as its park note predicted. Routed-records posted on #1668/#1669/#1674/#1680/#1682 so no park re-enumerates. One skipped sweep row (task:1333, relevant_kind: false) — benign, the documented malformed-line shape.
• Retraction re-check: 6 marker-mined problems scanned, 0 skipped. No mined premise had been retracted upstream.
• title-sync drift unchanged (8 rows: #432 #458 #654 #658 #664 #665 #763 #1226 — see logs/daily/2026-07-11.md for the verbatim rows).
• Transcript scope: 43 files had today's mtime; 28 survived the in-file-timestamp cross-check (15 were mtime-only bumps carrying no 2026-07-25 content — the documented hazard). ~117 distinct problems surfaced by 6 parallel miners.
• Filing volume is ~4× a normal night. 19 route-2 tasks against a 5-session infra cap means most sit at proposed for the watcher's proposed_infra_sweep to drain over the coming day. That is the designed pacing, but the queue will look crowded.

Living-docs drift

• living_docs.py check: FAIL — 199 findings (≈75 dead legacy anchors; 53 completed clean-results in no question's evidence; relates_to ⇄ evidence mismatches; spec-role-header State trailer stale vs 7 promoted results).
• Count identical to the 07-06/07-07/07-19/07-22/07-23/07-24 rows — materially the same drift.
• drift proposal already open (logged 2026-06-29/07-04/07-06/07-07); not re-proposing
• Step B: #713's declined living-docs proposal remains the only un-reconciled rejection; no context change today bearing on it — skipped silently. No parked-redundant on_hold follow-ups exist.

My thoughts

Highlighted results

• no results promoted today