Daily — 2026-07-26 (19 wf-fix tasks merged; main red 15.5h; 153 problems swept)

What happened

• Last night's 23-task filing wave fully drained. All 19 route-2 workflow-fix tasks filed by the 2026-07-25 sweep (#1690–#1704, #1709–#1712) reached completed and merged to main today; the 4 route-3 held tasks (#1705–#1708) correctly stayed at proposed. 20 epm:merged markers, 19 epm:done, 133 commits on main.
• main went red at 15:03Z and stayed red. #1701 merged a new workflow_lint check at 14:18Z; #1698's merge 45 min later carried a branch snapshot of workflow_lint.py predating it and deleted the function, while the test importing it survived. Full-suite pytest has aborted at collection (rc=2) ever since — ~15.5 h, still red at sweep time. One session diagnosed it correctly at 16:57Z and left it as a prose note.
• #1689 (the only running experiment) hit a compute wall. Three same-pod crash-fix relaunches (R11/R12/R13) got it to the fit_ladder phase, which is a 1,210,608-evaluation serial dense-SVD loop: 5 h 14 m elapsed, zero durable output, all four H100s at 0%. Measured extrapolation is 56–112 days. Filed as a held decision.
• Codex twin reviewers were absent all day (sentinel to 2026-08-06, 18 days in), so every review gate on all 19 merged tasks ran single-Claude.
• No clean-results promoted today. Fleet at sweep time: 1 running, 111 awaiting_promotion, 18 proposed, 2 plan_pending, 5 blocked.

Applied workflow improvements

Route 1 — self-applied (3)

1. Target: .claude/skills/issue/SKILL.md — what: fix the stale latest-marker --kind flag to the real --prefix — commit: 9a3dc22055 Why: six independent miner groups hit this. task.py latest-marker --help shows --prefix only; the documented --kind is an argparse error. In one session the failure was masked by a pipe and a status advance briefly rested on an unverified marker read. Verified: task.py latest-marker 1689 --prefix epm:run-launched returns the marker (rc=0); grep -rn 'latest-marker[^|]*--kind' .claude/ CLAUDE.md → clean. Applied edit:

   -  `task.py latest-marker <N> --kind epm:fact-candidates`, then branches
   -  on session mode.
   +  `task.py latest-marker <N> --prefix epm:fact-candidates`, then branches
   +  on session mode. (The flag is `--prefix`; `latest-marker` has no
   +  `--kind` option — see `task.py latest-marker --help`.)
   

2. Target: .claude/skills/issue/SKILL.md — what: restate the Step 10d lint-gate wall as 30–40 min under fleet load — commit: ddca58e0ed Why: the spec said "~9-12+ min"; measured today at 32 min (#1690), 37 min (#1694), ~30 min (#1711). Any wall-time-derived fence sized off the old figure is 3× too small. Applied edit:

   -  both). The two leg pairs + TG legs total ~9-12+ min, so the executable
   +  both). The two leg pairs + TG legs total ~9-12+ min on an IDLE VM, but
   +  **30-40 min under typical fleet load (3+ concurrent gates)** — measured
   +  2026-07-26: #1690 32 min, #1694 37 min, #1711 ~30 min. Size any
   +  wall-time-derived fence off the LOADED range, not the idle one. So the
   

3. Target: .claude/skills/daily/SKILL.md — what: inline no-pipe reminder at the commit recipe — commit: 9a3e969481 Why: guard_piped_git_push.sh fired across at least five sessions today, and blocked this run's own first commit. The rule lived ~30 lines from the recipe block it governs; it now sits in it. Applied edit:

   +# Run BARE — never `| tail`/`| head`/`| grep`. `guard_piped_git_push.sh`
   +# BLOCKS a piped git commit/push/merge (it masks the exit code, and SIGPIPEs
   +# the pre-commit hook mid-run, #1584). Need the text? `> /tmp/out 2>&1` then
   +# read the file. Same for the `git push` below.
    git commit -m "logs: daily stub for YYYY-MM-DD" -- logs/daily/...
   

Lint gate after all three: --check-references PASS, --check-asks PASS, no-flags PASS (WARNs unchanged from the pre-run baseline). All three verified on origin/main.

Route 2 — filed for independent review (22)

Each lands through the full planner / critic / code-review / test-verdict pipeline. #1713, #1720, #1726, #1728 already dispatched to planning; the rest sit at proposed for the watcher's infra sweep.

1. filed for review #1713 — URGENT, main is red. Restore check_inline_round_duty_mirror to scripts/workflow_lint.py (deleted by #1698's stale branch snapshot 45 min after #1701 added it; the test importing it survived, so full-suite pytest aborts rc=2 and Step 9c cannot classify) + close the lost-update class.
2. filed for review #1714 — spec-freshness sync manufactures self-inconsistent trees (skips a branch-edited file but syncs its derived artifacts; races the 30-min gate).
3. filed for review #1715 — draft-PR recipe compares against stale local main and omits --title, so gh pr create exits 1 non-interactively.
4. filed for review #1716 — code-reviewer has no full-ruleset ruff policy-pin duty; round-1 PASSes on diffs Step 9c then fails (~42 min lost on #1710).
5. filed for review #1717 — select_step9c_tests.py ergonomics: --map-files ignores --json, stderr NOTEs corrupt --json, uncommitted edits silently degrade the selection.
6. filed for review #1718 — AGENT_SPEC_SIZE_GRANDFATHER dict is a chronic 3-way merge-conflict hotspot (~107 min of gate wall-clock on one branch).
7. filed for review #1719 — single-flight pgrep probes self-match and are not issue-scoped; one silently skipped a required gate leg behind a green exit 0.
8. filed for review #1720 — make the implementer's NOT-RUN escape pre-emptive for known-slow tests (~51 min burned on 4 timeout-killed runs for zero information).
9. filed for review #1721 — planner grounding: run grep-answerable claims, quote line numbers, don't drop required edits, trace embedded shell control flow.
10. filed for review #1722 — Step 10d durable markers: wrong merge SHA, empty field from a broken $( ), splitlines()[0] IndexError on an empty note.
11. filed for review #1723 — CRON-TEARDOWN precedes the Step 10d merge (33-min merge ran unmonitored); a task went completed 83 min before its merge landed.
12. filed for review #1724 — verify_task_body.py returns a bare FAIL for every kind: infra task; should short-circuit to a non-applicable verdict.
13. filed for review #1725 — Step 4 has no root-divergence probe, and the landed post-merge sync sits below every epm:merged post.
14. filed for review #1726 — long fit loops must checkpoint per cell and log progress (#1689 ran 5 h 14 m with no output line and no artifact).
15. filed for review #1727 — flag ungated smoke/full variable pairs in review; an unconditional smoke list silently collapsed a 21-condition experiment to 1.
16. filed for review #1728 — reuse check (l) needs a call-shape bind (a runtime assert passed a signature check and crashed after ~13 h); renames need a whole-tree grep.
17. filed for review #1729 — verify_plan c38 false-positives on a backticked pytest node id (reproduced live).
18. filed for review #1730 — gcloud missing from the non-login PATH silently reports zero GCE instances — a silent zero on a spend surface.
19. filed for review #1731 — harden the /daily miner brief (JSONL traps; require probed-vs-inferred labelling).
20. filed for review #1732 — settle whether consistency-checker runs for a parentless kind: infra task (two opposite treatments on the same day).
21. filed for review #1734 — state the minimum Step-2 plan-review floor; one task skipped the planner, verify_plan, the fact-checker and the entire critic ensemble.
22. filed for review #1735 — the closed-sibling probe that shipped today suppressed 21 of 24 of tonight's filings, matching mostly on a shared hot target file or a generic title word (main, runs, step). All 21 were false positives. Found by this run's own dry-run; the batch was filed with --retry-suspects after eyeballing each.

Route 3 — filed as tracked needs-human (3)

23. #1733 — #1689 fit_ladder cannot finish. 1,210,608 evaluations (126 pairs × 4 layers × 2 arms × 1201 draws), each a 4-fit inner CV plus a dense 3584×3584 SVD. 5 h 47 m elapsed, zero durable output, all four H100s at 0%. Measured extrapolation 56–112 days; even 1 s/eval gives ~14 days. Carve-out: spends money, changes scientific meaning, destructive. Needs your call tonight.
24. #1736 — 23 stashes on the shared repo root, oldest 2026-05-27, from ≥8 distinct tasks. Carve-out: dropping a stash is destructive and not automatable safely.
25. #1737 — should /daily cap its nightly route-2 filing volume? 23 filed last night against a 5-slot cap (all 19 route-2 drained within a day). Carve-out: scope preference.

Other problems & notes

Sweep scope: 24 main transcripts, ~8,320 in-window messages, mined by 10 parallel agents → 153 distinct problems. 3 self-applied (route 1), 25 filed (22 route-2 + 3 route-3), 2 already-tracked, remainder noted below. Zero were dropped.

Already tracked — not re-filed:

• Codex twin outage → #1472 (blocked). 9 of 10 miner groups flagged it independently; the sentinel reads CODEX_QUOTA_LIVE until=2026-08-06T13:26:00+00:00, detected 2026-07-08. #1472 already carries the analysis and the recommendation: the OAuth token claims chatgpt_subscription_active_until = 2026-05-24, i.e. the ChatGPT Pro subscription appears to have lapsed — a ~5-minute check that would restore cross-family review ~11 days early, not a quota you have to wait out. Freshness note posted on #1472 tonight. This is the highest-leverage item on the list for you.
• Untracked scripts/issue823_single_split_protocol.py → #1686 (proposed). Untracked at the repo root since Jul 15; every Step 9c gate fleet-wide runs against a dirty baseline ledger because of it (4 miner groups; verified ?? now). Disposition needs a human — deleting another task's uncommitted draft is destructive.

Not filed by deliberate judgment:

• 6 open daily-held tasks lack the needs-human tag (#1038, #1067, #1141, #1217, #1331, #1472) so the PM's Needs you block never enumerates them. Checked the git history before treating this as a bug: all six were created with the tag and had it removed by your deliberate mass remove-tag on 2026-07-17 14:43 (commits 1bd90f800e, cbfcb10dc3, 136b105722). That is your decision, so nothing was re-tagged — but the consequence is worth knowing: #1472's 5-minute Codex fix has been invisible for 9 days.
• #1689 fit_ladder emits no per-cell progress line, so a 5 h phase is indistinguishable from a wedge. The live #1689 session owns scripts/issue1689_*; filing a fix against a file a live implementer owns would collide, so the general class was filed instead (longfit-percell-checkpoint-and-progress) and the issue-specific edit left to that session.
• task.py latest-marker --kind flag. Six groups tripped on this; the doc was the actual defect and is fixed in route 1 below. Adding a real --kind flag to task.py is a convenience, not a failure, so it is noted rather than filed.
• #1710 shipped with 1 of its 4 filed incident shapes undiagnosed, its plan-prescribed verification probe silently substituted, and the completion audit closed with "file it if it recurs" — no follow-up filed. Suggested action: make a [/]-partial completion-audit ask a hard trigger to file. Held out of tonight's batch only to bound its size; worth filing next sweep if it recurs.

Smaller items, one line each (all verified, none filed):

• py-spy is unusable on the RunPod pods (uv venv has no pip, no ptrace) while gotchas.md prescribes it as the first hang probe — and line 56 of the same file says it is unusable. Self-contradiction; cost a lost stack dump on the #1689 diagnosis.
• pod.py config --list prints "Total: 22 pods, 57 GPUs" while exactly 1 pod is live — a standing misread hazard for any burn claim.
• plan_patch.py has --anchor-file/--replace-file but no --verify-contains-file; 4 argparse errors, then the mandated edit-success gate was dropped.
• Goal-snapshot capture produced a silently EMPTY snapshot ($( inside a python regex → command substitution); the "no snapshot ⇒ mismatch" rule never fired.
• choom used as a command prefix without sudo — the permission failure killed a launch.
• guard_log_dump.sh blocked a cat of a zero-byte log (name match, no size check).
• guard_trigger_dense_read.sh blocks reading its own source, on a task whose subject is that guard.
• The using-superpowers plugin skill auto-loaded inside an autonomous EPS /issue session.
• Autocompact fired mid-merge-conflict-recovery in one session (recovered).
• $MAIN_SHA does not survive across Bash tool calls; merge diagnostics silently produced nothing.
• Two agent specs are over the byte WARN budget (interpretation-critic.md 35,253; planner.md 38,368; FAIL at 40,000).
• Parked workflow-fix sweep (Step C): 0 candidates to route. 1 skipped row (relevant_kind: false, the benign #1333 malformed-line shape). unmatched_record_fps advisory non-empty (263 refs / 121 sources) — already fixed by #1703, which merged today.
• retraction re-check: 25 filing candidates scanned, 0 skipped (no premise was retracted upstream between mining and composing).
• title-sync drift unchanged (8 rows: #432 #458 #654 #658 #664 #665 #763 #1226 — see logs/daily/2026-07-25.md). Read-only sweep exited 0.
• staged-set audit: clean on every commit this run — no foreign staged files.

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 07-06/07-07/07-19/07-22/07-23/07-24/07-25 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