Daily — 2026-07-11 (24 infra fixes merged; 11 new review filings)

What happened

• 24-task infra drain completed end-to-end. #1251–#1274 (the 07-10 nightly's daily-fix: filings plus workflow-fix tasks) all ran through autonomous /issue --auto sessions and merged to main — ~680 commits on 07-11 UTC, PRs #1007–#1030. Zero human input needed, zero API errors, zero refusals, zero blocked tasks. Highlights landed: verify_plan c7/c20/c26–c30 escape hardening (#1262/#1258/#1264), watcher boot-death lane (#1267), conftest hermeticity guards (#1265), LESSONS.md byte-budget ratchet (#1269), trigger-dense return-text contract (#1252), per-task routed-record stream grouping (#1274).
• Codex quota outage all day (sentinel live until 2026-08-06T13:26Z): every review round in all 24 sessions ran single-Claude per the #1204 pre-spawn skip; no fabricated twin verdicts found anywhere. Cross-family review diversity is effectively off fleet-wide until the quota resets.
• Dominant friction: Step 10d merge contention under fleet churn. First-attempt PR merge failed in at least 10 of 24 sessions (merge-conflicts / can't-be-rebased shapes); every one recovered via the documented classify-and-retry ladder at a cost of ~8–25 min each. Worst episode: origin/main transiently LOST task #1272's folder entirely during concurrent merges — #1271's session restored it via the prescribed scratch-worktree root-sync recovery (merge f56a2f93f9). Guard-1 false-positive fix filed tonight (#1280).
• No experiments, no pods, no promotions. Pure infra day. #928 is still holding at followups_running (last marker 2026-07-08) — see problems below.

Applied workflow improvements

No route-1 (trivial mechanical) fixes were self-applied tonight — every derivable fix was a behavior/logic change, so all routed to independent review. 11 route-2 filings (all filed via daily_drive_filings.py; #1275 spawned a session at filing time, #1276–#1285 are dispatch-pending on the watcher proposed_infra_sweep backstop — see logs/daily/filings-2026-07-11/filed.jsonl):

1. filed for review #1280: Step 10d Guard 1 computes FOREIGN tasks/ paths from a two-endpoint diff — 33 false positives on #1271; switch to three-dot merge-base diff (formal candidate, confidence high; parked on #1271 and MISSED by the Step C sweep — surfaced by the transcript miner).
2. filed for review #1281: sweep_parked_wf_candidates.py _row_is_parked misses mid-note park declarations (Routing: parked after prose) — the exact reason #1280's candidate was missed tonight.
3. filed for review #1277: verify_plan c4 N/A escape is a bare doc-global regex (self-escaping, same class as the c7 bug #1262 fixed) — migrate to _standalone_na_declared (formal candidate from #1262, fp 791d8ecf6fd1).
4. filed for review #1276: verify_plan c25 arm-(a) exemption is all-or-nothing — one doc-wide entity-content N/A line exempts every shell-tagged fence at once (parked on #1258).
5. filed for review #1275: orchestrator posts Codex twin verdicts from the output FILE on trigger-dense rounds instead of paging the findings body into context (parked on #1252).
6. filed for review #1279: edit-time PreToolUse guard for .claude/rules/LESSONS.md closing the direct-to-main bypass of the two blocking lint gates (parked on #1269).
7. filed for review #1278: consolidate the duplicated _stub_fleet_mutating_passes watcher-test helper across two test files (parked on #1265).
8. filed for review #1283: filer-side title-prefix WARN in file_infra_task.py for wf-fix-tagged filings, complementing #1273's driver-side guard (prose follow-up recorded on #1273).
9. filed for review #1282: infra-mode critic briefs cite the lens-reference heading verbatim — #1265's Alternatives critic grepped a translated title and found no span, reviewing from the brief's inline text instead of the canonical rubric.
10. filed for review #1284 (non-workflow, daily-auto-filed only): scripts/issue1092_bridge_refit.py + issue1092_figures.py heavy imports before/without load_dotenv — the pre-existing red on main that taxed two Step 9c gate runs today with pristine-oracle rounds.
11. filed for review #1285 (non-workflow, daily-auto-filed only): remove the stale duplicate tasks/reviewing/1196 folder from main (#644/#1253 remediation; canonical is tasks/completed/1196 — surfaced by #1274's session, left out of its scope).

Other problems & notes

• The Step C parked-candidate sweep silently skipped #1271's formal Guard-1 candidate — its note opens with a root-sync recovery record and only declares Routing: parked at the end, matching neither of _row_is_parked's two accept patterns. Recovered tonight because the transcript miner independently surfaced it; both the candidate (#1280) and the sweep fix (#1281) are filed. Routed-record posted on #1271.
• Dedup: #1258's second parked candidate (c7 bare-phrase escape) was NOT filed — #1262 (completed today) already migrated c7 to _standalone_na_declared (verify_plan.py:881, PR #1018). Routed-record with the dedup note posted on #1258.
• Retraction re-check: 7 marker/task-note-mined problems scanned, 0 skipped for retraction, 1 skipped as deduped (the c7 item above). No transcript-mined filing had a suppressing tail correction.
• Bang-backtick boot-killer residue (already closed): sessions 3753d39e/6312bbb4 died at 03:13/03:14 UTC with zero assistant rows — the `if !` code span in issue SKILL.md executed as a dynamic shell command at skill load. Fixed the same night (commit f75e1b4c13) and the lint check shipped today as #1266; both tasks completed via respawned sessions. No new filing (deduped against closed #1266).
• origin/main transiently lost task #1272's folder during concurrent Step 10d merges (~18:35Z); #1271's session restored it via the documented scratch-worktree recovery and the one-folder-per-task guard read clean afterward. No filing — the recovery machinery worked verbatim; #1280 attacks the false-positive strip pressure that feeds these episodes.
• guard_repo_root_branch.sh tripped twice on quoted git-verb text (a post-marker --note quoting SKILL.md git citations in #1268 16:07Z; a read-only git log -S pickaxe payload in #1253 08:05Z). Both recovered in under a minute — the hook's own error message teaches the --file bypass, and the hook header documents this exact fail-closed trade-off as deliberate. NOT filed (documented design decision, not a gap). Worth knowing: the #1253 session evaded via a -G 'c.eckout...' regex obfuscation, which works but is ugly.
• #928 stuck at followups_running since 2026-07-08 (last marker is another session's coordination claim, 07-08T14:21Z; tag followup-auto). The holding session appears dead — an idle followups_running is a stuck follow-up round, not a healthy pause. Suggested action: re-drive it (uv run python scripts/spawn_session.py spawn-issue --issue 928 --auto) or have the PM triage why the watcher's crash-recovery didn't.
• VM root disk sub-floor dip at 07:27Z (17–19 GiB free, below the 60 GiB sub-floor sentinel; largest attributed cache: issue-816 worktree hf_dl 1.5 GB). Recovered to 58 GiB free by end of day without intervention. No action.
• Triage-observer info rows ×3 (#1255, #1258, #1266): pre-dispatch external-marker triage line missing/none against a non-empty candidate window. Non-gating duty misses; no filing (the observer exists to track recurrence — watch for escalation).
• Transient one-offs, all self-resolved in-session, no action: #1263 heredoc plan-patcher AssertionError (rewritten in 30 s); #1267 inline-probe IndexError on an empty note (guarded retry 12 s later); #1270 recursive grep swept .claude/worktrees/ full checkouts (re-scoped same turn); #1255 implementer truncation-blind figure check caught by round-1 code review and fixed with regression tests (the pipeline working as designed).
• Title-sync drift sweep (Step D, first run in a daily): 8 WARN rows (read-only; which side is fresher is a human call — remediate via task.py set-title <N> "<H1 text>" or re-set-body):
  • #432 (completed): H1 'Expanding contrastive negatives from 2 to 9 moves software engineer from rank 26 to rank 7' vs title 'Expanding contrastive negatives from 2 to 9 lifts the trained source persona from rank 26 …' [grandfathered v3/v2]
  • #458 (completed): H1 vs title diverge in the tail (same lead text) [grandfathered v3/v2]
  • #654 (awaiting_promotion): H1 'Matched delimiter-position residual cosine sits above the within-tier shuffled floor…' vs title 'Query influence on the context-conditioned residual stream: per-layer context-end vs query…' [grandfathered v3/v2]
  • #658 (awaiting_promotion): H1 "…fails on Qwen2.5-7B for the four safety-relevant behavi…" vs title "…fails on Qwen2.5-7B — a mean answer-side activation sum…"
  • #664 (awaiting_promotion): H1 'The marker leakage gate never installs…' vs title 'Marker leakage gate never installs…'
  • #665 (awaiting_promotion): H1 vs title diverge in the tail (same lead text) [grandfathered v3/v2]
  • #763 (awaiting_promotion): H1 vs title diverge in the tail (same lead text)
  • #1226 (completed): no H1 found — a sentinelled clean-result must open with # <title>
• Daily stub push: first git push was rejected (non-fast-forward); sync_repo_root.py recovered and confirmed the push landed. No residue.

Living-docs drift

• living_docs.py check FAILs with 199 findings — drift hash 078285c14444, byte-identical to the open proposal (dead legacy anchors, relates_to ⇄ evidence mismatches incl. #722/#763/#494–#501, ~54 unlinked completed clean-results, 7 stale spec-role-header State-trailer rows).
• drift proposal already open (logged 2026-07-07, same hash 078285c14444); not re-proposing. The standing proposal remains: re-synthesize open_questions.md (living-docs updater/backfill) — needs your ok; not auto-applied.
• Parked-proposal re-consideration (Step B): the one un-reconciled living-docs rejection (#713) has had no context change (no results promoted today, no open_questions.md change since 07-07) and no followup-parked-redundant markers exist on on_hold tasks — nothing became a re-proposal candidate.

My thoughts

Highlighted results

• no results promoted today