EPS
← All docs·Activity

Daily — 2026-07-19 (28 infra merges; 3 experiments re-parked; 18 filed + 1 applied)

updated: 2026-09-03

What happened

  • Workflow-fix wave landed: ~28 kind: infra tasks completed + squash-merged today (#1523–#1536, #1538–#1548, #1550–#1553 band — the 07-18 /daily's 21 filings plus in-flight parks), including the ruff full-ruleset pin restore (#1532), verified-at-filing clause (e) (#1527), the commit-site HEAD guard + FF-recipe root-cause fix (#1530), and the trigger-dense poll-digest discipline (#1546).
  • #1482 (SAE k-resample line): shuffle-null free-analysis folded, then TWO cheap-band rounds (k-resample-noise-floor — after a determinism-gate crash-fix + a re-tokenization-seam diagnosis correction — and perdirection-shrinkage-control); headline upgraded MODERATE→HIGH; cheap cap 2/2 consumed; re-parked at awaiting_promotion.
  • #1417: registered-selector-refit round drained + a second milder-rude-render round (attempt-1 failed its register gate, attempt-2 passed); both cheap rounds consumed; parked at awaiting_promotion (19:08).
  • #1481 (contrastive-vs-positive-only containment grid) cleared its full critic pipeline and parked at awaiting_promotion; #1489 likewise parked, but its Step 10d merge was deferred after 3 lint-gate blocks (PR #1261 open — re-merges idempotently on the next /issue 1489 invocation).
  • #1538 (guard grep-pattern FP fix) completed only on its THIRD session — two orchestrator refusal-wedges en route. #1540 completed its pipeline but the merge turn was refusal-killed, leaving PR #1312 stranded (see notes; recovery spawned tonight).
  • Fleet condition: Codex org quota exhausted — sentinel live until 2026-08-06; every doubled review site ran single-Claude all day (#1204 pre-spawn skip worked everywhere, zero wasted composer spawns). Also: the my-goat tmux fix held overnight and the stranded old-socket tmux server was killed (~85 GiB RAM freed).
  • No clean-results promoted today.

Applied workflow improvements

  1. Target: .claude/rules/planner-section-reference.md, .claude/rules/upload-policy.md — what: sync the stale singular staging-probe phrase to the #1543-ratified "once per (source-family × staged consumer) pair (#928, #1481)" wording (parked #1543 code-reviewer candidate, fp f2178ea9e923) — commit: b4c516a41e Why: #1543's code-reviewer (round 1, 2026-07-19T08:17Z) found the 6th sibling site its plan's line-based grep missed (1-file staging probe wraps across lines at planner-section-reference.md:720-721; upload-policy.md:310 carried the pre-#1481 singular pointer). The behavior was already ratified by #1543's full pipeline (merged 041383903d); this is the lagging doc-sync only. Applied edit: 
    - probe + consumer-open gate before production (#928).
+ probe + consumer-open gate before production, once per
+ (source-family × staged consumer) pair (#928, #1481).
    (and the matching per-pair phrase + #1481 citation at upload-policy.md:310-311) Verification: grep -n "staged consumer) pair" confirms both sites; lint gate re-run green (see below).
  2. Filed for review #1554: hook-block worktree-local git -C <wt> merge main (LOCAL-main form; #1530's deferred D6, blast-radius concerns carried into the body).
  3. Filed for review #1555: artifact-reuse.md check (g) parity probe once per reused adapter recipe-class (parked #1543 candidate, confidence low).
  4. Filed for review #1556: poll_pipeline.py structural digest instead of raw log_tail_excerpt on trigger-dense workloads (parked #1546 candidate).
  5. Filed for review #1557: verify_plan.py check 31 — a NEW pin test on a workflow-surface glob must name its select_step9c_tests.py WORKFLOW_INVARIANT registration (parked #1546 candidate).
  6. Filed for review #1558: extend the followup-parked-by-cap note to the Step 9b cheap-band (C2) cap park path (parked #1548 candidate).
  7. Filed for review #1559: guard_tmp_tmux_sweep.sh — exempt (or document) single-file rm of a provably-dead tmux socket (mined; guard false-positive, override flow worked).
  8. Filed for review #1560: stop branch-era workflow_lint.py+tests redding Step 10d gates — 3 incidents today (#1489 3× block + abandoned merge; #1482 TG collection ERROR; #1417 inline-gate false block). Strongest recurring candidate of the day; body engages the recorded "deliberately specs-only" design rationale.
  9. Filed for review #1561: workflow_lint.py _hub_verify_bare_hits false-positives on monkeypatch attribute references (4 noise waivers on #1482).
  10. Filed for review #1562: prescribe committed-script (never inline -c) workload commands for ad-hoc probe dispatches (#1482 burned a GCE create+cancel on a quoting error).
  11. Filed for review #1563: extend trigger-dense vocabulary neutralization to ORCHESTRATOR turns on guard-hook tasks (two #1538 orchestrator refusal-wedges).
  12. Filed for review #1564: watcher audit pass for completed-without-epm:merged (the #1540 stranded-merge class — no automation currently sees it).
  13. Filed for review #1565: task_workflow.py re-resolve the task folder path under flock at marker-append time (duplicate tasks/proposed/1524 husk from a cached pre-move path).
  14. Filed for review #1566: guard_repo_root_branch.sh false-positive on git literals inside post-marker --note payloads (mask quoted payloads; #1413 masking pass is the precedent).
  15. Filed for review #1567: guard deny-shape tests pin their git context — 23 transient failures from a concurrent shared-root off-main state cost a 28-min gate re-run (#1528).
  16. Filed for review #1568: HF_ROUTING_FROZEN_SNAPSHOT staleness race — treat round-untouched flagged files as pre-existing WARN at merge-gate time (#1547 gate block).
  17. Filed for review #1569: implementers probe origin/main for the next free enumerated check id at implement time (#1550/#1551 both added c40 same day; PR conflict + renumber round).
  18. Filed for review #1570: merge=union gitattribute for tests/sparse_cones.txt (conflicted in two merges today; extends the #787/#1534 union-merge mechanism).
  19. Filed for review #1571 (wf_fix: false, library code): orchestrate/hub.py upload ergonomics — file-count throughput WARN (>~2k files/commit → pack/shard recipe) + retry-wrapper signature docs (31k-file upload crawl killed + repacked on #1481; two TypeErrors).

Other problems & notes

  • #1540 fix stranded off main (most serious dropped handoff today): the Step 10d merge turn was refusal-killed at 14:49 UTC right after epm:done; task reads completed, PR #1312 sat DRAFT ~9h with commit 50d50f0ab2 only on origin/issue-1540 — invisible to every watcher lane. Action taken tonight: recovery session spawned (spawn-issue --issue 1540 --auto) to re-drive the merge; durable fix filed as #1564.
  • Two orchestrator refusal-wedges on #1538 (sessions at 12:10 and 13:58 UTC, "violative cyber content" on a guard-hook grep-pattern diff) plus two reviewer refusal-kills — the watcher respawn + rung-b2 sonnet pin recovered it and a third session merged the task; prevention filing is #1563. One further refusal-kill hit a #1531 implementer (recovered on the rung-(b) retry).
  • Deduped against a landed fix: the parked #1516 + #1527 ruff-policy candidates (test_ruff_policy red on pristine main, 8 offenders) were FIXED same-day by #1532 (commit cb86ce3340); verified clean at re-check with ruff check --select C901,RUF001,RUF002,RUF003. Before the fix landed it cost ≥5 sessions a pristine-oracle strip each.
  • Retraction re-check: 8 marker-mined parked candidates scanned, 0 retracted upstream, 2 deduped against the landed #1532 fix. Transcript-mined filings were confirmed against the current tree at compose time.
  • #1489 merge deferred, self-recovering: PR #1261 open with epm:merge-failed posted after 3 lint-gate blocks (branch-era lint-script class → filing #1560); the merge retries idempotently on the next /issue 1489 invocation — worth nudging a session if it sits.
  • Codex quota outage until 2026-08-06 — all 5 doubled review sites ran single-Claude all day across ~30 merges. Already tracked in blocked daily-held #1472 (review diversity during Codex outage); no new filing. Worth knowing that today's whole infra wave shipped with zero cross-family review.
  • #1549 (daily-held: reap stranded old-socket tmux server) is now RESOLVED by today's work — the 0da7071f session killed the -L old server (702 processes, ~85 GiB freed) and removed the socket; PM can archive #1549.
  • #1482 session wedge: the owning session went dark 14:41→17:03 UTC waiting on a vstream until-loop (bg task returned killed); the watcher respawn recovered cleanly — the detached-phase pid/log breadcrumbs worked as designed. Mid-round the prior session's killed free-analysis implementer also handed off cleanly via a surviving setsid-detached phase.
  • Filing-quality note: #1531's filed diagnosis ("shadowing earlier anchor") was wrong — the clarifier reproduced the failure and re-anchored instead; the existing verified-at-filing context-read clause covers this, the filer under-applied it. Similarly #1530's filed premise (task.py stranding commits) was disproven by reflog forensics; that session shipped the real root-cause fix (FF-to-local-main recipe) instead.
  • Recurrences of documented rules (no new filing): ToolSearch-before-TaskOutput miss (#1531 session; the reminder already sits at SKILL.md:4497); 6+ guard_piped_git_push.sh firings across 3 sessions (hook absorbing a compose-time habit CLAUDE.md already bans); one blocked sleep && chain (#1417 session); GitHub mergeability-recompute lag on two merges (recipe already documented at SKILL.md Step 10d); issue1481_cjk_audit.py pre-existing lint red hit 3 sessions before #1552 waived it same-day (21:13 UTC).
  • Interactive session self-correction (e1f57cc3): a theory-assumptions audit claimed assumption A6 was "never run"; a committed eval_results/issue_658/inline_a3_5a_coherence/ artifact refuted it and the session corrected itself — the standing verify-before-asserting rule's never-run scan must include sibling issues' committed eval_results (it does; the audit under-applied it).
  • Driver sha-scanner annotations: several filing bodies cite 8-hex SESSION ids (labeled as sessions); the #1467 WARN-only scanner annotated them — no mis-cited commits shipped. One truly unresolved token (950079a7, cited as a commit by the miner for the #1524 husk) rides annotated into #1565 for the spawned session to verify.
  • Title-sync drift sweep (Step D): 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).
  • Minor self-corrected errors in aggregate (no filings): Edit anchor misses ×4 (plan files), oversized Reads ×3 (guard script 41K tokens, bg-task output 48K), inline one-liner tracebacks ×5, one PushNotification soft-fail ("Remote Control inactive", chat fallback used), one 429 upload retry, one stale /tmp path Read, one SendMessage to an ended subagent.

Living-docs drift

  • living_docs.py check: FAIL — 199 findings, materially the SAME drift as the still-open re-synthesis proposals of 06-29 (1736fca0804f), 07-04 (d2a80db46b3b), 07-06 (b009cef9ecd9), 07-07: ~75 dead legacy anchors (a1/b1/c1-class relates_to with no matching question), ~54 completed clean-results in no question's evidence, the #494–#501/#722/#763 relates_to⇄evidence mismatches, and the stale spec-role-header State trailer (7 newer linked results).
  • drift proposal already open (logged 2026-06-29, extended 07-04/07-06/07-07); not re-proposing — event row appended to .claude/cache/nightly-consolidation-events.jsonl.
  • Parked-proposal re-consideration (Step B): no genuine epm:living-docs-update-rejected markers awaiting reconciliation (the #713 grep hits were note-text echoes, not markers) and no on_hold parked-redundant follow-ups — nothing to re-surface.

My thoughts

Highlighted results

  • no results promoted today