EPS
← All docs·Activity

Daily — 2026-07-27 (22 wf-fix tasks merged-1; NEW main red ~17h; a lost urgent park; 90 problems swept)

updated: 2026-09-03

What happened

  • The wf-fix wave kept draining: 22 tasks completed today (#1713–#1735), 21 merged. #1713 (the urgent check_inline_round_duty_mirror restore) merged at 09:43Z and closed yesterday's 15.5h main-red. Every review gate ran single-Claude all day (Codex quota sentinel live until 2026-08-06).
  • A NEW main-red opened at 13:59Z and is still live at sweep time (~17h). #1721's squash merge (028b45ff44) grew .claude/agents/planner.md 39,371 → 40,900 bytes, over the 40,000-byte agent-spec FAIL ratchet — no-flags workflow_lint + the full-suite pytest are red on main. #1718 hit it at its own Step 10d and is blocked with finished work stranded on its branch (mutually-blocked). At least 7 later sessions classified the red as pre-existing; none emitted the mandatory routable urgent park.
  • #1718 DID park a mechanically-routable urgency: main-red candidate at 14:38Z — and it was invisible to every router. The Step C sweep's park predicate requires a literal "parked" token; the note says "URGENT-PARK … raised", so _row_is_parked = False and both the watcher's #1681 urgent router and the nightly sweep were blind. This run routed it manually: #1740 (urgent planner.md shrink, session spawned) + #1741 (sweep predicate fix), routed-record posted on #1718.
  • #1689 (the only long-running experiment) had a rough day: the r15b relaunch ran serial numpy on CPU at width 1 with the 4×H100 pod at 0% GPU for ~14h (3 idle escalations; compute-deviation ratio 32×) until Thomas ordered vectorization — the R16 torch port measured ~10×/pair and now saturates all 4 GPUs, but the wall is still ~44h and the draws were descoped 1000/40 → 10/2 without a plan amendment (held as #1759). A 16-way uv run resolution storm also FUSE-wedged /workspace (pod stop/resume + UV_NO_SYNC=1 relaunch recovered it).
  • Two new experiments spawned late: #1738 (running on GCE, phase-0 manifest build) and #1739 (multi-behavior real-corpus datagen — lost 9 subagents to refusal/thrash before converging on the content-opaque decomposition; completed its rounds).
  • No clean-results promoted today. Fleet at sweep time: 3 running (#1689, #1738, #1739), 111 awaiting_promotion, 21 proposed, 2 plan_pending.

Applied workflow improvements

Route 1 — self-applied (6 fixes, 5 commits)

  1. Target: .claude/rules/gotchas.md — what: 3 grouped additions: (a) MooseFS FUSE read-wedge trigger + prevention — N-way parallel uv run resolution storm on a stop/resumed pod; export UV_NO_SYNC=1 before fan-out (#1689); (b) Reddit-dump field semantics — string-typed booleans ('False' is truthy) + selftext not content (#1739 kept 0/800 twice); (c) SSH-MCP ssh_execute kills at 30 s locally regardless of timeout param (probed in mcp-ssh-manager/src/index.js), MooseFS find splits into <30 s sub-calls, strace absent from the pod image — commit: c8174178f9 Why: miners J P4/P5, B P7, I P14 — all probed as gaps (0 hits) tonight. Verified: greps confirmed absence pre-edit; lint gate PASS post-edit.
  2. Target: .claude/skills/issue/SKILL.md — what: 2 grouped clarifications: (a) the Step 10d pre-push lint gate is an INLINE recipe — the phantom step10d_lint_gate.sh helper does not exist (a session invoked it, 40 s lost); (b) mcp__happy__change_title exact call shape (deferred MCP tool, not a Skill; title required — both misfires occurred today) — commit: edc9905be4
  3. Target: .claude/rules/persona-vectors-recipe.md — what: the "except logits" carve-out is the judge MODEL only; the trait JSON eval_prompt rubric stays verbatim (a planning chat swapped the rubric; Thomas reverted it) — commit: 01553783ee
  4. Target: .claude/skills/issue/SKILL.md — what: 9a-ter inline gate lead-line: scripts/inline_lint_gate.py is the ONLY certifying entrypoint; manual component runs don't satisfy guard_root_code_commit.sh (~4 min lost tonight) — commit: 46e4b7212d
  5. Target: .claude/rules/gotchas.md — what: ugrep dies on bounded multibyte-class repetitions — silent no-data on background scans; use rg/grep — commit: 9b113bff2d

Lint gate after all five commits: --check-references PASS, --check-asks PASS. All pushed to origin/main.

Route 2 — filed for independent review (22)

Each lands through the full planner / critic / code-review / test-verdict pipeline. #1740 dispatched immediately (urgent); the rest spawn as the 5-session infra cap allows, with the watcher's proposed_infra_sweep as backstop.

  1. filed for review #1740 — URGENT, main is red ~17h. Shrink .claude/agents/planner.md below the 40,000-byte ratchet (relocate per-scenario content to rules). Manual routing of #1718's lost urgent park (fp 06bc0203d759); unblocks #1718's stranded merge.
  2. filed for review #1741 — the Step C sweep's park predicate misses URGENT-PARK notes (no "parked" token) — the #1681 router failed end-to-end on its first real firing; probed _row_is_parked = False on the actual #1718 row.
  3. filed for review #1742 — make the Step 9c urgent-park emission MECHANICAL (step9c_baseline.py prints URGENT-PARK-REQUIRED: <node>); 7 sessions skipped the prose mandate on its first live day.
  4. filed for review #1743 — code-reviewer verdict post via --file + read-back verify (round 1 falsely claimed its marker was posted; ~9 min + duplicate spawn).
  5. filed for review #1744 — implementer pin-sweep deleted-literal grep gets its own numbered sub-step (buried mid-paragraph; missed → ~70 min round-2 bounce on #1723).
  6. filed for review #1745 — pre-persist plan-header retitle to stop c40 header-version churn (2 sessions, ~4 wasted persist+verify cycles).
  7. filed for review #1746 — Step 9c gate degrades on collection ImportError (--continue-on-collection-errors / ledger-aware deselect) instead of aborting the whole run (3 sessions × ~40 min today).
  8. filed for review #1747 — Step 5a spec-freshness still syncs LOCAL main (regressed specs into #1724's worktree); mirror the Step 10d origin/main fix.
  9. filed for review #1748 — trigger-dense datagen briefs start from the content-opaque decomposition #1739 converged on after ~5h of subagent attrition.
  10. filed for review #1749 — crash-fix relaunch rounds re-run the compute-character statement (r15b relaunched as serial CPU on a 4×H100 pod, silently).
  11. filed for review #1750 — pod-side descope posts a marker + rewrites the dispatch handle's sentinel path (r15b: undocumented descope + dead completion wiring).
  12. filed for review #1751 — Step 10d surfaces sync_repo_root "stash: KEPT" lines instead of summarizing "guard clean".
  13. filed for review #1752 — poll_pipeline escalates in KIND after ≥3 identical gpu-idle escalations on one phase (width-re-eval demand, not a 4th identical note).
  14. filed for review #1753 — Step 10d certifies the LANDING tree (size caps, cap bumps, sync attribution) — the #1721/#1727/#1719 moving-main race family.
  15. filed for review #1754 — recursion-guard brief guidance: subagents emit candidate blocks normally, the orchestrator PARKS them (2 sessions briefed "prose only", defeating the sweep).
  16. filed for review #1755 — teammate idle handling branches on the brief's report channel (3/3 final-text teammates ate pointless SendMessage nudges).
  17. filed for review #1756 — compose marker-note files via the Write tool, never a Bash heredoc (git verbs in prose trip the repo-root guard's argv scan).
  18. filed for review #1757 — act on the consistency-checker's edit-locus WARN (hold merge until the named live sibling lands; the predicted #1718 conflict materialized).
  19. filed for review #1758 — --retry-suspects prints a zero-match notice.
  20. filed for review #1762 — user-directed inline dispatch arms the detached pod-side handoff by default + states the lifecycle line in the ack.
  21. filed for review #1763 — Step 10d merge-SHA subject check fetches before comparing (false "fatal: bad object" on #1735).
  22. filed for review #1764 — sudo -n tail for root-owned GCE workload-log probes (the #1738 diagnostic was permission-blind).

Other problems & notes

  • Route 3 — held for you (filed as tracked needs-human tasks):
    • #1759 — #1689 Phase-D disposition: headline CIs descoped to 10 bootstrap / 2 null draws (pre-registered 1000/40) + L19-only, ~44h more wall on the 4×H100 pod. Accept-with-caveat vs re-bootstrap (now ~10× cheaper post-vectorization) + keep vs downsize the pod. (Scientific-meaning + money carve-outs.)
    • #1760 — #1719's autonomous reset --hard + FORCE-PUSH of its own issue branch during a moving-main recovery: policy says force-push is user-ask-only — ruling wanted on the own-issue-branch case. (Destructive/policy carve-out.)
    • #1761 — untracked scripts/issue823_single_split_protocol.py (Jul 15) dirties every step9c baseline ledger; observer escalated daily ~12 days. Commit, move, or delete — ownership unclear. (Destructive carve-out.)
    • already tracked in #1736 (shared: stashes/triage): the 5 stranded repo-root stashes (oldest Jul 2; stash@{0} re-flagged by every sync incl. this run's own push recovery) — freshness note posted, duplicate filing suppressed (#1483).
  • Skipped as already-landed (verified at compose time, not filed):
    • code-reviewer long-loop persistence gate — landed TODAY via #1726 (verdict-routing block confirmed in code-reviewer.md); miner B P1's suggested duty exists now.
    • plan-compute-sizing "measured basis includes the draw battery" — the rule already says exactly this (§ Per-cell fit phases); #1689's plan violated it (compliance, not a gap).
    • verify_plan planned-wall check — c26 already parses basis/wall table rows (L4331); miner J's 0-hit probe was wrong. The #1689 plan predates it; no filing.
  • Verify-before-asserting recurrences (compliance, rules exist — no filing): 3 wrong artifact-coverage claims in one interactive session (Thomas corrected each; the #1092 sibling artifacts existed), and a compute-destructive "terminate the instruct leg" recommendation withdrawn after Thomas pushed back (point estimates are computed pre-draw-loop, full quality).
  • Register recurrences: unglossed jargon drew "what is this" ×2 ("U-whitened", "spread-matched labeled mixture") + "per-row grain"; sycophancy construct misread (trait, not wrong-agreement) and synthetic-vs-REAL eval sets each cost a correction round in the #1739 planning chat.
  • Guards earned their keep: ≥17 hook-blocks across sessions today (piped git/gh, repo-root checkout/pull/reset, sleep-chains, cat-log, heredoc-prose) — all self-recovered; this run's own first commit and one push were piped-guard-blocked too (logged with mild embarrassment). One SKIP= hook bypass during #1727's merge recovery is folded into #1753's scope.
  • Session wedge: the #1728 session went silent ~2.5h after implementer dispatch (16 unexecuted queue-operations); watcher respawn recovered it. Harness-level, cause unobservable — pattern-track only.
  • #1738 in flight: phase-0 manifest build at ~99% CPU 4.4h with the log frozen ~50 min at session end — heartbeat posted; #1764 covers the blind probe. Watch it.
  • Step C (parked wf-fix sweep): enumerator returned 0 candidates — which is itself tonight's headline bug (#1741): the one real park (#1718) was invisible. Routed manually; routed-record posted on #1718 (filed_task: #1740). 1 benign skipped row (relevant_kind: false, the #1333 shape); unmatched_record_fps advisory unchanged (263 refs — the #1703 fix landed yesterday; advisory-only).
  • Step B: 0 un-reconciled living-docs rejections; 0 parked-redundant follow-ups.
  • 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.
  • retraction re-check: 5 marker-mined problems scanned, 0 skipped (the #1689 deviation-v4 continue_as_is and R16 GPU port were incorporated into the filings' framing, not retractions).
  • staged-set audit: clean on every commit this run — no foreign staged files.
  • Push race note: one direct git push was remote-rejected mid-run (concurrent fleet commits); sync_repo_root.py recovered and confirmed the push landed.

Living-docs drift

  • living_docs.py check: FAIL — 199 findings (dead legacy anchors; completed clean-results in no question's evidence; relates_to ⇄ evidence mismatches; spec-role-header State trailer stale vs 7 promoted results).
  • drift_hash 078285c14444 — identical to the 07-07/07-22/07-23/07-24 rows.
  • drift proposal already open (logged 2026-07-07); not re-proposing. Event-stream row appended for tonight.

My thoughts

Highlighted results

  • no results promoted today