EPS
← All docs·Activity

Daily — 2026-07-06 (infra wave — ~15 fixes landed;

updated: 2026-07-07

What happened

  • Infra/workflow-fix wave landed. Nearly all of yesterday's 10 daily-filed fixes completed end-to-end through the full pipeline today: #1075 (verify_plan GPU-basis), #1076 (test regressions on main), #1077 (step9c dirty-oracle fallback), #1078 (paper_palette n>8), #1079 (BLAS thread pins), #1080 (inline-run pod-safety markers), #1081 (crash-fix fix-engaged contract), #1082 (PM negative-existence search recipe), #1083 (REGISTRY drift repair), #1084 (sentinel-replay dedupe), #1085 (Step 10d guard-block recovery contract), #1086 (verify_plan c12/c18 false-positives), plus #1088, #1089, #709 (RunPod glob-sentinel poll), #1091, #1093–#1095. #1096 (judge max_tokens rule) is mid-pipeline.
  • #1073 (decoding-regime robustness of the context→answer map) completed its run — one crash-fix round on a reused bundle missing the prompts key — and parked at awaiting_promotion (queue now at 88).
  • #1090 (persona-vectors-style content-behavior datagen): design rewritten twice under Thomas's direction (standardized persona-vectors pos/neg behavior definitions, plain-text elicitation, Claude as generator); plan self-composed after 2 planner refusal-kills; GPU phase launched, now at interpreting. #1092 (realistic sparse-crossed prefix×query corpus, prefix-map vs context-map) planned (v5, 85 GPU-h) and running.
  • #1074 (stage-split harmful-compliance follow-up): vLLM engine-memory crash fixed via crash-fix rounds, Batch-API judging launched — then the orchestrator went refusal-wedged for ~2h (every wake refused); the watcher respawn recovered it and the round continued (still followups_running). #825 follow-up (base-separator-control) survived a GCP spot preemption + respawn; #833 follow-up rounds (nonverbatim-profile ablation, emission-rate baseline) completed → awaiting_promotion.
  • Interactive: day-long #779 writeup + analysis session — fitter fair comparison (ridge vs KRR vs MLP) rerun at n=10K with a widened λ grid on a fresh 1×H100, per-direction/persona-vector reconstruction analyses, Prefix/Query/Context/Answer terminology standardized, dashboard rebuilt; plus a 4-teammate literature deep-dive on characterizing the linear context→answer map, and an inline #813 rank-spectrum free analysis.

Applied workflow improvements

  1. Target: scripts/guard_repo_root_branch.sh — what: block message now states the deny skipped the ENTIRE compound command (regenerate earlier clauses' outputs before retrying) — commit: d66f723bde Why: #1085's session parked this exact candidate under the recursion guard: blocked sessions retry only the offending clause and fail on missing intermediate files (#813/#1056 incident class). Message-text only, no logic change; 270 guard tests pass. Applied edit: 
    -For marker-note text mentioning git commands, use --file <path.md> instead of --note; for commit messages, use git commit -F <file>." >&2
+For marker-note text mentioning git commands, use --file <path.md> instead of --note; for commit messages, use git commit -F <file>.
+NOTE: this deny blocked your ENTIRE compound command — earlier clauses did NOT run either; regenerate any files/state those clauses were meant to produce before retrying the safe form (incident class #813/#1056)." >&2
  2. Target: .claude/skills/issue/SKILL.md — what: bare-push reminder comment at the Step 10d push snippet (agents in 6 sessions today improvised git push | tail during merge recovery; the hook blocked each) — commit: 44484f0134 Applied edit: 
    +    # Run every push / gh pr command BARE — never piped through tail/grep/head
+    # (guard_piped_git_push.sh blocks the pipe; a pipe masks the exit code).
+    # This applies to IMPROVISED recovery commands too, not just this snippet.
     git -C "$WT" push origin issue-<N> \
  3. filed for review #1097: Step 10d lint verdict is consumed (rm'd) BEFORE gh pr merge runs — a rebase-refusal/conflict orphans the PASS verdict; #1082 hand-recreated the gate file (echo pass > verdict), defeating consume-once. Fix: consume after the merge attempt returns.
  4. filed for review #1098: guard_repo_root_branch.sh false-positive blocks pod-side destructive git inside ssh <host> '...' args (#779 had to move a pod reset into a script).
  5. filed for review #1099: verify_plan c12 ASCII-x boundary + _C18_NEG_DEFER_RE negation gaps (routed from #1086's recursion-guard-parked candidates; sketches already written).
  6. filed for review #1100: catch task.py lifecycle commits stranded on never-merged issue branches at commit time (the root cause of the 15-problem REGISTRY drift #1083 repaired; parked by #1083).
  7. filed for review #1101: GCP success-path poweroff didn't fire after [phase=done] — eps-issue-1074 billed 13 extra minutes until manual teardown; poller should auto-teardown a done-phase RUNNING instance past a short grace.
  8. filed for review #1102: extend the digest-only raw-completion read rule (analyzer.md + CLAUDE.md item (d)) to real-world-corpus (LMSYS-class) rollout text — 2 analyzer refusal-kills on #1073 before the retry brief carried it.
  9. filed for review #1103: on-policy-completions.md — multi-behavior datagen defines every behavior in the standardized persona-vectors pos/neg shape (the #906 failure Thomas caught; encodes his #1090 direction as a standing rule).
  10. filed for review #1104: refusal-wedged orchestrator ticks — keep issue-tick wake prompts vocabulary-thin for harmful-content-family tasks + document orchestrator-refusal-wedge recovery (#1074 lost ~2h of wakes).
  11. filed for review #1105: pin the Step 10d guard-block recovery-contract phrases in tests/test_step10d_guard3.py (#1085's parked prose follow-up).

Other problems & notes

  • Spurious usage-policy refusal storm — ≥17 kills today, concentrated on harmful-content-family tasks: #1074 (~12: planner ×2, fact-checker, critics, reviewers ×2, analyzer re-fold ×2, tick wakes; orchestrator fully wedged 20:36–22:33 UTC until the watcher respawn), #1090 (2 planner kills → orchestrator self-composed the plan), #1073 (2 analyzer kills from paging raw LMSYS text), #833 (1 implementer final-report kill). Recovery ladders worked everywhere; prevention filings: #1102, #1104.
  • #1074 Phase D (Batch-API judging of the two mixed cells) was dispatched but unharvested when the wedged session died; the watcher respawn (session a2343e65) resumed the round and self-composed the interpretation — task still followups_running at day end. Suggested action: none (watcher owns it); check it advanced by morning.
  • Piped git push/gh pr blocked in 6 sessions (#1074/#1075/#1076/#1078/#833 ×2 + merge-recovery improvisations in #1081/#1082) — guard_piped_git_push.sh caught every one; route-1 fix 2 adds the snippet-side reminder.
  • Foreground sleep-chains blocked 5× (#779 ×2, #833, #1079, one Codex-wait) — down from 15+ on 07-03; harness guard + existing rule sufficient, note only.
  • Already fixed same-day (no action): (a) pod-779 watcher auto-stops mid-bootstrap misdiagnosed as a flaky host — the pod-safety pre-launch-signals rule with the #779 citation landed in CLAUDE.md today; (b) "did we run X?" coverage misses (assistant twice wrongly said the token-axis map / separator control were never run — they were, in #922/#931; and needed two re-steers to find #813's per-example-vs-averaged-map follow-up) — the negative-existence search recipe incl. follow-up-label greps landed in research-pm.md via #1082 today.
  • #779 inline analysis started CPU-serial — Thomas had to explicitly order "reroute to GPU. Also vectorize and parallelize as much as possible"; the vectorize-first + compute-character pre-launch rules already cover this — recurrence worth watching, note only. Also: ridge λ pinned at the grid ceiling (1000) at every n → full rerun with a widened grid (experiment code; consider fail-loud/auto-widen on boundary-selected hyperparameters as a habit).
  • #779 session left eval_results/issue_779/ + figures/issue_779/ modified/untracked at repo root (fair_comparison.json, ffc figures, batch2_edges_smoke.json) — not committed by /daily since the session may still be live; the owning session should commit per the end-of-session rule.
  • Edit-before-Read harness slips: 10 failed Edits in one humanize batch on #833 (a fix belongs in the GLOBAL ~/.claude/skills/humanize/SKILL.md — out of this skill's project-scoped target list; deferred to /memory-sleep: "Read the scratch copy before the first Edit"), plus 2 one-off Write/Edit slips (#1074, /daily 07-05).
  • Concurrent-edit collisions: 4× "File has been modified since read" on the shared issue779 driver while the orchestrator and the ffc-implementer teammate edited it simultaneously — single-writer-per-file discipline when pairing with teammate sessions.
  • Step 10d rebase-refusal + tasks/ rename-conflict recovery hit #1081 and #1082 (~20–30 min each; documented #1041 shape, squash fallback worked). The gate-orphaning half is filed as #1097.
  • Misc one-offs, all self-recovered: reused-bundle key-schema crash on #1073 (already filed as #1087 yesterday); step9c ledger/venv missing in a fresh sparse worktree (#1086, clean-oracle fallback worked); a cp racing a concurrent status transition (#1083); task.py correctly refusing set-status approved during a followups_running hold (#1074); guard blocking cd "$WT" && git checkout ... compound forms (self-corrected to git -C); assorted shell/jq/heredoc quoting slips; a 33K-token whole-file Read rejected; an splitlines()[0] IndexError on an empty note.
  • No route-3 judgment calls were held today (the one candidate — #1074 wedge recovery — had a concrete derivable fix and went to route 2 as #1104; the #906/#1090 datagen design question was already decided by Thomas in-session).

Living-docs drift

  • living_docs.py check FAILED with 199 findings (hash b009cef9ecd9): ~75 relates_to pointers to dead legacy anchors (a1–e5-era), ~54 completed clean-results absent from any question's evidence (incl. recent #722, #763, #494–#501 line), dozens of question-evidence ⇄ relates_to mismatches, and the spec-role-header State trailer stale vs 7 promoted results.
  • Still-open prior proposals cover this same drift family: 2026-06-29 (1736fca0804f) and 2026-07-04 (d2a80db46b3b); today's snapshot has grown from 203 → 199 findings (4 registry inconsistencies repaired by #1083).
  • Proposal: re-synthesize open_questions.md (run the living-docs updater/backfill) to reconcile the drift above — needs your ok; not auto-applied. (Extends the still-open 06-29/07-04 proposals; logged to the dedup event-stream as b009cef9ecd9.)
  • Parked-proposal re-consideration pass: clean — no un-reconciled living-docs rejections, no parked-redundant follow-ups eligible for revival.

My thoughts

Highlighted results

  • no results promoted today