Daily — 2026-06-12 (heavy autonomous pipeline day; ~30 workflow fixes auto-applied in-flight, 0 promotions)

What happened

• Fully autonomous, very high-throughput day — 2384 commits across the fleet. Multiple per-issue sessions ran the whole /issue pipeline concurrently: #613, #621, #622, #627, #628, #633, #483, #545, #522. #633 reached completed (reviewing → completed, test-verdict PASS, Step 10d auto-merge). #613 landed a clean-result ("Contrastive negatives drag the source marker implant down ~5×"), #621 landed one ("Rank-1 LoRA marker implants are ungated writes"), #627 iterated to interpretation v4.
• The workflow-fix-on-bug protocol carried the day. ~30 epm:workflow-fix-candidate → applied cycles fired and auto-merged in-flight across #522, #533, #545, #554, #572, #578, #580, #597, #598, #607, #610, #612, #613, #616, #621, #622, #627, #628, #633 — verifier check-count rot removed, verify_plan.py c11 dry-run smoke gate, bootstrap_pod fixes, dispatcher CVD-pin lint, watcher pass-level fixes, mentor-update-slides repairs, and more. Most fixable problems today were fixed the moment they surfaced.
• Two new always-on rules merged from Thomas's directives: .claude/rules/on-policy-completions.md (on-policy-first training completions — "all training behaviors as much on policy as possible", #616) and the existing contrastive-negatives sibling. #619 was filed (proposed) to capture the "think through LoRA placement" human task.
• Agent-memory was pruned (deliberate workflow-improver pass): codex-critic 200k→106k (-47%), critic 273k→198k (-27%); live state snapshotted pre-prune then re-pruned. Failure-lessons were consolidated as they landed — #621's TrainerCallback.on_log never sees direct wandb.log metrics → gotchas.md; #542's HF 5xx mid-upload retry → experimenter + experiment-implementer memories.
• Two tasks ended the day blocked (both legitimately): #545 (cached agent-defs — see notes) and #615 (clarifier blocking ambiguities — needs Thomas's answers).

Applied workflow improvements

1. Target: .claude/skills/mentor-update-slides/SKILL.md — what: added an Output Rules "one idea per slide — split, don't cram" overflow rule — commit: b9976069f Why: Thomas, slides session 2026-06-12 ([2184e3ee]): "some slides are overflowing and need to be separated" — and again "I think some slides are overflowing and need to be separated". The skill had no general overflow/splitting guidance: only a per-example "too long → put a 1-2 line excerpt on the slide and the full text in the Appendix" note buried in the qualitative-example bullet. Nothing told the deck author to split a dense slide. Applied edit:

     - Include confidence on every claim slide.
   + - **One idea per slide — split, don't cram.** If a slide's content (bullets,
   +   table rows, a figure plus its caption and qualitative example) does not fit
   +   comfortably at a readable font size, SPLIT it across multiple slides
   +   (`Result X (1/2)` / `Result X (2/2)`) or move the overflow detail to the
   +   Appendix — never shrink the text or pack a slide wall-to-wall. A mentor reads
   +   the deck live in a meeting; a dense slide goes unread. Rules of thumb: a table
   +   over ~6-7 rows, a bullet list over ~5 lines, or a slide carrying both a figure
   +   and a long table is a split-or-Appendix signal. (Thomas, 2026-06-12: slides
   +   were overflowing and needed to be separated.)
     - Preserve existing log and appendix sections when updating an existing deck.
   

   Lint gate (workflow_lint.py --check-references + --check-asks): PASS.

Other problems & notes

• #545 blocked — cached agent-defs after Fable 5 was removed upstream. Held: not a live file bug. Thomas ([71cef8cf]): "All of my fable 5 sessions died because fable 5 got removed." 25 .claude/agents/*.md files had been pinned to model: "claude-fable-5[1m]"; when that model id disappeared, every spawn died with 'claude-fable-5[1m]' at ~500ms. The on-disk files were already restored to a valid claude-opus-4-7[1m] in-flight (commits 00566584c, 1d8198350, "restored 25 .claude/agents/*.md"). The block is purely a session-runtime cache (the Agent SDK reads agent defs once at session start), so the live files are clean — verified: no fable-5 in any live agent/skill/rule/script/workflow.yaml/CLAUDE.md (only an illustrative --model alias example in spawn_session.py:691). Suggested action: respawn #545 with spawn_session.py spawn-issue --issue 545 --auto (no GPU is running; nothing is burning). The deeper fragility — agent files pinning a specific model that can vanish upstream — is a genuine judgment call about agent-model strategy (omit the per-agent pin and inherit? maintain a rotting allowlist guard?) that's yours, not auto-fixable.
• #615 blocked — clarifier blocking ambiguities (failure_class: data). Held: the Step 1 clarifier gate working as designed; "Autonomous paper pipeline: proposal-in → autonomous progress (design)" needs your answers on the blocking ambiguities. The watcher correctly did NOT auto-respawn (status not ACTIVE) but has flagged it ALIVE-BUT-STALLED for ~60min. Suggested action: open the session and answer the clarifier, or archive if superseded.
• 17 real stop_reason:refusal + 22 real Request rejected (429) today (precise signatures, after excluding the CLAUDE.md-quote false matches — the raw grep -l counts of 42/29 were inflated by CLAUDE.md being loaded into every transcript). Both are known classes with documented recovery (CLAUDE.md "Spurious usage-policy refusals" + "429 token-pacing"); the pipeline completed, so recovery worked. No single derivable fix — noting the volume for awareness; if refusals keep climbing, a thinner-wording pass on the most trigger-dense subagent prompts is the lever.
• Methodology jargon on mentor slides. Thomas ([2184e3ee]): "What does this mean: 10 held-out contexts + 20% of cells quarantined for predictor scoring" and "make it clear this is validation/test set". Borderline: the existing plain-English-condition-names rule + the "explain obscure concepts intuitively" memory cover the spirit, but neither explicitly bans methodology jargon ("quarantined cells", "held-out contexts") in mentor-facing prose. One instance — suggested action: if this recurs, extend the mentor-slides plain-English rule to methodology terms, not just condition codes.
• Stray 0-byte file **Figure.** in the repo root (created Jun 10, untracked). Junk from a botched redirect. Held: deleting an untracked file isn't git revert-able (carve-out: destructive/irreversible). Suggested action: rm './**Figure.**' when convenient — it's empty.
• Uncommitted working-tree edits to scripts/persona.sh (40+/28-) sitting on main. Not touched — could be an in-progress manual edit; committing unknown working-tree changes is not the daily's call. Suggested action: review and commit or discard.

My thoughts

Highlighted results

• no results promoted today