EPS
← All docs·Activity

Daily — 2026-06-20 (quiet research/synthesis day; no experiments committed)

updated: 2026-06-23

What happened

  • Quiet day on the project: no commits, no task-state changes in the repo since 2026-06-18 evening. All of today's activity was research-thinking + result-synthesis chat (5 sessions, all in the 22:38–23:27 PT window).
  • PM session (/pm): restarted #651 (HF-Hub-504 block — see Other notes), archived #655, and surveyed the follow-up landscape across the 52 awaiting_promotion results + 19 on_hold parked follow-ups. Meta-read: the base behavioral prior beats every hand-designed geometry predictor (#518/#537/#545/#605/#603/#649 all null on cosine/JS), with two surviving positive signals — #623's behavior-direction alignment (ρ=0.73) and #559's base-self-scoring. Spawned several background follow-ups (extend #623's predictor across behaviors + to leakage; generalize #559's base-self-scoring predictor).
  • Formalization sessions (two, same prompt): formalizing the "context C → behavioral profile" framework — C induces a state that deterministically maps a query to answer A, with behavior B = measurement μ_B(A). Removed a CLAUDE.md rule (anchor math to Dan's Slack message verbatim) at Thomas's request.
  • Result-synthesis sessions: summarized the rank-1 read/write LoRA-geometry cluster (#604, #621, #650, #653, #568, #637); deep-read #604/#621/#650/#653; set up an explicit same-issue follow-up on #653 (rewrites its "diffuse but install-limited" headline → folds onto #653, not a child).

Applied workflow improvements

  • no workflow-fixable problems found today — the day's one mechanical recurring gotcha (Beeper MCP limit ≤ 20) lives in the global slack skill, which is /memory-sleep's domain, not the project workflow surface this skill is scoped to. It's logged below. Project lint baseline (workflow_lint.py --check-references) PASSes; nothing in .claude/ / CLAUDE.md / scripts/ needed a fix tonight.

Other problems & notes

  • Beeper MCP rejected limit > 20 twice (session 3ec18b3f, the formalization session pulling Dan's Slack message): MCP error -32603: 500 … "Number must be less than or equal to 20" … path: [limit]. Recurring, mechanical, and concrete — but the owning file is the global ~/.claude/skills/slack/SKILL.md (no project workflow surface references Beeper). Suggested action (route to /memory-sleep): add an explicit "Beeper list/search calls cap limit at 20 — never pass higher" guard to the global slack skill.
  • #651 didn't auto-restart (session 730d8e82, Thomas: "Why didn't it automatically get restarted"). Not a bug — by design: the watcher's capacity-retry pass only re-drives blocked tasks whose latest failure is failure_class: infra + reason: no_compute_available. #651's block was a transient HF-Hub 504 (different reason) and the in-session 5-retry HF recovery cap was already exhausted, so it stayed human-gated. Restarted manually. Suggested action (judgment call — held): decide whether transient external-dependency blocks (HF 504) with an exhausted retry budget should self-heal after a longer backoff, or stay human-gated as now. This alters auto-recovery behavior (blast radius across the fleet), so it's a design call for Thomas, not an auto-fix.
  • "An automatic followup should've run" on #653 (session 33364133). Not a bug — the free-analysis (0-GPU) follow-up did fire at Step 9a-ter; the GPU-backed follow-up machinery (proposer + cheap-band auto-run) lives at Step 10, which only runs after task.py promote 653, and #653 has been parked at awaiting_promotion since 10:37Z. Handled correctly via an explicit same-issue follow-up. Note: the promotion-gate dependency is a recurring source of "why didn't the follow-up run" confusion; no concrete one-line fix, but worth keeping in mind if it recurs.
  • Concurrent repo-root committer swept the rule-removal edit (session 3ec18b3f): the "Remove that rule" CLAUDE.md edit was swept into a parallel session's commit before this session's git add ran, so its own commit found nothing staged. Documented concurrent-committer behavior; the session verified the rule was gone + pushed on origin/main. Net result correct, no fix needed.

My thoughts

Highlighted results

  • no results promoted today