EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 2 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Move the Step 10 CRON-TEARDOWN and the completed / epm:done transition to AFTER the Step 10d merge posts epm:merged, so the merge window keeps re-drive coverage and the durable record never claims done while the branch is unmerged.

Workflow gap

Bug observed: Step 10 step 6 tears down the /issue-tick backstop cron and applies the terminal completed status before Step 10d runs, so the entire merge — up to 33 minutes with conflict recovery and two background lint-gate waits — executes with no re-drive coverage and with epm:done already posted against an unmerged branch.
Why it is a workflow gap: the ordering is written into .claude/skills/issue/SKILL.md Step 10 step 6, Step 10d is absent from the § CRON-TEARDOWN exit-site list, and two sibling sessions read the spec two opposite ways within the same hour.
Confidence (emitter): high
verified-at-filing: grep -n 'Run CRON-TEARDOWN before applying the terminal status' .claude/skills/issue/SKILL.md → 1 hit (L9320, presence confirmed); awk 'NR>=9711 && NR<=12320 && /CRON-TEARDOWN/{c++} END{print c+0}' .claude/skills/issue/SKILL.md → 0 occurrences of CRON-TEARDOWN anywhere inside Step 10d (L9711–12320, the whole auto-merge step) — the absence-of-guard evidence; grep -n 'at \completed` (Step 10 auto-complete)' .claude/skills/issue/SKILL.md→ 1 hit (L4693) confirming the exit-site list names Step 10 but not Step 10d;grep -n 'long-phase-heartbeat' .claude/skills/issue/SKILL.md` → 2 hits (L4808, L6364), neither inside Step 10d (2026-07-26)

Evidence

Session 7ce3a81f: 08:41:13.376Z CronList {} → 08:41:17.400Z CronDelete {"id": "07ebadfd"} → 08:41:20.282Z RESULT: No scheduled jobs., then Step 10d began at [08:41:52] TOOL_USE Bash: … Guard 0: agent-memory pre-commit … and the merge landed MERGED mergedAt=2026-07-26T09:13:46Z. The measured cost is 33 minutes of merge — including a shape-2 merge-conflict recovery and two roughly 12-minute background lint-gate waits — with no cron armed. No loss was realized on this run; the session survived.
Sibling session #1695 the same day deferred CRON-TEARDOWN to 09:25:24Z, after the merge landed — the opposite ordering, applied within 11 minutes of 7ce3a81f's. The spec is being read two ways.
Session e3b70618 (task #1709): epm:done posted at 2026-07-26T14:48:35Z, epm:merged at 2026-07-26T16:11:00Z — 83 minutes during which the durable record read completed + done with the branch unmerged. The epm:merged note reads **Merge attempts:** 3 (2 blocked by workflow-lint gate; the third PASS came after dropping the ….
Session 8380a48c shows the same shape at smaller scale: epm:done 10:16:41Z → epm:merged 10:51:09Z, a 34-minute window.
That window is exactly the state the watcher's completed_unmerged pass (#1564/#1653) flags and, on a second flagged interval with no live owner, respawns /issue <N> --auto against. The pass runs hourly with a day-capped respawn — far slower than the 45-minute tick the teardown just removed. Its recovery row is already documented at .claude/skills/issue/SKILL.md L12458. On both 2026-07-26 sessions the owning session was alive and did merge, so no spurious respawn was observed; the ordering nonetheless makes a long-tail merge indistinguishable from a stranded one.
unverified hypothesis — verify at plan time: that a session wedge or refusal-death inside the uncovered merge window would in fact have left the task completed + unmerged. No such death occurred in either transcript; the claim is inferred from the ordering, not observed.

Proposed change

In .claude/skills/issue/SKILL.md Step 10 step 6 (L9320): move CRON-TEARDOWN out of the pre-terminal-status position. Keep the /issue-tick backstop armed through Step 10d and tear it down at the Step 10d exit site, once epm:merged has been posted (or once Step 10d has terminally failed and routed to its epm:merge-failed handling).
Reorder Step 10 so the terminal completed status and the epm:done marker (step 7) follow the Step 10d merge rather than precede it, so epm:done and epm:merged cannot be separated by a long merge window. If the reorder is judged too invasive for the awaiting_promotion experiment path (which merges at Step 9b and reaches Step 10 already merged), scope the reorder to the code-change path that actually merges at Step 10d.
Add Step 10d to the § CRON-TEARDOWN exit-site list at .claude/skills/issue/SKILL.md ~L4691–4700, next to the existing at \completed` (Step 10 auto-complete)` bullet, and update the parallel prose at ~L4839.
Cheaper companion, and a useful backstop even if the reorder lands: on any Step 10d gate block or re-run, post an epm:progress [long-phase-heartbeat] step10d-merge attempt=<k> note, so the completed_unmerged audit in scripts/autonomous_session_watch.py can distinguish "merge in flight, attempt k" from "stranded". Alternatively have that pass treat an epm:progress newer than epm:done as liveness.
Keep the teardown idempotent and keep the not-found-is-success semantics; the existing § CRON-TEARDOWN procedure is unchanged, only its call site moves.
Add a contract pin so the two orderings cannot diverge again across sibling sessions.

Scope / surfaces

Primary target: .claude/skills/issue/SKILL.md
scripts/autonomous_session_watch.py (only if the completed_unmerged liveness-signal companion is taken)

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
uv run python scripts/workflow_lint.py passes (no-flags); ruff clean on touched files.
This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

sha-verify (filing-time, #1467): 7ce3a81f cited in commit context does NOT resolve as a commit in this repo at filing time — treat as a transcript/session reference, not a commit.
sha-verify (filing-time, #1467): 07ebadfd cited in commit context does NOT resolve as a commit in this repo at filing time — treat as a transcript/session reference, not a commit.
sha-verify (filing-time, #1467): e3b70618 cited in commit context does NOT resolve as a commit in this repo at filing time — treat as a transcript/session reference, not a commit.
sha-verify (filing-time, #1467): 8380a48c cited in commit context does NOT resolve as a commit in this repo at filing time — treat as a transcript/session reference, not a commit.
fingerprint: 4279aa5705ab
workflow_fix_target: .claude/skills/issue/SKILL.md
fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: I-P2, J-P11.