EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 3 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Close the two remaining shared-repo-root staleness windows in /issue: add a root-divergence probe at the Step 4 worktree cut, and move the Step 10d root sync ahead of the epm:merged post so a just-merged guard or spec fix is live at the root for the session's remaining root-side calls.

Workflow gap

Bug observed: the Step 4 worktree cut reads the shared root's LOCAL main with no divergence probe, and Step 10d's only root sync sits after epm:merged is posted — so a session can proceed against a diverged root and can be blocked by the very guard it just merged.
Why it is a workflow gap: .claude/skills/issue/SKILL.md Step 4 prescribes a rev-list --count main..issue-<N> pre-check against local main with no reconciliation path, and Step 10d's unconditional sync_repo_root.py is positioned inside the post-merge stale-task-folder guard, downstream of every epm:merged posting site.
Confidence (emitter): high
verified-at-filing: awk 'NR>=1968 && NR<=2100 && /is-ancestor|sync_repo_root|diverg/{c++} END{print c+0}' .claude/skills/issue/SKILL.md → 0 hits across the whole of Step 4 (absence-of-guard evidence); grep -n 'rev-list --count main..issue-<N>' .claude/skills/issue/SKILL.md → 1 hit (L2021), confirming the draft-PR pre-check reads local main; grep -n 'sync_repo_root' .claude/skills/issue/SKILL.md → 19 hits, the Step 10d unconditional pre-sync at L12127, versus the safe-case epm:merged success post at L11273 and the two artifact-confirmed posts at L12000 / L12068 — every epm:merged site precedes the sync in document order; git rev-parse --verify --quiet '9301f1a7bf^{commit}' → resolves, git log -1 --date=iso-strict 9301f1a7bf → 2026-07-26T04:03:24-07:00 task #1694: Step 10d apply-verification + post-merge guard pre-sync (#1465) (2026-07-26)

Evidence

The unconditional post-merge root sync is already LANDED and is excluded from this filing. Commit 9301f1a7bf (2026-07-26T04:03:24-07:00 = 11:03:24Z) added the Step 10d post-merge stale-task-folder guard pre-sync at L12127: uv run python "$REPO_ROOT/scripts/sync_repo_root.py" || echo "post-merge guard pre-sync: …", in a section headed Post-merge stale-task-folder guard (runs after EVERY merge form lands). Filing a second "add sync_repo_root after gh pr merge" change would duplicate it. What remains open is its POSITION and the Step 4 probe.
Session 2b779905 merged #1699 at commit d79fa07b0e (git rev-parse resolves; 2026-07-26T06:01:52-07:00 = 13:01:52Z), whose whole content is: bare ruff check uses pyproject.toml per-file-ignores that relax scripts/*, so the implementer must also run pytest tests/test_ruff_policy.py::test_live_workflow_helpers_clean_under_full_ruleset. It merged server-side and the shared root stayed behind=2 (confirmed in session 891b2cc6 row 445: ahead=29 behind=2).
Session 891b2cc6 spawned its implementer at 13:39:39Z — 38 minutes after that merge — reading the STALE .claude/agents/implementer.md. Its round-1 report: "**Lint:** uv run ruff check scripts/autonomous_session_watch.py … && uv run ruff format --check … — All checks passed; 3 files already formatted." (no pin test). The 26-minute Step 9c gate then FAILed: "FAIL count=1 … NEW: tests/test_ruff_policy.py::test_live_workflow_helpers_clean_under_full_ruleset — the FULL-ruleset ruff check (no per-file-ignores) surfaces 2 payload errors" (RUF003 and C901). Measured cost: roughly 30 minutes of wall clock for a second full 26:24 gate, plus one extra implementer round and one extra verdict cycle.
Session 7ce3a81f, 09:14:11.489Z: 25 seconds after the PR squash-merged, task.py post-marker 1710 epm:merged --note "…recovery: git merge origin/main into branch…" was blocked — "PreToolUse:Bash hook error: [scripts/guard_repo_root_branch.sh]: BLOCKED: 'git merge (branch merge on the shared root)' would move the SHARED repo-root tree off main…" — because the pre-fix guard was still on disk at the root. The session's own read: "Guard tripped on --note prose containing \"git merge\" (my own fix hasn't propagated to the shared root yet — pre-fix guard is running here). Using --file route." Cost: 1 blocked call, 1 cancelled sibling call, a Write and a re-issue, roughly 35 s; no data loss.
Session c0319d9e, 12:02:42Z–12:03:04Z: at the Step 4 worktree cut, git rev-list --count main..issue-1701 returned 2 before the implementer had committed anything. Investigating, the session found local main and origin/main mutually non-ancestral — "origin/main is NOT ancestor of main\nmain is NOT ancestor of origin/main" — then moved straight on: "Draft PR will open post-implementer commits. Dispatching implementer now." A grep over every Bash command in that session found 0 sync_repo_root.py invocations. No cost realized this run (Step 10d merged server-side, bypassing the root); the latent cost is that the next session's root push is rejected.

Proposed change

In .claude/skills/issue/SKILL.md Step 4a, before the draft-PR pre-check at L2021, add a root-divergence probe: if git -C "$REPO_ROOT" merge-base --is-ancestor origin/main main fails AND git -C "$REPO_ROOT" merge-base --is-ancestor main origin/main also fails, run uv run python "$REPO_ROOT/scripts/sync_repo_root.py" — the sanctioned single-flight recovery — before proceeding. Fetch origin/main first so the probe reads a current ref.
Make the divergence outcome visible rather than silent: on a detected divergence, record it in the chat line and re-probe after the sync; a still-diverged root after one sync attempt is reported, not stepped over.
In Step 10d, run the root sync BEFORE the epm:merged post at L11273 (and before the artifact-confirmed posts at L12000 / L12068), not only inside the post-merge stale-task-folder guard at L12127. The guard's own pre-sync stays where it is — it is idempotent and single-flight, so the earlier call costs nothing when the root is already current.
Make the Step 10d epm:merged recipe use the task.py post-marker --file form by default. Merge-recovery notes routinely quote git verbs and the repo-root guard matches on note prose, so the --note form is the fragile default.
Note in Step 10d that a merged diff touching scripts/*guard* or .claude/hooks/* is not live at the shared root until the sync runs, so the landing session cannot exercise its own fix before then.
unverified hypothesis — verify at plan time: that placing the sync before epm:merged is safe on every Step 10d exit path (safe-case, merge-conflict recovery, artifact-confirmed / surgical-additive checkout). The sync pull-rebases the root and can move the canonical task-folder path, which the post-merge guard already re-resolves; whether an earlier call interacts with the epm:merged idempotency skip at L9735 was not checked at compose time.

Scope / surfaces

Primary target: .claude/skills/issue/SKILL.md
none

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): 2b779905 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): 891b2cc6 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): 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): c0319d9e 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: 845a0f74c1f5
workflow_fix_target: .claude/skills/issue/SKILL.md

/daily 2026-07-26 route-2 filing. Miner refs: F-P1, C-P2, I-P4.