EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 4 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

Make the Step 5a / Step 10d spec-freshness sync family-atomic (never sync a derived or import-partner artifact whose partner was skipped for branch-side edits) and move the Step 10d pre-gate re-sync out of the race with the ~30-minute lint gate.

Workflow gap

Bug observed: the per-path "branch-side edits — skipping" guard skips one member of a coupled family while syncing the rest from main, manufacturing a tree the branch never created — a synced markers.md against a skipped workflow.yaml (lint gate block), a synced pin test against a skipped scripts/workflow_lint.py (pytest collection ImportError, rc=2) — and, separately, the Step 10d pre-gate re-sync snapshots origin/main and then ~30 min of lint gate elapses before the merge, so the merge routinely returns CONFLICTING.
Why it is a workflow gap: the skip grain is PER-ITEM by design (fail-safe against the #535 clobber), but the SPECS list contains items that are derived from, or import from, other items in the same list — so a per-item skip is not fail-safe at the family level, and the sync's placement relative to the gate is fixed by the SKILL recipe.
Confidence (emitter): high
verified-at-filing: grep -n 'spec-freshness\|SAFE_SPECS\|branch-side\|SPECS=' .claude/skills/issue/SKILL.md → ~30 hits in the single named target, incl. the Step 5a block at L2236-2281 (SPECS= at L2249; the per-item skip at L2259-2275; the sync + commit at L2277-2281) and the Step 10d pre-gate re-sync bullet at L10240-10276. Presence claim confirmed per-target. Coupling verified: grep -n 'generated\|emit-tables' .claude/skills/issue/markers.md → L62-64 "The kinds table is auto-generated from … uv run python scripts/workflow_lint.py --emit-tables to regenerate"; grep -n 'markers.md' scripts/workflow_lint.py → L751 _REPO_ROOT / ".claude" / "skills" / "issue" / "markers.md". Landed-fix check: git log --oneline --since='7 days ago' -- .claude/skills/issue/SKILL.md shows no commit implementing family atomicity or a post-gate re-sync. Context binding: the paragraph at SKILL.md L2332-2337 already ACKNOWLEDGES the symptom ("a workflow test that FAILs inside a long-lived issue worktree … including a collection-time ImportError from a workflow_lint / rules-pin symbol — is worktree-staleness … cross-check at the repo root") but prescribes only a diagnosis habit, not family atomicity — the fix is NOT landed. (2026-07-26)

Evidence

Defect (i), derived-artifact shape. Session 0e2c3b21, 2026-07-26T11:50:30Z → 12:08:19Z: the Step 10d pre-push lint gate returned block on a markers.md auto-generated-table staleness. The branch had edited .claude/workflow.yaml, so workflow.yaml was skipped while .claude/skills — which CONTAINS the derived markers.md — was synced from origin/main, importing #1692's regenerated markers.md alongside the branch's pre-#1692 workflow.yaml. Session's own diagnosis: "The block is stale workflow.yaml: my branch predates #1692 (164b46cd5c) which added per-arm-resolution/PASS_PARTIAL prose to the epm:smoke-architecture-check marker AND regenerated markers.md. My spec-freshness syncs skipped workflow.yaml because I edited it (Edit D)." Cost: ~18 min (7 min diagnosis + 11 min gate re-run), one unplanned merge commit, forced merge-form change to --squash.
Defect (i), import-partner shape. Session 2de5253e, 2026-07-26T14:23:21Z → 14:27:50Z: the Step 5a sync pulled in new tests/test_* files from origin/main but SKIPPED .claude/agents, .claude/skills, and scripts/workflow_lint.py for branch-side edits. Result: "Collection ImportError on test_workflow_lint_inline_round_duty_mirror.py: the test file was synced from origin/main but the workflow_lint.py function it imports is only on newer main (per-file guard skipped syncing workflow_lint.py because the branch has its own edit to it). Classic \"stale-family\" from Step 5a." — collected 5039 items / 1 error, PYTEST_RC=2. Cost: one wasted Step 9c gate invocation + ~5 min, and it was the direct trigger of that session's first merge conflict.
Defect (i), compound shape. Session e3b70618, 2026-07-26T14:50:07Z → 16:10:28Z, on a one-line SPECS= widen in SKILL.md: the pre-gate re-sync committed 3 files from main onto the branch but skipped .claude/skills/issue/SKILL.md (branch-side edits) — exactly the file the freshly-synced check pins. Gate output: workflow_lint: check-inline-round-duty-mirror: .claude/skills/issue/SKILL.md has 0 occurrence(s) of anchor '(1) BEFORE any ridge', expected exactly 1 (part (a) count invariant), alongside spec-freshness: .claude/skills carries branch-side feature edits — skipping blind sync. Cost: 3 pre-push lint-gate runs (~34 + ~31 + ~7 ≈ 72 min of gate wall-time) plus ~45 min of orchestrator diagnosis for a 1-line diff; epm:merged records merge_attempts: 3.
Defect (ii), the gate race. Session 67cf175e: pre-gate re-sync at 15:54:22Z committed 3 workflow-surface files (156+/153-) from origin/main onto issue-1711; the lint gate then ran ~30 min in the background; during that window main advanced further on scripts/workflow_lint.py and the squash merge was refused at 16:25:49Z — "head-sync pre-check: parity at 0dd49d67571c (mergeable=CONFLICTING) … X Pull request superkaiba/explore-persona-space#1476 is not mergeable: the merge commit cannot be cleanly created." and "CONFLICT (content): Merge conflict in scripts/workflow_lint.py". Cost: ~47 min of Step 10d wall-clock (15:53 → 16:40), one extra full lint-gate run, one residual-conflict subagent.
The two defects compound: e3b70618's sync ALSO re-added 139 lines of workflow_lint.py that origin/main had removed hours earlier, and the near-miss silent revert was caught only by an ad-hoc post-gate git diff origin/main HEAD -- scripts/workflow_lint.py. The orchestrator dropped the sync commit (git reset --hard HEAD~1 in the worktree) and ran a third gate.

Proposed change

Family atomicity in the Step 5a block (.claude/skills/issue/SKILL.md, the SPECS= loop at L2249-2281). Declare the coupled families explicitly and make the skip transitive: .claude/workflow.yaml ↔ its derived .claude/skills/issue/markers.md and the SKILL.md generated-table block; scripts/workflow_lint.py ↔ :(glob)tests/test_workflow_lint*.py; .claude/hooks ↔ :(glob)tests/test_guard_*.py. When ANY member of a family is skipped for branch-side edits, skip the WHOLE family (or fail loud naming the skipped member and the members being withheld) rather than syncing the rest.
Derived-artifact alternative for the workflow.yaml family. Where the derived artifact is mechanically regenerable, an acceptable substitute for skipping is: sync the family, then run uv run python scripts/workflow_lint.py --emit-tables and fold the regenerated markers.md / SKILL.md tables into the SAME sync commit. Add a one-line note at the skip site naming markers.md as GENERATED FROM workflow.yaml so the coupling is visible where the decision is made.
Post-sync fast-forward assertion. After the sync commit, run git diff --stat origin/main HEAD -- <synced paths> and assert the sync is a pure fast-forward of the synced paths — refuse to sync a path whose branch content would be strictly ahead of origin/main (the 139-line re-add case), since main may have deliberately REMOVED the thing being "freshened".
Move the Step 10d pre-gate re-sync out of the race (.claude/skills/issue/SKILL.md L10240-10276). Either (a) run the re-sync AFTER the lint gate returns, immediately before the push + merge, or (b) keep the pre-gate placement and re-run it when the gate took longer than ~5 min, re-deriving freshness against a freshly-fetched origin/main immediately before committing. Option (a) is preferred — the gate's own landing-tree copy already builds from git archive origin/main, so a post-gate sync does not invalidate the verdict.
Both sites carry the same SPECS loop — Step 5a (L2236-2281) and the Step 10d pre-gate bullet (which instructs running the Step 5a block with origin/main) — so the family-atomicity change must land once and be inherited by both, not duplicated divergently.

Scope / surfaces

Primary target: .claude/skills/issue/SKILL.md
scripts/workflow_lint.py (only if the --emit-tables fold is chosen as the derived-artifact remedy; no behavior change expected)
tests/test_issue_skill_lint_family_sync.py (existing gate-region pins — the Step 10d section carries a negative assert whose region spans the pre-push gate block; any edit there must keep that pin green)

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.
The per-item skip must stay FAIL-SAFE — the #535 incident (blind sync clobbered a feature branch's own marker registrations) is why the guard exists; family atomicity must widen the skip, never narrow it into a clobber.
The Step 10d gate-region negative assert in tests/test_issue_skill_lint_family_sync.py bans writing a full-message grep-exclusion invocation into that section; keep the subject-scoped convention.
This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

fingerprint: 8a12cccef12a
workflow_fix_target: .claude/skills/issue/SKILL.md
fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: J-P1, H-P2, D-P3, G-P6.