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

Make the implementer's NOT-RUN escape PRE-EMPTIVE — when the Step 9c selector reports a selected file in its slow_tests_selected list, route it straight to NOT-RUN + defer to Step 9c with zero local attempts.

Workflow gap

Bug observed: Implementers attempted tests/test_workflow_lint.py locally under sub-600 s fences at least six times across three sessions on 2026-07-26; every attempt was timeout-killed and every session ended in the same NOT-RUN escape it could have taken at minute zero.
Why it is a workflow gap: .claude/agents/implementer.md frames the escape post-hoc ("if a mandatory-set file genuinely cannot finish in-turn … the existing NOT-RUN escape applies"), and cites a stale 319-771 s runtime, while the selector already publishes the machine-readable signal (slow_tests_selected, and a 2400 s surcharge for this exact file) that neither implementer spec reads.
Confidence (emitter): high
verified-at-filing: grep -n 'NOT-RUN\|test_workflow_lint\|slow_tests_selected' .claude/agents/implementer.md .claude/agents/experiment-implementer.md → implementer.md 3 hits (L174 post-hoc escape naming tests/test_workflow_lint.py at "319-771 s"; L197 generic NOT-RUN escape; L254 report field), experiment-implementer.md 1 hit (L953, generic NOT-RUN, file not named); slow_tests_selected and recommended_timeout_s → 0 hits in BOTH specs. grep -n 'SLOW_TESTS\|slow_tests_selected' scripts/select_step9c_tests.py → the selector emits "slow_tests_selected" in its --json payload (L1797) and carries SLOW_TESTS = {"tests/test_workflow_lint.py": 2400} (L740-746) whose comment records the wall growing "771 -> 1819 s max" and a measured 1188.62 s standalone. git log --oneline --since='7 days ago' -- .claude/agents/implementer.md .claude/agents/experiment-implementer.md → 6 commits, none making the escape pre-emptive (2026-07-26)

Evidence

Session 0e2c3b21 (implementer subagent agent-a2759a68efd2d1b7f), 09:32:52Z→10:23:44Z: four attempts on tests/test_workflow_lint.py (9,088 lines / 531 tests) — timeout 500 SIGTERM'd, timeout 850 inside a 900 s Bash (harness-converted to background, Terminated), a detached timeout 900 + poll loop, plus two intervening pgrep/sleep poll loops — before taking the documented NOT-RUN escape. Its own report: "rc=124 = timeout hit (900s exceeded …). Made it to ~54% (~380/700+ tests) before the 900s cap. Under this fleet load test_workflow_lint.py needs >900s … The result: NOT-RUN due to timeout." Measured cost ~51 min of implementer wall-time, 4 discarded runs, on a shared VM already running 3 concurrent Step 9c gates. Step 9c covered the file 40 min later with a 5310 s budget.
Session 8571eca6, 12:14:20Z: the #1698 implementer reported the same file "timed out at both 540s foreground and 1400s background under fleet load from 2 concurrent Step 9c gate runs on the shared VM", naming it NOT-RUN in the report.
Session a5a4b7bd (subagent a510618c), 07:48:25Z→07:57:25Z: timeout --kill-after=30s 540s uv run pytest tests/test_workflow_lint.py -x -q returned Exit code 143 Terminated; the implementer then ran a -k subset (29 passed / 535 deselected in 4 s) and deferred the remainder. Measured cost ~9 min on a killed run, and the implementer-side verification of a file it had just edited rested on 29 of 564 tests.
The failure was predictable at dispatch in every case: the file is the sole entry in the selector's SLOW_TESTS surcharge table, whose in-tree comment records a measured max of 1819 s — above every fence any of the three sessions could set, since the Bash tool cap is 600 s foreground.
Net information gained across all three sessions: zero. Every run ended at the same NOT-RUN escape, and Step 9c ran the file to completion in each case.

Proposed change

In .claude/agents/implementer.md § "Run tests — gate-matched scope (#1288)" (L174), replace the post-hoc sentence with a pre-emptive rule: any selected file appearing in the selector's --json slow_tests_selected list is routed to NOT-RUN + Step 9c deferral with ZERO local attempts; report it in (c) How to verify with the exact copy-pasteable command and the selector's recommended_timeout_s.
Drop the stale "319-771 s" figure — it is contradicted by the selector's own SLOW_TESTS comment (max 1819 s, 1188.62 s standalone) and reads as "worth one try". Cite the selector field, not a frozen number, so the spec cannot drift again.
Keep the existing distinction intact: a pin-sweep HIT left NOT-RUN stays presumptively blocker-adjacent for the code-reviewer; the pre-emptive route changes WHEN the escape is taken, not its downstream treatment.
Add the corollary that a -k subset is an acceptable local substitute ONLY when it is named explicitly in the epm:results marker with its deselected count, so a 29-of-564 subset can never read as a full local run.
Mirror the same sentence in .claude/agents/experiment-implementer.md (its NOT-RUN escape at L953 is generic and names no slow file).
unverified hypothesis — verify at plan time: a fleet-wide single-flight around Step 9c / pre-push lint-gate launches would remove the contention that pushed this file past its historical band (session 8571eca6 proposed it). That is a larger change than the pre-emptive escape and should be scoped separately; the escape is correct regardless of contention.

Scope / surfaces

Primary target: .claude/agents/implementer.md
.claude/agents/experiment-implementer.md (mirror sentence)

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

fingerprint: 9ba4711d501b
workflow_fix_target: .claude/agents/implementer.md
fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: H-P1, E-P15, J-P8.