EPS Dashboard

Overview / Motivation

Auto-filed by /daily 2026-07-20 from the transcript problem sweep. THREE same-day incidents of teammate/subagent idle-notification ambiguity causing duplicate or re-done work: (1) #1112 rankem — a second implementer (i1112-rankem-impl2) was spawned for M4/M5/B3 after the first self-reported "too deep to do them safely", but the first landed all three anyway ("the dispatcher half of M5+B3 got implemented twice by parallel sessions"; a cancel/push cross followed at 05:18); (2) #1112 xmethod — "Re-render hasn't landed — the implementer went idle without committing it. Doing the fix myself", while the implementer's commit landed in parallel (duplicate fix race); (3) #958 round — mixed-turn-fit-958 idled repeatedly, the orchestrator finished + committed its work, then "The agent just woke up on step 5 — standing it down before it races me."

Goal

Add a teammate-coordination rule to the orchestrator guidance: (a) ONE implementer per file set — never spawn a second implementer over the same file set while the first is live, without standing the first down and confirming its durable state first; (b) an idle notification is NOT a done/stall verdict — probe durable state (commits, files, markers) before re-doing or re-assigning a live teammate's work; (c) teammate reports must be delivered on the teammate channel, not plain text output (invisible to teammates).

Workflow gap

Bug observed: the incidents above — duplicated implementation, duplicate fix races, and an orchestrator racing a woken teammate — each traced to treating an idle ping as done/stalled or splitting work off a live implementer on its self-reported WIP state. The T1 orchestrator's own close-out: "Lesson absorbed: one implementer per file set, enforced by me rather than by agents self-reporting their WIP state."
Why it is a workflow gap: the adjacent standing rules cover a NARROWER shape — the ownership-check bullet (CLAUDE.md § Orchestrator vs subagent re-invocation) gates resuming/relaunching a RUN on a shared artifact path, and the #825 bullet gates Monitor false-DONE on empty results dirs — but nothing governs implementer division-of-labor or idle-ping semantics for teammates; 3 incidents in one day says the gap recurs.
Confidence (emitter): medium
verified-at-filing: grep -n 'live owner\|never launch a duplicate' CLAUDE.md → the ownership-check bullet exists (§ Orchestrator vs subagent re-invocation) and is scoped to resume/relaunch of runs on shared artifact paths; grep -n 'idle' CLAUDE.md .claude/rules/*.md → no hit governs teammate idle-notification semantics or one-implementer-per-file-set division (2026-07-21).

Proposed change (candidate diff sketch — refine in planning)

Add to CLAUDE.md § Orchestrator vs subagent re-invocation (or a new .claude/rules/teammate-coordination.md with an index row): the three duties in ## Goal, each with its incident citation. Planner decides placement (always-on bullet vs on-demand rule).

Scope / surfaces

Primary target: CLAUDE.md (§ Orchestrator vs subagent re-invocation) and/or a new .claude/rules/*.md + LESSONS.md index row

Constraints / invariants

Prose duty only. If a new rule file is added, update .claude/rules/LESSONS.md (workflow_lint.py --check-lessons-index).

Provenance

fingerprint: 1e3ca1fb3acb
workflow_fix_target: CLAUDE.md

Origin evidence (transcript-mined, sessions 24ae2158 + 7ce77e0f, 2026-07-20/21): quotes above; orchestrator self-reports at 05:00:10Z ("implemented twice by parallel sessions... Lesson absorbed"), 20:28:01Z ("went idle without committing it. Doing the fix myself"), 22:07:15Z ("standing it down before it races me").