EPS Dashboard

Overview / Motivation

Auto-filed by /daily 2026-08-02 (route 2: behavior/logic change -> independent review) from the nightly problem sweep (miners 8/6/4, sessions 472284ce/a0400dd4/75f66748, tasks #1902/#1345).

Goal

Durable coordination records (stand-downs, handoffs) reflect confirmed state, and no successor dispatches over a live owner.

Workflow gap

Bug observed: Three same-day coordination races: (a) session 472284ce stood down its ladder agent (SendMessage + PID kill) and wrote a durable 'STOOD DOWN' marker -- the agent had already entered production, COMPLETED, and committed (aa7d5c867a); the durable record was wrong for ~1.7h until a RECORD CORRECTION marker. (b) session a0400dd4's successor nearly double-dispatched duty B while the 'closing' prior terminal was still live and owning the round -- stopped only at the pre-dispatch queue check. (c) session 75f66748 spawned a ladder runner on #1902 while a live owner session already ran it; the runner's own probe caught it.
Why it is a workflow gap: The clause covers channels and durable-state probes for REDOING work, but not confirmation of a stand-down's effect before recording it, nor an explicit release protocol for session-to-session handoffs.
Confidence (emitter): medium (incidents probed by miners: marker sequences + git show of the completed commit)
verified-at-filing: grep -n 'stand-down' CLAUDE.md -> 1 section (clause (c): channel discipline only); 0 hits for 'ownership release'/'release marker'; the (b)/(c) probes cover re-doing a teammate's WORK, not the stood-down-but-completing race.

Proposed change (refine in planning)

extend the teammate-coordination clause: (i) a stand-down is not effective until CONFIRMED -- probe the agent's state/durable outputs after the SendMessage before writing an aborted/stood-down claim into a durable marker (SendMessage does not preempt an in-flight turn); (ii) session-to-session continuation handoffs post an explicit ownership-RELEASE marker before the successor dispatches; (iii) the ownership probe runs in the ORCHESTRATOR before spawning a runner on a shared artifact path, not only inside the spawned agent.

Scope / surfaces

Primary target: CLAUDE.md

Constraints / invariants

Workflow-surface rules apply; scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 (Provenance workflow_fix_target: line) -- it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

fingerprint: 6f8c4b1860f1
workflow_fix_target: CLAUDE.md