EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a prose follow-up surfaced on task #1768 (emitting agent: experiment-implementer lasttoken-repool-1768, completion report 2026-08-01; incident ~05:05Z).

Goal

Make the watcher pod-safety pass's inferred follow-up predicate safe on MULTI-ROUND issues: a sibling round's routine status transition must not flip protection off another round's healthy live pod.

Workflow gap

Bug observed: watcher pod-safety pass auto-stopped pod-1768-lt mid-fits because a sibling round epm:status-changed postdated the round epm:run-launched flipping the inferred predicate off on a multi-round issue
Why it is a workflow gap: the predicate compares the newest follow-up signal (epm:run-launched/epm:followup-scope/epm:free-analysis-followup-run) against the newest done-transition (epm:promoted/epm:status-changed) at ISSUE grain; on an issue running concurrent/serial follow-up rounds (three inline pods + a session loop on #1768 tonight), routine round transitions post epm:status-changed continually, so any pod whose run-launched predates the latest sibling transition loses inferred protection — only the timestamp-independent issue-wide keep-running tag shields it, and the standing inline-round guidance (this orchestrator's own briefs included) says NO tag to avoid blocking sibling teardowns (#1485). The two protections are mutually exclusive at current grain.
Consequence realized: pod-1768-lt auto-STOPPED at ~05:05Z with 142/216 fit cells done (timeline as reported by the emitting agent — unverified hypothesis — verify at plan time: exact marker timestamps 04:11:11Z / 22:34Z; the MECHANISM is confirmed against the predicate's own docstring). Recovered by tag+resume with no data loss, ~1h delay.
Confidence (emitter): high
verified-at-filing: grep -n "run-launched\|followup-scope\|free-analysis-followup-run\|keep-running" scripts/autonomous_session_watch.py → 8 hits (2026-08-01); per-target: scripts/autonomous_session_watch.py — L202-209 document exactly the newest-signal-vs-done-transition predicate (_SESSION_FOLLOWUP_SIGNAL_KINDS) and the "no keep-running tag (the explicit user override)" arm; context read: the predicate is issue-grain by construction, no per-pod matching exists (gap present, not landed-fixed). CLAUDE.md carries the inline-round pod-safety clause that prescribes marker-or-tag (the guidance half of the fix).

Proposed change (candidate diff sketch — refine in planning)

diff_sketch: | scripts/autonomous_session_watch.py (pod-safety pass):

predicate: newest follow-up signal on the ISSUE vs newest done-transition on the ISSUE

per-POD refinement: an epm:run-launched whose note names the pod (name/suffix — the
notes already do, e.g. "pod-1768-lt ... provisioned") shields THAT pod while the pod is
RUNNING and younger than a staleness ceiling, regardless of later sibling
status-transitions; issue-grain inference retained as fallback for unnamed launches CLAUDE.md (user-chat inline carve-out, pod-safety pre-launch signals) + .claude/skills/issue/SKILL.md 9a-ter:
document the multi-round exposure; until the per-pod predicate lands, concurrent inline
rounds SET keep-running at provision + remove at completion, with the surgical
--name-suffix teardown noted as the sibling-safe path under the issue-wide tag

Scope / surfaces

Primary targets: scripts/autonomous_session_watch.py, CLAUDE.md
Also check .claude/skills/issue/SKILL.md § 9a-ter pod-safety block and scripts/pod_lifecycle.py (#1485 tag semantics) for consistency; grep grep -rn "keep-running" scripts/ .claude/ CLAUDE.md and reconcile every guidance site.

Constraints / invariants

Workflow-surface only. The watcher's escalate-only and bounded-stop contracts unchanged; fail toward KEEP (a mis-shielded pod bills, a mis-stopped pod loses work — the wedged-owner escalation arm #1582 remains the billing backstop).
scripts/workflow_lint.py --check-asks passes; tests pinning watcher invariants updated alongside.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

workflow_fix_target: scripts/autonomous_session_watch.py,CLAUDE.md
fingerprint: 2cec7073efd3

Verbatim surfaced prose (agent completion report, 2026-08-01):

INCIDENT you should know about: pod-1768-lt was auto-STOPPED mid-fits at ~05:05Z, 142/216 cells done. Cause: a SIBLING round's epm:status-changed at 04:11:11Z is newer than my epm:run-launched (22:34Z), which flips the watcher pod-safety pass's inferred follow-up predicate off — only the timestamp-independent keep-running tag shields that window, and the brief specified NO tag. Recovered with no data loss: tagged, resumed, harvested 46 complete arms, relaunched (resume predicate skipped all 46), remaining 26 finished rc=0. Worth considering whether inline rounds on a multi-round issue should tag by default.