EPS Dashboard

Overview / Motivation

Auto-filed by the /daily 2026-07-12 problem sweep (transcript-mined; emitting session f7dd0696 / recovery session d2c117e4, task #1277).

Goal

Make the watcher recover a watcher-dispatched session that is refusal-killed on its BOOT turn — before the /issue-tick cron is armed — within minutes, not 12.5 hours.

Workflow gap

Bug observed: #1277's first session (spawned 06:54Z by proposed_infra_sweep) died on its boot turn at 06:55:41Z with the spurious "violates our Usage Policy" API error, BEFORE Step 0 armed the /issue-tick cron. It accumulated zero failed wake turns, so none of the five wedge triggers (#1241) fired; the 09:03Z session-stalled-alert lane is alert-only; recovery came only via the 12h stale-registration-unregister (19:13Z) + proposed_infra_sweep re-dispatch (19:23Z). Net: ~12.5h wall-clock stall on a 0-GPU-h infra task.
Why it is a workflow gap: the #1209 failed-turn-silence dead-wake trigger was built for exactly this freeze-below-counting-thresholds shape, but its predicate does not cover a pre-cron boot-turn death (no wake turns ever accumulate, and the tick cron that would generate them was never armed). The stalled-session lane that DOES notice the freeze (~2h) cannot respawn.
Confidence (emitter): medium
verified-at-filing: uv run python scripts/task.py view 1277 --json events timeline (2026-07-13): transcript f7dd0696 last row = api-error at 06:55:41Z; session-stalled-alert 09:03Z (alert-only); stale-registration-unregister 19:13Z; re-dispatch 19:23Z; completion 20:34Z. Behavioral gap — not grep-verifiable beyond the timeline.

Proposed change (candidate diff sketch — refine in planning)

Make the watcher's stalled/dead-wake machinery respawn-capable for this shape: when a registered autonomous session's transcript tail row is an api-error/refusal row, the task is at a pre-run status (proposed/planning/plan_pending/approved), no /issue-tick cron exists for it, and the transcript has been silent ≥ EPM_TICK_WEDGE_DEAD_SILENCE_MIN — stop the session and let the existing crash-recovery/infra-sweep re-dispatch, under the existing episode-belt + per-issue day cap (#1241). Alternatively extend the #1209 failed-turn-silence predicate to count a boot-turn api-error death as a failed turn even with zero wakes.

Scope / surfaces

Primary target: scripts/autonomous_session_watch.py (prompt-wedge / stalled-session lanes), possibly .claude/skills/issue-tick/SKILL.md docs.
Grep the workflow surface for the pattern before editing and update every hit; list them in the plan.

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
Respawn arms must stay bounded (episode belt + 3/issue/UTC-day cap, #1241) and fail toward keep.
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: 790fa08b599a
workflow_fix_target: scripts/autonomous_session_watch.py

Origin: /daily 2026-07-12 transcript sweep, sessions f7dd0696 (boot-turn refusal death, task #1277) + d2c117e4 (the 19:23Z recovery session). Related incidents: #1209 (turn-1 refusal death freeze), #1127, #1074, #1241 (wedge-trigger belts).