EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1298 (emitting agent: issue-1298 orchestrator, triaging the #928 followups_running 6-day freeze).

Goal

Add a bounded escalation rung to the watcher's stalled-MANUAL-session lane so an unactioned stalled-alert on a manual registration holding an ACTIVE task with an unrun followup scope eventually UNREGISTERS the manual registration (never stopping the user session), letting the orphan sweep re-drive the task.

Workflow gap

Bug observed: A live-but-wedged manual (user-driven, manual-issue-928.json) session held task #928 at followups_running mid-round for 6 days (2026-07-08 -> 07-14) with an unrun followup scope (prefix-based-mapping-arms, plan v7 posted 07-08 07:43Z, never approved). The stalled-session lane alerted at +2.3h (07-08 10:03Z, "NOT auto-respawned (manual user-driven session; alert-only by design)") but no automatic recovery path ever fired; recovery required a human spawn_session.py stop + autonomous re-drive on 07-14 21:04Z (the round then completed cleanly by 07-15 01:38Z).
Why it is a workflow gap: Manual-session alert-only is deliberate (#505 — never auto-respawn a user-driven session), but there is no bounded escalation when the alert goes unactioned. The only unregister path, decide_stale_registration (#845), gates on TRANSCRIPT idle >= 12h of a LIVE session — any transcript activity in the live wedged session (user interaction, in-session cron ticks) defeats it indefinitely, so an ACTIVE task with an unrun followup scope and zero non-watcher progress has no watcher-side recovery. Alerts alone are insufficient ("sent != seen"); the #928 alert sat unactioned for 6 days across two /daily flags (07-11, 07-12) and a filed needs-human task (#1298).
Confidence (emitter): medium (the 6-day freeze + alert-only latch are marker-proven; the exact reason the 12h transcript-idle gate never fired is inferred from the live-on-07-14 deliberate-stop, not transcript-proven)
verified-at-filing: grep -n "alert-only by design" scripts/autonomous_session_watch.py -> 1 hit in 1 file (:11291, the stalled-manual branch — presence confirmed; the lane posts the alert and returns, no escalation); semantic absence probe: grep -n "unregister" scripts/autonomous_session_watch.py -> 20 hits, ALL in the 12h stale-registration pass (:988, :2647-2672, :19999-20023), zombie/GC, and docstring contexts — none in the stalled-alert lane (:11291-11388); landed-fix history: git log --oneline --since='2026-07-08' -- scripts/autonomous_session_watch.py -> 20 commits, none adding a stalled-manual escalation (boot-death lanes #1267/#1287 fire only on boot-dead/boot-refused transcripts; the #1127/#1209 wedge lanes require failed wake turns) (2026-07-17).

Proposed change (candidate diff sketch — refine in planning)

  # scripts/autonomous_session_watch.py, stalled-session lane (~:11291)
  reason = "manual user-driven session; alert-only by design"
+ # Escalation rung (#928 6-day freeze): after >= K (default 3) consecutive
+ # stalled-alert confirmations spanning >= EPM_STALLED_MANUAL_ESCALATE_H
+ # (default 24) hours on the SAME manual registration, with (a) zero
+ # non-watcher task events since the first confirmation and (b) an unrun
+ # followup scope OR an otherwise ACTIVE status, UNREGISTER the manual
+ # registration (delete manual-issue-<N>.json; NEVER stop the session) and
+ # post a loud [stalled-manual-escalation] marker + push. The orphan sweep
+ # then re-drives the ACTIVE task on its next pass. Every unresolvable
+ # input fails toward keep (alert-only, today's behavior).

Scope / surfaces

Primary target: scripts/autonomous_session_watch.py
Grep the workflow surface for the pattern before editing (grep -rln 'alert-only by design' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan. Companion docs: .claude/rules/background-automation.md (stalled-session lane description) and the watcher docstring pass inventory (the #1225 lint).

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
NEVER stop or respawn the user-driven session itself (#505 stands) — the escalation acts on the REGISTRATION only, exactly like the stale-registration pass; the orphan sweep's existing re-drive machinery does the recovery.
Fail toward keep on every unresolvable input (the watcher-wide convention).
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; docstring pass-count lint (#1225) stays consistent.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

workflow_fix_target: scripts/autonomous_session_watch.py
fingerprint: 29eaf4438ca3

target_file: scripts/autonomous_session_watch.py bug_observed: A live-but-wedged manual (user-driven) session held task #928 at followups_running mid-round for 6 days (2026-07-08 to 07-14) with an unrun followup scope; the stalled-session lane alerted at +2.3h but is alert-only for manual registrations and no automatic recovery fired why_workflow_gap: Manual-session alert-only is deliberate (#505), but there is no bounded escalation when the alert goes unactioned: decide_stale_registration's transcript-idle>=12h gate is defeated indefinitely by any transcript activity in the live wedged session, so an ACTIVE task with an unrun followup scope and zero non-watcher progress has no watcher-side recovery proposed_change: Add an escalation rung to the stalled-manual-session lane: after >=K consecutive stalled-alert confirmations spanning >=24h on the same manual registration with zero non-watcher task progress and an unrun followup scope, UNREGISTER the manual registration (never stop the user session) so the orphan sweep re-drives the ACTIVE task diff_sketch: |

after >=3 consecutive stalled-alert confirmations spanning >=24h on the
same manual registration (zero non-watcher task events; unrun followup
scope / ACTIVE status): delete manual-issue-.json (never stop the
session), post [stalled-manual-escalation] marker + push; orphan sweep
re-drives. Fail toward keep on unresolvable inputs. confidence: medium related_task: #1298