Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #689 (the deferred D.3 follow-up of the plan; emitting agent: planner — surfaced in v3 §D.3 + a verbatim <!-- workflow-fix-candidate v1 --> block at the plan tail).

Goal

Add a RUNNING-but-no-port predicate to scripts/autonomous_session_watch.py's pod-safety pass that reuses backend_poll._is_runpod_async_wedge_failure (from #689 fix (b)) so a wedge on a poll-dead task is still caught.

Workflow gap

• Bug observed: The watcher's pod-safety pass does not recognize a RunPod RUNNING-but-unreachable (no-port) pod on an ACTIVE task — it is a billing leak the pass currently misses. The poller-side fix (b) in #689 auto-recovers the common case, but a task whose poll loop has died still needs the watcher backstop (the residual failure mode #689 explicitly leaves open).
• Why it is a workflow gap: autonomous_session_watch.py's pod-safety reconciliation pass is a workflow-surface backstop for pod-safety; it should compose the #689 fix-(b) wedge predicate (backend_poll._is_runpod_async_wedge_failure) so a wedge on a poll-dead task is still caught. Without this, a session whose backend_poll.py chain has died will leave its RUNNING-but-no-port pod billing indefinitely until the next interactive /issue <N> resume.
• Confidence (emitter): medium (the predicate is small, but the file is ~10,628 lines and the pod-safety pass needs grep-locating; sequencing carefully matters).

Proposed change (candidate diff sketch — refine in planning)

# in the pod-safety pass (grep-locate it in the ~10.6k-line file):
+ # #664 RunPod wedge backstop (composes #689 fix (b)): a managed RUNNING pod
+ # with null runtime.ports for >K on an ACTIVE task is an unreachable billing
+ # leak the poller may have missed if its poll loop died. Surface/alert; gate
+ # any auto-stop on per-cell inputs-on-HF + keep-running, mirroring fix (b).
+ if _runpod_running_no_port(pod) and task_active(issue) and not keep_running(issue):
+     ... alert / compose with fix (b) recovery ...

Scope / surfaces

• Primary target: scripts/autonomous_session_watch.py
• Grep the workflow surface for the pattern before editing:

  grep -rln 'pod-safety\|RUNNING.*no.*port\|TRANSIENT_CAPACITY_REASONS' scripts/autonomous_session_watch.py CLAUDE.md
  

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• The new predicate must COMPOSE with the existing #689 fix (b) detection — call backend_poll._is_runpod_async_wedge_failure directly, never re-define it (recursion-guard exempt: this is the watcher-side BACKSTOP for the case where the poller's _maybe_escalate_runpod_wedge never ran because the poll loop is dead).
• Gate any auto-stop on the same inputs-on-HF + keep-running checks as fix (b) so the watcher's backstop is symmetric with the poller path.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
• 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: 7d213f536fe9

target_file: scripts/autonomous_session_watch.py bug_observed: The watcher's pod-safety pass does not recognize a RunPod RUNNING-but-unreachable (no-port) pod on an ACTIVE task — it is a billing leak the pass currently misses (the poller-side fix (b) in #689 auto-recovers the common case, but a task whose poll loop has died still needs the watcher backstop). why_workflow_gap: autonomous_session_watch.py's pod-safety reconciliation pass is a workflow-surface backstop for pod-safety; it should compose the #689 fix-(b) wedge predicate (backend_poll._is_runpod_async_wedge_failure) so a wedge on a poll-dead task is still caught. proposed_change: Add a RUNNING-but-no-port predicate to the watcher's pod-safety pass that reuses backend_poll's #664 wedge detection, gated on the same per-cell inputs-on-HF + keep-running exemptions as fix (b), composing with the existing no_compute_available capacity-retry pass. diff_sketch: |

in the pod-safety pass (grep-locate it in the ~10.6k-line file):

• #664 RunPod wedge backstop (composes #689 fix (b)): a managed RUNNING pod

• with null runtime.ports for >K on an ACTIVE task is an unreachable billing

• leak the poller may have missed if its poll loop died. Surface/alert; gate

• any auto-stop on per-cell inputs-on-HF + keep-running, mirroring fix (b).

• if _runpod_running_no_port(pod) and task_active(issue) and not keep_running(issue):

• ... alert / compose with fix (b) recovery ...
  

confidence: medium related_task: #689