Description
/wr-itil:work-problems Step 5 dispatches each iteration (and the Step 0b/0c/0d pre-flights) to a claude -p subprocess wrapped in a backgrounded poll loop inside a single Bash call. That poll loop assumes the orchestrator process can foreground-wait for up to WORK_PROBLEMS_IDLE_TIMEOUT_S (default 3600s / 60min). In the interactive Claude Code harness the Bash tool caps a single foreground call at 10 minutes, so any iter or pre-flight that runs longer is SIGTERM'd at ~10min with a 0-byte JSON file (the P147 stuck-before-emit class), and no iter work lands.
Observed 2026-07-03: the Step 0b /wr-itil:review-problems pre-flight subprocess died at exactly 10min (exit 143, empty JSON), forcing the P358 non-blocking revert-and-proceed path (reverted the partial docs/problems/.upstream-cache.json + docs/audits/inbound-discovery-log.md writes). The same ceiling blocks full /wr-itil:manage-problem iter dispatch, so the loop cannot make forward progress on any ticket whose iter exceeds 10min when orchestrated foreground.
Symptoms
- Step 0b review-problems pre-flight: exit 143, 0-byte JSON, at ~600s wall-clock (2026-07-03).
- Any manage-problem iter dispatched as a single foreground Bash call is capped at 10min regardless of
WORK_PROBLEMS_IDLE_TIMEOUT_S.
Workaround
Dispatch long subprocesses via the harness-level background primitive (Bash tool run_in_background: true) which runs detached across turns and re-invokes the orchestrator on completion, rather than a single foreground Bash call with an in-call poll loop. The Step 5 poll-loop shape is designed for a host that permits unbounded foreground shell waits; the interactive Claude Code harness is not that host.
Affected plugin / component
@windyroad/[email protected] (work-problems SKILL Step 5 dispatch)
Frequency
Every iter/pre-flight exceeding 10min when work-problems is orchestrated foreground from an interactive Claude Code session.
Versions
Cross-reference
Reported from the downstream windy-road content project, tracked locally as P110 in its docs/problems/ directory. Observed 2026-07-03 when the Step 0b review-problems pre-flight died at exactly 10min (exit 143, 0-byte JSON) under the harness foreground Bash ceiling.
Description
/wr-itil:work-problemsStep 5 dispatches each iteration (and the Step 0b/0c/0d pre-flights) to aclaude -psubprocess wrapped in a backgrounded poll loop inside a single Bash call. That poll loop assumes the orchestrator process can foreground-wait for up toWORK_PROBLEMS_IDLE_TIMEOUT_S(default 3600s / 60min). In the interactive Claude Code harness the Bash tool caps a single foreground call at 10 minutes, so any iter or pre-flight that runs longer is SIGTERM'd at ~10min with a 0-byte JSON file (the P147 stuck-before-emit class), and no iter work lands.Observed 2026-07-03: the Step 0b
/wr-itil:review-problemspre-flight subprocess died at exactly 10min (exit 143, empty JSON), forcing the P358 non-blocking revert-and-proceed path (reverted the partialdocs/problems/.upstream-cache.json+docs/audits/inbound-discovery-log.mdwrites). The same ceiling blocks full/wr-itil:manage-problemiter dispatch, so the loop cannot make forward progress on any ticket whose iter exceeds 10min when orchestrated foreground.Symptoms
WORK_PROBLEMS_IDLE_TIMEOUT_S.Workaround
Dispatch long subprocesses via the harness-level background primitive (Bash tool
run_in_background: true) which runs detached across turns and re-invokes the orchestrator on completion, rather than a single foreground Bash call with an in-call poll loop. The Step 5 poll-loop shape is designed for a host that permits unbounded foreground shell waits; the interactive Claude Code harness is not that host.Affected plugin / component
@windyroad/[email protected] (work-problems SKILL Step 5 dispatch)
Frequency
Every iter/pre-flight exceeding 10min when work-problems is orchestrated foreground from an interactive Claude Code session.
Versions
@windyroad/[email protected]@windyroad/[email protected]Cross-reference
Reported from the downstream windy-road content project, tracked locally as P110 in its
docs/problems/directory. Observed 2026-07-03 when the Step 0b review-problems pre-flight died at exactly 10min (exit 143, 0-byte JSON) under the harness foreground Bash ceiling.