Bug ‚Äî Claude Desktop scheduled-task runner leaks claude-code subprocesses

Summary

The local-agent-mode scheduled-task runner in Claude Desktop 1.5354.0 (using bundled claude-code/2.1.121) does not tear down spawned runner subprocesses after a scheduled task completes. Each fire leaks one disclaimer wrapper + one claude child, both stuck in interruptible sleep (S) on stdio. After ~24 hours of normal cron activity (~7 fires/hr), ~200 stuck pairs accumulate. Once the renderer's internal subprocess cap is reached, the dispatcher silently swallows further fires ‚Äî nextRunAt advances past missed slots without spawning a runner.

Repro

1. Configure any local-agent-mode scheduled task with a frequent cron (e.g. every 15 min).
2. Leave Claude Desktop running for 24h.
3. Observe ps -ax | grep 'claude-code/2.1.121/claude.app/Contents/MacOS/claude' | wc -l climbing monotonically.
4. After enough accumulation, scheduled tasks stop firing entirely while the desktop UI shows no error.

The bug reproduces independently of:

• Which model the task uses (observed across claude-opus-4-7, claude-sonnet-4-6, claude-sonnet-4-5-20250929, default)
• Permission mode (observed across bypassPermissions and default)
• The skill content (no skills in the affected setup spawn background processes; verified by grep -rE 'nohup|disown|setsid|& *$' ~/.claude/scheduled-tasks/)
• User-configured hooks (no Stop / PostToolUse hook configured)

Process signature

Each leaked pair has:

Outer (disclaimer wrapper):

• argv0: /Applications/Claude.app/Contents/Helpers/disclaimer
• argv: includes the full path to claude and all its flags
• ppid: Claude.app/Contents/MacOS/Claude (main process)

Inner (claude runner):

• argv0: /Users/<user>/Library/Application Support/Claude/claude-code/2.1.121/claude.app/Contents/MacOS/claude
• ppid: the outer disclaimer process
• relevant flags: --output-format stream-json --input-format stream-json --permission-prompt-tool stdio --allow-dangerously-skip-permissions --include-partial-messages --replay-user-messages --settings {}
• state: S (interruptible sleep)
• RSS: ~80 MB, stable

The combination of --input-format stream-json and --permission-prompt-tool stdio means the child reads from stdin until EOF. The hang signature is consistent with the parent (Claude Helper Renderer hosting the local-agent-mode runner) failing to close the child's stdin pipe when the skill's Stop event fires.

Environment

• Claude Desktop version: 1.5354.0
• claude-code build: 2.1.121 (bundled in Claude.app)
• macOS: 25.1.0 (Darwin)
• Scheduled-tasks feature: ccdScheduledTasksEnabled: true + coworkScheduledTasksEnabled: true

Impact

• All scheduled tasks silently stop firing after the cap is hit
• No error surfaces in the Claude Desktop UI, log files, or notifications
• Recovery requires manually killing the entire pile (pkill -f claude-code/2.1.121) ‚Äî no in-app reset
• The watchdog/heartbeat task pattern Anthropic users rely on (e.g. a daily cron-heartbeat task that pings if any other task is stale) doesn't catch this because the heartbeat itself is one of the silently-skipped tasks

Workaround in place locally

A janitor launchd agent (com.alex.argus.claude-janitor, every 10 min) reaps claude-code/2.1.121 processes whose parent's command contains Claude.app AND whose elapsed time exceeds 30 min. Hard-coded to build 2.1.121 so it auto-no-ops once a fixed build ships. Source available on request.

What would help

1. Fix: have the runner explicitly close child stdio (or kill -9 the child) when the skill's Stop event fires ‚Äî preferably with a 30-second hard-deadline regardless of pipe state.
2. Telemetry: surface a UI warning when pending child-runner count exceeds, say, 20. Right now the failure is invisible until the user manually ps-greps.
3. Recovery: an in-app "Restart scheduled-tasks runner" button or auto-reaper for orphaned local-agent-mode subprocesses older than 30 min.

Local diagnosis report

Full investigation, evidence, and ruled-out hypotheses: ~/Documents/Claude by Alex/Reports/2026-05-01 stuck-cron-process-diagnosis.md