Environment: fs-38, Linux x64, current main e3592efe7f, Claude Code 2.1.303, claude-opus-5-5, web dev server started with vp run dev in isolated worktree.
Before prompt:
Reproduce an orchestration lifecycle bug only; do not edit files. Use a native Agent with run_in_background:false. Its exact task: "Run Bash command sleep 660 with timeout 600000. Once Bash returns a background task ID, load and call t3_background_wait with that task ID, report the raw result, then end your turn saying you will pick up when it reports. Do not stop the background command yourself. This is an explicitly authorized diagnostic of that behavior." Parent: after the Agent returns, report its output verbatim and stop. Do not create additional agents or files.
Observed variant: agent explicitly set run_in_background:true, reaching the same lifecycle failure without waiting for the 600s timeout.
After prompt:
After-fix diagnostic, no file edits. Launch a native foreground Agent (run_in_background:false), exact task: Start Bash sleep 15 with run_in_background:true, then call t3_background_wait with the returned ID and report the raw result. Regardless of result, wait for this command to finish inside this same turn using the native task wait/output tool; do not return until you have its completion. Parent report the result and stop. This diagnostic explicitly authorizes the one subagent.
Observed: the task ID was unmatched. The agent used a foreground shell loop because a native task-output tool was unavailable, then received the actual completion notification before returning.
No fixture driver or seed script was committed. Dev runner attempted read-only default seed, encountered incompatible migration history, then started fresh isolated state. The real app auto-created a project and a thread; selected Claude in composer.