claude-code - 💡(How to fix) Fix Subagent liveness/heartbeat: Task-spawned agents can die silently with no signal [1 comments, 2 participants]

• A Codex review subagent (pid 13224) died mid-review before the result harvest stage. The orchestrating session showed it as in-flight for ~1.5 hours before the user noticed. No error, no exit signal surfaced.

TL;DR

Implementing a subagent liveness tracking mechanism in the harness, such as tracking last-output timestamps and emitting a SubagentStalled event, can help detect and handle silent subagent deaths.

Guidance

• Consider adding a last_output_timestamp field to the Task agent data structure to track the last time each subagent produced output.
• Introduce a configurable timeout window (e.g., 30 minutes) to determine when a subagent is considered stalled.
• Design an event or status flag (e.g., SubagentStalled) to notify the orchestrator when a subagent has not produced output within the specified window.
• Evaluate exposing the subagent liveness information in TaskList and TaskGet output to simplify detection for orchestrators.

Example

No code snippet is provided due to the high-level nature of the issue, but an example implementation might involve modifying the Task agent data structure and adding event emission logic.

Notes

The proposed solution requires changes to the harness and may involve additional complexity, such as handling edge cases (e.g., subagents with variable output frequencies) and ensuring backwards compatibility.

Recommendation

Apply a workaround, such as the existing subagent-heartbeat convention, until a native harness solution is implemented, as it provides an immediate, albeit user-land, solution to the problem.