Fullsend waits for hard timeout after Claude emits successful result

Summary

A code-agent run completed its implementation, committed the change, validated its structured output, and emitted a successful Claude stream-json result. Fullsend nevertheless kept reporting Agent running until the exact 35-minute hard timeout because an auto-backgrounded Jest task was still marked as running.

The GitHub Actions job ultimately succeeded because Fullsend recovered the output and commit during post-processing, but it wasted about 21.5 minutes after the model had already completed.

Run: https://github.com/redhat-developer/rhdh-agentic/actions/runs/29257824641/job/86842795543

Related: #19 (auto-backgrounded test/retry behavior)

Timeline

• 14:27:10.530Z — agent process started
• 14:33:34.923Z — agent ran yarn test with a 120-second Bash timeout
• 14:35:35.149Z — Claude Code auto-backgrounded the command as task b3jjxrrgf
• 14:37:38.237Z — TaskOutput timed out with status: running, even though the output showed all 5 tests passed
• Agent said Let me stop the task and proceed, but did not invoke TaskStop
• 14:40:29.648Z — final fullsend-check-output tool call started
• 14:40:29.950Z — validation returned successfully (~302 ms)
• 14:40:38.636Z — Claude emitted its final answer
• Raw output.jsonl contains a terminal type: "result", subtype: "success", terminal_reason: "completed", and duration_ms: 807613
• 15:02:10.559Z — Fullsend reached the exact 35-minute limit and reported Agent exited with code -1

That is roughly 21m32s of idle waiting after Claude had completed successfully.

Misleading progress annotation

The last visible tool annotation was:

Notice: Bash: fullsend-check-output (13m19s, 80 tools)

This looks like fullsend-check-output ran for 13m19s. It actually ran for about 302 ms. 13m19s is cumulative session elapsed time, and 80 tools is the cumulative tool count.

Suggested wording:

Bash: fullsend-check-output (13m19s elapsed, tool #80)

Root cause

The immediate leak was the test command:

cd .../plugins/agent-common && yarn test 2>&1

The tests finished, but the Backstage/Jest command remained alive in watch-like mode. Claude Code auto-backgrounded it when the Bash timeout elapsed. The task was never stopped, so the Claude CLI process did not exit even after writing a successful terminal result.

Expected behavior

Once Fullsend receives a terminal stream-json record such as:

{
  "type": "result",
  "subtype": "success",
  "terminal_reason": "completed"
}

it should treat the agent as completed, terminate/clean up any remaining child or background processes, and proceed to extraction/post-processing without waiting for the hard timeout.

Proposed fixes

1. In the Fullsend runtime, stop waiting for process exit after receiving a terminal successful result; terminate the CLI process/process group and clean up background children.
2. Ensure timeout cleanup kills the whole process group, not only the foreground wrapper.
3. Update code-agent guidance to run Jest non-interactively, for example CI=true yarn test --watchAll=false.
4. Require agents to invoke TaskStop for any verification command that was auto-backgrounded before finishing.
5. Clarify progress annotations so cumulative elapsed time is not presented like individual tool duration.

Acceptance criteria

• A successful terminal result advances immediately to artifact extraction even if a child test process remains alive.
• Background child processes are cleaned up.
• The run does not wait until the configured agent timeout in this scenario.
• Tool annotations distinguish total session elapsed time from per-tool duration.

Inspect the Fullsend runtime handling of Claude stream-json records and process cleanup, using the linked Actions run and raw output.jsonl as the reproduction evidence. Verify behavior when a successful terminal result arrives while a background Jest process remains alive. Done means extraction starts immediately, child processes are cleaned up, the hard timeout is avoided, and tool annotations distinguish elapsed session time from per-tool duration.