[BUG] TaskOutput incorrectly returns <retrieval_status>timeout</retrieval_status> immediately upon background agent completion

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

When using the Claude Code agentic framework to spawn multiple background sub-agents, there is a serious synchronization bug between the background task runner and the TaskOutput tool.

Specifically, exactly at the millisecond a background sub-agent successfully finishes its execution, the blocking TaskOutput tool returns a <retrieval_status>timeout</retrieval_status> error to the main orchestrator LLM instead of returning the correct result payload.

What Should Happen?

Expected Behavior
The TaskOutput tool should cleanly resolve and return the actual <result> payload generated by the completed sub-agent to the main LLM.

Actual Behavior
The TaskOutput tool instantly resolves with a false error: <retrieval_status>timeout</retrieval_status>.
Subsequently, the background runner's queue processor asynchronously injects a successful <task-notification> (containing the true result) into the conversation history.
This creates a race condition / context confusion where the main model simultaneously sees a timeout tool failure and a successful injected notification, often leading the orchestrator to incorrectly assume the entire task failed.

Error Messages/Logs

Used 6 tools
Used 6 tools
TaskOutput
<retrieval_status>timeout</retrieval_status>
TaskOutput
<retrieval_status>timeout</retrieval_status>
TaskOutput
<retrieval_status>timeout</retrieval_status>
TaskOutput
<retrieval_status>timeout</retrieval_status>
TaskOutput
<retrieval_status>timeout</retrieval_status>
TaskOutput
<retrieval_status>timeout</retrieval_status>
Show less

Steps to Reproduce

Initiate several background sub-agents to perform code generation tasks concurrently.
The main orchestrator uses the TaskOutput tool with block: true and a long timeout (e.g., 600000ms) to wait for the sub-agent's response.
Observe the outputs when the sub-agent successfully completes its work (e.g., writing the requested files to disk) within the initial timeout limit.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Desktop App for Windows Version 1.1.7714 (3bd6f6)

Platform

Other

Operating System

Windows

Terminal/Shell

Other

Additional Information

_No response_