[BUG] run_in_background=true Task agents silently lose all output unless Write tool explicitly called

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 spawning Task agents with run_in_background=true, the agent's response is silently discarded after completion. The output file referenced in the task notification exists but is empty (0 bytes). All agent work is lost without any error or warning.

Steps to reproduce:

1. Spawn a Task agent with run_in_background: true and a prompt requesting analysis
2. Wait for task-notification showing status: completed
3. Read the output file - it's empty despite agent completing "successfully"

Workarounds attempted (all failed):

• Adding explicit "You MUST use the Write tool" instructions
• Adding "Your work will be LOST if you only return text" warnings
• Background agents don't reliably follow Write tool instructions

Current workaround: Avoid run_in_background=true entirely; use blocking execution.

What Should Happen?

Background agent responses should be captured to the output file, same as blocking execution. Alternatively, if this is intentional behavior, there should be a clear warning when spawning background agents that their conversational output will not be captured.

Error Messages/Logs

No errors are surfaced. That's part of the problem - silent data loss.

  Task tool result:
    agentId: a6bc4c0
    output_file: C:\Users\...\tasks\a6bc4c0.output
    status: completed

  Reading output file:
    (empty - 0 bytes)

  Agent completes with success status but output file contains nothing.

Steps to Reproduce

Steps to reproduce

1. Spawn a Task agent with run_in_background: true:

Task({
subagent_type: "general-purpose",
run_in_background: true,
prompt: "Analyze these files and report findings: [file list]"
})

1. Wait for task-notification showing status: completed
2. Read the output file referenced in the notification:

Read(output_file path from notification)

1. Output file is empty (0 bytes) - all agent analysis is lost

Note: Adding explicit Write tool instructions to the prompt ("You MUST use Write tool to save output") does not reliably fix this - background agents don't consistently follow these instructions.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.2

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

• Spawned 5 parallel agents with run_in_background: true - all 5 completed but produced zero output
• Agents were instructed to write to specific file paths with warnings like "Your work will be LOST if you only return text without calling Write tool" - still no output
• The designated output directory was never created, confirming agents didn't call Write tool despite explicit instructions
• This appears to be a fundamental issue with how background agent responses are handled, not a prompt engineering problem
• Blocking execution (run_in_background: false or omitted) works correctly and captures agent responses
• Discovered in Claude Code CLI session on 2026-01-08