[BUG] TaskOutput deprecation note causes agents to Read full sub-agent conversation history, triggering autocompact thrashing

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 a background local_agent task completes, the TaskOutput tool's deprecation message instructs the agent to Read the task's .output file directly:

DEPRECATED: Prefer using the Read tool on the task's output file path instead.
Background tasks return their output file path in the tool result, and you receive
a <task-notification> with the same path when the task completes — Read that file directly.

This guidance does not distinguish between task types. For local_bash tasks, the .output file contains stdout/stderr — reading it is appropriate.
For local_agent tasks, however, the .output file contains the full raw conversation history JSON of the sub-agent, which can be 80K+ characters.

The Agent tool already returns a summarized result directly when the task completes. But following the deprecation note, the agent reads the .output file anyway — injecting the entire sub-agent conversation dump into the main agent's context window. When multiple sub-agents run in the same session, this causes rapid context saturation and triggers autocompact thrashing.

What Should Happen?

The deprecation guidance should distinguish between task types:

• local_bash tasks: Reading the .output file is appropriate — it contains command stdout/stderr.
• local_agent tasks: The Agent tool result already contains the sub-agent's summarized output. The agent should not Read the .output file — it

contains the full conversation history JSON and will inflate the context unnecessarily.

Suggested fix to the TaskOutput prompt:

DEPRECATED: Prefer using the Read tool on the task's output file path instead.

• For bash tasks: Read the output file path to get stdout/stderr.
• For agent tasks: use the Agent tool result directly. Do NOT Read the .output file —

it contains the full sub-agent conversation history and will overflow your context window.

Error Messages/Logs

SDK AssistantMessage carries error=invalid_request | text_len=257 
 SDK API error (fatal): error=invalid_request | 
 raw={'content': [TextBlock(text='Autocompact is thrashing: the context refilled to the limit 
 within 3 turns of the previous compact, 3 times in a row. A file being read or a tool output 
 is likely too large for the context window. Try reading in smaller chunks, or use /clear to 
 start fresh.')]} 
 
 The oversized Read that triggered the issue: 
 tool_call Read | input: {'file_path': '/tmp/claude-1000/.../tasks/aa8bd3ced112e5ae5.output'} 
 tool_result Read | ok | output: {"parentUuid":null,"isSidechain":true,"promptId":"747dfb10...", 
 "type":"user","message":{"role":"user","content":"In the repo at ... <- 87,286 chars 
 
 Comparison — Agent tool result vs .output file: 
 Agent tool return value : 4,087 chars (summarized result, appropriate) 
 .output file via Read : 87,286 chars (full conversation history JSON, 21x larger) 
 
 3 compacts each saving ~167K tokens, all triggered within the same turn: 
 SDK compact triggered: count=1 pre_tokens=167199 tokens_saved=167199 text=SDK auto compact 
 SDK compact triggered: count=2 pre_tokens=167883 tokens_saved=167883 text=SDK auto compact 
 SDK compact triggered: count=3 pre_tokens=167202 tokens_saved=167202 text=SDK auto compact

Steps to Reproduce

1. Use Claude Code (v2.1.92) with a long-running agent session.
2. Launch a sub-agent using the Agent tool with run_in_background: true.
3. When task_notification fires, observe the main agent immediately calling Read on the path /tmp/claude-1000/.../tasks/<task_id>.output.
4. The file read returns the full sub-agent conversation history JSON (~87K chars per sub-agent).
5. With 3+ sub-agents in one session, context fills to the 800K token compact threshold.
6. After 3 consecutive compacts where context refills within a few turns each time, the SDK raises invalid_request: Autocompact is thrashing.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.92

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_