[MODEL] Main assistant pre-fetches data and restricts custom agents from using their tools, causing incorrect analysis

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Subagent behaved unexpectedly

What You Asked Claude to Do

User prompt:
use code-reviewer agent to review PR 1119

Context: The code-reviewer is a custom agent configured with the following tools:

• tools: Glob, Grep, Read, WebFetch, TodoWrite, WebSearch, BashOutput, KillBash, mcp__ide__getDiagnostics, Bash
• bash_allow:
• "gh pr view:*"
• "gh pr diff:*"
• "gh api:*"
• "gh pr checks:*"

The agent is specifically instructed to look for truncated diffs and to load each file's full contents when truncated as follows in the agent Markdown file.

...
**IMPORTANT: Handle Large PRs Correctly**

For large PRs, `gh pr diff` output may be truncated. ALWAYS follow this workflow to avoid making incorrect assumptions:

1. **Fetch PR Information**: Use `gh pr view` to understand the PR context
2. **Get File List**: Use `gh pr view <pr-number> --json files --jq '.files[].path'` to get ALL changed files
3. **Check for Truncation**: When using `gh pr diff`, look for truncation indicators like `... [XXXX lines truncated] ...`
4. **Read Key Files Directly**:
   - If diff is truncated, DO NOT make assumptions about missing files
   - Use the Read tool to examine important files directly:
     - Database migrations (look for <redacted>)
     - Test files (look for  <redacted>)
     - Configuration files
     - Core implementation files
...
 ```

### What Claude Actually Did

The main Claude Code assistant executed this pattern:

  1. Pre-fetched the PR diff itself:
  gh pr diff 1119
    - Output was ~50KB and got truncated at "... [471 lines truncated] ..."
    - Migration file content was in the truncated section
    - Full implementations were cut off
  2. Passed truncated data to the code-reviewer agent with this instruction:
  prompt: "Review PR #1119 titled <redacted title> by <redacted author>

  **Full PR Diff:**
  [... 50KB of truncated diff ...]

  Do NOT read any local files - work only with the diff provided above."

  3. Result - Agent made multiple false claims:
    - ❌ "BLOCKING: Missing database migration" - FALSE (migration file existed but was in truncated section)
    - ❌ "<redacted> syntax errors" - FALSE (misread the diff format)
    - ❌ "This code won't compile" - FALSE (code compiled successfully)
    - ❌ "Missing index on composite query" - FALSE (index was in migration file that was truncated)
  4. Wasted resources:
    - Generated detailed analysis of 8 "critical blocking issues"
    - 4 of those issues were completely incorrect
    - User had to manually verify and correct the agent's analysis
    - Consumed tokens analyzing incomplete/truncated data

### Expected Behavior

The main Claude Code assistant should have simply invoked the custom agent with a high-level task and trusted the agent to use its own tools:

  Task(
    subagent_type: "code-reviewer",
    prompt: "Review PR #1119. Check for security issues, missing tests, database migrations, 
             indexes, and code quality problems."
  )

  Then the code-reviewer agent would:
  1. Use its own gh pr diff or gh pr view commands (it can handle large output)
  2. Provide accurate analysis based on complete, verified information

  Why this matters:

  - Custom agents are designed for autonomy: Users create agents with specific tools precisely so they can explore the codebase themselves
  - Agents handle large data better: Unlike bash output (which truncates), agents can read files incrementally, search systematically, and decide what they
   need
  - Verification capability: Agents can verify their own claims by reading source files, searching for patterns, etc.
  - No truncation issues: Reading files directly avoids bash output limits

  The anti-pattern to avoid:

  ❌ Don't do this:
  1. Main assistant runs: gh pr diff
  2. Gets truncated output
  3. Passes truncated data to agent
  4. Tells agent: "Do NOT read local files"

  ✅ Do this instead:
  1. Main assistant invokes agent with high-level task
  2. Agent uses its own tools to fetch/explore
  3. Agent works with complete data
  4. Agent verifies its own claims

### Files Affected

```shell

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. prompt: use code-reviewer agent to review PR 1119
2. model requests permission to check out the branch using a git command, I hit ESC to interrupt
3. prompt: Use gh command to find the diffs. Don't disrupt my workspace.

Sometimes the Claude Code assistant passes the (truncated) data to the subagent.

Claude Model

Sonnet

Relevant Conversation

[The code-reviewer subagent is instructed to look for truncated diffs then 

Claude said: I'll use the gh command to fetch the PR information and review it without touching your workspace.

[Then it used the gh command to fetch the diffs and passed the truncated diffs to the code-reviewer subagent.]

Impact

Medium - Extra work to undo changes

Claude Code Version

2.0.28 (Claude Code)

Platform

Anthropic API

Additional Context

This affects all custom agents with Read/Grep/Glob/Bash tools:

• Code review agents
• Codebase exploration agents
• Refactoring agents
• Any agent designed to autonomously explore and analyze code

Root cause: The main assistant's system prompt likely lacks guidance on:

• When to pre-fetch data vs let agents fetch it
• Never restricting agent tool usage with "Do NOT read files" instructions
• Trusting custom agents to use their configured tools

Suggested fix location: Main Claude Code assistant's Task tool documentation/guidance should include:

• "Don't pre-fetch what agents can fetch themselves"
• "Never tell agents not to use their configured tools"
• "Pass high-level tasks, not pre-processed data"
• "Only provide data when the agent lacks tools to fetch it"

Environment:

• Claude Code Version: 2.0.x (affects all versions with Task tool)
• Agent Type: Custom user-created agents
• Platform: All platforms