[BUG] 2.1.27 has sub agents reading the user level claude.md

Resolved 💬 4 comments Opened Jan 31, 2026 by blackholegalaxy Closed Feb 4, 2026

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?

My global ~/.claude/CLAUDE.md define orchestration rule to delegate to sub agents. Then in project i have a CLAUDE.md describing general project rules. The skills and agents are inside the project too.

When main thread boots a sub agent, it fails cause sub-agent reads my user level CLAUDE.md and want to boot a sub-agent itself.

It's also wasting a lot of tokens from start since global claude.md is very detailed.

What Should Happen?

The sub-agents should NOT be able to read user level claude.md. It WAS NOT the behavior we got previously.
Correct version: 2.1.19. Downgrading solves this issue and sub agent doesn't behave as orchestrators anymore.

We HAVE to be able to define somewhere a context that is ONLY accessible from main thread.

Error Messages/Logs

⏺ qa-engineer(Wave 1 Mobile QA Testing)
  ⎿  Done (0 tool uses · 24.6k tokens · 11s)

⏺ I see the issue - I delegated to a qa-engineer agent, but that agent is also an orchestrator following the same orchestration rules. Let me
  check what agents are actually available for this work.

⏺ Read 3 files (ctrl+o to expand)

⏺ Perfect! The qa-engineer agent is configured as a WORKER who can execute Chrome MCP testing directly. The previous agent invocation got
  confused because it had orchestrator context. Let me launch a fresh qa-engineer agent with clean, focused instructions:

⏺ qa-engineer(Wave 1 Mobile QA Testing)
  ⎿  Done (0 tool uses · 24.2k tokens · 15s)

⏺ The qa-engineer agent is still acting as an orchestrator. This is because the global CLAUDE.md instructions are being inherited by all agents.
  I need to directly use the QA skill instead. Let me check available skills:

Steps to Reproduce

  1. Write a user level claude.md telling that he must act as main frame orchestrator and delegate any task to sub agents. Sub agents and skills are in the project
  2. Write a sub agent like a qa-engineer that is meant to just read a file for example
  3. Task the main frame to do a qa pass.
  4. The qa engineer fails while trying to boot a sub agent

Claude Model

Sonnet (default) and Opus. Not tested with haiku.

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.19

Claude Code Version

2.1.27

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

Tested on various terminal and shells. Still same behavior.

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗