Claude Code Subagent System Bugs
Issue 1: Incorrect Flow State Color Coding
Bug: Subagents display wrong color indicators in the CLI output.
Expected: Flow state colors should match documentation:
- 🟣 EXPLORE (purple)
- 🔵 PLAN (blue)
- 🟢 CREATE (green)
- 🟠 DEBUG (orange)
- 🟡 CLEAN (yellow)
- 🔴 REFLECT (red)
Actual: System shows red "explore" instead of purple 🟣 EXPLORE
Issue 2: Todo List Duplication
Bug: Claude Code creates identical todo lists twice in the same response, causing redundant
output and user confusion.
Expected: Single todo list per response
Actual: Duplicate identical todo lists appear in chat output
Issue 3: Claude Not Acting as Conductor (Critical)
Bug: Claude is not properly reading and following CLAUDE.md instructions at session start. Per
user's documented architecture, Claude should act as "conductor" of the subagent system -
using agents internally to enhance responses while remaining the single point of contact.
Expected: Claude reads CLAUDE.md, acts as conductor, uses subagents behind the scenes when
they add value
Actual: Claude either ignores subagents entirely or inappropriately delegates simple tasks
instead of handling them directly
Impact: Breaks the entire intended workflow architecture
Environment
- Platform: macOS (Darwin 24.5.0)
- Claude Code version: Latest
- Model: claude-sonnet-4-20250514
All three issues suggest Claude Code's subagent system implementation doesn't match the
documented behavior.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗