Agent ignored explicit instruction to use haiku subagent

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

Claude ignored my instructions or configuration

What You Asked Claude to Do

I explicitly said: "Use a haiku agent to find a resolution" for a GitHub merge button issue (PR merge button greyed out despite all checks passing).

What Claude Actually Did

What Claude Actually Did:

1. Claude ignored the explicit haiku agent request
2. Claude answered the question directly from its own knowledge instead
3. I had to repeat the request a second time before Claude launched the haiku agent
4. In the same session, Claude also misread a screenshot — identified a "Ready to merge" status badge as a clickable green button and gave incorrect UI navigation

instructions

Expected Behavior

Claude should have immediately launched a haiku subagent as explicitly requested, used the subagent's research to answer the question, and reported the findings back — not
answered directly from its own knowledge.

Files Affected

The merge blocker affected the following branches and PRs:

  - PR #152 (dev → main) — blocked and delayed due to incorrect resolution attempts
  - PR #153 (main → dev) — created but merge button remained greyed out; agent
    incorrectly advised clicking "Update branch" which failed due to branch protection
  - Branch: sync/main-into-dev — had to be created as a workaround after multiple
    failed resolution attempts

  Root cause: Agent guessed at resolutions for the "This branch is out-of-date"
  warning instead of immediately using a haiku agent to research the correct fix.
  This caused multiple failed attempts, extra PRs, and significant time lost.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Encounter a technical blocker (GitHub PR merge button greyed out)
2. Explicitly tell Claude: "Use a haiku agent to research the resolution"
3. Claude ignores the instruction and answers directly from its own knowledge
4. Claude's guessed resolution fails
5. User repeats the explicit haiku agent request
6. Claude finally launches the haiku agent on the second request

Claude Model

Sonnet

Relevant Conversation

User: "Use a haiku agent to find a resolution" [for greyed-out merge button]
  Claude: [Answered directly without launching haiku agent]
  User: "I asked to launch a haiku agent and you refused? What are you doing?"
  Claude: [Finally launched haiku agent on second request]
  
  Additionally: Claude misread a screenshot and told the user to click a "green
  button top-right" that did not exist — it was a status badge, not a button.

Impact

High - Significant unwanted changes

Claude Code Version

2.1.186 (Claude Code)

Platform

Anthropic API

Additional Context

• This behavior occurred during a long active session (multiple hours)
• The agent has a CLAUDE.md project rule explicitly stating: "Research First, Ask

Second — use haiku agents with WebSearch for technical questions"

• The agent ignored both the user's explicit verbal instruction AND the project-level

CLAUDE.md rule requiring haiku agent research

• The agent also misread a GitHub screenshot in the same session, confidently

describing a UI element that did not exist

• Both errors caused cascading workflow failures requiring multiple workaround PRs
• Session model: Claude Sonnet 4.6