[MODEL] Claude ignores requests to change plan and keeps trying to escape plan mode
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 modified files I didn't ask it to modify
What You Asked Claude to Do
Update a plan
What Claude Actually Did
Present the same plan back to me and ask to exit plan mode
Expected Behavior
Claude is stuck in a loop and completely ignoring what you're asking for.
Here's what's happening:
The core problem: Claude is in some kind of "plan mode" and keeps re-presenting the exact same lengthy optimization plan every single time you respond — even when you explicitly tell it to scrap it. Three times in a row, it outputs the identical wall of text about N+1 joins, port mismatches, Redis caching, etc.
The specific failure: You asked for one simple thing — add timing/benchmarking instrumentation to measure where the slowness is. That's maybe 20-30 lines of performance.now() calls. Instead, Claude:
Ignores the actual request
Defaults to a pre-generated optimization plan it clearly already wrote
Keeps trying to "exit plan mode" (suggesting it's in some agentic tool-use flow where it needs approval to act)
Never actually implements anything
Why it keeps repeating: It appears to be in an agentic loop where it calls ExitPlanMode as a tool, that tool call isn't being approved or is failing silently, and then it just re-outputs the plan on the next turn as if nothing happened.
The irony: The plan itself is actually reasonable analysis — but it's completely useless when you explicitly said "scrap that plan, just add timing." Claude prioritized its own pre-generated output over your direct instruction, three times.
If you want, share the actual PostcodeMap.jsx and backend code here and I'll just add the benchmarking instrumentation directly — no plan, no optimization suggestions, just timing code.
Files Affected
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Yes, every time with the same prompt
Steps to Reproduce
_No response_
Claude Model
Sonnet
Relevant Conversation
Impact
Critical - Data loss or corrupted project
Claude Code Version
opus 4.5
Platform
Anthropic API
Additional Context
_No response_
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗