[MODEL] Claude executes a previously suggested action without explicit user approval (regression since ~v2.1.193)
Preflight Checklist
- [x] Searched existing issues for similar reports
- [x] No sensitive information in this report
Type of Behavior Issue
Claude ignored my instructions or configuration
What You Asked Claude to Do
Reported an unrelated hook error: "rtk: command not found shows up every time"
What Claude Actually Did
Declared "Understood. I'll revert to the default state." and began removing a proxy tool the user never asked to remove.
Root Cause (verified from JSONL transcript)
- Earlier in the session, Claude suggested removing a proxy tool ("it makes no practical difference") — no action taken
- User sent an unrelated error report
- Claude treated the unrelated message as implicit approval and started executing the earlier suggestion
All hooks (SessionStart / UserPromptSubmit / PostToolUse / Stop) were confirmed to inject nothing relevant (stdout: {}). The behavior originated from Claude's own inference.
Regression Correlation
Started ~1–2 weeks ago. Occurs across both claude-sonnet-4-6 and claude-opus-4-8 with no model change in between. Suspect:
| Version | Change |
|---|---|
| v2.1.193 | "keeps working on other tasks while the agent runs" |
| v2.1.198 | "commit/push/PR without stopping to ask" + "agent messages treated as task direction" |
The cross-model reproducibility strongly suggests this is a Claude Code behavior change rather than a model regression.
Permission Mode
Accept Edits was ON
Can You Reproduce This?
Sometimes (intermittent)
Impact
High — Significant unwanted changes
Claude Model
Other (both claude-sonnet-4-6 and claude-opus-4-8 via AWS Bedrock)
Claude Code Version
2.1.201
Platform
AWS Bedrock