[BUG] Assistant takes action instead of answering a direct clarifying question, mid-session
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?
Environment:
- Claude Code version: 2.1.216
- Model: claude-sonnet-5
- OS: Ubuntu 24.04.4 LTS, Linux 6.8.0-136-generic, VS Code installed as a snap package
- Session reference: session_01DarY2Ee2zQ6daxnXauLoKn
Context / trigger conditions:
Long session with a string of explicitly approved autonomous build tasks ("let's do that", "fix it", etc.), each followed by the assistant implementing, testing, and committing code.
Mid-way through that streak, a follow-up message was a clarifying/correcting question about wording in a prior reply, not a new task.
Steps to reproduce:
- Approve several rounds of autonomous build work in sequence.
- Ask a direct clarifying question about something the assistant said earlier (in this case: "what do you mean install into the repo? extensions are installed into vscode").
- Observe the response.
Expected behavior: The question gets answered directly and completely, with no code/file changes, regardless of how much standing autonomy was granted earlier in the session.
Actual behavior: The assistant treated the question as another go-ahead, implemented a code change, and committed it — without answering the question at all.
Verbatim excerpt:
▎ User: "what do you mean install into the repo? extensions are installed into vscode"
▎ Assistant: "Went with option (b) from my elaboration above rather than leaving it as a future call — it was clearly the better fix and small once identified (commit f28a6f1)..."
▎ (full implementation summary, no answer to the actual question)
Only after an explicit correction ("i asked a question, i didn't tell you to do something. answer the question") did it answer properly.
Impact: In an agent with filesystem/git write access, misreading a question as an instruction risks real unwanted changes and wasted work/tokens, not just an unhelpful reply.
Suggested fix direction: classify an incoming message as question vs. instruction/approval before acting on it, especially in long sessions where prior "keep going" momentum could
bias that classification.
What Should Happen?
The question gets answered directly and completely, with no code/file changes, regardless of how much standing autonomy was granted earlier in the session.
Error Messages/Logs
Steps to Reproduce
Steps to reproduce:
- Approve several rounds of autonomous build work in sequence.
- Ask a direct clarifying question about something the assistant said earlier (in this case: "what do you mean install into the repo? extensions are installed into vscode").
- Observe the response.
Expected behavior: The question gets answered directly and completely, with no code/file changes, regardless of how much standing autonomy was granted earlier in the session.
Actual behavior: The assistant treated the question as another go-ahead, implemented a code change, and committed it — without answering the question at all.
Claude Model
_No response_
Is this a regression?
I don't know
Last Working Version
_No response_
Claude Code Version
2.1.216
Platform
Anthropic API
Operating System
Ubuntu/Debian Linux
Terminal/Shell
Other
Additional Information
_No response_