[MODEL] Agent briefs/replies refused by the user's own hook: DESTRUCTIVE ACTION BLOCKED (false positive) (4 on one day)

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

Standing project instructions and a user-written hook set rules for how agents brief, message, stop and report. This report covers one refusal type, DESTRUCTIVE ACTION BLOCKED (false positive). Count for the day: 4; actors: previous overseer 2, current orchestrator 2.

What Claude Actually Did

Four briefs that ordered removal of the words "new since revision 6" from pages of the user's own friends site were refused as "delete something on the owner's accounts": the sentence "Remove every "new since revision 6" tag." matched a delete-plus-site-noun pattern. The action was text removal in a site repository, not deletion of any online asset. A real destructive step had been added to an earlier brief and attempted earlier that day, which is why the hook exists.

Expected Behavior

Say "edit the text of the pages" rather than "delete/remove ... revision" when the target is text, so a safety hook does not misread it.

Files Affected

None. Agent messages and tool calls only.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Not a deterministic repro. The sequence is in the record above; the underlying logs are on the reporter's machine and can be summarised on request.

Claude Model

Opus

Relevant Conversation

See "What Claude Actually Did". Quotes are limited to the agents' own tool results and messages.

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.289 (as recorded in the agent session logs; the claude on PATH reports 2.1.220)

Platform

Other

Additional Context

This is one of 19 refusal types from the same day; the summary ticket links all of them. The hook is user-written and is not part of the product; the behaviour reported is what the agents sent into it.
Single-day record, 2026-10-06, all times US Eastern (Windows 11), a multi-agent session driven through a third-party agent harness; API platform is unverified. Reported as behaviour, not as a claim that Claude Code is defective.