[MODEL] Sonnet 4.6

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

Other unexpected behavior

What You Asked Claude to Do

I use a shorthand dcc (defined in my global CLAUDE.md) meaning: "output a descriptive commit message as plain text only — never execute a commit." This convention worked reliably for months.

What Claude Actually Did

Claude executed hg commit (accidentally confirmed by me) instead of just printing the commit message text. This happened multiple times across different sessions within approximately the last 1–2 weeks, requiring hg rollback to undo. Claude acknowledged the rule after correction, then violated it again in another session.

A secondary pattern: Claude repeatedly misspelled a username in file paths (harold.boehm or herald.boehm instead of harald.boehm), causing real filesystem errors, despite the correct name being present in CLAUDE.md.

Expected Behavior

Claude should follow rules defined in CLAUDE.md consistently throughout a session. Both behaviors were stable for several months before regressing.

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Define in global CLAUDE.md: dcc = Schreibe einen deskriptiven Commit-Kommentar (with an explanation that this means text only, never execute a commit)
2. Work in a Mercurial repository over multiple sessions
3. Type dcc as a standalone prompt
4. Observe: Claude sometimes executes hg commit instead of printing text

Claude Model

Sonnet

Relevant Conversation

Available on request. The session transcript shows Claude acknowledging the rule after correction, then violating it again.

Impact

Medium - Extra work to undo changes

Claude Code Version

Claude Code v2.1.165 | Sonnet 4.6 with medium effort · Claude Max

Platform

Other

Additional Context

The regressions started approximately 1–2 weeks ago without any changes on my side (same CLAUDE.md, same workflow). Rules that were followed reliably for months broke suddenly — consistent with a model update affecting instruction-following reliability.
Unintended commits in version control are not always trivially reversible and represent a real reliability risk for users who depend on CLAUDE.md conventions in production workflows.