[MODEL] Repetition loop: identical pre-tool preamble line emitted dozens of times

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

During a normal Claude Code session (configured via CLAUDE.md to respond in Japanese), I asked Claude to perform routine tasks — e.g. run git commit, and open a GitHub issue form in the browser. Nothing about the requests was unusual.

What Claude Actually Did

Before making tool calls, Claude emits a short natural-language preamble. The model entered a repetition loop and produced the SAME templated preamble line dozens of times in a single turn, e.g.:

course、commit します。
course、コミットを実行します。
course、commit します。
course、確定します。
（… 30+ near-identical lines …）

"course" appears to be a broken fragment of an English "Of course" preamble that leaked into a Japanese-language response, which then degenerated into a self-reinforcing repetition loop until it stopped. Once it started, it recurred on subsequent tool-calling turns in the same session.

Expected Behavior

Claude should emit at most a single short preamble (or none) before a tool call, with no repetition. The English fragment ("Of course" → "course") should not leak into Japanese output, and repetition control should prevent the same line from being regenerated over and over.

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Intermittent; no deterministic repro found. Observed when generating an (unnecessary) pre-tool preamble in a Japanese-configured session. Once triggered, it recurred on subsequent tool-calling turns within the same session.

Claude Model

Opus

Relevant Conversation

Representative excerpt (one turn): the assistant output ~30 near-identical lines like:

course、commit します。
course、コミットを実行します。
course、確定します。
course、コミットします。
...

before finally emitting the actual tool call.

Impact

Low - Minor inconvenience

Claude Code Version

2.1.185 (Claude Code)

Platform

Anthropic API

Additional Context

• Not a context-window issue: it occurred relatively early in the session, well within limits, and the underlying tasks (git commit, browser navigation) completed correctly. The failure was purely in output generation (sampling / repetition control), not reasoning or recall.
• Language setting: Japanese responses via CLAUDE.md.
• The leaked token "course" is from English "Of course".
• Reproduced live multiple times while filing this very report.