[Bug] Anthropic API Error: Reasoning extraction safeguard false positive on consecutive tool calls

**Bug Description**
Fable 5 safeguards false-positive, error detail [reasoning_extraction], on two consecutive Agent-tool builder dispatches. The prompt was a normal software-build brief (read a gate record, edit FastAPI code, run pytest, commit) with reporting instructions like "return a short summary, do not paste file bodies". Same prompt ran on Sonnet without a flag. Request IDs: req_011CeDgxtc5NNpyUpNAqBwjE and req_011CeDh2SSLMUvwmRscbjAfS.

**Environment Info**
- Platform: win32
- Terminal: null
- Version: 2.1.237
- Feedback ID: 0a1d729f-a9d5-4e49-ae3b-07ee84951d53

**Errors**
```json
[]
```

No repository files or tests are named. Start by reproducing the two consecutive Agent-tool builder dispatches with the reported prompt and compare the reasoning_extraction safeguard behavior across the two request IDs. Done means the normal consecutive-tool-call flow no longer produces a false positive, with a regression check covering that sequence.

索引モデルが issue の本文から書いたものです。