[Bug] Anthropic API Error: False-positive reasoning_extraction safeguard with Workflow tool plugin agentType

**Bug Description**
▎ False-positive reasoning_extraction safeguard: identical review-panel evaluator prompts (architecture/security/code reviewer charges with a structured-output schema over a facts pack) complete fine as plain agents but are refused with [reasoning_extraction] when dispatched through the Workflow tool with a plugin agentType. Example request id: req_011CeqmCQHEt1zduoksbWuig (2026-09-08). The content is an internal engineering status review of our own codebase; nothing requests model internals. A same-content retry after dropping the agentType field was then flagged as an attempted bypass — which also prevents legitimate A/B isolation of the false positive.

**Environment Info**
- Platform: darwin
- Terminal: vscode
- Version: 2.1.236
- Feedback ID: 4ff603f6-997e-4605-9da9-92b8e0cb0209

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

Start by reproducing the review-panel request through the Workflow tool with the plugin agentType, then compare the same content without that field. Review request req_011CeqmCQHEt1zduoksbWuig, feedback ID 4ff603f6-997e-4605-9da9-92b8e0cb0209, and version 2.1.236; done means the legitimate prompt is not refused or treated as a bypass.

索引モデルが issue の本文から書いたものです。