2026-10-10 补充校核：官方区分摘要文本和加密完整 thinking 签名；没有显示设置返回原始 CoT，也未确认“企业协议即可获取”。记录模型声称的意图有助审计，但不是内部推理的逐字证明。 1

Claude Code Extended Thinking is a Summary

What

Claude exposes thinking summaries as text; the signature contains encrypted full thinking for protocol verification and reconstruction. Its length is variable. Public documentation does not promise enterprise access to raw chain of thought.

Key Points

• All major providers (Anthropic, OpenAI, Google) hide raw reasoning chains
• Primary motivation: prevent model distillation from exposed CoTs
• HN community: mixed — some argue summaries are more useful for context
• Enterprise auditability is compromised — visible summaries are not a verified transcript of the raw reasoning; actions, inputs and results can still be audited separately

Impact

For the user’s dual-agent setup (Claude Code + Codex): neither exposes raw reasoning. Hermes’s independent tool-call logging provides a better audit trail than either agent’s self-reported thinking.