Agents & MCP
2 min read · Hacker News
Claude Code Extended Thinking Logs Are Encrypted Signatures, Not Raw Agent Reasoning
Analysis of local Claude Code session logs reveals that 'extended thinking' blocks are stored as encrypted 600-character signatures. The local system receives only a lossy summary of the model's thinking process, while the actual reasoning trace remains locked by Anthropic.
Why it matters. You cannot rely on local Claude Code session files to build reliable, high-fidelity audit trails of agent behavior.