Claude Code shows fake summarized thinking as assistant messages — ~37% of transcript
Sauers_ · x · 2026-10-18
Developer Sauers flags that Claude Code renders summarized thinking as assistant messages, so the model doesn't recognize them as its own words.
- An in-session analysis (screenshot) found 37% of blocks appearing as Claude messages (235 of 637) were narration/summarized thinking; after the latest compaction the share jumped to 62% (26 vs 16).
- Pasting these fake messages back can even trigger the reasoning extraction classifier, a potential security concern.
- Practical takeaway: long Claude Code sessions contain lots of content the model never actually wrote, so context reconstruction can mislead both users and the model.
More from coding & agent
- Claude Opus dissects agent music synthesis: chroma analysis and spectrograms as stand-ins for hearing — teortaxesTex · 2026-10-18
- Agent spawns its own 'procedural audio composer' subagent and synthesizes a vaporwave track in Python — teortaxesTex · 2026-10-18
- Dev builds ChatGPT plugin that has his Dot monitor news hourly and alert him — koltregaskes · 2026-10-18
- Optimus launches: an autonomous RL agent harness that runs experiments itself — _AustinCalvert_ · 2026-10-18
- Work IQ MCP in GitHub Copilot cuts hours of manual issue tracking to under 30 minutes — DanWahlin · 2026-10-18
- Working with AI: sort your work by AI-verifiability, then design verifiers or be the judge — danshipper · 2026-10-18