internalcot now requires a detailed visible reasoning trace while $internalcot mode is active.
Highlights:
- Requires a note before the first substantive tool call or answer.
- Records intermediate derivations, calculations, alternatives, failed approaches, uncertainty, evidence, and checks instead of concise checkpoints.
- Requires continued notes between reasoning phases and final verification before answering.
- Adds explicit non-substitution and retry/recovery instructions inspired by reliable agent-tool contracts.
- Strengthens the API observation prompt and tool schema to match the original experiment.
- Credits and links Can Bölük’s original Python proof of concept: https://pasta.can.ac/omegiligox.py
Original post: https://x.com/_can1357/status/2087228354399265125