internalcot

internalcot

internalcot asks a model to put its working notes in a visible function call before answering. The notes are model-authored scratchpad text, not access to hidden reasoning tokens.

Install

npm install --global internalcot

Set OPENAI_API_KEY before running the CLI. OPENAI_BASE_URL can point the OpenAI SDK at a compatible gateway.

Usage

internalcot "Solve this problem"
cat problem.md | internalcot
internalcot --model gpt-5.6-sol --effort none "Check this proof"

Scratchpad output goes to stderr; the final answer goes to stdout so it can be redirected or piped.

internalcot "Check this proof" >answer.md 2>scratchpad.txt

Run internalcot --help for all options.

How it works

The CLI owns the Responses API request. It sets native reasoning effort to none by default, requires an internalcot function call on the first turn, displays that call's arguments, and then returns the tool result so the model can produce its final answer.

This does not reveal private or hidden chain-of-thought. It elicits a separate, observable scratchpad whose usefulness and faithfulness should be evaluated independently.

Development

npm install
npm run check
npm test
npm run build
npm link

Publishing

The package is public and unscoped. Before the first release:

npm login
npm whoami
npm run prepublishOnly
npm pack --dry-run
npm publish

Check that internalcot is still available on npm immediately before publishing.