Changelog
All notable changes to this project are documented here. The format follows Keep a Changelog, and the project uses Semantic Versioning.
[Unreleased]
Changed
- The docs now live at https://keeljourney.dev.
- Reframed the project as a learning journey with the tagline "Learning by building, one port at a time". The one-file agent is Destination 1.
- Docs are organized by destination: a journey home page, then a section per destination.
bench/baseline.shneedsMODELS, andbench/setup.shpulls only the Ollama models it lists.
Added
- MIT license.
- GitHub Actions workflow that type-checks and tests every push and pull request.
npm run typecheck, powered by TypeScript..nvmrcpinning Node 22.18.0.- Package metadata: description, repository, keywords and license.
- Documentation site built with MkDocs Material and published to GitHub Pages.
bench/swebench_lite.py: runs Keel on SWE-bench Lite and grades the patches, with resumable runs and automatic Node download.- Benchmark instructions for Terminal-Bench 4.0 and Aider Polyglot through Harbor.
- Logo and banner, and a README header with badges.
PROXY_HOSTfor the Bedrock proxy, so benchmark containers on Linux can reach it.
Fixed
- Keel no longer fails with
other side closedwhen a command outlasts the model server's keep-alive timeout, as with the Bedrock proxy or vLLM. bench/setup.shandbench/baseline.shwork on Linux, andbench/swebench_lite.pyusesKEEL_BASE_URLlike the Harbor runs.
[0.1.0] - 2026-11-20
Added
agent.ts: a single-file coding agent that works through a bash tool over any OpenAI-compatible chat API.- Text adapter that reads one fenced
bashblock per reply, so it works the same on every model and server. - Canonical
PartandMessagetypes with anAdapterinterface for supporting new APIs. - Completion signalled by a command that prints
COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT. - Test suite that runs the real loop against a scripted fake model server.
- Harbor adapter (
bench/keel_agent.py) for Terminal-Bench and SWE-bench. - Optional AWS Bedrock proxy (
bedrock/).