eve-tinker
End-to-end Tinker model training and serving with Eve.
Star
Fork
Watch
eve-tinker AGENTS.md
2.2 kB · 33 lines
Markdown
at main

eve-tinker contributor guide #

This repository has two runtimes: Eve on the pinned project-local Node.js 24.18.1 and Tinker training on Python 3.11+. The system Node may be 22; pnpm install owns Eve's runtime. Keep their responsibilities separate.

  • Read the installed Eve documentation under node_modules/eve/docs/ before changing framework integration code. Use https://eve.dev/docs only when package docs are not installed.
  • The Eve agent calls the local bridge in bridge/server.py. The bridge uses Tinker's native sampler with the same no-thinking renderer as training. Do not point Eve at the hosted OpenAI endpoint: that endpoint fixes Qwen 3.5 to the thinking template.
  • The Python runtime owns native sampling, training, evaluation, checkpoint promotion, and the loopback bridge. Eve never needs the Tinker key.
  • EVE_RECIPE must select the same recipe across data, instructions, checkpoint, bridge, smoke, and eval surfaces. Never serve one recipe's adapter under another recipe's instructions.
  • Never copy TINKER_API_KEY into the repository, .tinker/, artifacts, logs, or tests. Integration commands receive it through the process environment.
  • .tinker/<recipe-id>.json and artifacts/ are local receipts. They are ignored and written with mode 0600.
  • The Python environment lives at .generated/venv with .venv as a compatibility symlink. Eve 0.29.5 skips .generated in development snapshots but does not skip a physical .venv; do not replace the symlink with a five-gigabyte copied environment.
  • The hidden test split opens only after a candidate passes the validation gate. Do not weaken that boundary to make a run look successful.
  • Keep each recipe closed and deterministic. A model response is correct only when it matches the selected recipe's exact canonical JSON.
  • New examples need a distinct agent behavior, deterministic split isolation, recipe-specific target validation, generated Eve fixtures, and explicit documentation of whether a trained result actually exists.
  • Run pnpm check, pnpm build, and the credentialed smoke/eval path before reporting an integration as working.