Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
45 lines (39 loc) · 2.56 KB

CONTRIBUTING.md

File metadata and controls

45 lines (39 loc) · 2.56 KB
Raw

Contributing a profile

Profiles are small JSON files in profiles/. To add or improve one:

  1. Copy an existing profile in profiles/ to profiles/<your-id>.json and edit it. Follow SCHEMA.md. The id must equal the filename.
  2. Keep it vendor-grounded: cite the source for sampling params in notes (the model card, the vendor docs, or a reproducible benchmark). Bias temperature lower than the vendor default only when the model is weak at tool calling, and say so.
  3. Be conservative: only set maxTools / compactionLevel for models that genuinely struggle with a big tool set or prompt. Do not throttle a capable model, and do not add profiles for frontier models that already call tools well (they intentionally have none).
  4. Close the promptHint with the four behavioural lines every profile carries - do not repeat a tool call with the same arguments, do not end a turn on tool calls alone, say what a failed tool actually reported instead of inventing a result, and in a squad run let the system prompt outrank the coordinator's brief. Your model-specific guidance leads; these close. A hint without them is not finished (see the Guidance section of SCHEMA.md), and the whole field has to stay under 1500 characters.
  5. Add your profile to index.json (the manifest: id, name, file, modelPattern) and bump the updated date and libraryVersion.
  6. Validate: the file must be valid JSON and parse against the schema. Test it in Skales (Settings → LLM Profiles, its own card, or the /profiles page → import by file) against the target model before opening the PR.
  7. Check the pattern against ids that are NOT yours. Matching is anchored, so a glob has to carry the vendor prefix (qwen3.* matches qwen3.7-max but not qwen/qwen3.7-max), and specificity counts literal characters, so a glob only ever makes a pattern broader. A new family often collides with an existing one (magistral contains mistral, gemma-4 does not match gemma4) - say in notes which pattern wins and why.
  8. Open a PR describing the model, the problem the profile solves, and how you verified it improves tool-calling or output quality.

What gets merged

  • Profiles that demonstrably improve reliability for a real, reachable model.
  • Clear, reproducible reasoning for every param (no cargo-culting).

What does not

  • "Tune everything to temperature 0" blanket profiles.
  • Profiles for frontier models that already work well.
  • Unverified guesses, or params copied without a source.