Skip to content

CRUSH.md

Latest commit

 

History

History
51 lines (37 loc) · 2.2 KB

CRUSH.md

File metadata and controls

51 lines (37 loc) · 2.2 KB
Raw

Catwalk - AI Provider Database

Build/Test Commands

  • go run . - Build and run the main HTTP server on :8080
  • go run ./cmd/{provider-name} - Build and run a CLI to update the {provider-name}.json file
  • go test ./... - Run all tests

Code Style Guidelines

  • Package comments: Start with "Package name provides/represents..."
  • Imports: Standard library first, then third-party, then local packages
  • Error handling: Use fmt.Errorf("message: %w", err) for wrapping
  • Struct tags: Use json tags with omitempty for optional fields
  • Constants: Group related constants with descriptive comments
  • Types: Use custom types for IDs (e.g., InferenceProvider, Type)
  • Naming: Use camelCase for unexported, PascalCase for exported
  • Comments: Use //nolint:directive for linter exceptions
  • HTTP: Always set timeouts, use context, defer close response bodies
  • JSON: Use json.MarshalIndent for pretty output, validate unmarshaling
  • File permissions: Use 0o600 for sensitive config files
  • Always format code with gofumpt

Model cost fields

The cached cost fields in provider configs are counterintuitive:

  • cost_per_1m_in_cached = cache creation (write) price
  • cost_per_1m_out_cached = cache read price

Providers usually advertise a single discounted "cached" price (e.g. "$0.044/M cached") — that is the cache read price, so it goes in cost_per_1m_out_cached. Leave cost_per_1m_in_cached at 0 unless the provider explicitly prices cache writes (Anthropic-style).

Adding more provider commands

  • Create the ./cmd/{provider-name}/main.go file
  • Try to use the provider API to figure out the available models. If there's no endpoint for listing the models, look for some sort of structured text format (usually in the docs). If none of that exist, refuse to create the command, and add it to the MANUAL_UPDATES.md file.
  • Add it to .github/workflows/update.yml

Updating providers manually

Zai

For zai, we'll need to grab the model list and capabilities from https://docs.z.ai/guides/overview/overview.

That page does not contain the exact context_window and default_max_tokens though. We can grab the exact value from ./internal/providers/configs/openrouter.json.