Models & providers
Codesteward routes specialist stages through a model router:
- OpenAI
- Anthropic
- SpaceXAI
- OpenAI-compatible endpoints
- LiteLLM
Configuration surfaces
| Surface | Use |
|---|---|
| Host env | Single-tenant dogfood (OPENAI_API_KEY, …) |
| Org Models UI | Per-org encrypted provider keys + stage matrix |
| Platform runtime | Install-wide knobs (platform operators) |
Org admins set provider credentials and which model serves which stage (planning, specialists, judge, …). Per-stage “env: API key ref” columns are not used — keys live on the provider.
Langfuse
Optional dual-write tracing: org Langfuse project and/or platform Langfuse. Useful for cost and quality debugging without shipping prompts to a third SaaS review vendor.
Cost control tips
- Start with a single strong model for all stages
- Use cheaper models for nit-heavy or low-risk tiers
- Cap concurrent specialists (
STEW_MAX_CONCURRENT) - Prefer path-scoped stewardship on large monorepos