AGENTS.md

Rama main · réplica de lectura del repositorio original.

- Respond short, precise, and concise.
- If you didn't add a test, you didn't fix a bug. Every bug fix must include a reproducible test that fails without the fix and passes with it.
- When implement a big feature, create specs first then implement. Each spec is tightly scoped.
- You commit frequently, one small scope diff at a time. Push to main once a full batch of work is complete and verified.
- Leftover must be cleaned up using follow up commits.
- Target for high quality outcome, >90% coverage, each feature verified via an e2e test.
- Write in audience language, where documents aim users value and usage understanding, api surface aim builders precision, and code comments aim tech precision.
- Specs live in specs/
- Relevant latere projects and shared components (latere.ai/x/pkg/*) are in ../
- This repo owns the open-weights inference layer end to end: frozen weights in S3, serving on bare-metal k8s GPU nodes, OpenAI-compatible endpoints registered behind the Lux model gateway.
- Serving conventions follow the shared latere service contract: /livez, /readyz, /version, /metrics (latere.ai/x/pkg/health), Docker + k8s packaging.
- Every sentence is written for one reader (user, contributor, developer) and the register follows the reader; an error has one code, one fixed user sentence in `message`, and one developer detail in a separate field. The rule and the review checklist: https://github.com/latere-ai/pkg/blob/main/docs/writing/registers.md