Repository navigation
Platform tokens follow-ups: edge reject on inference routes, per-IP limiter, revoke on sign-out-all #167
Description
Activity
Additional follow-ups (from the API reference review, 2026-10-10):
- cloud-api: the
session_required401 message ("API keys only work for inference") is outdated since member sk- keys and nan_pat_ tokens work on GET /api/workspaces[/{uuid}] and /v1/runs*. Reword it (the website docs quote it verbatim in src/data/openapi.json; update both together). - Document the 429 / 500 bodies the /api auth layer can return in the API reference.
- cloud-api: the
Follow-up from T-167 (cloud-api #326): the per-IP limiter for rejected platform tokens (30/min, burst 5) also 429s VALID tokens from the same IP once the bucket is spent. Workspaces egress through their host's public IP, so a misbehaving agent in one member's workspace could briefly 429 another member's valid PAT calls from the same host (same property as the existing sk- limiter). Options: exempt the fleet egress IPs (devops fleet registry) from the IP bucket and rely on per-token limits there, or key the bucket by token prefix + IP.
Metadata
Metadata
Assignees
Labels
No labels
Follow-ups from cloud-api #323 review (#166).
api-edge-normaliser): rejectAuthorization: Bearer nan_pat_...(any whitespace, tabs incl.) on every non-platform route of api.nan.builders, so inference routes served directly by LiteLLM (/v1/chat/completions, /v1/responses, /v1/embeddings, ...) never even receive a PAT. LiteLLM already 401s them; this is defence in depth. Tests: byte-identical behavior for sk- keys; /v1/runs* passthrough unchanged.