[playgrndlabs-bot]

Requested by PlaygrndLabs · project thread

What

Follow-up to #24. Kimi logins and Vertex service accounts now route the same way Claude and ChatGPT logins already do:

• One candidate per login. Each Kimi login and Vertex service account is now its own candidate (id file:<name>, group oauth-served). It used to be a single candidate per provider that the Durable Object rotated through.
• Failover and cooldowns. A rate-limited, refused or failing login rests on CPA's cooldowns, and the next login serves the request.
  • The cooldowns are shared through the DO.
  • They show in Auth Files and can be cleared with the reset button.
• Routing fields. The auth-file fields priority, weight, prefix, disable-cooling and request-retry now apply to these logins.
• Tokens by name.
  • Backend.loginToken(provider, name?) fetches a named login's token.
  • Tokens are cached per provider:name in each isolate.
  • On a 401, only that login's token is dropped and refreshed (still at most once a minute per login).
• Fallback for older Durable Objects. A table with no logins list still gets one "next in turn" candidate per provider.

Security pass

• Login names. Names sent to the DO come from the server's own routing table and are never taken from the client. A client's prefix/ only filters the list.
• Lookup scope.
  • Account.loginToken(provider, name) only looks among that provider's enabled files.
  • A name belonging to another provider is refused with a generic 503, so no tokens or file details leak (covered by a test).
• Token isolation. Cached tokens are keyed per provider and login, so one login's token is never sent with another's request.
• Limits. The existing caps still hold: 10 upstream attempts and 120 s of waiting per request (Free plan's 50 subrequests). Refreshes after a 401 are still limited to once a minute per login.
• No new data stored. Nothing new is logged or stored beyond the cooldown rows Route across logins and keys with CPA's strategies, retries and cooldowns #24 already keeps (id, model, reason, status, until).
• Not ported. No cloaking, impersonation or TLS fingerprinting.

Tests

npm run verify: typecheck, lint and 227 tests all pass. New tests:

• Two Kimi logins: the rate-limited one cools (file:k1.json, model-scoped) and the other serves the request. Usage is counted per login.
• Two Vertex logins: a 401 refreshes only the refused token. team/ picks the login with that prefix.
• A Kimi login that can't get a token is skipped for the next one.
• DO level: Vertex appears in routing().logins, its token can be fetched by name, and a name from another provider is refused.

---

Generated by Claude Code