One request. Clear boundaries.

Draft catalog · web port3003

GET /api/catalog/preview returns anonymous versioned reference metadata only. It accepts no query parameters; other methods return405. The response is no-store and grants no model invocation, price approval or tenant authority.

Separate from gateway /v1/models. Prices remain non-billable draft data, not accounting configuration.

Gateway · port4100

• GET /v1/models: project key with models:read.
• POST /v1/chat/completions: project key with inference:write.
• GET /healthz: public liveness, not readiness or available capacity.

Send a Bearer project key from a trusted backend. Never ship keys in browser JavaScript.

Control · port4000

Project reads, key metadata and audit are scoped by organization and project:

/v1/organizations/{organizationId}
/projects/{projectId}
/projects/{projectId}/api-keys
/projects/{projectId}/audit-events

Combine the organization prefix with one project suffix. User tokens and workload keys follow distinct authentication paths; live membership and scopes remain authoritative.

Discover your workspace

GET /v1/console/workspaces uses a signed user access token and returns organizations and projects visible to that identity. Workload keys cannot discover user workspaces.

Choose an explicit organization/project pair from the response before reading project resources. An empty result means no visible workspace; an unavailable service does not. Discovery metadata is a snapshot, not permission to perform a later action. The control service checks authority on each request.

Use secure credential injection in a trusted terminal or backend. Never paste an access token into browser JavaScript, command arguments or shell history.

Completion request

{
  "model": "pacinfrax/atlas-instruct-preview",
  "messages": [
    {
      "role": "user",
      "content": "Explain a project boundary."
    }
  ],
  "max_completion_tokens": 64,
  "stream": false
}

Supported message roles: system, developer, user, assistant. Messages contain text only. Unknown fields are rejected; tool calls, media and embeddings are not implemented in this contract.

Request bounds

• 1–256 messages; each text contains 1–32768 characters.
• Combined text at most131072 characters.
• Output limit1–4096; use either max_tokens or max_completion_tokens, never both.
• Temperature0–2; optional signed32-bit integer seed.

Model context limits add another check. Local token counting is approximate; do not treat it as a provider tokenizer guarantee.

Streaming safely

Set stream: true for SSE. Consume bounded chat chunks, terminal usage and the final [DONE] marker. EOF without a complete terminal sequence is not success.

Cancellation must close the response reader. It does not prove a remote GPU job was canceled or usage reconciled. Never automatically replay a POST after a timeout: a server-owned execution may already have occurred.

Diagnose without leaking

Gateway errors use an error object with type, code, message and request_id; HTTP status still matters. Store sanitized request IDs, not tokens, headers or prompts.

401 requires credential repair;403 requires permission review;429 requires admission/backoff handling. Unsupported models return404. Do not silently switch keys or models to bypass a denial.