Harness gateway
Model discovery, protocol support, output limits, and streaming.
Use the endpoint, defaultModel, and authorized models returned by team resolution. Cloud requests use a public ID returned there, such as the fictional s46/example-published-model; the gateway maps it to the signed backend alias. Do not send backend aliases as public model IDs or hard-code an identity-only candidate. See model assurance status for the current Devstral and DeepSeek state.
Model POST routes require a bearer token in production and whenever registry verification is enabled, including development. The gateway enforces independent current registry verification and the caller's team policy. Explicitly sync-disabled development and default airplane mode retain local/mock behavior.
Chat completions
Omit model to use the team default. For a complete discovery-driven call, see the quickstart. OpenAI SDK clients can use the returned gateway endpoint plus /v1 as their base URL and select an ID from the team's authorized list.
Responses & Codex
Both routes accept text input, message history, function calls/results, function tools, and text/JSON output formats. Send full history on each request: stored responses, nonempty previous_response_id, background execution, automatic truncation, and non-null reasoning are unsupported. Multimodal input and built-in/server tools are unsupported.
Anthropic Messages
Supports text, system prompts, function tools, tool_choice, tool_use, and text tool_result blocks. Images, documents, thinking blocks/signatures, server tools, and context-management features are unsupported.
These are compatibility subsets, not complete vendor APIs. Unsupported Responses or Anthropic features return 400 unsupported_feature.
Output limits
The verified cloud cap is the smaller positive catalog value of maxOutputTokens and contextWindow, or the sole positive value. Omitted output limits default to that cap; excessive values are rejected, not clamped. Without either limit, the gateway supplies no catalog-derived default.
| Protocol | Accepted output-limit fields |
|---|---|
| Chat Completions | max_tokens, max_completion_tokens |
| Responses / Codex | max_output_tokens, max_tokens, max_completion_tokens |
| Anthropic Messages | max_tokens |
Limits must be positive integers no greater than 2,147,483,647 and the effective catalog limit. Multiple supplied limits must agree. Invalid, conflicting, protocol-inappropriate limits and cloud n_predict return 400 invalid_request. Runtime reasoning metadata does not imply that the compatibility adapters accept reasoning controls.
Streaming
Set stream: true for text/event-stream:
- Chat emits completion chunks and
[DONE]on success. - Responses/Codex emit Responses-style events, ending with
response.completedorresponse.incompletewhen output is limited. - Anthropic emits Messages-style events through
message_stopon success.
An upstream failure before streaming returns 502 upstream_error. After headers are sent, inspect stream errors rather than HTTP status: chat emits an error object, Responses emits response.failed, and Anthropic emits an error event. Do not treat a truncated stream as successful completion.
Model discovery
Discovery does not require authentication. When verification is enabled, lists contain independently verified servable catalog entries; they are not the caller's team allow-list. Use team resolution for authorized selection.
OpenAI/Codex return { object: "list", data: [...] }, with each entry's catalog details under metadata, including format, runtime, contextWindow, and maxOutputTokens. Anthropic returns { data: [{ id, type: "model", display_name }], has_more: false } without runtime metadata. See Models and the OpenAPI schemas.
Policy & availability failures
| Status | Code | Meaning |
|---|---|---|
403 | model_not_allowed | Model is not currently authorized and available for this team. |
503 | model_registry_unavailable | Independent registry verification is unavailable or no longer current. |
503 | model_unavailable | No configured serving upstream is available. |
502 | upstream_error | Upstream request or response failed. |
See Errors for authentication, validation, and rate-limit failures. /healthz and /readyz are described in Operations.
Airplane-mode workers
Local worker discovery uses ID local-llamacpp and states not_configured, unavailable, unhealthy, model_missing, or ready. Local GGUF models such as s46/devstral-small-2-24b use llama.cpp; native safetensors cloud packages are a separate serving path. Local model IDs map to configured backend aliases, and local output limits are clamped to local configuration.
Upstream behavior
Verified cloud Chat, Responses, Codex, and Anthropic calls normalize through one Chat Completions upstream. Dedicated Anthropic/Codex passthroughs and arbitrary-model proxy/mock behavior are only used when verification is disabled. Caller authorization headers and cookies are not forwarded upstream.
For upstream settings and policy setup, see the environment reference and policy runbook.