Skip to main content
A model connection specifies wire protocol, base URL, exact model ID, and credential. A vendor name is not a protocol; one vendor may expose multiple interfaces. Use a management identity and an endpoint/model you are authorized to access. Console Infrastructure → LLM offers the same configure/verify/enable flow. The examples read deployment-specific values from private environment configuration.

Configure Chat Completions

The base can end in /v1; do not append /chat/completions. Responses should not append /responses. Configure the appropriate root for Messages. See protocol identifiers.

What connection validation actually sends

All three model connection schemas accept exactly base_url and model in config. The URL must use HTTPS without embedded username/password, query parameters or a fragment. The model ID is a nonempty string of at most 200 characters. Put credentials in the connection credential field, not in the URL or config. Each validation request has a 30-second timeout and rejects redirects. It checks protocol shape, not that the answer literally equals OK. Responses may pass validation with an incomplete response, so ready is not proof of task completeness or sufficient output budget. No validation request exercises streaming, Tool calls or multi-turn conversation. For an illustrative base https://model.example/service/v1/, Chat Completions targets https://model.example/service/v1/chat/completions. A root ending in an operation such as /chat/completions is not automatically corrected. Configure the service’s supported API root; do not guess a path from the vendor name. This example domain is not a usable endpoint. The creation snippets create a new connection every time. Persist the returned connection ID and use update/validate/enable for later changes. If creation’s response is lost, inspect existing records before rerunning the create snippet. A friendly display name is not a unique request key.

Complete a task with a recorded endpoint

For a concrete Chat Completions starting point, the November 18 matrix records https://api.minimaxi.com/v1, model MiniMax-M3, and built-in Pi 1.5.2 / OpenCode 1.2.27 through independent Python and TypeScript tasks. Supply your own authorized credential and these values to the configuration example only if this is your intended service. This is historical evidence, not a current account check. Save the enabled model connection ID, choose the exact Harness and a verified Sandbox combination, then use the first Session programs with that selection. Keep their persistent journal, artifact-content checks, close and retention checks intact. Run both clients independently. The November 20 first-Session evidence uses a different configured Messages endpoint and does not prove this Chat configuration was rerun by that tutorial.

Use and verify

Pass the connection ID as model_connection_id with a compatible Harness. Verification sends a bounded model request and can consume usage. It does not establish complete Agent Tool or continuation behavior. Original keys remain in the control plane. Guests receive bounded grants or native placeholders through protected access, not account keys in ordinary environment. Do not repair networking by putting the original key into prompts. Check URL, protocol, model ID, account permissions, and quota, then distinguish control-plane upstream access from Sandbox credential ingress. model_response_invalid means an unexpected validation response shape; HTTP 200 alone does not establish compatibility. Revalidate/re-enable after edits and account for pinned Session configuration. Complete a real task, required Tools/continuation/streaming, and cleanup. Historical model evidence helps select a starting combination but does not replace fresh evidence. Sources: Chat Completions adapter, Responses. Next: protocol compatibility, choose a combination, network ingress.