Decision models

About

The gateway has dedicated routes for structured decisions. These requests contain state or input and typed questions, not chat messages. The gateway sends the body to a provider that speaks its format and returns the provider’s JSON response. It does not translate System One requests into OpenAI Decisions requests or the reverse.

Decision calls use your organization’s provider credentials, or a key supplied for that call by an authenticated internal caller. They never fall back to a platform operator’s key or a license’s managed-service token. A Generic System One endpoint can be keyless when it is explicitly configured; known vendors still require your key.

When to use

• Route Jev or another System One model through the gateway rather than calling its upstream directly.
• Send an OpenAI Decisions request without converting it into a chat prompt.
• Use an organization-owned decision endpoint with the gateway’s access controls, limits and request logs.
• Route decision traffic through an outbound HTTP proxy while keeping existing chat connections direct.

For reusable Jev evaluations in the Future AGI UI, see Jev evaluators. You do not need to call these routes manually to use Jev in the evaluator editor or Instant eval.

How to

curl "$GATEWAY_URL/v1/decisions/models" \
  -H "Authorization: Bearer $AGENTCC_API_KEY"

curl "$GATEWAY_URL/v1/systemone" \
  -H "Authorization: Bearer $AGENTCC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "typesafe/jev-latest",
    "state": {"text": "The assistant answered the question and confirmed the next step."},
    "questions": {
      "q1": {
        "type": "score",
        "instructions": "Rate whether the assistant provided a clear next step.",
        "criteria": ["No clear next step", "Clear next step"]
      }
    }
  }'

Routes and formats

The two POST routes are not chat-completions routes. A System One-only provider cannot answer a chat request.

Supported dialects

For System One, the gateway identifies TypeSafe and Perplexity from the base URL’s host, then the provider ID. Another explicit endpoint uses the Generic dialect. A provider configured with the wrong api_format is refused rather than treated as a compatible decision provider.

Generic means a server that actually implements the System One body and response format. Merely serving OpenAI-compatible chat is not enough. It can include a compatible self-hosted server or a compatible endpoint under a longer base path, such as a Foundry resource path. The configured base path and API path prefix are used to build the decision URL.

Keyless Generic System One

An enabled organization provider can omit credentials only when it is a Generic System One provider with an explicit usable base URL and a model list. The endpoint may implement its own network access controls; keyless does not mean public access is safe.

This exception does not make TypeSafe, Perplexity or OpenAI keyless. A Generic endpoint is still subject to the organization’s URL checks and model access rules. See Private and local provider URLs.

Which key is used

Ordinary organization keys cannot send x-agentcc-provider-key. The internal header is removed before request headers are retained for logs or spans. Internal callers never borrow a config.yaml provider’s API key, even if it serves the same vendor. The authentication-disabled case is for a gateway whose operator is the customer, not a way for tenants to use a platform-owned key.

Outbound proxies and host checks

Every decision call honors HTTP_PROXY, HTTPS_PROXY and NO_PROXY in the gateway’s environment. This applies to both formats and to providers from config.yaml, organization providers, and internal calls carrying the customer’s key. Set these variables on the gateway process or service, not only on the Future AGI backend. The Helm chart’s global.proxy supplies them.

Chat providers still connect directly. A proxy that permits decision traffic does not by itself make chat traffic work behind a restricted firewall.

For an organization-owned endpoint, the gateway validates the destination even when a proxy is used. Private-address policy and the blocks on loopback, link-local and cloud metadata addresses still apply. DNS is resolved before each HTTP/1.1 request, including reuse of a pooled connection. A DNS failure or refused address fails closed rather than asking the proxy to bypass the check. Configure the proxy’s own destination policy too.

Decision calls never follow upstream redirects, so a redirect cannot carry the provider key or request content to another destination. Trusted operator-configured providers have different URL rules from organization providers; see Private and local provider URLs.

Errors

Errors use the gateway’s JSON error envelope. Useful codes include:

A decision request body is limited to the smaller of the gateway’s configured body limit and 32 MiB. Upstream response bodies are capped at 10 MiB. The gateway does not retry a decision automatically.

Next steps