Update engine constructors, token counters, and Responses settings to use one explicit API connection.
Inference engines and token counters now require a connection as their first argument. Settings factories resolve credentials and endpoints; payload constructors consume the result. There are no compatibility wrappers for the removed constructor options.
Previously, an engine could resolve credentials itself:
// Before
engine, err := openai.NewOpenAIEngine(s, openai.WithBearerTokenSource(source))
Resolve the binding at the source boundary, then pass it explicitly. These snippets assume imports and surrounding error handling; source is a host-owned token source, or nil for the selected settings key.
// After
c, err := connectionfactory.New(s, source)
if err != nil {
return err
}
engine, err := openai.NewOpenAIEngine(c, s)
Apply the same argument order to the other constructors:
responses, err := openai_responses.NewEngine(c, s)
claudeEngine, err := claude.NewClaudeEngine(c, s)
geminiEngine, err := gemini.NewGeminiEngine(c, s)
counter, err := openai_responses.NewTokenCounter(c, s)
claudeCounter, err := claude.NewTokenCounter(c, s)
Each line is an independent example: use a connection with the selected protocol's routes and authentication. Nil connections fail immediately. Keep observer and payload options on the engine; move custom authentication, headers, routes and HTTP policy into connection construction. Remove engine WithConnection, bearer options, and token-counter connection/auth options.
Existing generic factory callers can continue using factory.NewEngineFromSettings(s), or factory.NewStandardEngineFactory(factory.WithBearerTokenSource(source)). Factories resolve the connection before constructing the payload engine. A host that already owns a connection can still pass factory.WithConnection(c) to the factory.
Rename the protocol and its Go constant to open-responses / types.ApiTypeOpenResponses. The old openai-responses alias is removed. Responses no longer searches OpenAI or legacy slots for credentials, URLs, or outbound policy.
Declare the canonical pair, even when calling OpenAI itself:
inference_settings:
chat:
api_type: open-responses
engine: gpt-5-mini
api:
api_keys:
open-responses-api-key: <key for this endpoint>
base_urls:
open-responses-base-url: https://api.openai.com/v1
Move relevant policy entries to the open-responses key too. Keep credentials in your existing trusted configuration source; do not put real keys into committed examples. Ordinary openai chat configuration is unchanged. Pi's external catalog protocol name openai-responses is still supported and translates to the canonical internal protocol; it is not a settings fallback.
Remove WithRequestTransport. For Codex, construct its canonical connection directly:
c, err := openaicodex.NewConnection(source, openaicodex.Options{}, hostClient)
if err != nil {
return err
}
engine, err := openai_responses.NewEngine(c, s)
Other integrations use connection.New with explicit routes and authentication. Pi performs this binding itself. Embedding constructors already take connections; this change does not redesign direct rerank or decision convenience constructors.
Search callers for the removed options and alias, compile the application, then run its ordinary tests. Verify at least one configured request per protocol with a recording transport. Confirm the intended endpoint receives authentication and a foreign redirect is refused. Run socket-based tests in an environment that permits local test servers.
| Problem | Cause | Fix |
|---|---|---|
| Missing constructor argument | Old direct constructor call | Pass the resolved connection first |
| Missing Responses key or URL | Reliance on another settings slot | Declare the complete canonical pair |
| Unknown operation | Connection lacks a protocol route | Register the payload client's routes |
geppetto-credential-scope: connection ownership, route tables and trust limits.use-renewable-bearer-credentials: host-owned credential renewal.