fix: Google/Gemini provider support for openai-completions

[okuyam2y]

Problem

When using Gemini models via Google's OpenAI-compatible endpoint (generativelanguage.googleapis.com), three issues prevent normal operation in openai-completions.ts:

1. 400 errors from unsupported params

Google's endpoint rejects store and stream_options parameters. These are sent because Google is not in the isNonStandard list in detectCompat().

Fix: Add isGoogle detection and include it in isNonStandard.

2. Multi-turn tool calling broken with Gemini 2.5

Gemini 2.5 Flash/Pro use extra_content.google.thought_signature on tool calls for multi-turn continuity. Currently not captured from streaming response nor echoed back in convertMessages(), breaking the second tool call turn.

Fix: Capture from choice.delta.tool_calls[].extra_content.google.thought_signature during streaming, echo back as extra_content.google in assistant messages.

3. Crash on unknown finish_reason

Gemini returns MALFORMED_FUNCTION_CALL as finish_reason, which throws in the exhaustive switch in mapStopReason().

Fix: Return "error" from the default case instead of throwing.

---

Happy to submit a PR once approved.

[ntcho]

Running into the 400 error too for endpoint https://generativelanguage.googleapis.com/v1beta/openai. Adding 1 solved it for me for now, but it'd be great to have this fixed.

[okuyam2y]

Thanks for confirming! Glad the workaround works for you too.

[ntcho]

I ended up patching the model-registry. Kinda hacky but this made it work much better. From a gist I wrote earlier today:

~/node_modules/@mariozechner/pi-coding-agent/dist/core/model-registry.js

// Separate providers into "full replacement" (has models) vs "override-only" (no models)
const replacedProviders = new Set();
const overrides = new Map();
for (const [providerName, providerConfig] of Object.entries(config.providers)) {

+    if (providerConfig.baseUrl.includes("generativelanguage.googleapis.com/v1beta/openai")) {
+        providerConfig.baseUrl = providerConfig.baseUrl.replace(/\/openai/, '') // remove OpenAI compat
+        providerConfig.api = "google-generative-ai"; // override api type
+    }

    if (providerConfig.models && providerConfig.models.length > 0) {
        // Has custom models -> full replacement
        replacedProviders.add(providerName);
    }
    ...

Note

This patch will make Pi to use google-generative-ai with generativelanguage.googleapis.com/v1beta, instead of openai-completions with generativelanguage.googleapis.com/v1beta/openai which has a lot of quirks that will result in 400 errors.

[okuyam2y]

Nice find! Switching to the native Google API at the model-registry level is much cleaner than my workaround (appending /1 to the endpoint). Going to try switching to your approach.