Fix OpenClaw Google model API when routing through cllama

[mostlydev]

Problem

Pods can now declare direct Gemini models like google/gemini-3-flash-preview behind cllama, but the OpenClaw driver still compiles models.providers.google.api as google-generative-ai whenever the provider prefix is google.

That is correct for direct-to-Google runtime wiring, but it is wrong when cllama is enabled. In the cllama case the provider base URL is rewritten to http://cllama:8080/v1, and the proxy only exposes Google through the OpenAI-compatible /v1/chat/completions surface.

Reproduction

1. Configure an OpenClaw service with:
   • x-claw.cllama: passthrough
   • x-claw.models.primary: google/gemini-3-flash-preview
2. Run claw up -d.
3. Observe generated OpenClaw config under models.providers.google:
   • baseUrl: http://cllama:8080/v1
   • api: google-generative-ai
4. Send a message to the agent.
5. The request fails with 404 page not found from the provider path.

Direct probes against Google's OpenAI-compatible endpoint succeed with the same key and model. The failure is specific to the runner->cllama request shape.

Root cause

internal/driver/openclaw/config.go calls defaultModelAPIForProvider(provider) even inside the cllama rewrite branch. For google, that returns google-generative-ai, which causes OpenClaw to speak Google-native request semantics to a cllama endpoint that expects OpenAI-compatible semantics.

Expected behavior

When OpenClaw rewrites a provider to cllama, the compiled api should match the proxy surface, not the upstream vendor-native surface. For google/* models behind cllama, that means openai-completions.

Scope

• Fix OpenClaw config generation for google/* + cllama.
• Add regression coverage in internal/driver/openclaw/config_test.go.
• Check whether any other driver has the same provider-API rewrite bug before closing.

[mostlydev]

Starting implementation review on branch . I’m validating the OpenClaw cllama rewrite path, checking other drivers for the same provider-API mismatch, and will open a PR once the focused tests pass.

[mostlydev]

Ready for review in PR #129.

What changed:

• OpenClaw now uses the cllama proxy API surface for google/* models when cllama is enabled.
• Direct non-cllama Google behavior remains unchanged.
• Regression tests cover both the proxied and direct cases.

Validation:

• go test -run 'TestGenerateConfig(CllamaGoogleUsesOpenAICompletions|DirectGoogleKeepsNativeAPI|CllamaRewritesProviderBaseURL|CllamaRewritesAllModelProviders)' -count=1 ./internal/driver/openclaw
• go test -count=1 ./internal/driver/openclaw
• go test -count=1 ./internal/driver/...

Audit note: the same provider-API mismatch does not exist in the other cllama-aware drivers.