Add a provider

gmat-copilot ships adapters for Anthropic, OpenAI, Ollama, and GitHub Models, but the provider abstraction is open: any backend that can turn a prompt into text can generate drafts. A provider is just an object matching one protocol — there is nothing to subclass.

Implement the protocol

A provider needs a name, a reachable() check, and a complete() that returns a Completion:

from gmat_copilot.providers import Completion


class MyProvider:
    """Generate through some other backend."""

    name = "myco"

    def reachable(self) -> bool:
        # True when a call could succeed now — e.g. a credential or host is configured.
        return True

    def complete(
        self,
        prompt: str,
        *,
        model: str,
        temperature: float = 0.0,
        max_tokens: int = 1024,
    ) -> Completion:
        text = my_backend_call(prompt, model=model, temperature=temperature, max_tokens=max_tokens)
        return Completion(text=text, provider=self.name, model=model, usage={})

Provider is a runtime-checkable protocol, so matching the shape is enough — isinstance(MyProvider(), Provider) is True without inheritance. Populate usage with whatever integer token counts your backend reports (it may be left empty).

Use it

Pass the provider straight to draft. When you supply provider, the model argument is the bare model name handed to it (no provider: prefix):

from gmat_copilot import draft

result = draft(
    "a 500 km circular Earth orbit at 51.6 degrees; propagate one day and report altitude",
    provider=MyProvider(),
    model="my-model-name",
)

print(result.provider, result.model)   # "myco" "my-model-name"
print(result.lint.clean)

Everything downstream is unchanged: the draft is still retrieval-grounded, validated by the lint gate, and returned as a CopilotResult.

Notes

• The four built-in selectors (anthropic:, openai:, ollama:, github:) are what the CLI resolves; a custom provider is a library-level extension you wire in through the provider= argument.
• Raise ProviderError for a missing credential or an unreachable backend so failures surface clearly, consistent with the built-in adapters.
• To record a custom provider's output for deterministic replay, wrap it in a RecordingProvider; see the provider surface and the evaluation suite.