Providers

The LLM boundary. Module: regista/providers/.

One protocol, four implementations. An adapter's whole job is normalization — wire format, tool-call shape, usage extraction, retries:

class Provider(Protocol):
    name: str    # recorded in session.start
    model: str   # chosen where the provider is constructed — never defaulted
    async def complete(self, request: ModelRequest) -> ModelResponse: ...
    def stream(self, request: ModelRequest) -> AsyncIterator[ProviderDelta | ModelResponse]: ...

regista's internal message model is a superset of the Anthropic shape (role + typed content blocks, with tool_use/tool_result/thinking first-class). Rich→flat translation is easy; flat→rich is lossy — so adapters translate into the rich shape, never out of it.

AnthropicProvider

Official SDK. Blocks map 1:1; the system prompt carries a prompt-cache breakpoint; cache read/write tokens land in Usage and flow into cost; thinking signatures round-trip. SDK errors become ProviderError with an honest retryable flag.

AnthropicProvider("claude-sonnet-4-6")           # reads ANTHROPIC_API_KEY

OpenAICompatProvider

Raw httpx against any /v1/chat/completions endpoint — OpenAI, Ollama, vLLM, LM Studio — with local exponential backoff. The lossy direction, made explicit: thinking blocks drop on the way out; each tool result becomes its own role="tool" message; invalid JSON tool arguments become {"raw_arguments": ...} so dispatch fails as error-data instead of killing the session.

OpenAICompatProvider("gpt-4o", api_key=...)
OpenAICompatProvider("llama3.1", base_url="http://localhost:11434/v1")   # no key needed

FakeProvider

Scripted responses; records every request it receives. Public API — it's how you unit-test your own agents and how contributors test loop changes. ~90% of regista's own suite runs on it.

ReplayProvider

Serves a recorded trace, hash-verifying each request — see deterministic replay. Because it's just another provider, replay needed no special support anywhere else in the harness.