Record & Replay

Cassette wrappers let you record real LLM responses once and replay them deterministically — perfect for tests and notebooks that must run offline and fast.

Record

Wrap a real client; every response is written to a JSONL cassette:

from mmar_carl import (
    RecordingLLMClient, OpenAICompatibleClient, OpenAIClientConfig, ReasoningContext,
)

real = OpenAICompatibleClient(OpenAIClientConfig(
    base_url="https://openrouter.ai/api/v1", api_key="sk-or-v1-...", model="qwen/qwen3-coder",
))
rec = RecordingLLMClient(real, "cassette.jsonl", overwrite=True)

ctx = ReasoningContext(outer_context="...", api=rec)
await chain.execute_async(ctx)   # makes real calls, records them

Replay

Swap in the player — no API key, no network:

from mmar_carl import PlayingLLMClient

play = PlayingLLMClient("cassette.jsonl")
ctx = ReasoningContext(outer_context="...", api=play)
await chain.execute_async(ctx)   # reads from the cassette, ~instant

How matching works

The cassette key is sha256(method + prompt/messages + model + temperature). If a run asks for a key that isn’t in the cassette, PlayingLLMClient raises CassetteMissError (with a prompt preview) — so a drifted prompt fails loudly instead of silently calling a real API.

Client	Constructor
RecordingLLMClient	(real_client, cassette_path, overwrite=False)
PlayingLLMClient	(cassette_path, model=None, temperature=None)

See also

• LLM clients · Retries