Models

OpenAIModel speaks the OpenAI Responses API. Wrappers nest: Retrying retries ModelUnavailable, fallback tries the next model, Metered refuses the next call once receipts pass a cap.

from agentsparty.llm.compose import Metered, Retrying, Unavailable, fallback

primary = Unavailable('primary down')
backup = Unavailable('backup down')
model = fallback(
    Retrying(primary, attempts=2),
    Metered(backup, tokens=8_000),
)
print(type(model).__name__)

Each wrapper is itself a LanguageModel. The order is the nesting you wrote: there is no registry and no hook list.

OpenAIModel

Construct the adapter with a model id and an AsyncOpenAI client. Pass max_retries=0 and a finite timeout on the client so retry policy lives in Retrying and a hung transport cannot stall the session. See security.

from openai import AsyncOpenAI
from agentsparty import OpenAIModel
from agentsparty.llm.compose import Metered, Retrying, fallback

client = AsyncOpenAI(max_retries=0, timeout=30.0)
primary = OpenAIModel('gpt-5.6-luna', client)
backup = OpenAIModel('gpt-5.6-luna', client)
model = fallback(
    Retrying(primary, attempts=2),
    Metered(backup, tokens=8_000),
)
print(type(model).__name__)

An Agent takes the resulting model. The protocol never sees the wrapper stack; a transformer maps a request to an answer to that same request.

Wrappers

Unavailable is the honest value for "no model configured": putting it in front of a fallback chain changes nothing.

A worked composition lives in examples/online/model_composition.py. The participant that consumes a model is an Agent; see participants and roles. See agentsparty.llm.openai.OpenAIModel.