Streaming

Pass on_text_delta to Agent and it's called with each chunk of text as the model generates it, instead of only getting the assembled reply at the end:

from pycodeloop.core.agent import Agent
from pycodeloop.providers import GenericProvider

agent = Agent(
    provider=GenericProvider.from_json("templates/openai.json"),
    on_text_delta=lambda chunk: print(chunk, end="", flush=True),
)

agent.run("explain what this codebase does")

GenericProvider streams real SSE chunks for the default OpenAI chat-completions request/response shape — a config like templates/openai.json or templates/ollama.json that doesn't override response_shape/response_paths. A config with a custom response shape (e.g. templates/anthropic.json, which sets response_shape: "anthropic") still works with on_text_delta, but delivers the full text as a single chunk once the request completes rather than token-by-token. If on_text_delta is None (the default), providers skip streaming entirely and return the full response in one call.

In the CLI

pycodeloop run and the chat always stream — the reply prints as it arrives, matching a terminal coding agent's feel rather than waiting on a spinner.