Token usage

Every ProviderResponse carries a Usage(input_tokens, output_tokens). Agent accumulates it across the whole run and exposes both the per-turn delta and the running total through the on_usage hook:

from pycodeloop.core.agent import Agent

agent = Agent(
    provider=provider,
    on_usage=lambda turn, total, elapsed: print(
        f"turn: {turn.input_tokens} in / {turn.output_tokens} out "
        f"(total: {total.input_tokens} in / {total.output_tokens} out) — {elapsed:.1f}s"
    ),
)

agent.run("...")
print(agent.usage)  # Usage(input_tokens=..., output_tokens=...) — cumulative for this Agent

agent.usage persists across multiple agent.run(...) calls on the same Agent instance — it's a running session total, not reset per call.

In the CLI

pycodeloop run and the chat print the running total after every turn:

🤖 1.2k in / 340 out