hermes-agent/agent/chat_completion_headers.py

"""Non-streaming chat completions that still feed response headers to the credits capture."""

from __future__ import annotations

import logging

logger = logging.getLogger(__name__)


def _chat_completion_capturing_headers(agent, client, api_kwargs: dict):
    """Non-streaming ``chat.completions.create`` that also feeds the response headers to the
    rate-limit / credits capture, as ``_chat_stream_created`` does for streams.

    The parsed ``ChatCompletion`` drops headers, so without this a non-streaming turn never
    observes ``x-nous-credits-*`` (usage-window ``spend_source`` / gauges go stale). The SDK's
    ``with_raw_response`` returns the same request with headers attached; a client without it keeps
    the plain call. Usage-window plans
    only: every other provider, and a Nous session not yet known to be on a window plan (the
    session-start seed or a stream header establishes it), keep the plain call; headers are
    captured only when the response itself is the usage-window kind. A capture failure never
    fails the turn.
    """
    # `is True`: a Mock agent's invented attribute is not a window plan.
    state = getattr(agent, "_credits_state", None)
    if getattr(agent, "provider", "") != "nous" or getattr(state, "is_usage_window", False) is not True:
        return client.chat.completions.create(**api_kwargs)
    raw_api = getattr(getattr(getattr(client, "chat", None), "completions", None), "with_raw_response", None)
    if raw_api is None:
        return client.chat.completions.create(**api_kwargs)
    raw = raw_api.create(**api_kwargs)
    from agent.credits_tracker import response_is_usage_window
    if not response_is_usage_window(raw):
        return raw.parse()
    for capture in (agent._capture_rate_limits, agent._capture_credits,
                    agent._capture_nous_model_switch, agent._check_openrouter_cache_status):
        try:
            capture(raw)
        except Exception:
            logger.debug("non-streaming header capture failed", exc_info=True)
    return raw.parse()