src/inference/__init__.py

"""Inference backends — dispatches to OpenRouter (UI server) or Codex (local).

Backend selection (`active_backend`):
  * On a Hugging Face Space (``SPACE_ID``/``SPACE_HOST`` set) -> ALWAYS ``openrouter``.
    A Space can't hold codex's interactive OAuth and the codex CLI isn't installed there.
  * Otherwise -> env ``INFERENCE_BACKEND`` (``openrouter`` | ``codex``), default ``openrouter``.
    Local computing sets ``INFERENCE_BACKEND=codex`` in ``.env`` to bill the ChatGPT
    subscription via the codex CLI instead of the OpenRouter API.

Both backends expose the same surface, so callers (ui/cache.py, ui/app.py) are unchanged.
"""
from __future__ import annotations

import os

_BACKENDS = {"openrouter", "codex"}


def active_backend() -> str:
    """Resolve the inference backend for this process (see module docstring)."""
    if os.environ.get("SPACE_ID") or os.environ.get("SPACE_HOST"):
        return "openrouter"
    name = os.environ.get("INFERENCE_BACKEND", "openrouter").strip().lower()
    return name if name in _BACKENDS else "openrouter"


def backend_label() -> str:
    """Human label for the active backend (UI health line)."""
    return "Codex" if active_backend() == "codex" else "OpenRouter"


def _backend():
    if active_backend() == "codex":
        from inference import codex as module
    else:
        from inference import openrouter as module
    return module


def run_inference(
    system_prompt: str, user_prompt: str, *, model: str | None = None, timeout: float | None = None
) -> str:
    """Run one completion through the active backend. ``timeout`` is forwarded only when
    given, so each backend applies its own default (OpenRouter 180s, Codex 900s)."""
    kwargs: dict = {"model": model}
    if timeout is not None:
        kwargs["timeout"] = timeout
    return _backend().run_inference(system_prompt, user_prompt, **kwargs)


def available_models() -> list[str]:
    return _backend().available_models()


def default_model() -> str:
    return _backend().default_model()


def healthcheck() -> tuple[bool, str]:
    return _backend().healthcheck()


__all__ = [
    "active_backend",
    "backend_label",
    "available_models",
    "default_model",
    "healthcheck",
    "run_inference",
]