__init__.py 3.5 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778
  1. """Meta Model API (Muse Spark) provider profile — https://api.meta.ai/v1.
  2. Bundled from albertodepaola/hermes-meta-provider; rides entirely on
  3. ProviderProfile hooks (zero core edits). The reasoning dial is emitted as a
  4. top-level ``reasoning_effort`` kwarg — not ``extra_body.reasoning``, whose
  5. emission is gated by a core host allowlist a third-party plugin must not edit.
  6. """
  7. import os
  8. from typing import Any
  9. from agent.reasoning_effort import META_AI_EFFORTS, clamp_effort
  10. from providers import register_provider
  11. from providers.base import ProviderProfile
  12. class MetaAIProfile(ProviderProfile):
  13. """Meta Model API — top-level reasoning_effort, self-contained."""
  14. # Non-chat model prefixes excluded from the agent picker. The live
  15. # /v1/models catalog includes image-generation and transcription models
  16. # that are not suitable for agentic chat.
  17. _NON_CHAT_PREFIXES = ("muse-image-", "muse-voice-")
  18. def fetch_models(
  19. self,
  20. *,
  21. api_key: str | None = None,
  22. base_url: str | None = None,
  23. timeout: float = 8.0,
  24. ) -> list[str] | None:
  25. """Fetch and filter the live catalog, excluding non-chat models."""
  26. live = super().fetch_models(api_key=api_key, base_url=base_url, timeout=timeout)
  27. if live is None:
  28. return None
  29. return [
  30. m for m in live
  31. if not any(m.startswith(p) for p in self._NON_CHAT_PREFIXES)
  32. ]
  33. def build_api_kwargs_extras(
  34. self, *, reasoning_config: dict | None = None, supports_reasoning: bool = False, **context: Any
  35. ) -> tuple[dict[str, Any], dict[str, Any]]:
  36. """Ignores the core ``supports_reasoning`` gate (host-allowlist driven); Muse Spark always
  37. accepts ``reasoning_effort``. Muse 400s on ``none``: disabled/"none" -> ``minimal``
  38. (closest to off); unset/bespoke levels -> ``medium``."""
  39. rc = reasoning_config or {}
  40. effort = str(rc.get("effort") or "").strip().lower()
  41. if rc.get("enabled") is False or effort == "none":
  42. mapped = "minimal"
  43. else:
  44. clamped = clamp_effort(effort, META_AI_EFFORTS)
  45. mapped = clamped if clamped in META_AI_EFFORTS else "medium"
  46. return {}, {"reasoning_effort": mapped}
  47. meta_ai = MetaAIProfile(
  48. name="meta-ai", aliases=("meta", "muse", "muse-spark", "model-api", "msl"), display_name="Meta Model API",
  49. description="Meta Muse Spark family (Meta Superintelligence Labs)",
  50. signup_url="https://developer.meta.com/ai/",
  51. # MODEL_API_KEY is Meta's documented env var; the aliases are conveniences.
  52. env_vars=("MODEL_API_KEY", "META_API_KEY", "META_MODEL_API_KEY", "META_BASE_URL"),
  53. base_url=os.getenv("META_BASE_URL", "").strip() or "https://api.meta.ai/v1", auth_type="api_key",
  54. # Responses API engages Muse prompt caching (0 cached tokens on chat/completions vs
  55. # 93-99% hits on /v1/responses); the hook above still covers custom non-api.meta.ai base URLs.
  56. api_mode="codex_responses",
  57. # Natively multimodal, but only on user turns: an image envelope inside a role:tool
  58. # message 400s "content did not match any supported type".
  59. supports_vision=True, supports_vision_tool_messages=False,
  60. # See #101668.
  61. default_aux_model="muse-spark-1.2-contributor",
  62. # Muse spends completion budget on hidden reasoning first; low caps can finish with empty content.
  63. default_max_tokens=16384,
  64. # Single safety-net entry, shown only when the live /v1/models fetch fails.
  65. fallback_models=("muse-spark-1.2",),
  66. )
  67. register_provider(meta_ai)