__init__.py 6.5 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138
  1. """OpenCode provider profiles (Zen + Go).
  2. Both route api_mode per model in core; these profiles carry the
  3. chat_completions reasoning translations (GLM-5.2, Kimi K2, DeepSeek, Ox Alpha).
  4. """
  5. from typing import Any
  6. from agent import reasoning_effort as re_
  7. from hermes_cli.version_info import hermes_user_agent
  8. from providers import register_provider
  9. from providers.base import LazyHeaders, ProviderProfile
  10. from datetime import UTC
  11. # Attribution headers (same values as OpenRouter / Vercel / Fireworks); via
  12. # default_headers so they survive model switches and credential rotation.
  13. _ATTRIBUTION_HEADERS = {
  14. "HTTP-Referer": "https://hermes-agent.nousresearch.com",
  15. "X-Title": "Hermes Agent",
  16. "User-Agent": lambda: hermes_user_agent("HermesAgent"),
  17. }
  18. def _flat_model_name(model: str | None) -> str:
  19. """Bare OpenCode model ID, tolerating aggregator prefixes."""
  20. return (model or "").strip().rsplit("/", 1)[-1].lower()
  21. # Version-less DeepSeek ids that still carry the thinking/effort knobs on this wire: the retired
  22. # ``deepseek-reasoner`` alias and the canonical ``deepseek-flash`` (2026-09 Flash refresh), for
  23. # which the Go relay honours the same top-level ``reasoning_effort``/``thinking`` contract.
  24. _THINKING_CAPABLE_IDS: frozenset[str] = frozenset({"deepseek-reasoner", "deepseek-flash"})
  25. def _is_deepseek_thinking_model(model: str | None) -> bool:
  26. m = _flat_model_name(model)
  27. return (m.startswith("deepseek-v") and not m.startswith("deepseek-v3")) or m in _THINKING_CAPABLE_IDS
  28. def _is_glm_5_2_model(model: str | None) -> bool:
  29. """GLM-5.2 across alias spellings (glm-5.2 / glm-5-2 / glm-5p2)."""
  30. m = _flat_model_name(model)
  31. return any(token in m for token in ("glm-5.2", "glm-5-2", "glm-5p2"))
  32. class OpenCodeGoProfile(ProviderProfile):
  33. """OpenCode Go - model-specific reasoning controls."""
  34. # The relay's default max_tokens (262144) exceeds what Xiaomi accepts for
  35. # mimo-v2.5-pro and 400s; keys are normalized via _flat_model_name().
  36. _MODEL_MAX_TOKENS: dict[str, int] = {"mimo-v2.5-pro": 131072}
  37. def get_max_tokens(self, model: str | None) -> int | None:
  38. cap = self._MODEL_MAX_TOKENS.get(_flat_model_name(model))
  39. return self.default_max_tokens if cap is None else cap
  40. def fetch_account_usage(self, *, base_url: str | None = None, api_key: str | None = None):
  41. """Go subscription windows for /usage via ``/zen/go/v1/usage`` (anomalyco/opencode#16513).
  42. Percent-based payload ``{"usage": {"rolling"|"weekly"|"monthly": {"percent", "resetsAt"}}}``.
  43. Literal endpoint, NOT the runtime base_url: that one loses its /v1 suffix in
  44. anthropic_messages mode and /usage only exists under /v1.
  45. """
  46. from datetime import datetime, timezone
  47. import httpx
  48. from agent.account_usage import AccountUsageSnapshot, AccountUsageWindow
  49. from hermes_cli.runtime_provider import resolve_runtime_provider
  50. runtime = resolve_runtime_provider(requested=self.name, explicit_base_url=base_url, explicit_api_key=api_key)
  51. token = str(runtime.get("api_key", "") or "").strip()
  52. if not token:
  53. return None
  54. with httpx.Client(timeout=10.0) as client:
  55. response = client.get("https://opencode.ai/zen/go/v1/usage",
  56. headers={"Authorization": f"Bearer {token}", "Accept": "application/json"})
  57. response.raise_for_status()
  58. usage = (response.json() or {}).get("usage") or {}
  59. windows = []
  60. for key, label in (("rolling", "Rolling window"), ("weekly", "Weekly"), ("monthly", "Monthly")):
  61. window = usage.get(key) or {}
  62. if window.get("percent") is None:
  63. continue
  64. reset_raw = str(window.get("resetsAt") or "").replace("Z", "+00:00")
  65. reset_at = datetime.fromisoformat(reset_raw) if reset_raw else None
  66. windows.append(AccountUsageWindow(label=label, used_percent=float(window["percent"]), reset_at=reset_at))
  67. return AccountUsageSnapshot(provider=self.name, source="go_usage_api",
  68. fetched_at=datetime.now(UTC), windows=tuple(windows))
  69. def build_api_kwargs_extras(
  70. self, *, reasoning_config: dict | None = None, model: str | None = None, **context
  71. ) -> tuple[dict[str, Any], dict[str, Any]]:
  72. if _is_glm_5_2_model(model):
  73. # Native reasoning_effort knob (high/max); server default when unset/disabled.
  74. effort = re_.requested_effort(reasoning_config)
  75. if effort is None or effort == "none":
  76. return {}, {}
  77. clamped = re_.clamp_effort(effort, re_.GLM52_EFFORTS, re_.GLM52_OVERRIDES)
  78. return {}, {"reasoning_effort": clamped if clamped in re_.GLM52_EFFORTS else "high"}
  79. if _flat_model_name(model).startswith("kimi-k2"):
  80. if not isinstance(reasoning_config, dict):
  81. return {}, {}
  82. return re_.thinking_toggle_extras(reasoning_config, re_.KIMI_K2_EFFORTS)
  83. if _is_deepseek_thinking_model(model):
  84. return re_.thinking_toggle_extras(reasoning_config, re_.DEEPSEEK_V4_EFFORTS, re_.DEEPSEEK_V4_OVERRIDES)
  85. return {}, {}
  86. class OpenCodeZenProfile(ProviderProfile):
  87. """OpenCode Zen - model-specific reasoning controls."""
  88. def build_api_kwargs_extras(
  89. self, *, reasoning_config: dict | None = None, model: str | None = None, **context
  90. ) -> tuple[dict[str, Any], dict[str, Any]]:
  91. return re_.ox_alpha_reasoning_extras(reasoning_config, model)
  92. opencode_zen = OpenCodeZenProfile(
  93. name="opencode-zen", aliases=("opencode", "opencode_zen", "zen"), env_vars=("OPENCODE_ZEN_API_KEY",),
  94. base_url="https://opencode.ai/zen/v1", default_headers=LazyHeaders(_ATTRIBUTION_HEADERS),
  95. default_aux_model="gemini-3-flash",
  96. )
  97. opencode_go = OpenCodeGoProfile(
  98. name="opencode-go", aliases=("opencode_go", "go", "opencode-go-sub"), env_vars=("OPENCODE_GO_API_KEY",),
  99. base_url="https://opencode.ai/zen/go/v1", default_headers=LazyHeaders(_ATTRIBUTION_HEADERS),
  100. default_aux_model="glm-5",
  101. # The Go relay's upstream validates tool content as a strict string: list-type tool
  102. # content (native vision embeds) 422s with ``messages.N.tool.content.str Input should
  103. # be a valid string`` (Console Go, #104731) or 400s ``text is not set`` (MiMo, #47026),
  104. # and the rejected row stays in history so every later call dies too. Images in user
  105. # messages are fine, so vision itself keeps working via the text-summary downgrade.
  106. supports_vision_tool_messages=False,
  107. )
  108. register_provider(opencode_zen)
  109. register_provider(opencode_go)