f678ed8299
Four chat_completions profiles (kimi-coding, deepseek, opencode-go's Kimi K2 and DeepSeek branches, actual) each hand-rolled the same extra_body.thinking / top-level reasoning_effort translation, and the copies had already drifted in small ways (kimi's `.get("enabled", True)`, deepseek's separate effort parsing). agent.reasoning_effort.thinking_toggle_extras is now the single implementation: the Moonshot default emits effort XOR toggle (both is an HTTP 400), and always_emit_toggle=True covers DeepSeek's contract where the toggle must ride on every request to dodge the reasoning_content echo trap. actual keeps its two contract-specific lines (reasoning_config None -> nothing; effort "none" -> disabled toggle plus reasoning_effort="none", which the relay accepts as a real level) and delegates the rest. ox_alpha_reasoning_extras moves alongside so opencode-free imports it like any other helper instead of reaching into the zen plugin's module through sys.modules and swallowing every exception into ({}, {}) - a failure there previously silently dropped the user's effort setting. No wire behavior changes; tests/plugins/model_providers/test_thinking_toggle_parity.py pins the XOR invariant across the matrix and zen/free parity.
104 lines
4.6 KiB
Python
104 lines
4.6 KiB
Python
"""OpenCode provider profiles (Zen + Go).
|
|
|
|
Both route api_mode per model in core; these profiles carry the
|
|
chat_completions reasoning translations (GLM-5.2, Kimi K2, DeepSeek, Ox Alpha).
|
|
"""
|
|
|
|
from typing import Any
|
|
|
|
from agent import reasoning_effort as re_
|
|
from hermes_cli import __version__ as _HERMES_VERSION
|
|
from providers import register_provider
|
|
from providers.base import ProviderProfile
|
|
|
|
# Attribution headers (same values as OpenRouter / Vercel / Fireworks); via
|
|
# default_headers so they survive model switches and credential rotation.
|
|
_ATTRIBUTION_HEADERS = {
|
|
"HTTP-Referer": "https://hermes-agent.nousresearch.com",
|
|
"X-Title": "Hermes Agent",
|
|
"User-Agent": f"HermesAgent/{_HERMES_VERSION}",
|
|
}
|
|
|
|
|
|
def _flat_model_name(model: str | None) -> str:
|
|
"""Bare OpenCode model ID, tolerating aggregator prefixes."""
|
|
return (model or "").strip().rsplit("/", 1)[-1].lower()
|
|
|
|
|
|
# Version-less DeepSeek ids that still carry the thinking/effort knobs on this wire: the retired
|
|
# ``deepseek-reasoner`` alias and the canonical ``deepseek-flash`` (2026-09 Flash refresh), for
|
|
# which the Go relay honours the same top-level ``reasoning_effort``/``thinking`` contract.
|
|
_THINKING_CAPABLE_IDS: frozenset[str] = frozenset({"deepseek-reasoner", "deepseek-flash"})
|
|
|
|
|
|
def _is_deepseek_thinking_model(model: str | None) -> bool:
|
|
m = _flat_model_name(model)
|
|
return (m.startswith("deepseek-v") and not m.startswith("deepseek-v3")) or m in _THINKING_CAPABLE_IDS
|
|
|
|
|
|
def _is_glm_5_2_model(model: str | None) -> bool:
|
|
"""GLM-5.2 across alias spellings (glm-5.2 / glm-5-2 / glm-5p2)."""
|
|
m = _flat_model_name(model)
|
|
return any(token in m for token in ("glm-5.2", "glm-5-2", "glm-5p2"))
|
|
|
|
|
|
class OpenCodeGoProfile(ProviderProfile):
|
|
"""OpenCode Go - model-specific reasoning controls."""
|
|
|
|
# The relay's default max_tokens (262144) exceeds what Xiaomi accepts for
|
|
# mimo-v2.5-pro and 400s; keys are normalized via _flat_model_name().
|
|
_MODEL_MAX_TOKENS: dict[str, int] = {"mimo-v2.5-pro": 131072}
|
|
|
|
def get_max_tokens(self, model: str | None) -> int | None:
|
|
cap = self._MODEL_MAX_TOKENS.get(_flat_model_name(model))
|
|
return self.default_max_tokens if cap is None else cap
|
|
|
|
def build_api_kwargs_extras(
|
|
self, *, reasoning_config: dict | None = None, model: str | None = None, **context
|
|
) -> tuple[dict[str, Any], dict[str, Any]]:
|
|
if _is_glm_5_2_model(model):
|
|
# Native reasoning_effort knob (high/max); server default when unset/disabled.
|
|
effort = re_.requested_effort(reasoning_config)
|
|
if effort is None or effort == "none":
|
|
return {}, {}
|
|
clamped = re_.clamp_effort(effort, re_.GLM52_EFFORTS, re_.GLM52_OVERRIDES)
|
|
return {}, {"reasoning_effort": clamped if clamped in re_.GLM52_EFFORTS else "high"}
|
|
if _flat_model_name(model).startswith("kimi-k2"):
|
|
if not isinstance(reasoning_config, dict):
|
|
return {}, {}
|
|
return re_.thinking_toggle_extras(reasoning_config, re_.KIMI_K2_EFFORTS)
|
|
if _is_deepseek_thinking_model(model):
|
|
return re_.thinking_toggle_extras(reasoning_config, re_.DEEPSEEK_V4_EFFORTS, re_.DEEPSEEK_V4_OVERRIDES)
|
|
return {}, {}
|
|
|
|
|
|
class OpenCodeZenProfile(ProviderProfile):
|
|
"""OpenCode Zen - model-specific reasoning controls."""
|
|
|
|
def build_api_kwargs_extras(
|
|
self, *, reasoning_config: dict | None = None, model: str | None = None, **context
|
|
) -> tuple[dict[str, Any], dict[str, Any]]:
|
|
return re_.ox_alpha_reasoning_extras(reasoning_config, model)
|
|
|
|
|
|
opencode_zen = OpenCodeZenProfile(
|
|
name="opencode-zen", aliases=("opencode", "opencode_zen", "zen"), env_vars=("OPENCODE_ZEN_API_KEY",),
|
|
base_url="https://opencode.ai/zen/v1", default_headers=dict(_ATTRIBUTION_HEADERS),
|
|
default_aux_model="gemini-3-flash",
|
|
)
|
|
|
|
opencode_go = OpenCodeGoProfile(
|
|
name="opencode-go", aliases=("opencode_go", "go", "opencode-go-sub"), env_vars=("OPENCODE_GO_API_KEY",),
|
|
base_url="https://opencode.ai/zen/go/v1", default_headers=dict(_ATTRIBUTION_HEADERS),
|
|
default_aux_model="glm-5",
|
|
# The Go relay's upstream validates tool content as a strict string: list-type tool
|
|
# content (native vision embeds) 422s with ``messages.N.tool.content.str Input should
|
|
# be a valid string`` (Console Go, #104731) or 400s ``text is not set`` (MiMo, #47026),
|
|
# and the rejected row stays in history so every later call dies too. Images in user
|
|
# messages are fine, so vision itself keeps working via the text-summary downgrade.
|
|
supports_vision_tool_messages=False,
|
|
)
|
|
|
|
register_provider(opencode_zen)
|
|
register_provider(opencode_go)
|