GLM-5 and Kimi K2 reasoning parameter handling

Fixes on record

Merged to main. A stamp counts the issues a fix is linked to.

• #58884 feat(providers): GLM-5.2 native reasoning_effort controls (salvage #51108)

  merged to main

  files: plugins/model-providers/opencode-zen/__init__.py, plugins/model-providers/zai/__init__.py, tests/plugins/model_providers/test_opencode_go_profile.py, tests/plugins/model_providers/test_zai_profile.py

  Show the diff

  • plugins/model-providers/opencode-zen/__init__.py

    @@ -31,6 +31,12 @@ def _is_deepseek_thinking_model(model: str | None) -> bool:
         return m == "deepseek-reasoner"
     
     
    +def _is_glm_5_2_model(model: str | None) -> bool:
    +    """Detect 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."""
     
    @@ -55,6 +61,21 @@ def build_api_kwargs_extras(
             extra_body: dict[str, Any] = {}
             top_level: dict[str, Any] = {}
     
    +        if _is_glm_5_2_model(model):
    +            # GLM-5.2 on OpenCode Go uses its native OpenAI-compatible
    +            # reasoning_effort knob, which has exactly two enabled levels:
    +            # high and max. Map Hermes' richer scale onto those; leave the
    +            # server default alone when reasoning is disabled or unset.
    +            if not isinstance(reasoning_config, dict):
    +                return extra_body, top_level
    +            if reasoning_config.get("enabled") is False:
    +                return extra_body, top_level
    +            effort = (reasoning_config.get("effort") or "").strip().lower()
    +            if not effort or effort == "none":
    +                return extra_body, top_level
    +            top_level["reasoning_effort"] = "max" if effort in {"xhigh", "max"} else "high"
    +            return extra_body, top_level
    +
             if _is_kimi_k2_model(model):
                 # Kimi K2 on OpenCode Go uses Moonshot's native wire shape:
                 # extra_body.thinking (binary toggle) + top-level reasoning_effort

  • plugins/model-providers/zai/__init__.py

    @@ -16,6 +16,12 @@
     is omitted so the server default applies, matching prior behavior.  GLM
     models before 4.5 (e.g. ``glm-4-9b``) don't accept ``thinking`` and are left
     untouched.
    +
    +GLM-5.2 additionally exposes a native ``reasoning_effort`` knob with exactly
    +two enabled levels — ``high`` and ``max`` — on the OpenAI-compatible endpoint
    +(per Z.AI / BigModel docs).  Hermes' richer effort scale is collapsed onto
    +those two so the user's effort preference actually reaches the model instead
    +of being silently dropped.
     """
     
     from __future__ import annotations
    @@ -40,8 +46,44 @@ def _model_supports_thinking(model: str | None) -> bool:
         return (major, minor) >= (4, 5)
     
     
    +def _is_glm_5_2(model: str | None) -> bool:
    +    """Detect GLM-5.2 across the alias spellings providers use.
    +
    +    Covers the canonical ``glm-5.2`` plus the ``glm-5-2`` / ``glm-5p2``
    +    variants seen on relays (Fireworks ``glm-5p2``, etc.) and any
    +    vendor-prefixed form (``z-ai/glm-5.2``, ``zai-org-glm-5-2``).
    +    """
    +    m = (model or "").strip().lower()
    +    if not m:
    +        return False
    +    return any(token in m for token in ("glm-5.2", "glm-5-2", "glm-5p2"))
    +
    +
    +def _glm_5_2_reasoning_effort(reasoning_config: dict | None) -> str | None:
    +    """Map Hermes reasoning effort onto GLM-5.2's native ``high``/``max``.
    +
    +    GLM-5.2 only supports two enabled effort levels. ``xhigh``/``max``
    +    request the top tier; everything else that is enabled requests ``high``
    +    (its minimum thinking level). When reasoning is explicitly disabled, or
    +    no effort preference is supplied, the server default is left untouched.
    +    """
    +    if not isinstance(reasoning_config, dict):
    +        return None
    +    if reasoning_config.get("enabled") is False:
    +        return None
    +
    +    effort = (reasoning_config.get("effort") or "").strip().lower()
    +    if not effort or effort == "none":
    +        return None
    +
    +    if effort in {"xhigh", "max"}:
    +        return "max"
    +    # low / medium / minimal / high all clamp to GLM-5.2's minimum: high.
    +    return "high"
    +
    +
     class ZaiProfile(ProviderProfile):
    -    """Z.AI / GLM — extra_body.thinking enabled/disabled."""
    +    """Z.AI / GLM — extra_body.thinking on/off + GLM-5.2 reasoning_effort."""
     
         def build_api_kwargs_extras(
             self, *, reasoning_config: dict | None = None, model: str | None = None, **context
    @@ -49,7 +91,7 @@ def build_api_kwargs_extras(
             extra_body: dict[str, Any] = {}
             top_level: dict[str, Any] = {}
     
    -        if not _model_supports_thinking(model):
    +        if not _model_supports_thinking(model) and not _is_glm_5_2(model):
                 return extra_body, top_level
     
             # Only emit when the user expressed a preference; omitting the field
    @@ -58,6 +100,11 @@ def build_api_kwargs_extras(
                 enabled = reasoning_config.get("enabled") is not False
                 extra_body["thinking"] = {"type": "enabled" if enabled else "disabled"}
     
    +        if _is_glm_5_2(model):
    +            effort = _glm_5_2_reasoning_effort(reasoning_config)
    +            if effort is not None:
    +                top_level["reasoning_effort"] = effort
    +
             return extra_body, top_level
     
     

  • tests/plugins/model_providers/test_opencode_go_profile.py

    @@ -122,13 +122,68 @@ def test_xhigh_and_max_normalize_to_max(self, opencode_go_profile):
                 assert top_level == {"reasoning_effort": "max"}
     
     
    +class TestOpenCodeGoGLM52Reasoning:
    +    """GLM-5.2 uses its native high/max reasoning_effort knob on OpenCode Go."""
    +
    +    def test_high_maps_to_high(self, opencode_go_profile):
    +        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True, "effort": "high"},
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {}
    +        assert top_level == {"reasoning_effort": "high"}
    +
    +    def test_low_and_medium_clamp_up_to_high(self, opencode_go_profile):
    +        for effort in ("low", "medium", "minimal"):
    +            extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
    +                reasoning_config={"enabled": True, "effort": effort},
    +                model="glm-5.2",
    +            )
    +            assert extra_body == {}
    +            assert top_level == {"reasoning_effort": "high"}
    +
    +    def test_xhigh_and_max_map_to_max(self, opencode_go_profile):
    +        for effort in ("xhigh", "max"):
    +            extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
    +                reasoning_config={"enabled": True, "effort": effort},
    +                model="z-ai/glm-5.2",
    +            )
    +            assert extra_body == {}
    +            assert top_level == {"reasoning_effort": "max"}
    +
    +    def test_disabled_leaves_server_default(self, opencode_go_profile):
    +        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": False, "effort": "high"},
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {}
    +        assert top_level == {}
    +
    +    def test_no_config_leaves_server_default(self, opencode_go_profile):
    +        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
    +            reasoning_config=None,
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {}
    +        assert top_level == {}
    +
    +    @pytest.mark.parametrize("model", ["glm-5-2", "glm-5p2"])
    +    def test_alias_spellings_recognized(self, opencode_go_profile, model):
    +        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True, "effort": "max"},
    +            model=model,
    +        )
    +        assert top_level == {"reasoning_effort": "max"}
    +
    +
     class TestOpenCodeGoModelGating:
    -    """Other OpenCode Go models must not receive Kimi/DeepSeek controls."""
    +    """Other OpenCode Go models must not receive Kimi/DeepSeek/GLM controls."""
     
         @pytest.mark.parametrize(
             "model",
             [
                 "glm-5.1",
    +            "glm-5",
                 "qwen3.6-plus",
                 "minimax-m2.7",
                 "deepseek-v3.1",

  • tests/plugins/model_providers/test_zai_profile.py

    @@ -1,4 +1,4 @@
    -"""Unit tests for the Z.AI / GLM provider profile's thinking-mode wiring.
    +"""Unit tests for the Z.AI / GLM provider profile's reasoning wiring.
     
     Z.AI's GLM-4.5-and-later chat models default to thinking-mode ON when the
     request omits ``thinking``.  Before the profile emitted the parameter,
    @@ -6,6 +6,10 @@
     Z.AI route — users who turned thinking off kept burning thinking tokens on
     every turn (the desktop "thinking reverts to medium" report).
     
    +GLM-5.2 additionally exposes a native ``reasoning_effort`` knob with two
    +enabled levels (high / max) on the OpenAI-compatible ``/api/paas/v4``
    +endpoint; the Hermes effort scale is collapsed onto those.
    +
     These tests pin the profile's wire-shape contract so Z.AI requests stay
     correctly shaped without going live.
     """
    @@ -61,12 +65,102 @@ def test_explicitly_disabled_sends_disabled_marker(self, zai_profile):
             assert top_level == {}
     
         def test_no_effort_levels_leak_to_top_level(self, zai_profile):
    -        """GLM has no effort knob — never emit ``reasoning_effort``."""
    +        """Non-5.2 GLM models have no effort knob — never emit
    +        ``reasoning_effort`` for them (GLM-5.2 is the exception, below)."""
             for effort in ("minimal", "low", "medium", "high", "xhigh"):
    -            _, top_level = zai_profile.build_api_kwargs_extras(
    -                reasoning_config={"enabled": True, "effort": effort}, model="glm-5.2"
    -            )
    -            assert top_level == {}
    +            for model in ("glm-5", "glm-5.1", "glm-4.6"):
    +                _, top_level = zai_profile.build_api_kwargs_extras(
    +                    reasoning_config={"enabled": True, "effort": effort}, model=model
    +                )
    +                assert top_level == {}
    +
    +
    +class TestZaiGLM52ReasoningEffort:
    +    """GLM-5.2's native ``reasoning_effort`` knob (two enabled levels)."""
    +
    +    def test_high_maps_to_high(self, zai_profile):
    +        extra_body, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True, "effort": "high"},
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {"thinking": {"type": "enabled"}}
    +        assert top_level == {"reasoning_effort": "high"}
    +
    +    @pytest.mark.parametrize("effort", ["low", "medium", "minimal"])
    +    def test_lower_efforts_clamp_up_to_high(self, zai_profile, effort):
    +        """GLM-5.2's minimum thinking level is high — lower Hermes levels
    +        clamp onto it."""
    +        extra_body, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True, "effort": effort},
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {"thinking": {"type": "enabled"}}
    +        assert top_level == {"reasoning_effort": "high"}
    +
    +    @pytest.mark.parametrize("effort", ["xhigh", "max"])
    +    def test_strong_efforts_map_to_max(self, zai_profile, effort):
    +        extra_body, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True, "effort": effort},
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {"thinking": {"type": "enabled"}}
    +        assert top_level == {"reasoning_effort": "max"}
    +
    +    def test_disabled_sends_no_effort(self, zai_profile):
    +        """Disabled reasoning still sends the thinking-off marker but never
    +        an effort level."""
    +        extra_body, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": False, "effort": "high"},
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {"thinking": {"type": "disabled"}}
    +        assert top_level == {}
    +
    +    def test_no_config_leaves_server_default(self, zai_profile):
    +        extra_body, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config=None,
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {}
    +        assert top_level == {}
    +
    +    def test_no_effort_sends_no_effort_level(self, zai_profile):
    +        """Enabled but no effort preference → thinking marker only; the
    +        server picks its default effort."""
    +        extra_body, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True},
    +            model="glm-5.2",
    +        )
    +        assert extra_body == {"thinking": {"type": "enabled"}}
    +        assert top_level == {}
    +
    +    @pytest.mark.parametrize(
    +        "model",
    +        [
    +            "z-ai/glm-5.2",
    +            "glm-5-2",
    +            "glm-5p2",
    +            "accounts/fireworks/models/glm-5p2",
    +            "zai-org-glm-5-2",
    +        ],
    +    )
    +    def test_alias_spellings_recognized(self, zai_profile, model):
    +        _, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True, "effort": "max"},
    +            model=model,
    +        )
    +        assert top_level == {"reasoning_effort": "max"}
    +
    +    @pytest.mark.parametrize(
    +        "model",
    +        ["glm-5.1", "glm-5", "glm-4.7", "glm-4-9b", "", None],
    +    )
    +    def test_non_glm_5_2_models_get_no_effort(self, zai_profile, model):
    +        _, top_level = zai_profile.build_api_kwargs_extras(
    +            reasoning_config={"enabled": True, "effort": "high"},
    +            model=model,
    +        )
    +        assert top_level == {}
     
     
     class TestZaiModelGating:
    @@ -110,7 +204,7 @@ def test_non_thinking_models_emit_nothing(self, zai_profile, model):
     
     
     class TestZaiFullKwargsIntegration:
    -    """End-to-end: the transport's full kwargs carry the thinking marker."""
    +    """End-to-end: the transport's full kwargs carry the reasoning wiring."""
     
         def test_disabled_reaches_the_wire(self, zai_profile):
             from agent.transports.chat_completions import ChatCompletionsTransport
    @@ -139,3 +233,18 @@ def test_no_preference_keeps_wire_clean(self, zai_profile):
                 provider_name="zai",
             )
             assert "thinking" not in kwargs.get("extra_body", {})
    +
    +    def test_glm_5_2_effort_reaches_top_level(self, zai_profile):
    +        from agent.transports.chat_completions import ChatCompletionsTransport
    +
    +        kwargs = ChatCompletionsTransport().build_kwargs(
    +            model="glm-5.2",
    +            messages=[{"role": "user", "content": "ping"}],
    +            tools=None,
    +            provider_profile=zai_profile,
    +            reasoning_config={"enabled": True, "effort": "max"},
    +            base_url="https://api.z.ai/api/paas/v4",
    +            provider_name="zai",
    +        )
    +        assert kwargs["reasoning_effort"] == "max"
    +        assert kwargs["extra_body"]["thinking"] == {"type": "enabled"}