Hermes-agent

tests/plugins/model_providers/test_deepseek_profile.py

@@ -0,0 +1,207 @@
+"""Unit tests for the DeepSeek provider profile's thinking-mode wiring.
+
+DeepSeek V4 (and the legacy ``deepseek-reasoner``) expects every request to
+carry an explicit ``extra_body.thinking`` parameter.  Omitting it makes the
+server default to thinking-mode ON, which then enforces the
+``reasoning_content``-must-be-echoed-back contract on subsequent turns and
+breaks the conversation with HTTP 400 (#15700, #17212, #17825).
+
+These tests pin the profile's wire-shape contract so DeepSeek requests stay
+correctly shaped without going live.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+
+@pytest.fixture
+def deepseek_profile():
+    """Resolve the registered DeepSeek profile.
+
+    Going through ``providers.get_provider_profile`` keeps the test honest —
+    if someone later replaces the registered class with a plain
+    ``ProviderProfile``, every assertion below collapses.
+    """
+    # ``model_tools`` triggers plugin discovery on import, which is what
+    # registers the DeepSeek profile in the global provider registry.
+    import model_tools  # noqa: F401
+    import providers
+
+    profile = providers.get_provider_profile("deepseek")
+    assert profile is not None, "deepseek provider profile must be registered"
+    return profile
+
+
+class TestDeepSeekThinkingWireShape:
+    """``build_api_kwargs_extras`` produces DeepSeek's exact wire format."""
+
+    def test_v4_pro_default_enables_thinking_without_effort(self, deepseek_profile):
+        """No reasoning_config → thinking enabled, server picks default effort."""
+        extra_body, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config=None, model="deepseek-v4-pro"
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {}
+
+    def test_v4_pro_enabled_with_high_effort(self, deepseek_profile):
+        extra_body, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "high"},
+            model="deepseek-v4-pro",
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {"reasoning_effort": "high"}
+
+    @pytest.mark.parametrize("effort", ["low", "medium", "high"])
+    def test_standard_efforts_pass_through(self, deepseek_profile, effort):
+        _, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": effort},
+            model="deepseek-v4-pro",
+        )
+        assert top_level == {"reasoning_effort": effort}
+
+    @pytest.mark.parametrize("effort", ["xhigh", "max", "MAX", "  Max  "])
+    def test_xhigh_and_max_normalize_to_max(self, deepseek_profile, effort):
+        _, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": effort},
+            model="deepseek-v4-pro",
+        )
+        assert top_level == {"reasoning_effort": "max"}
+
+    def test_explicitly_disabled_sends_disabled_marker(self, deepseek_profile):
+        """``reasoning_config.enabled=False`` → ``thinking.type=disabled``.
+
+        The crucial bit is that the parameter is *sent* at all — DeepSeek
+        defaults to thinking-on when ``thinking`` is absent.
+        """
+        extra_body, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": False}, model="deepseek-v4-pro"
+        )
+        assert extra_body == {"thinking": {"type": "disabled"}}
+        # No effort when disabled — DeepSeek rejects it.
+        assert top_level == {}
+
+    def test_disabled_ignores_effort_field(self, deepseek_profile):
+        """Effort silently dropped when thinking is off."""
+        _, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": False, "effort": "high"},
+            model="deepseek-v4-pro",
+        )
+        assert top_level == {}
+
+    def test_unknown_effort_omits_top_level(self, deepseek_profile):
+        """Garbage effort → omit reasoning_effort so DeepSeek applies its default."""
+        _, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "garbage"},
+            model="deepseek-v4-pro",
+        )
+        assert top_level == {}
+
+    def test_empty_effort_omits_top_level(self, deepseek_profile):
+        _, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": ""},
+            model="deepseek-v4-pro",
+        )
+        assert top_level == {}
+
+
+class TestDeepSeekModelGating:
+    """V4 family + ``deepseek-reasoner`` get thinking; V3 stays untouched."""
+
+    @pytest.mark.parametrize(
+        "model",
+        [
+            "deepseek-v4-pro",
+            "deepseek-v4-flash",
+            "deepseek-v4-future-variant",
+            "deepseek-reasoner",
+            "DEEPSEEK-V4-PRO",  # case-insensitive
+        ],
+    )
+    def test_thinking_capable_models_emit_thinking(self, deepseek_profile, model):
+        extra_body, _ = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config=None, model=model
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+
+    @pytest.mark.parametrize(
+        "model",
+        [
+            "deepseek-chat",         # V3 alias
+            "deepseek-v3-0324",      # explicit V3
+            "deepseek-v3.1",         # V3 minor revisions
+            "",                       # bare/unknown
+            None,                     # missing
+            "deepseek-unknown",      # unrecognized
+        ],
+    )
+    def test_non_thinking_models_emit_nothing(self, deepseek_profile, model):
+        extra_body, top_level = deepseek_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "high"}, model=model
+        )
+        assert extra_body == {}
+        assert top_level == {}
+
+
+class TestDeepSeekFullKwargsIntegration:
+    """End-to-end: the transport's full kwargs match DeepSeek's live wire format.
+
+    The live test harness in ``tests/run_agent/test_deepseek_v4_thinking_live.py``
+    sends ``{"reasoning_effort": "high", "extra_body": {"thinking": {"type":
+    "enabled"}}}``.  Confirm the transport produces that exact shape when wired
+    through the registered DeepSeek profile.
+    """
+
+    def test_full_kwargs_match_live_wire_shape(self, deepseek_profile):
+        from agent.transports.chat_completions import ChatCompletionsTransport
+
+        kwargs = ChatCompletionsTransport().build_kwargs(
+            model="deepseek-v4-pro",
+            messages=[{"role": "user", "content": "ping"}],
+            tools=None,
+            provider_profile=deepseek_profile,
+            reasoning_config={"enabled": True, "effort": "high"},
+            base_url="https://api.deepseek.com/v1",
+            provider_name="deepseek",
+        )
+        assert kwargs["model"] == "deepseek-v4-pro"
+        assert kwargs["reasoning_effort"] == "high"
+        assert kwargs["extra_body"] == {"thinking": {"type": "enabled"}}
+
+    def test_v3_chat_full_kwargs_omit_thinking(self, deepseek_profile):
+        from agent.transports.chat_completions import ChatCompletionsTransport
+
+        kwargs = ChatCompletionsTransport().build_kwargs(
+            model="deepseek-chat",
+            messages=[{"role": "user", "content": "ping"}],
+            tools=None,
+            provider_profile=deepseek_profile,
+            reasoning_config={"enabled": True, "effort": "high"},
+            base_url="https://api.deepseek.com/v1",
+            provider_name="deepseek",
+        )
+        assert "reasoning_effort" not in kwargs
+        assert "extra_body" not in kwargs or "thinking" not in kwargs.get("extra_body", {})
+
+
+class TestDeepSeekAuxModel:
+    """DeepSeek aux model is set on the profile so users stop seeing the
+    bogus 'No auxiliary LLM provider configured' warning (#26924).
+
+    Pinned at the profile layer rather than the legacy
+    `_API_KEY_PROVIDER_AUX_MODELS_FALLBACK` dict — new providers are
+    expected to set `default_aux_model` on `ProviderProfile`, and the
+    fallback dict only exists for providers that predate the profiles
+    system.
+    """
+
+    def test_profile_advertises_deepseek_chat(self, deepseek_profile):
+        assert deepseek_profile.default_aux_model == "deepseek-chat"
+
+    def test_consumer_api_returns_deepseek_chat(self):
+        from agent.auxiliary_client import _get_aux_model_for_provider
+        assert _get_aux_model_for_provider("deepseek") == "deepseek-chat"
+
+    def test_consumer_api_returns_non_empty(self):
+        from agent.auxiliary_client import _get_aux_model_for_provider
+        assert _get_aux_model_for_provider("deepseek") != ""

tests/plugins/model_providers/test_kimi_profile.py

@@ -0,0 +1,130 @@
+"""Unit tests for the Kimi/Moonshot provider profile's reasoning wiring.
+
+Moonshot's OpenAI-compat endpoint (``api.moonshot.ai/v1``) treats
+``extra_body.thinking`` and a top-level ``reasoning_effort`` as mutually
+exclusive. The profile must send at most one of them — never both — so a
+request can't trip "cannot specify both 'thinking' and 'reasoning_effort'".
+
+This mirrors the kimi-k2 handling already shipped for the opencode-go relay
+(see ``tests/plugins/model_providers/test_opencode_go_profile.py``).
+"""
+
+from __future__ import annotations
+
+import pytest
+
+
+@pytest.fixture
+def kimi_profile():
+    """Resolve the registered Kimi profile via the provider registry.
+
+    Importing ``model_tools`` triggers plugin discovery, which registers the
+    Kimi profile. Going through ``get_provider_profile`` keeps the test honest:
+    if the registered class is ever swapped for a plain ``ProviderProfile`` the
+    assertions below collapse.
+    """
+    import model_tools  # noqa: F401
+    import providers
+
+    profile = providers.get_provider_profile("kimi-coding")
+    assert profile is not None, "kimi-coding provider profile must be registered"
+    return profile
+
+
+class TestKimiReasoningWireShape:
+    """``build_api_kwargs_extras`` never emits thinking + reasoning_effort together."""
+
+    def test_no_config_enables_thinking_without_effort(self, kimi_profile):
+        """No reasoning_config → thinking on, server picks the depth.
+
+        Regression guard: this path previously also sent
+        ``reasoning_effort="medium"``, pairing thinking + effort on every
+        default call.
+        """
+        extra_body, top_level = kimi_profile.build_api_kwargs_extras(reasoning_config=None)
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {}
+
+    @pytest.mark.parametrize("effort", ["low", "medium", "high"])
+    def test_explicit_effort_sends_effort_only(self, kimi_profile, effort):
+        extra_body, top_level = kimi_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": effort}
+        )
+        assert top_level == {"reasoning_effort": effort}
+        assert "thinking" not in extra_body
+
+    def test_enabled_without_effort_falls_back_to_thinking(self, kimi_profile):
+        extra_body, top_level = kimi_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True}
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {}
+
+    @pytest.mark.parametrize("effort", ["", "garbage", "xhigh", "max"])
+    def test_unrecognized_effort_falls_back_to_thinking(self, kimi_profile, effort):
+        """Unknown/strong efforts aren't in Moonshot's low|medium|high set, so
+        we drop to the thinking toggle rather than sending an invalid effort."""
+        extra_body, top_level = kimi_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": effort}
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {}
+
+    def test_disabled_sends_thinking_disabled_only(self, kimi_profile):
+        extra_body, top_level = kimi_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": False}
+        )
+        assert extra_body == {"thinking": {"type": "disabled"}}
+        assert top_level == {}
+
+    def test_disabled_ignores_effort(self, kimi_profile):
+        extra_body, top_level = kimi_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": False, "effort": "high"}
+        )
+        assert extra_body == {"thinking": {"type": "disabled"}}
+        assert top_level == {}
+
+    @pytest.mark.parametrize(
+        "reasoning_config",
+        [
+            None,
+            {"enabled": True},
+            {"enabled": True, "effort": "high"},
+            {"enabled": True, "effort": "garbage"},
+            {"enabled": False},
+            {"enabled": False, "effort": "low"},
+        ],
+    )
+    def test_never_emits_both(self, kimi_profile, reasoning_config):
+        """The core invariant: thinking and reasoning_effort are never both set."""
+        extra_body, top_level = kimi_profile.build_api_kwargs_extras(
+            reasoning_config=reasoning_config
+        )
+        assert not ("thinking" in extra_body and "reasoning_effort" in top_level)
+
+
+class TestKimiFullKwargsIntegration:
+    """The transport's full kwargs carry at most one reasoning knob."""
+
+    def _build(self, kimi_profile, reasoning_config):
+        from agent.transports.chat_completions import ChatCompletionsTransport
+
+        return ChatCompletionsTransport().build_kwargs(
+            model="kimi-k2-turbo-preview",
+            messages=[{"role": "user", "content": "ping"}],
+            tools=None,
+            provider_profile=kimi_profile,
+            reasoning_config=reasoning_config,
+            base_url="https://api.moonshot.ai/v1",
+            provider_name="kimi-coding",
+        )
+
+    def test_explicit_effort_omits_thinking(self, kimi_profile):
+        kwargs = self._build(kimi_profile, {"enabled": True, "effort": "high"})
+        assert kwargs["reasoning_effort"] == "high"
+        assert "thinking" not in kwargs.get("extra_body", {})
+
+    def test_no_config_omits_effort(self, kimi_profile):
+        kwargs = self._build(kimi_profile, None)
+        assert "reasoning_effort" not in kwargs
+        assert kwargs["extra_body"] == {"thinking": {"type": "enabled"}}

tests/plugins/model_providers/test_minimax_profile.py

@@ -0,0 +1,120 @@
+"""Unit tests for the MiniMax provider profile.
+
+Three MiniMax provider profiles (`minimax` direct API, `minimax-cn` China direct
+API, `minimax-oauth` browser OAuth) all advertise a `default_aux_model` on
+their `ProviderProfile`. The previous M2.7 / M2.7-highspeed values were
+stale relative to the current frontier model (M3, released 2026-06-01) and
+inconsistent with the `_PROVIDER_MODELS["minimax"]` catalog top entry in
+`hermes_cli/models.py`.
+
+This file pins the new defaults so the choice is reviewable and any future
+revert shows up in a failing test rather than silent behavior drift.
+
+Refs:
+  - Issue #36196: M3 support request
+  - PR #36205 (closed unmerged): Csrayz's M3 + 1M context work
+  - PR #36212 (open): adds M3 to `_PROVIDER_MODELS["minimax"]` catalog
+  - PR #6082: M2.7-highspeed → M2.7 for aux model (half-price fix)
+  - Commit 773a0faca: same profile-layer fix pattern for `deepseek`
+"""
+
+from __future__ import annotations
+
+import pytest
+
+
+@pytest.fixture(params=["minimax", "minimax-cn", "minimax-oauth"])
+def minimax_profile(request):
+    """Resolve each registered MiniMax profile.
+
+    Going through ``providers.get_provider_profile`` keeps the test honest —
+    if someone later replaces the registered class with a plain
+    ``ProviderProfile``, every assertion below collapses.
+    """
+    import model_tools  # noqa: F401  -- triggers plugin discovery
+    import providers
+
+    profile = providers.get_provider_profile(request.param)
+    assert profile is not None, f"{request.param} provider profile must be registered"
+    return profile, request.param
+
+
+class TestMinimaxAuxModelM3:
+    """MiniMax profile aux model is the new frontier M3, not the stale M2.7.
+
+    The catalog top entry is ``MiniMax-M3`` in
+    ``hermes_cli.models._PROVIDER_MODELS['minimax']`` and the
+    user-facing ``model.default`` for a Token-Plan install is M3,
+    so pinning the aux default to the same model keeps the runtime
+    consistent (same auth, same billing pool, same rate limits, no
+    surprise 2x-cost highspeed variant). M3 was released 2026-06-01
+    — picking it as the aux default matches the forward-looking
+    catalog order rather than the pre-M3 era.
+    """
+
+    @pytest.mark.parametrize(
+        "provider_id,expected",
+        [
+            ("minimax", "MiniMax-M3"),
+            ("minimax-cn", "MiniMax-M3"),
+            # minimax-oauth sticks with M2.7: the OAuth / Coding Plan
+            # tier historically used -highspeed (PR #6082 collapsed that
+            # to plain M2.7 to avoid the 2x TPS surcharge). M3 is not on
+            # the OAuth/Coding Plan tier per platform docs as of this PR,
+            # so the safe choice is the cheapest generally-available
+            # M2.7 — matching PR #6082's intent.
+            ("minimax-oauth", "MiniMax-M2.7"),
+        ],
+    )
+    def test_profile_advertises_expected_aux_model(
+        self, provider_id, expected
+    ):
+        import model_tools  # noqa: F401
+        import providers
+
+        profile = providers.get_provider_profile(provider_id)
+        assert profile is not None
+        assert profile.default_aux_model == expected, (
+            f"{provider_id} default_aux_model drifted to "
+            f"{profile.default_aux_model!r}, expected {expected!r}"
+        )
+
+    def test_consumer_api_returns_non_empty_for_each_provider(self, minimax_profile):
+        from agent.auxiliary_client import _get_aux_model_for_provider
+
+        profile, provider_id = minimax_profile
+        resolved = _get_aux_model_for_provider(provider_id)
+        assert resolved != "", (
+            f"_get_aux_model_for_provider({provider_id!r}) returned empty — "
+            "the 'No auxiliary LLM provider configured' warning will fire on "
+            f"every {provider_id} session even though the profile advertises "
+            f"default_aux_model={profile.default_aux_model!r}"
+        )
+        assert resolved == profile.default_aux_model, (
+            f"_get_aux_model_for_provider({provider_id!r}) returned "
+            f"{resolved!r} but profile advertises {profile.default_aux_model!r} "
+            "— the consumer API and the profile have drifted out of sync"
+        )
+
+
+class TestMinimaxAuxModelNotHighspeed:
+    """Regression guard against re-introducing the M2.7-highspeed aux default.
+
+    PR #6082 collapsed the highspeed aux choice to plain M2.7 because the
+    highspeed variant costs 2x with no real benefit for compression / vision /
+    session-search aux tasks. None of the three MiniMax profiles should
+    silently re-introduce that 2x-cost path.
+    """
+
+    @pytest.mark.parametrize("provider_id", ["minimax", "minimax-cn", "minimax-oauth"])
+    def test_default_aux_model_is_not_highspeed(self, provider_id):
+        import model_tools  # noqa: F401
+        import providers
+
+        profile = providers.get_provider_profile(provider_id)
+        assert profile is not None
+        assert "highspeed" not in profile.default_aux_model.lower(), (
+            f"{provider_id} default_aux_model={profile.default_aux_model!r} "
+            "is a -highspeed variant — that costs 2x for the same model and "
+            "broke #4082 the first time. Revert to plain M2.7 or M3."
+        )

tests/plugins/model_providers/test_opencode_go_profile.py

@@ -0,0 +1,180 @@
+"""Unit tests for OpenCode Go reasoning-control wiring."""
+
+from __future__ import annotations
+
+import pytest
+
+
+@pytest.fixture
+def opencode_go_profile():
+    """Resolve the registered OpenCode Go provider profile."""
+    import model_tools  # noqa: F401
+    import providers
+
+    profile = providers.get_provider_profile("opencode-go")
+    assert profile is not None, "opencode-go provider profile must be registered"
+    return profile
+
+
+class TestOpenCodeGoKimiReasoning:
+    """Kimi K2 models use Moonshot's thinking + reasoning_effort shape on OpenCode Go."""
+
+    def test_high_effort_emits_thinking_and_effort(self, opencode_go_profile):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "high"},
+            model="kimi-k2.6",
+        )
+        assert extra_body == {}
+        assert top_level == {"reasoning_effort": "high"}
+
+    def test_disabled_emits_thinking_disabled_without_effort(self, opencode_go_profile):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": False},
+            model="kimi-k2.6",
+        )
+        assert extra_body == {"thinking": {"type": "disabled"}}
+        assert top_level == {}
+
+    def test_minimal_effort_enables_thinking_without_effort(self, opencode_go_profile):
+        # "minimal" is not a Moonshot-supported value — drop it, keep thinking on.
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "minimal"},
+            model="kimi-k2.6",
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {}
+
+    @pytest.mark.parametrize(
+        "effort",
+        [
+            "xhigh",
+            "max",
+        ],
+    )
+    def test_strong_efforts_clamp_to_high(self, opencode_go_profile, effort):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": effort},
+            model="moonshotai/kimi-k2.6",
+        )
+        assert extra_body == {}
+        assert top_level == {"reasoning_effort": "high"}
+
+    def test_low_and_medium_pass_through(self, opencode_go_profile):
+        for effort in ("low", "medium"):
+            extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+                reasoning_config={"enabled": True, "effort": effort},
+                model="kimi-k2.5",
+            )
+            assert extra_body == {}
+            assert top_level == {"reasoning_effort": effort}
+
+    def test_no_config_preserves_server_default(self, opencode_go_profile):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config=None,
+            model="kimi-k2.6",
+        )
+        assert extra_body == {}
+        assert top_level == {}
+
+
+class TestOpenCodeGoDeepSeekThinking:
+    """DeepSeek V4 models use DeepSeek-style thinking controls on OpenCode Go."""
+
+    def test_high_effort_emits_thinking_and_effort(self, opencode_go_profile):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "high"},
+            model="deepseek-v4-pro",
+        )
+        assert extra_body == {}
+        assert top_level == {"reasoning_effort": "high"}
+
+    def test_disabled_emits_thinking_disabled_without_effort(self, opencode_go_profile):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": False, "effort": "high"},
+            model="deepseek-v4-pro",
+        )
+        assert extra_body == {"thinking": {"type": "disabled"}}
+        assert top_level == {}
+
+    def test_no_config_emits_thinking_enabled_without_effort(self, opencode_go_profile):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config=None,
+            model="deepseek-v4-pro",
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {}
+
+    def test_minimal_effort_enables_thinking_without_effort(self, opencode_go_profile):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "minimal"},
+            model="deepseek-v4-pro",
+        )
+        assert extra_body == {"thinking": {"type": "enabled"}}
+        assert top_level == {}
+
+    def test_xhigh_and_max_normalize_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="deepseek/deepseek-v4-pro",
+            )
+            assert extra_body == {}
+            assert top_level == {"reasoning_effort": "max"}
+
+
+class TestOpenCodeGoModelGating:
+    """Other OpenCode Go models must not receive Kimi/DeepSeek controls."""
+
+    @pytest.mark.parametrize(
+        "model",
+        [
+            "glm-5.1",
+            "qwen3.6-plus",
+            "minimax-m2.7",
+            "deepseek-v3.1",
+            "deepseek-chat",
+            "",
+            None,
+        ],
+    )
+    def test_non_target_models_emit_nothing(self, opencode_go_profile, model):
+        extra_body, top_level = opencode_go_profile.build_api_kwargs_extras(
+            reasoning_config={"enabled": True, "effort": "high"},
+            model=model,
+        )
+        assert extra_body == {}
+        assert top_level == {}
+
+
+class TestOpenCodeGoFullKwargsIntegration:
+    """End-to-end transport kwargs include the profile-provided controls."""
+
+    def test_kimi_reasoning_reaches_extra_body_and_top_level(self, opencode_go_profile):
+        from agent.transports.chat_completions import ChatCompletionsTransport
+
+        kwargs = ChatCompletionsTransport().build_kwargs(
+            model="kimi-k2.6",
+            messages=[{"role": "user", "content": "ping"}],
+            tools=None,
+            provider_profile=opencode_go_profile,
+            reasoning_config={"enabled": True, "effort": "high"},
+            base_url="https://opencode.ai/zen/go/v1",
+        )
+        assert "extra_body" not in kwargs
+        assert kwargs["reasoning_effort"] == "high"
+
+    def test_deepseek_thinking_reaches_extra_body_and_top_level(
+        self, opencode_go_profile
+    ):
+        from agent.transports.chat_completions import ChatCompletionsTransport
+
+        kwargs = ChatCompletionsTransport().build_kwargs(
+            model="deepseek-v4-pro",
+            messages=[{"role": "user", "content": "ping"}],
+            tools=None,
+            provider_profile=opencode_go_profile,
+            reasoning_config={"enabled": True, "effort": "high"},
+            base_url="https://opencode.ai/zen/go/v1",
+        )
+        assert "extra_body" not in kwargs
+        assert kwargs["reasoning_effort"] == "high"