feat: support Kimi K3

nanobot/providers/openai_compat_provider.py

@@ -59,6 +59,7 @@ _DEFAULT_OPENROUTER_HEADERS = {
     "X-OpenRouter-Title": "nanobot",
     "X-OpenRouter-Categories": "cli-agent,personal-agent",
 }
+_KIMI_K3_MODEL = "kimi-k3"
 _KIMI_THINKING_MODELS: frozenset[str] = frozenset({
     "kimi-k2.5",
     "kimi-k2.6",
@@ -113,9 +114,9 @@ def _provider_prefix_key(name: str) -> str:
 
 
 def _requires_max_completion_tokens(model_name: str) -> bool:
-    """Return True for models that reject ``max_tokens`` (GPT-5 family, o-series)."""
+    """Return True for models that require ``max_completion_tokens``."""
     slug = _model_slug(model_name)
-    return "gpt-5" in slug or any(
+    return slug == _KIMI_K3_MODEL or "gpt-5" in slug or any(
         slug == p or slug.startswith((p + "-", p + ".")) for p in ("o1", "o3", "o4")
     )
 
@@ -716,9 +717,12 @@ class OpenAICompatProvider(LLMProvider):
     ) -> bool:
         """Return True when the model accepts a temperature parameter.
 
-        GPT-5 family and reasoning models (o1/o3/o4) reject temperature
-        when reasoning_effort is set to anything other than ``"none"``.
+        Kimi K3 uses a fixed temperature that should be omitted. GPT-5 family
+        and reasoning models (o1/o3/o4) reject temperature when
+        reasoning_effort is set to anything other than ``"none"``.
         """
+        if _model_slug(model_name) == _KIMI_K3_MODEL:
+            return False
         if reasoning_effort and reasoning_effort.lower() != "none":
             return False
         name = model_name.lower()
@@ -788,6 +792,17 @@ class OpenAICompatProvider(LLMProvider):
                 semantic_effort = "minimal"
 
         wire_effort = reasoning_effort
+        slug = _model_slug(model_name)
+        if slug == _KIMI_K3_MODEL and semantic_effort is not None:
+            # K3 always reasons and currently accepts only the top-level
+            # reasoning_effort="max". Preserve disabled/default semantics by
+            # omitting the field; normalize older enabled presets to "max" so
+            # switching from a K2.x model does not send an unsupported value.
+            if semantic_effort in ("none", "minimal"):
+                wire_effort = None
+            else:
+                semantic_effort = "max"
+                wire_effort = "max"
         if spec and spec.name == "dashscope" and semantic_effort == "minimal":
             # DashScope accepts none/minimum/low/medium/high/xhigh; "minimal" 400s.
             wire_effort = "minimum"
@@ -825,7 +840,6 @@ class OpenAICompatProvider(LLMProvider):
         # Only send thinking controls when reasoning_effort is explicit so
         # omitting the config preserves each provider's default.
         if reasoning_effort is not None:
-            slug = _model_slug(model_name)
             thinking_enabled = semantic_effort not in ("none", "minimal")
             for thinking_style in _thinking_styles_for(spec, model_name):
                 if not thinking_enabled and slug in _KIMI_ALWAYS_THINKING_MODELS:

nanobot/webui/settings_api.py

@@ -121,7 +121,7 @@ _IMAGE_GENERATION_ASPECT_RATIOS = {
     "2:3",
     "21:9",
 }
-_CONTEXT_WINDOW_TOKEN_OPTIONS = {65_536, 200_000, 262_144}
+_CONTEXT_WINDOW_TOKEN_OPTIONS = {65_536, 200_000, 262_144, 1_048_576}
 _MODEL_CONFIGURATION_SLUG_RE = re.compile(r"[^a-z0-9_-]+")
 _ENV_REF_RE = re.compile(r"\$\{([A-Za-z_][A-Za-z0-9_]*)\}")
 
@@ -634,7 +634,9 @@ def _parse_context_window_tokens(value: str | None) -> int | None:
     except ValueError:
         raise WebUISettingsError("context_window_tokens must be an integer") from None
     if parsed not in _CONTEXT_WINDOW_TOKEN_OPTIONS:
-        raise WebUISettingsError("context_window_tokens must be 65536, 200000, or 262144")
+        raise WebUISettingsError(
+            "context_window_tokens must be 65536, 200000, 262144, or 1048576"
+        )
     return parsed
 
 
@@ -706,6 +708,10 @@ def _reasoning_effort_values_for(provider_name: str, model: str) -> list[str]:
         return list(_DEFAULT_REASONING_EFFORT_VALUES)
 
     model_lower = (model or "").lower()
+    if model_lower.rsplit("/", 1)[-1] == "kimi-k3":
+        # K3 always reasons and currently exposes only its default/max effort.
+        return ["", "max"]
+
     implicit = getattr(spec, "implicit_reasoning_models", ())
     if implicit and any(pat in model_lower for pat in implicit):
         # Reasoning is always on; only "Default" makes sense.

tests/providers/test_litellm_kwargs.py

@@ -1540,6 +1540,43 @@ def test_kimi_k25_no_extra_body_when_reasoning_effort_none() -> None:
     assert "extra_body" not in kw
 
 
+def test_kimi_k3_uses_native_defaults() -> None:
+    """K3 omits fixed sampling params and uses the non-deprecated token limit field."""
+    kw = _build_kwargs_for("moonshot", "kimi-k3", reasoning_effort=None)
+
+    assert "temperature" not in kw
+    assert "max_tokens" not in kw
+    assert kw["max_completion_tokens"] == 1024
+    assert "reasoning_effort" not in kw
+    assert "extra_body" not in kw
+
+
+def test_kimi_k3_uses_top_level_max_reasoning_effort() -> None:
+    """K3 uses top-level reasoning_effort=max, never the K2.x thinking body."""
+    kw = _build_kwargs_for("moonshot", "kimi-k3", reasoning_effort="max")
+
+    assert kw["reasoning_effort"] == "max"
+    assert "temperature" not in kw
+    assert "extra_body" not in kw
+
+
+def test_kimi_k3_normalizes_legacy_enabled_reasoning_effort() -> None:
+    """Switching a K2.x preset to K3 must not send unsupported effort values."""
+    kw = _build_kwargs_for("moonshot", "kimi-k3", reasoning_effort="high")
+
+    assert kw["reasoning_effort"] == "max"
+    assert "extra_body" not in kw
+
+
+def test_kimi_k3_omits_disabled_reasoning_effort() -> None:
+    """K3 cannot disable reasoning, so disabled effort falls back to its default."""
+    kw = _build_kwargs_for("moonshot", "kimi-k3", reasoning_effort="none")
+
+    assert "reasoning_effort" not in kw
+    assert "temperature" not in kw
+    assert "extra_body" not in kw
+
+
 def test_kimi_k25_thinking_enabled_with_openrouter_prefix() -> None:
     """OpenRouter-style model names like moonshotai/kimi-k2.5 must trigger thinking.
 

tests/webui/test_settings_api.py

@@ -15,6 +15,7 @@ from nanobot.webui.settings_api import (
     _docs_version,
     _model_catalog_kind,
     _oauth_provider_status,
+    _reasoning_effort_values_for,
     create_model_configuration,
     login_oauth_provider,
     provider_models_payload,
@@ -40,6 +41,10 @@ def test_docs_version_uses_released_versions_and_falls_back_for_dev() -> None:
     assert _docs_version("0.2.3+editable") == "latest"
 
 
+def test_kimi_k3_only_offers_supported_reasoning_effort_values() -> None:
+    assert _reasoning_effort_values_for("moonshot", "kimi-k3") == ["", "max"]
+
+
 def test_settings_payload_includes_versioned_docs(
     tmp_path,
     monkeypatch: pytest.MonkeyPatch,
@@ -357,13 +362,13 @@ def test_update_model_configuration_accepts_context_window_options(
     payload = update_model_configuration(
         {
             "name": ["codex"],
-            "context_window_tokens": ["262144"],
+            "context_window_tokens": ["1048576"],
         }
     )
 
-    assert payload["agent"]["context_window_tokens"] == 262144
+    assert payload["agent"]["context_window_tokens"] == 1048576
     saved = load_config(config_path)
-    assert saved.model_presets["codex"].context_window_tokens == 262144
+    assert saved.model_presets["codex"].context_window_tokens == 1048576
 
 
 def test_update_context_window_rejects_unknown_values(
@@ -376,7 +381,7 @@ def test_update_context_window_rejects_unknown_values(
 
     with pytest.raises(
         WebUISettingsError,
-        match="context_window_tokens must be 65536, 200000, or 262144",
+        match="context_window_tokens must be 65536, 200000, 262144, or 1048576",
     ):
         update_agent_settings({"context_window_tokens": ["128000"]})
 

webui/src/components/settings/SettingsView.tsx

@@ -222,7 +222,7 @@ type ProviderApiType = "auto" | "chat_completions" | "responses";
 type ProviderForm = { apiKey: string; apiBase: string; apiType: ProviderApiType };
 type CustomMcpTransport = "stdio" | "streamableHttp" | "sse";
 
-const CONTEXT_WINDOW_TOKEN_OPTIONS = [65_536, 200_000, 262_144] as const;
+const CONTEXT_WINDOW_TOKEN_OPTIONS = [65_536, 200_000, 262_144, 1_048_576] as const;
 const DEFERRED_MODEL_LIST_PROVIDERS = new Set([
   "aihubmix",
   "atomic_chat",
@@ -2797,7 +2797,13 @@ function ModelsSettings({
               options={CONTEXT_WINDOW_TOKEN_OPTIONS.map((tokens) => ({
                 value: String(tokens),
                 label:
-                  tokens === 262_144 ? "256K" : tokens === 200_000 ? "200K" : "64K",
+                  tokens === 1_048_576
+                    ? "1M"
+                    : tokens === 262_144
+                      ? "256K"
+                      : tokens === 200_000
+                        ? "200K"
+                        : "64K",
               }))}
               onChange={(value) =>
                 setForm((prev) => ({

webui/src/tests/settings-view.test.tsx

@@ -1602,6 +1602,7 @@ describe("SettingsView Apps catalog", () => {
     expect(screen.getByRole("button", { name: "64K" })).toBeInTheDocument();
     expect(screen.getByRole("button", { name: "200K" })).toBeInTheDocument();
     expect(screen.getByRole("button", { name: "256K" })).toBeInTheDocument();
+    expect(screen.getByRole("button", { name: "1M" })).toBeInTheDocument();
   });
 
   it("keeps the default model distinct from the active named configuration", async () => {