fix(providers): kimi_coding rejects custom temperature, skip sending

Upstream returns HTTP 400 `invalid temperature: only 1 is allowed for
this model` for every model on the kimi_coding endpoint. Treat it like
the o1/o3/o4 / gpt-5-mini family — omit `temperature` from the request
body so the server applies its mandatory default.

Two new tests:
  - kimi_coding strips temperature even when ChatRequest.Options sets it
  - other providers still receive the override (negative control)

Author: raihan0824

internal/providers/openai_extra_headers_test.go

@@ -88,3 +88,40 @@ func TestOpenAIProvider_ExtraHeaders_NoOpWhenEmpty(t *testing.T) {
 		t.Errorf("ExtraHeaders after empty calls = %v, want nil", got)
 	}
 }
+
+// TestKimiCoding_TemperatureSkipped reproduces the upstream rejection
+// `invalid temperature: only 1 is allowed for this model`. When the provider
+// is kimi_coding, the request body must omit temperature entirely so the
+// upstream applies its mandatory default.
+func TestKimiCoding_TemperatureSkipped(t *testing.T) {
+	p := NewOpenAIProvider("kimi-coding", "sk-fake", "https://api.kimi.com/coding/v1", "kimi-k2-turbo-preview").
+		WithProviderType("kimi_coding")
+
+	body := p.buildRequestBody("kimi-k2-turbo-preview", ChatRequest{
+		Messages: []Message{{Role: "user", Content: "hi"}},
+		Options:  map[string]any{OptTemperature: 0.7},
+	}, true)
+
+	if _, present := body["temperature"]; present {
+		t.Errorf("temperature must not be sent to kimi_coding; got body[temperature]=%v", body["temperature"])
+	}
+}
+
+// TestKimiCoding_TemperatureSentForOtherProviders is the negative control —
+// without provider_type=kimi_coding, a temperature option still flows through.
+func TestKimiCoding_TemperatureSentForOtherProviders(t *testing.T) {
+	p := NewOpenAIProvider("openai", "sk-fake", "https://api.openai.com/v1", "gpt-4o-mini")
+
+	body := p.buildRequestBody("gpt-4o-mini", ChatRequest{
+		Messages: []Message{{Role: "user", Content: "hi"}},
+		Options:  map[string]any{OptTemperature: 0.7},
+	}, true)
+
+	got, ok := body["temperature"]
+	if !ok {
+		t.Fatal("temperature must be sent for non-kimi providers")
+	}
+	if got != 0.7 {
+		t.Errorf("temperature = %v, want 0.7", got)
+	}
+}

internal/providers/openai_request.go

@@ -184,6 +184,12 @@ func (p *OpenAIProvider) buildRequestBody(model string, req ChatRequest, stream
 		// Note: gpt-5.X flagship models (gpt-5.1, gpt-5.4, gpt-5.5) DO support temperature;
 		// only the mini/nano reasoning variants reject it.
 		skipTemp := strings.HasPrefix(capabilityModel, "gpt-5-mini") || strings.HasPrefix(capabilityModel, "gpt-5-nano") || strings.HasPrefix(capabilityModel, "o1") || strings.HasPrefix(capabilityModel, "o3") || strings.HasPrefix(capabilityModel, "o4")
+		// Kimi Coding rejects any temperature override — `invalid temperature: only
+		// 1 is allowed for this model`. Skip sending so the upstream applies its
+		// own default (1). Matches the model-locked behavior of o1/o3/o4.
+		if p.providerType == "kimi_coding" {
+			skipTemp = true
+		}
 		if !skipTemp {
 			body["temperature"] = v
 		}