fix: retry custom models without temperature

src/lib/llm-client.test.ts

@@ -130,6 +130,40 @@ describe("streamChat — buffered streaming responses", () => {
     expect(onDone).not.toHaveBeenCalled()
   })
 
+  it("retries a custom endpoint without temperature when the provider rejects it", async () => {
+    mockHttpFetch
+      .mockResolvedValueOnce(new Response(JSON.stringify({
+        error: { message: "Unsupported parameter: temperature" },
+      }), { status: 400 }))
+      .mockResolvedValueOnce(new Response([
+        openAiSseToken("retried"),
+        "data: [DONE]",
+      ].join("\n\n"), { status: 200 }))
+    const onToken = vi.fn()
+    const onDone = vi.fn()
+    const onError = vi.fn()
+
+    await streamChat(
+      customStreamingCfg,
+      [{ role: "user", content: "hi" }],
+      { onToken, onDone, onError },
+      undefined,
+      { temperature: 0.1, max_tokens: 512 },
+    )
+
+    expect(mockHttpFetch).toHaveBeenCalledTimes(2)
+    expect(JSON.parse(String(mockHttpFetch.mock.calls[0][1]?.body))).toMatchObject({
+      temperature: 0.1,
+      max_tokens: 512,
+    })
+    const retryBody = JSON.parse(String(mockHttpFetch.mock.calls[1][1]?.body))
+    expect(retryBody.temperature).toBeUndefined()
+    expect(retryBody.max_tokens).toBe(512)
+    expect(onToken).toHaveBeenCalledWith("retried")
+    expect(onDone).toHaveBeenCalledTimes(1)
+    expect(onError).not.toHaveBeenCalled()
+  })
+
   it("cancels a still-open response body after an SSE endpoint error", async () => {
     let bodyCancelled = false
     const body = new ReadableStream<Uint8Array>({

src/lib/llm-client.ts

@@ -142,6 +142,25 @@ export function isReasoningOnlyResponseError(err: unknown): boolean {
   return /^Model produced [\d,]+ characters of reasoning \/ chain-of-thought, but no actual response content\./.test(message)
 }
 
+function shouldRetryWithoutTemperature(
+  config: LlmConfig,
+  status: number,
+  errorDetail: string,
+  requestOverrides?: RequestOverrides,
+): boolean {
+  if (config.provider !== "custom" || requestOverrides?.temperature === undefined) return false
+  if (status !== 400 && status !== 422) return false
+  const detail = errorDetail.toLowerCase()
+  return detail.includes("temperature") && (
+    detail.includes("unsupported") ||
+    detail.includes("not support") ||
+    detail.includes("unknown") ||
+    detail.includes("not allowed") ||
+    detail.includes("only") ||
+    detail.includes("invalid")
+  )
+}
+
 export async function streamChat(
   config: LlmConfig,
   messages: import("./llm-providers").ChatMessage[],
@@ -261,6 +280,10 @@ export async function streamChat(
     } catch {
       // ignore body read failure
     }
+    if (shouldRetryWithoutTemperature(config, response.status, errorDetail, requestOverrides)) {
+      const { temperature: _temperature, ...retryOverrides } = requestOverrides ?? {}
+      return streamChat(config, messages, callbacks, signal, retryOverrides)
+    }
     if (
       response.status === 404 &&
       (config.provider === "azure" ||