fix(ai): recognize LiteLLM context overflow errors

packages/ai/src/provider-error.ts

@@ -42,6 +42,8 @@ const patterns = [
   // Hugging Face Text Generation Inference, e.g. Together
   /`inputs` tokens \+ `max_new_tokens` must be <= \d+/i,
   /context window exceeds limit/i,
+  // LiteLLM's ContextWindowExceededError
+  /context window exceeded/i,
   /exceeded model token limit/i,
   /context[_ ]length[_ ]exceeded/i,
   /context length is only \d+ tokens/i,

packages/ai/test/provider-error.test.ts

@@ -21,6 +21,8 @@ describe("provider error classification", () => {
       "Prompt exceeds max length",
       "The input is too long for the selected model.",
       "This text request is too large for the model's configured input limit.",
+      // LiteLLM's own check, before it calls the model.
+      "litellm.ContextWindowExceededError: litellm.BadRequestError: litellm._pre_call_checks: Context Window exceeded for given call. No models have context window large enough for this call.",
     ]
 
     expect(messages.every(isContextOverflow)).toBe(true)

packages/core/test/session-runner.test.ts

@@ -5952,6 +5952,26 @@ describe("SessionRunnerLLM", () => {
     })
   }
 
+  scenario("compacts when the provider cuts an oversized prompt to its window and replies with nothing", function* (s) {
+    // Xiaomi MiMo answers an oversized prompt with HTTP 200, finish "length", no output, and a prompt count that fills
+    // the window instead of rejecting it.
+    s.currentModel = recoveryModel
+    yield* s.llm.push(
+      TestLLM.complete({ reason: { normalized: "length" }, usage: { inputTokens: 199_994, nonCachedInputTokens: 199_994, outputTokens: 0 } }),
+      TestLLM.text("## Objective\n- summary", "summary"),
+      TestLLM.text("Done", "done"),
+    )
+
+    yield* s.runPrompt("Large prompt")
+
+    expect(s.requests).toHaveLength(3)
+    expect(userTexts(s.requests[1]).at(-1)).toContain("Summarize only what the user and the assistant said and did")
+    expect((yield* s.messages).filter((message) => message.type === "compaction")).toMatchObject([
+      { status: "completed" },
+    ])
+    expect((yield* s.context).at(-1)).toMatchObject(Expected.assistant({ finish: "stop" }, [Expected.text("Done")]))
+  })
+
   for (const tools of ["hosted", "unfinished"] as const) {
     scenario(`caps output token limit continuations and resets for new input (${tools})`, function* (s) {
       const toolEvents =