fix(errors): recover from Z.AI token-limit errors (#111744)

* fix(errors): classify Z.AI token-limit overflow

* fix(zai): own token overflow classification

* fix(zai): classify live prompt overflow errors

---------

Co-authored-by: Altay <altay@hey.com>

extensions/zai/index.ts

@@ -385,6 +385,10 @@ export default definePluginEntry({
         }),
       ],
       resolveDynamicModel: (ctx) => resolveGlm5ForwardCompatModel(ctx),
+      matchesContextOverflowError: ({ errorMessage }) =>
+        /\b(?:tokens? in request more than max tokens? allowed|prompt exceeds max(?:imum)? length)\b/i.test(
+          errorMessage,
+        ),
       ...buildProviderReplayFamilyHooks({
         family: "openai-compatible",
         dropReasoningFromHistory: false,

packages/ai/src/utils/overflow.test.ts

@@ -57,6 +57,8 @@ describe("provider overflow messages", () => {
     "Error: 400 Input length (265330) exceeds model's maximum context length (262144).",
     "Provider returned error: Input length 131393 exceeds the maximum allowed input length of 131,040 tokens.",
     "Input length 131393 exceeds maximum allowed input length of 131040 token",
+    "code 1210: tokens in request more than max tokens allowed",
+    "code 1261: Prompt exceeds max length",
   ])("detects %s", (text) => {
     expect(isContextOverflow(errorMessage(text), 262_144)).toBe(true);
   });

packages/ai/src/utils/overflow.ts

@@ -33,7 +33,9 @@ export function isConfiguredContextSizeOverflowError(errorMessage: string): bool
  * - Kimi For Coding: "Your request exceeded model token limit: X (requested: Y)"
  * - Cerebras: "400/413 status code (no body)"
  * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
- * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
+ * - z.ai: May return "tokens in request more than max tokens allowed" (code 1210),
+ *   "Prompt exceeds max length" (code 1261), or accept overflow silently; handled via the
+ *   error patterns or usage.input > contextWindow
  * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
  *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
  *   input filling the context window.
@@ -56,6 +58,8 @@ const OVERFLOW_PATTERNS = [
   /greater than the context length/i, // LM Studio
   /context window exceeds limit/i, // MiniMax
   /exceeded model token limit/i, // Kimi For Coding
+  /tokens? in request more than max tokens? allowed/i, // Z.AI / Zhipu GLM error 1210
+  /prompt exceeds max(?:imum)? length/i, // Z.AI / Zhipu GLM error 1261
   /too large for model with \d+ maximum context length/i, // Mistral
   CONFIGURED_CONTEXT_SIZE_OVERFLOW_RE, // DS4 server
   /model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text
@@ -115,10 +119,12 @@ function resolveContextInputTokens(message: AssistantMessage): number | undefine
  * - llama.cpp: "exceeds the available context size"
  * - LM Studio: "greater than the context length"
  * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
+ * - z.ai: "tokens in request more than max tokens allowed" or "Prompt exceeds max length"
  *
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
- *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ *   sometimes returns rate limit errors instead of the explicit overflow error above. Pass
+ *   contextWindow param to detect silent overflow.
  * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
  *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
  * - Ollama: May truncate input silently for some setups, but may also return explicit

src/plugin-sdk/test-helpers/provider-runtime-contract.ts

@@ -800,6 +800,26 @@ export function describeZAIProviderRuntimeContract(load: ProviderRuntimeContract
       });
     });
 
+    it("owns Z.AI token-limit overflow classification", () => {
+      const provider = requireProviderContractProvider("zai");
+
+      expect(
+        provider.matchesContextOverflowError?.({
+          errorMessage: "code 1210: tokens in request more than max tokens allowed",
+        }),
+      ).toBe(true);
+      expect(
+        provider.matchesContextOverflowError?.({
+          errorMessage: "code 1261: Prompt exceeds max length",
+        }),
+      ).toBe(true);
+      expect(
+        provider.matchesContextOverflowError?.({
+          errorMessage: "code 1210: invalid request parameters",
+        }),
+      ).toBe(false);
+    });
+
     it("owns usage auth resolution", async () => {
       const provider = requireProviderContractProvider("zai");
       await expect(