fix(ai-judge): classify provider aborts after timeout as timeout

packages/pi-permission-ai-judge/src/model.ts

@@ -7,7 +7,13 @@ import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 import { buildJudgeContext, MAX_REASON_CODE_POINTS, REPORT_VERDICT_TOOL_NAME } from "./prompt";
 import type { BashJudgmentEvidence } from "./evidence";
 
-const DEFAULT_TIMEOUT_MS = 15_000;
+// 60s: glm-5.2 at the user's default `thinking: high` profile was observed
+// both finishing in seconds and racing the former 15s deadline (one verdict
+// landed at exactly 15.003s and was mislabeled `aborted`); high-variance
+// reasoning latency needs the wider bound. The judge is the first chain
+// link, so its wait delays the human prompt by at most this much.
+// PIEXTENSIO-11 calibrates a final value from cohort data.
+const DEFAULT_TIMEOUT_MS = 60_000;
 // Reasoning-token aware cap. Providers that bill chain-of-thought inside
 // completion tokens (observed on zai glm-5.2 despite `thinking: disabled`:
 // 669 reasoning + 70 output for one verdict) exhausted a 256-token budget
@@ -203,12 +209,19 @@ export async function requestStructuredVerdict(
             requestController.signal,
         );
 
-        if (shutdownSignal.aborted || response.stopReason === "aborted") {
+        if (shutdownSignal.aborted) {
             return failure("aborted");
         }
+        // Check timeout before the provider's abort stopReason: a provider
+        // returning `aborted` after our own deadline hit is a timeout, not a
+        // session-shutdown abort. Mislabeling it starves PIEXTENSIO-11's
+        // timeout calibration of exactly the rows it needs.
         if (timeoutController.signal.aborted) {
             return failure("timeout");
         }
+        if (response.stopReason === "aborted") {
+            return failure("aborted");
+        }
         if (response.stopReason === "error") {
             return failure("model_error");
         }

packages/pi-permission-ai-judge/test/model.test.ts

@@ -14,6 +14,7 @@ const metadata = {
 function response(
     content: AssistantMessage["content"],
     output = 12,
+    stopReason: AssistantMessage["stopReason"] = "toolUse",
 ): AssistantMessage {
     return {
         role: "assistant",
@@ -21,6 +22,7 @@ function response(
         api: metadata.api,
         provider: metadata.provider,
         model: metadata.model,
+        stopReason,
         usage: {
             input: 10,
             output,
@@ -35,7 +37,6 @@ function response(
                 total: 0,
             },
         },
-        stopReason: "toolUse",
         timestamp: Date.now(),
     };
 }
@@ -297,4 +298,24 @@ describe("requestStructuredVerdict", () => {
         });
         expect(complete).not.toHaveBeenCalled();
     });
+
+    it("classifies a provider abort arriving after the deadline as timeout", async () => {
+        const complete = vi.fn(
+            (_context: Context, signal: AbortSignal) =>
+                new Promise<AssistantMessage>((resolve, reject) => {
+                    signal.addEventListener("abort", () =>
+                        // Provider surfaces the client abort as an `aborted`
+                        // stopReason after our 1ms deadline already fired.
+                        resolve(response([], 12, "aborted")),
+                    );
+                }),
+        );
+        const timedOut = await requestStructuredVerdict(
+            ready(complete as never),
+            evidence,
+            new AbortController().signal,
+            1,
+        );
+        expect(timedOut).toMatchObject({ code: "timeout" });
+    });
 });