feat(ai-judge): capture model per permission request

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

@@ -1,5 +1,7 @@
 import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 import { getAgentDir } from "@earendil-works/pi-coding-agent";
+import type { Model } from "@earendil-works/pi-ai";
+import type { ModelRegistry } from "@earendil-works/pi-coding-agent";
 import {
     getPermissionsService,
     PERMISSIONS_READY_CHANNEL,
@@ -20,7 +22,9 @@ const REVIEW_SCHEMA_VERSION = 1;
 interface RootSession {
     readonly getSessionId: () => string;
     readonly expectedSessionId: string;
-    readonly model: ModelAvailability;
+    /** Current-model probe: reads the live model at each authorize call. */
+    readonly getModel: () => Model<any> | undefined;
+    readonly modelRegistry: ModelRegistry;
     readonly shutdown: AbortController;
     /** Opaque per-runtime identity for cohort segmentation. */
     readonly judgeRuntimeId: string;
@@ -181,10 +185,16 @@ export default function permissionAiJudge(pi: ExtensionAPI): void {
                         return { kind: "defer" };
                     }
 
-                    // `captured.model` is the session-start snapshot. Config and
+                    // Per-request model capture (PIEXTENSIO-3 cat.3): an
-                    // model-select support are deliberately outside this slice.
+                    // in-request switch does not change an in-flight attempt;
+                    // a between-request switch affects the next attempt. The
+                    // probe reads the live current model here, not at start.
+                    const availability = createModelAvailability(
+                        captured.getModel(),
+                        captured.modelRegistry,
+                    );
                     const result = await requestStructuredVerdict(
-                        captured.model,
+                        availability,
                         evidence,
                         captured.shutdown.signal,
                         captured.config.timeoutMs,
@@ -266,7 +276,8 @@ export default function permissionAiJudge(pi: ExtensionAPI): void {
         root = {
             getSessionId: () => ctx.sessionManager.getSessionId(),
             expectedSessionId: sessionId,
-            model: createModelAvailability(ctx.model, ctx.modelRegistry),
+            getModel: () => ctx.model,
+            modelRegistry: ctx.modelRegistry,
             shutdown: new AbortController(),
             judgeRuntimeId: crypto.randomUUID(),
             config: loadJudgeConfig({ agentDir: getAgentDir() }),

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

@@ -145,6 +145,61 @@ afterEach(() => {
 });
 
 describe("AI judge lifecycle", () => {
+    it("captures the model per request: a between-request switch changes the next call", async () => {
+        let authorize: Authorizer["authorize"] | undefined;
+        const service = {
+            registerAuthorizer: vi.fn((_name, callback) => {
+                authorize = callback;
+                return vi.fn();
+            }),
+            checkPermission: vi.fn(),
+            getToolPermission: vi.fn(),
+        } as unknown as PermissionsService;
+        publishPermissionsService(service);
+        publishedService = service;
+
+        // A mutable "current model" the session switches mid-run.
+        let currentModel = {
+            id: "model-a",
+            provider: "test-provider",
+            api: "openai-codex-responses",
+        } as Model<any>;
+        const seen: string[] = [];
+        const complete = vi.fn(async (model: Model<any>) => {
+            seen.push(model.id);
+            return modelResponse();
+        });
+        const ctx = {
+            hasUI: true,
+            sessionManager: { getSessionId: () => "session-root" },
+            get model() {
+                return currentModel;
+            },
+            modelRegistry: { complete },
+            ui: { notify: vi.fn() },
+        } as unknown as ExtensionContext;
+
+        const harness = createFakePi();
+        extension(harness.pi);
+        harness.start(ctx);
+        harness.ready();
+
+        const log = {
+            review: vi.fn(),
+            debug: vi.fn(),
+        };
+        await authorize!(ask(), { checkPermission: vi.fn(), getToolPermission: vi.fn() }, log);
+        currentModel = {
+            id: "model-b",
+            provider: "test-provider",
+            api: "openai-codex-responses",
+        } as Model<any>;
+        await authorize!(ask(), { checkPermission: vi.fn(), getToolPermission: vi.fn() }, log);
+
+        expect(seen).toEqual(["model-a", "model-b"]);
+        harness.shutdown();
+    });
+
     it("calls the current model once, records metadata, and still defers in Shadow", async () => {
         let authorize: Authorizer["authorize"] | undefined;
         const dispose = vi.fn();