fix(server): surface Claude safety model fallback notices instead of dropping them (#8853)

Darahaas Yajamanyam committed 2 months ago

Commit a5bbad91 parent 1963ca02 files · +34 −10Browse files

Unified · Split

2 files

apps/server/src/provider/Layers/ClaudeAdapter.test.ts +26 −9
apps/server/src/provider/Layers/ClaudeAdapter.ts +8 −1

apps/server/src/provider/Layers/ClaudeAdapter.test.ts+26−9

     const harness = makeHarness();
     return Effect.gen(function* () {
       const adapter = yield* ClaudeAdapter;
-      const runtimeEvents: Array<ProviderRuntimeEvent> = [];
+      const runtimeEventsFiber = yield* adapter.streamEvents.pipe(
-      const runtimeEventsFiber = yield* Stream.runForEach(adapter.streamEvents, (event) =>
+        Stream.takeUntil(
-        Effect.sync(() => runtimeEvents.push(event)),
+          (event) =>
-      ).pipe(Effect.forkChild);
+            event.type === "session.state.changed" && event.payload.reason === "api_retry:3/10",
+        ),
+        Stream.runCollect,
+        Effect.forkChild,
+      );
       yield* adapter.startSession({
         threadId: THREAD_ID,
           uuid: "tu",
         },
         { type: "system", subtype: "commands_changed", session_id: "session", uuid: "cc" },
-        { type: "system", subtype: "model_refusal_fallback", session_id: "session", uuid: "mrf" },
         { type: "system", subtype: "local_command_output", session_id: "session", uuid: "lco" },
         { type: "system", subtype: "plugin_install", session_id: "session", uuid: "pi" },
         { type: "system", subtype: "memory_recall", session_id: "session", uuid: "mr" },
       ]) {
         harness.query.emit(message as unknown as SDKMessage);
       }
+      // Safety model-fallback notices DO surface as a warning row.
+      harness.query.emit({
+        type: "system",
+        subtype: "model_refusal_fallback",
+        trigger: "refusal",
+        direction: "retry",
+        original_model: "claude-fable-5",
+        fallback_model: "claude-opus-4-8",
+        request_id: "req_test",
+        api_refusal_category: "cyber",
+        api_refusal_explanation: null,
+        content: "Safeguards flagged this message. Switched to Opus 4.8.",
+        session_id: "session",
+        uuid: "mrf",
+      } as unknown as SDKMessage);
       // High-priority notifications DO surface as a warning row.
       harness.query.emit({
         type: "system",
         session_id: "session",
         uuid: "retry",
       } as unknown as SDKMessage);
-      yield* Effect.yieldNow;
+      const runtimeEvents = Array.from(yield* Fiber.join(runtimeEventsFiber));
-      yield* Effect.yieldNow;
       const warnings = runtimeEvents.filter((event) => event.type === "runtime.warning");
-      // Exactly three warnings: the high-priority notification, the
+      // Exactly four warnings: the fallback notice, high-priority notification,
       // warning-level informational note, and the refusal. Nothing else.
       assert.deepEqual(
         warnings.map((event) => event.payload.message),
         [
+          "Safeguards flagged this message. Switched to Opus 4.8.",
           "context window nearly full",
           "Stop hook prevented continuation",
           "The request was declined by the API.",
           event.payload.reason.startsWith("api_retry:"),
       );
       assert.equal(heartbeat?.type, "session.state.changed");
-      runtimeEventsFiber.interruptUnsafe();
     }).pipe(
       Effect.provideService(Random.Random, makeDeterministicRandomService()),
       Effect.provide(harness.layer),

apps/server/src/provider/Layers/ClaudeAdapter.ts+8−1

           yield* emitRuntimeWarning(context, message.text, message);
         }
         return;
+      case "model_refusal_fallback":
+        // A safety fallback switched the model mid-session (e.g. Fable 5
+        // retried on Opus 4.8 after a flagged request). The CLI ships the
+        // user-facing notice in `content`; surface it like high-priority
+        // notifications so the rest of the session isn't silently served
+        // by a different model.
+        yield* emitRuntimeWarning(context, message.content, message);
+        return;
       // Inner protocol/UX details with no T3 surface today — consumed
       // deliberately so they don't masquerade as unknown-subtype warnings.
       // `background_tasks_changed` is a roster snapshot ({tasks: [...]}); the
       // source. `control_request_progress` is a liveness heartbeat for an
       // in-flight control request. `worker_shutting_down` is a Remote
       // Control worker notice; the session close path reports the outcome.
-      case "model_refusal_fallback":
       case "local_command_output":
       case "plugin_install":
       case "commands_changed":

fix(server): surface Claude safety model fallback notices instead of dropping them (#8853)

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code