fix(#66595): [BUG] Fable 5 safety classifier auto-switches to Opus mid-session during authorized, in-scope security testing

src/commands/disengage.js

@@ -0,0 +1,9 @@
+// Add a new command to reset the authorized engagement context
+class DisengageCommand {
+  execute(input) {
+    const model = this.getModel();
+    if (model instanceof Fable5) {
+      model.setAuthorizedEngagement(false);
+    }
+  }
+}

src/commands/engage.js

@@ -0,0 +1,9 @@
+// Add a new command to set the authorized engagement context
+class EngageCommand {
+  execute(input) {
+    const model = this.getModel();
+    if (model instanceof Fable5) {
+      model.setAuthorizedEngagement(true);
+    }
+  }
+}

src/index.js

@@ -0,0 +1,3 @@
+// Register the new plugin
+const authorizedEngagementPlugin = new AuthorizedEngagementPlugin();
+this.plugins.register(authorizedEngagementPlugin);

src/models/fable5.js

@@ -0,0 +1,38 @@
+// Add a new property to store the authorized engagement context
+class Fable5 {
+  constructor() {
+    this.authorizedEngagement = false;
+  }
+
+  // ...
+
+  // Update the safety classifier to check for authorized engagement context
+  safetyClassifier(input) {
+    if (this.authorizedEngagement) {
+      // If authorized engagement context is set, do not auto-switch to Opus
+      return false;
+    }
+    // Otherwise, perform the regular safety classification
+    // ...
+  }
+
+  // Add a new method to set the authorized engagement context
+  setAuthorizedEngagement(context) {
+    this.authorizedEngagement = context;
+  }
+}
+
+// Update the model selection logic to respect the authorized engagement context
+class ModelSelector {
+  // ...
+
+  selectModel(input) {
+    const currentModel = this.getCurrentModel();
+    if (currentModel instanceof Fable5 && currentModel.authorizedEngagement) {
+      // If Fable 5 is the current model and authorized engagement context is set, do not switch to Opus
+      return currentModel;
+    }
+    // Otherwise, perform the regular model selection logic
+    // ...
+  }
+}

src/plugins/authorizedEngagement.js

@@ -0,0 +1,16 @@
+// Add a new plugin to provide the authorized engagement context
+class AuthorizedEngagementPlugin {
+  constructor() {
+    this.authorizedEngagement = false;
+  }
+
+  // ...
+
+  setAuthorizedEngagement(context) {
+    this.authorizedEngagement = context;
+  }
+
+  getAuthorizedEngagement() {
+    return this.authorizedEngagement;
+  }
+}