@vellumai/assistant 0.11.5-staging.1 → 0.11.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.11.5-staging.1",
+  "version": "0.11.5",
   "license": "MIT",
   "type": "module",
   "exports": {

package/src/__tests__/config-loader-backfill.test.ts

@@ -49,6 +49,7 @@ afterAll(() => {
 });
 
 import {
+  CODE_DEFAULT_PROFILE_ENTRIES,
   getEffectiveProfile,
   getEffectiveProfiles,
   getEffectiveProfilesForProvider,
@@ -164,6 +165,14 @@ const LEGACY_HATCH_PROFILE_NAMES = [
   "custom-cost-optimized",
 ] as const;
 
+/**
+ * The managed Balanced model. These tests assert that resolution serves the
+ * code catalog rather than a workspace body, so they follow the catalog
+ * instead of restating its pin.
+ */
+const MANAGED_BALANCED_MODEL = CODE_DEFAULT_PROFILE_ENTRIES.balanced
+  .model as string;
+
 function createProviderConnectionsDb(): DrizzleDb {
   const sqlite = new Database(":memory:");
   sqlite.exec("PRAGMA journal_mode=WAL");
@@ -669,7 +678,7 @@ describe("loadConfig startup behavior", () => {
     expect(raw.llm.profiles).toEqual({});
     // Default content resolves from the code catalog via the effective view.
     const effectiveBalanced = getEffectiveProfile(raw.llm.profiles, "balanced");
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
     expect(effectiveBalanced?.provider).toBe("vellum");
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
   });
@@ -997,7 +1006,7 @@ describe("loadConfig startup behavior", () => {
     // Resolution ignores the drifted body: a managed-source entry contributes
     // only label/status/topP, everything else comes from the catalog.
     const effectiveBalanced = getEffectiveProfile(raw.llm.profiles, "balanced");
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
   });
 
@@ -1077,7 +1086,7 @@ describe("loadConfig startup behavior", () => {
     expect(raw.llm.profiles.balanced).toEqual(drifted);
     expect(raw.llm.activeProfile).toBe("balanced");
     const effectiveBalanced = getEffectiveProfile(raw.llm.profiles, "balanced");
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
     expect(effectiveBalanced?.maxTokens).toBe(32000);
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
     // The catalog body carries no topP and the entry has none, so the
@@ -1109,7 +1118,7 @@ describe("loadConfig startup behavior", () => {
     expect(raw.llm.profiles.balanced).toEqual(edited);
     const effectiveBalanced = getEffectiveProfile(raw.llm.profiles, "balanced");
     // Content is served from the catalog...
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
     // ...with the user's label and status overlaid.
     expect(effectiveBalanced?.label).toBe("My Default");
     expect(effectiveBalanced?.status).toBe("disabled");
@@ -1134,7 +1143,7 @@ describe("loadConfig startup behavior", () => {
     expect(raw.llm.profiles.balanced).toEqual(stub);
     const effectiveBalanced = getEffectiveProfile(raw.llm.profiles, "balanced");
     expect(effectiveBalanced?.label).toBe("My Default");
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
   });
 
   test("off-platform boot preserves user-toggled status on a managed stub", () => {
@@ -1156,7 +1165,7 @@ describe("loadConfig startup behavior", () => {
     expect(effectiveBalanced?.status).toBe("disabled");
     // Content still comes from the catalog — only label/status/topP are
     // workspace-owned.
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
   });
 
   test("boot preserves a user-edited topP override on a managed stub", () => {
@@ -1180,7 +1189,7 @@ describe("loadConfig startup behavior", () => {
     expect(effectiveBalanced?.topP).toBe(0.5);
     // Content still comes from the catalog — topP is workspace-owned, the
     // rest is code-owned.
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
   });
 
   test("effective balanced profile carries no topP override by default", () => {
@@ -1235,7 +1244,7 @@ describe("loadConfig startup behavior", () => {
     expect(raw.llm.profiles.balanced).toBeUndefined();
     const effectiveBalanced = getEffectiveProfile(raw.llm.profiles, "balanced");
     expect(effectiveBalanced?.label).toBe("Balanced");
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
     // Status is unset — the default resolves active.
     expect("status" in (effectiveBalanced ?? {})).toBe(false);
   });
@@ -1314,7 +1323,7 @@ describe("loadConfig startup behavior", () => {
     // overlay boot. The overlay-set label is what shows through the
     // effective view.
     expect(mainAgentConfig.provider).toBe("vellum");
-    expect(mainAgentConfig.model).toBe("gpt-5.6-luna");
+    expect(mainAgentConfig.model).toBe(MANAGED_BALANCED_MODEL);
 
     const raw = JSON.parse(readFileSync(CONFIG_PATH, "utf-8"));
     expect(raw.llm.profiles.balanced).toEqual({
@@ -1345,7 +1354,7 @@ describe("loadConfig startup behavior", () => {
     );
     expect(effectiveBalanced?.provider).toBe("vellum");
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
-    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
+    expect(effectiveBalanced?.model).toBe(MANAGED_BALANCED_MODEL);
     expect(effectiveBalanced?.maxTokens).toBe(32000);
     expect(effectiveBalanced?.thinking).toEqual({
       enabled: true,

package/src/__tests__/managed-profile-guard.test.ts

@@ -1090,7 +1090,7 @@ describe("code-owned default profiles — wire view and write normalization", ()
     });
     const response = (await getRoute.handler({})) as Record<string, any>;
     const wireBalanced = response.llm.profiles.balanced;
-    expect(wireBalanced.model).toBe("gpt-5.6-luna");
+    expect(wireBalanced.model).toBe("accounts/fireworks/models/glm-5p2");
     expect(wireBalanced.provider).toBe("vellum");
     expect(wireBalanced.provider_connection).toBeUndefined();
     expect(wireBalanced.status).toBe("disabled");
@@ -1337,7 +1337,9 @@ describe("code-owned default profiles — echoes over stale on-disk bodies", ()
     });
     const response = (await getRoute.handler({})) as Record<string, any>;
     // The wire view serves catalog content, not the stale body.
-    expect(response.llm.profiles.balanced.model).toBe("gpt-5.6-luna");
+    expect(response.llm.profiles.balanced.model).toBe(
+      "accounts/fireworks/models/glm-5p2",
+    );
     const result = await patchRoute.handler({
       body: { llm: { profiles: response.llm.profiles } },
     });

package/src/__tests__/voice-scoped-grant-consumer.test.ts

@@ -135,6 +135,8 @@ function createMockSession(opts?: {
       };
     },
     handleSecretResponse: () => {},
+    // The image-bearing profile pin reads the leg's history.
+    getMessages: () => [],
     runAgentLoop: async (
       _content: string,
       _messageId: string,

package/src/__tests__/voice-session-bridge.test.ts

@@ -74,6 +74,8 @@ function makeStreamingSession(events: AssistantEvent[]): Conversation {
       }
     },
     handleConfirmationResponse: () => {},
+    // The image-bearing profile pin reads the leg's history.
+    getMessages: () => [],
     abort: () => {},
   } as unknown as Conversation;
 }
@@ -133,6 +135,8 @@ function makePersistingStreamingSession(
     },
     handleConfirmationResponse: () => {},
     abort: () => {},
+    // The image-bearing profile pin reads the leg's history.
+    getMessages: () => [],
   } as unknown as Conversation &
     PersistUserMessageContext & {
       callSessionId?: string;
@@ -263,6 +267,8 @@ describe("voice-session-bridge", () => {
       abort: () => {
         abortCalled = true;
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -302,6 +308,8 @@ describe("voice-session-bridge", () => {
           onEvent(event);
         }
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -348,6 +356,8 @@ describe("voice-session-bridge", () => {
           onEvent(event);
         }
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -393,6 +403,8 @@ describe("voice-session-bridge", () => {
       abort: () => {
         abortCalled = true;
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -430,6 +442,8 @@ describe("voice-session-bridge", () => {
       setTurnChannelContext: (ctx: unknown) => {
         capturedTurnChannelContext = ctx;
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -586,6 +600,8 @@ describe("voice-session-bridge", () => {
           capturedTrustContext = ctx;
         }
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -629,6 +645,8 @@ describe("voice-session-bridge", () => {
           capturedPrompt = prompt;
         }
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -690,6 +708,8 @@ describe("voice-session-bridge", () => {
           capturedPrompt = prompt;
         }
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -782,6 +802,8 @@ describe("voice-session-bridge", () => {
         });
       },
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -870,6 +892,8 @@ describe("voice-session-bridge", () => {
         });
       },
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -946,6 +970,8 @@ describe("voice-session-bridge", () => {
         handleConfirmationCalls.push({ requestId, decision });
       },
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -1014,6 +1040,8 @@ describe("voice-session-bridge", () => {
         handleConfirmationCalls.push({ requestId, decision });
       },
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -1080,6 +1108,8 @@ describe("voice-session-bridge", () => {
         handleConfirmationCalls.push({ requestId, decision });
       },
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -1164,6 +1194,8 @@ describe("voice-session-bridge", () => {
         } as AssistantEvent);
       },
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
   }
 
@@ -1303,6 +1335,8 @@ describe("voice-session-bridge", () => {
         handleSecretCalls.push({ requestId, value, delivery });
       },
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);
@@ -1354,6 +1388,8 @@ describe("voice-session-bridge", () => {
       runAgentLoop: async () => {},
       handleConfirmationResponse: () => {},
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation & { forcePromptSideEffects: boolean };
 
     injectDeps(() => session);
@@ -1413,6 +1449,8 @@ describe("voice-session-bridge", () => {
       runAgentLoop: async () => {},
       handleConfirmationResponse: () => {},
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation & {
       forcePromptSideEffects: boolean;
       callSessionId?: string;
@@ -1479,6 +1517,8 @@ describe("voice-session-bridge", () => {
       runAgentLoop: async () => {},
       handleConfirmationResponse: () => {},
       abort: () => {},
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation & {
       forcePromptSideEffects: boolean;
       callSessionId?: string;
@@ -1540,6 +1580,8 @@ describe("voice-session-bridge", () => {
       abort: () => {
         abortCalled = true;
       },
+      // The image-bearing profile pin reads the leg's history.
+      getMessages: () => [],
     } as unknown as Conversation;
 
     injectDeps(() => session);

package/src/calls/__tests__/voice-session-bridge.test.ts

@@ -31,6 +31,14 @@ mock.module("../../daemon/conversation-store.js", () => ({
   getOrCreateConversation: async () => fakeConversation,
 }));
 
+// Vision capability of the image pin's target profile. Install-dependent in
+// production (a BYO provider resolves the profile key through its own column
+// of the intent matrix), so it is scripted rather than read from a catalog.
+let pinProfileSupportsVision = true;
+mock.module("../../plugin-api/vision-support.js", () => ({
+  doesSupportVision: () => pinProfileSupportsVision,
+}));
+
 // Conversation-CRUD doubles for the teardown transcript-hygiene pass. The
 // real module is spread so every other export keeps its production behavior;
 // only the functions the hygiene pass (and discard) touch are recorded.
@@ -133,6 +141,7 @@ interface FakeConversation {
     opts?: { decisionContext?: string },
   ) => void;
   runAgentLoop: (...args: unknown[]) => Promise<void>;
+  getMessages: () => Array<{ role: string; content: unknown[] }>;
   abort: (reason?: unknown) => void;
   loadFromDb: () => Promise<void>;
   toolsDisabledDepth: number;
@@ -149,6 +158,8 @@ function makeFakeConversation(opts: {
   onPersist?: (attempt: number) => void;
   /** Workspace root; pass empty to model a missing boundary. */
   workingDir?: string;
+  /** In-memory history the profile pin reads; undefined models a text-only call. */
+  messages?: Array<{ role: string; content: unknown[] }>;
 }) {
   const waitForIdleCalls: WaitForIdleCall[] = [];
   const confirmationDecisions: Array<{ requestId: string; decision: string }> =
@@ -206,6 +217,7 @@ function makeFakeConversation(opts: {
       confirmationDecisions.push({ requestId, decision });
     },
     runAgentLoop: () => (opts.runAgentLoop ?? (async () => {}))(),
+    getMessages: () => opts.messages ?? [],
     abort: () => {},
     loadFromDb: async () => {
       opts.events?.push("loadFromDb");
@@ -2165,3 +2177,105 @@ describe("transcript hygiene (teardown pass)", () => {
     expect(crudLog.deletes).toContain("assistant-row-1");
   });
 });
+
+describe("startVoiceTurn image-bearing profile pin", () => {
+  /** A persisted user message carrying a photo taken mid-call. */
+  const PHOTO_HISTORY = [
+    { role: "user", content: [{ type: "text", text: "here's a photo:" }] },
+    {
+      role: "user",
+      content: [
+        { type: "text", text: "here's a photo:" },
+        { type: "image", source: { type: "base64", data: "abc" } },
+      ],
+    },
+  ];
+
+  beforeEach(() => {
+    pinProfileSupportsVision = true;
+  });
+
+  async function runOptionsFor(opts: {
+    messages?: Array<{ role: string; content: unknown[] }>;
+    turn?: Record<string, unknown>;
+  }): Promise<Record<string, unknown>> {
+    const fake = makeFakeConversation({
+      processing: false,
+      ...(opts.messages ? { messages: opts.messages } : {}),
+    });
+    fakeConversation = fake.conversation;
+    let runOptions: Record<string, unknown> = {};
+    fake.conversation.runAgentLoop = async (...args: unknown[]) => {
+      runOptions = args[2] as Record<string, unknown>;
+    };
+    await startVoiceTurn({ ...makeTurnOptions(), ...(opts.turn ?? {}) });
+    return runOptions;
+  }
+
+  test("a text-only call keeps the call-site profile", async () => {
+    const runOptions = await runOptionsFor({});
+
+    expect(runOptions.overrideProfile).toBeUndefined();
+    expect(runOptions.forceOverrideProfile).toBeUndefined();
+  });
+
+  test("an image in history pins the image-capable profile", async () => {
+    // `callAgent`'s balanced profile carries no guarantee that its model takes
+    // an image, and a model that rejects one fails the whole leg.
+    const runOptions = await runOptionsFor({ messages: PHOTO_HISTORY });
+
+    expect(runOptions.overrideProfile).toBe("latency-optimized");
+    // callAgent is not `mainAgent`, so an unforced override would sit below
+    // the call-site profile and never apply.
+    expect(runOptions.forceOverrideProfile).toBe(true);
+  });
+
+  test("an image nested in a tool result counts too", async () => {
+    const runOptions = await runOptionsFor({
+      messages: [
+        {
+          role: "user",
+          content: [
+            {
+              type: "tool_result",
+              contentBlocks: [{ type: "image", source: { data: "abc" } }],
+            },
+          ],
+        },
+      ],
+    });
+
+    expect(runOptions.overrideProfile).toBe("latency-optimized");
+  });
+
+  test("a front-door leg is left alone — its own call site pins it", async () => {
+    const runOptions = await runOptionsFor({
+      messages: PHOTO_HISTORY,
+      turn: { routingLeg: "front-door" },
+    });
+
+    expect(runOptions.overrideProfile).toBeUndefined();
+    expect(runOptions.callSite).toBe("voiceFrontDoor");
+  });
+
+  test("no pin when the pin target can't take an image either", async () => {
+    // Fireworks: `latency-optimized` resolves to a text-only model while
+    // `balanced` is vision-capable, so pinning would break the very turn the
+    // pin exists to save.
+    pinProfileSupportsVision = false;
+
+    const runOptions = await runOptionsFor({ messages: PHOTO_HISTORY });
+
+    expect(runOptions.overrideProfile).toBeUndefined();
+    expect(runOptions.forceOverrideProfile).toBeUndefined();
+  });
+
+  test("an explicit routing pin wins over the image pin", async () => {
+    const runOptions = await runOptionsFor({
+      messages: PHOTO_HISTORY,
+      turn: { overrideProfile: "quality-optimized" },
+    });
+
+    expect(runOptions.overrideProfile).toBe("quality-optimized");
+  });
+});

package/src/calls/voice-session-bridge.ts

@@ -35,8 +35,9 @@ import {
   updateMessageContent,
 } from "../persistence/conversation-crud.js";
 import { VOICE_ESCALATION_CONTINUATION_MESSAGE_KIND } from "../plugin-api/constants.js";
+import { doesSupportVision } from "../plugin-api/vision-support.js";
 import { pinnedListeningLanguage } from "../providers/speech-to-text/provider-catalog.js";
-import type { ContentBlock } from "../providers/types.js";
+import type { ContentBlock, Message } from "../providers/types.js";
 import { broadcastMessage } from "../runtime/assistant-event-hub.js";
 import { DAEMON_INTERNAL_ASSISTANT_ID } from "../runtime/assistant-scope.js";
 import { getCurrentSeq } from "../runtime/assistant-stream-state.js";
@@ -67,6 +68,40 @@ import {
 
 const log = getLogger("voice-session-bridge");
 
+/**
+ * Profile an image-bearing voice leg is pinned to.
+ *
+ * The latency-class profile is the one voice already leans on (it fronts every
+ * turn through `voiceFrontDoor`); `callAgent`'s `balanced` profile carries no
+ * guarantee that its model takes images, and a model that rejects an image
+ * fails the whole leg rather than degrading it.
+ *
+ * Whether THIS profile takes images is an install-level question, not a
+ * constant: a BYO provider resolves the key through its own column of the
+ * intent matrix, and on Fireworks that lands on a text-only model while its
+ * `balanced` column is vision-capable. Pinning there would break the exact
+ * turns this pin exists to save, hence the capability check at the call site.
+ */
+const VOICE_IMAGE_PROFILE = "latency-optimized";
+
+/**
+ * Does this conversation's history carry an image?
+ *
+ * Images persist inline and are re-sent on every later turn, so one photo
+ * taken mid-call makes every remaining turn of that call an image turn -- the
+ * check is over the whole history, not just this turn's own content.
+ */
+function conversationCarriesImage(messages: readonly Message[]): boolean {
+  return messages.some((message) =>
+    message.content.some(
+      (block) =>
+        block.type === "image" ||
+        (block.type === "tool_result" &&
+          block.contentBlocks?.some((nested) => nested.type === "image")),
+    ),
+  );
+}
+
 /**
  * Front-door decision rule with the registry-derived capability digest. The
  * front-door leg runs toolless (see the `toolsDisabledDepth` bracket in
@@ -1657,6 +1692,23 @@ export async function startVoiceTurn(
         conversation.toolsDisabledDepth++;
         frontDoorToolsSuppressed = true;
       }
+      // Resolved once here rather than inside the options literal below, so
+      // the history scan happens once per leg. A front-door leg is skipped:
+      // its own call site already resolves to the same profile. The
+      // capability check comes before the scan because it is the cheaper of
+      // the two and it decides whether the pin is worth anything at all.
+      const carriesImage =
+        opts.routingLeg !== "front-door" &&
+        doesSupportVision(VOICE_IMAGE_PROFILE) &&
+        conversationCarriesImage(conversation.getMessages());
+      if (carriesImage) {
+        log.info(
+          { turnId, routingLeg: opts.routingLeg ?? null },
+          "Voice leg carries an image; pinning the image-capable profile",
+        );
+      }
+      const profilePin =
+        opts.overrideProfile ?? (carriesImage ? VOICE_IMAGE_PROFILE : null);
       await conversation.runAgentLoop(persistedContent, messageId, {
         onEvent: (msg: AssistantEvent) => {
           if (msg.type === "assistant_turn_start") {
@@ -1737,11 +1789,13 @@ export async function startVoiceTurn(
         // strong escalation profile. `forceOverrideProfile` floats it above the
         // callAgent call-site layers (callAgent is not `mainAgent`, so the
         // override would otherwise sit below the call-site profile).
-        ...(opts.overrideProfile != null
-          ? {
-              overrideProfile: opts.overrideProfile,
-              forceOverrideProfile: true,
-            }
+        //
+        // An explicit routing pin wins; failing that, a leg whose history
+        // carries an image is pinned to a profile whose model takes one. A
+        // front-door leg needs neither: its own call site already resolves
+        // there.
+        ...(profilePin != null
+          ? { overrideProfile: profilePin, forceOverrideProfile: true }
           : {}),
       });
       if (lastError) {

package/src/config/__tests__/balanced-model-experiment.test.ts

@@ -31,8 +31,8 @@ import {
 // profile that runs another.
 
 const FLAG = BALANCED_MODEL_EXPERIMENT_FLAG_KEY;
-const SHIPPED_MODEL = "gpt-5.6-luna";
-const GLM_MODEL = "accounts/fireworks/models/glm-5p2";
+const SHIPPED_MODEL = "accounts/fireworks/models/glm-5p2";
+const GLM_MODEL = SHIPPED_MODEL;
 
 function setArm(value: boolean | string): void {
   setOverridesForTesting({ [FLAG]: value });
@@ -79,7 +79,7 @@ describe("balanced-model experiment arms", () => {
     expect(entry?.source).toBe("managed");
   });
 
-  test("glm-5p2 repoints mainAgent at GLM 5.2 within the model's output cap", () => {
+  test("glm-5p2 resolves mainAgent to GLM 5.2 within the model's output cap", () => {
     setArm("glm-5p2");
     const resolved = resolveCallSiteConfig(
       "mainAgent",
@@ -93,7 +93,7 @@ describe("balanced-model experiment arms", () => {
     const cap = catalogMaxOutputTokens(upstream as string, GLM_MODEL);
     expect(cap).toBeDefined();
     expect(resolved.maxTokens).toBeLessThanOrEqual(cap as number);
-    // Only the model moves: the shipped token budget stands.
+    // The shipped token budget stands.
     expect(resolved.maxTokens).toBe(
       CODE_DEFAULT_PROFILE_ENTRIES.balanced.maxTokens as number,
     );
@@ -111,7 +111,7 @@ describe("balanced-model experiment arms", () => {
   });
 
   test("the other default profiles are untouched by an arm", () => {
-    setArm("glm-5p2");
+    setArm("terra");
     for (const key of [
       "quality-optimized",
       "cost-optimized",
@@ -239,10 +239,10 @@ describe("balanced-model experiment boundaries", () => {
   });
 
   test("an install with no defaultProvider still picks up the arm", () => {
-    setArm("glm-5p2");
+    setArm("terra");
     expect(
       resolveDefaultProfileForProvider(undefined, "balanced", null)?.model,
-    ).toBe(GLM_MODEL);
+    ).toBe("gpt-5.6-terra");
   });
 });
 

package/src/config/__tests__/default-profile-catalog.test.ts

@@ -57,12 +57,12 @@ describe("getEffectiveProfiles", () => {
     }
   });
 
-  test("the managed Balanced profile routes GPT-5.6 Luna through OpenAI", () => {
+  test("the managed Balanced profile routes GLM 5.2 through Fireworks", () => {
     const balanced = CODE_DEFAULT_PROFILE_ENTRIES.balanced;
-    expect(balanced.model).toBe("gpt-5.6-luna");
+    expect(balanced.model).toBe("accounts/fireworks/models/glm-5p2");
     expect(resolveRoutingIdentity(balanced.provider, balanced.model)).toEqual({
       connectionName: "vellum",
-      expectedProvider: "openai",
+      expectedProvider: "fireworks",
     });
   });
 

package/src/config/default-profile-catalog.ts

@@ -73,7 +73,7 @@ type ProfileImpls = Record<DefaultProfileKey, DefaultProfileTemplate>;
  */
 const VELLUM_PROFILE_IMPLS: ProfileImpls = {
   balanced: {
-    model: "gpt-5.6-luna",
+    model: "accounts/fireworks/models/glm-5p2",
     provider: "vellum",
     source: "managed",
     label: "Balanced",
@@ -130,8 +130,7 @@ const VELLUM_PROFILE_IMPLS: ProfileImpls = {
     // provisioned in every environment, and it alone selects the upstream:
     // `provider` below is the provider-agnostic managed sentinel, so
     // `getManagedUpstream` resolves the real upstream from the model's catalog
-    // owner. Sharing `balanced`'s model is deliberate: the split is effort
-    // and thinking, and this model is the one live-voice TTFT drives validated.
+    // owner. This model is the one live-voice TTFT drives validated.
     model: "gpt-5.6-luna",
     provider: "vellum",
     source: "managed",
@@ -165,8 +164,8 @@ const VELLUM_PROFILE_IMPLS: ProfileImpls = {
  * arm is remote input, and an object lookup would resolve `constructor` or
  * `toString` to an inherited `Object.prototype` member instead of missing.
  *
- * `glm-5p2` is text-only: Balanced does not pass `doesSupportVision` on that
- * arm, so image input routes through the image-fallback captioning plugin.
+ * `glm-5p2` names the same model as the shipped pin and stays in the table so
+ * the arm keeps its meaning if the shipped pin moves again.
  */
 const BALANCED_EXPERIMENT_MODELS = new Map<string, string>([
   ["terra", "gpt-5.6-terra"],
@@ -206,7 +205,7 @@ const CHATGPT_PROFILE_IMPLS: ProfileImpls = {
     source: "managed",
     label: "Balanced",
     description: "Good balance of quality, cost, and speed",
-    // Matches the vellum column's Balanced (same model): the Codex path
+    // Matches the vellum column's Balanced token budget: the Codex path
     // sends no max_output_tokens, so this only sizes internal budgeting.
     maxTokens: 32000,
     effort: "high",

package/src/config/feature-flag-registry.json

@@ -228,14 +228,6 @@
       "description": "Enable self-hosted assistant configuration.",
       "defaultEnabled": false
     },
-    {
-      "id": "web-remote-ingress",
-      "scope": "client",
-      "key": "web-remote-ingress",
-      "label": "Web Remote Ingress",
-      "description": "Show the pair-a-device card in web settings. Visibility only; pairing itself is always available.",
-      "defaultEnabled": false
-    },
     {
       "id": "paired-devices-ui",
       "scope": "client",
@@ -324,14 +316,6 @@
       "description": "Persist chat secret redactions as sentinel markers instead of the legacy HTML marker, enriched with the vault identity when the value byte-matches a credential revealed this turn. Also gates the `credentials reveal --for-chat` channel and the live-stream swap that replaces an echoed reveal plaintext with its sentinel before emission. Redaction itself is NOT gated: a route-proven reveal plaintext stays out of persisted rows in both modes. Client chip rendering is gated on daemon version, not on this flag.",
       "defaultEnabled": false
     },
-    {
-      "id": "assistant-switcher",
-      "scope": "client",
-      "key": "assistant-switcher",
-      "label": "Assistant Switcher",
-      "description": "Show the Switch Assistant section in General settings, linking to the assistant chooser. Covers local clients, signed-in cloud-hub users, remote-gateway origins, and native mobile shells; remote-gateway desktop browsers hand off to the hub chooser with a self-registration param. Also makes the chooser's selection authoritative for the app's active assistant on every surface including the platform hub, a job that previously belonged to multi-platform-assistant alone.",
-      "defaultEnabled": false
-    },
     {
       "id": "marketing-pricing-takeover",
       "scope": "client",
@@ -377,7 +361,7 @@
       "scope": "assistant",
       "key": "experiment-balanced-model-2026-08-06",
       "label": "Experiment: Balanced Profile Model",
-      "description": "Multivariate experiment on the model the managed Balanced inference profile resolves to. control = gpt-5.6-luna (the shipped pin); terra = gpt-5.6-terra; glm-5p2 = accounts/fireworks/models/glm-5p2. Only the managed (vellum) implementation moves: ChatGPT-subscription and BYOK installs run the provider the user chose and are outside the experiment. Rollout and targeting are managed in the LaunchDarkly dashboard; any value that is not an arm name resolves to control.",
+      "description": "Multivariate experiment on the model the managed Balanced inference profile resolves to. control = the shipped pin (accounts/fireworks/models/glm-5p2); terra = gpt-5.6-terra; glm-5p2 = accounts/fireworks/models/glm-5p2. Only the managed (vellum) implementation moves: ChatGPT-subscription and BYOK installs run the provider the user chose and are outside the experiment. Rollout and targeting are managed in the LaunchDarkly dashboard; any value that is not an arm name resolves to control.",
       "defaultEnabled": "control",
       "values": ["control", "terra", "glm-5p2"]
     },