@vellumai/assistant 0.11.1-staging.2 → 0.11.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.11.1-staging.2",
+  "version": "0.11.1",
   "license": "MIT",
   "type": "module",
   "exports": {

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

@@ -650,7 +650,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
     expect(effectiveBalanced?.provider).toBe("vellum");
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
   });
@@ -978,7 +978,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
   });
 
@@ -1058,7 +1058,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
     expect(effectiveBalanced?.maxTokens).toBe(32000);
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
     // The catalog body carries no topP and the entry has none, so the
@@ -1090,7 +1090,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
     // ...with the user's label and status overlaid.
     expect(effectiveBalanced?.label).toBe("My Default");
     expect(effectiveBalanced?.status).toBe("disabled");
@@ -1115,7 +1115,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
   });
 
   test("off-platform boot preserves user-toggled status on a managed stub", () => {
@@ -1137,7 +1137,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
   });
 
   test("boot preserves a user-edited topP override on a managed stub", () => {
@@ -1161,7 +1161,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
   });
 
   test("effective balanced profile carries no topP override by default", () => {
@@ -1216,7 +1216,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
     // Status is unset — the default resolves active.
     expect("status" in (effectiveBalanced ?? {})).toBe(false);
   });
@@ -1295,7 +1295,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(mainAgentConfig.model).toBe("gpt-5.6-luna");
 
     const raw = JSON.parse(readFileSync(CONFIG_PATH, "utf-8"));
     expect(raw.llm.profiles.balanced).toEqual({
@@ -1326,7 +1326,7 @@ describe("loadConfig startup behavior", () => {
     );
     expect(effectiveBalanced?.provider).toBe("vellum");
     expect(effectiveBalanced?.provider_connection).toBeUndefined();
-    expect(effectiveBalanced?.model).toBe("accounts/fireworks/models/glm-5p2");
+    expect(effectiveBalanced?.model).toBe("gpt-5.6-luna");
     expect(effectiveBalanced?.maxTokens).toBe(32000);
     expect(effectiveBalanced?.thinking).toEqual({
       enabled: true,

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

@@ -1067,7 +1067,7 @@ 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("accounts/fireworks/models/glm-5p2");
+    expect(wireBalanced.model).toBe("gpt-5.6-luna");
     expect(wireBalanced.provider).toBe("vellum");
     expect(wireBalanced.provider_connection).toBeUndefined();
     expect(wireBalanced.status).toBe("disabled");
@@ -1314,9 +1314,7 @@ 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(
-      "accounts/fireworks/models/glm-5p2",
-    );
+    expect(response.llm.profiles.balanced.model).toBe("gpt-5.6-luna");
     const result = await patchRoute.handler({
       body: { llm: { profiles: response.llm.profiles } },
     });

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

@@ -55,6 +55,15 @@ describe("getEffectiveProfiles", () => {
     }
   });
 
+  test("the managed Balanced profile routes GPT-5.6 Luna through OpenAI", () => {
+    const balanced = CODE_DEFAULT_PROFILE_ENTRIES.balanced;
+    expect(balanced.model).toBe("gpt-5.6-luna");
+    expect(resolveRoutingIdentity(balanced.provider, balanced.model)).toEqual({
+      connectionName: "vellum",
+      expectedProvider: "openai",
+    });
+  });
+
   test("the managed Quality profile routes GPT-5.6 Sol through OpenAI", () => {
     const quality = CODE_DEFAULT_PROFILE_ENTRIES["quality-optimized"];
     expect(quality.model).toBe("gpt-5.6-sol");

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

@@ -70,7 +70,7 @@ export type DefaultProfileTemplate = Omit<
  */
 const VELLUM_PROFILE_IMPLS: Record<ProfileMatrixKey, DefaultProfileTemplate> = {
   balanced: {
-    model: "accounts/fireworks/models/glm-5p2",
+    model: "gpt-5.6-luna",
     provider: "vellum",
     source: "managed",
     label: "Balanced",

package/src/plugin-api/vision-support.test.ts

@@ -161,18 +161,19 @@ describe("doesSupportVision", () => {
 
 describe("doesSupportVision with a BYO default provider", () => {
   test("judges a default profile against the default provider's column model", () => {
-    // Managed/vellum column: balanced → glm-5p2 (text-only, supportsVision:
-    // false). Anthropic column: balanced → claude-sonnet-4-6 (supportsVision:
-    // true). The judged model must be the one the BYO install actually runs.
+    // Managed/vellum column: cost-optimized → deepseek-v4-flash (text-only,
+    // supportsVision: false). Anthropic column: cost-optimized carries
+    // intent "latency-optimized" → claude-haiku-4-5 (supportsVision: true).
+    // The judged model must be the one the BYO install actually runs.
     setMockConfig({}, { provider: "anthropic" });
-    expect(doesSupportVision(profile("balanced"))).toBe(true);
+    expect(doesSupportVision(profile("cost-optimized"))).toBe(true);
   });
 
   test("without a default provider, a default profile judges the managed column", () => {
-    // Null-reduction: no defaultProvider resolves balanced through the
-    // vellum column (glm-5p2, text-only).
+    // Null-reduction: no defaultProvider resolves cost-optimized through the
+    // vellum column (deepseek-v4-flash, text-only).
     setMockConfig({});
-    expect(doesSupportVision(profile("balanced"))).toBe(false);
+    expect(doesSupportVision(profile("cost-optimized"))).toBe(false);
   });
 
   test("mix arms naming a default profile resolve through the same column", () => {
@@ -181,7 +182,7 @@ describe("doesSupportVision with a BYO default provider", () => {
         "mix-profile": {
           mix: [
             { profile: "text-arm", weight: 0.5 },
-            { profile: "balanced", weight: 0.5 },
+            { profile: "cost-optimized", weight: 0.5 },
           ],
         },
         "text-arm": {
@@ -191,8 +192,8 @@ describe("doesSupportVision with a BYO default provider", () => {
       },
       { provider: "anthropic" },
     );
-    // The "balanced" arm resolves to the anthropic column's vision-capable
-    // model, so the mix can route to vision.
+    // The "cost-optimized" arm resolves to the anthropic column's
+    // vision-capable model, so the mix can route to vision.
     expect(doesSupportVision(profile("mix-profile"))).toBe(true);
   });