@vellumai/assistant 0.11.4 → 0.11.5-staging.2

package/src/cli/commands/skills.help.ts

@@ -6,21 +6,8 @@ export const skillsHelp: CliCommandHelp = {
   name: "skills",
   description: "Browse and install skills from the Vellum catalog",
   helpText: `
-Manage skills from the Vellum catalog. Skills extend the assistant's
-capabilities with pre-built workflows and tools.
-
-Examples:
-  $ assistant skills list
-  $ assistant skills list --json
-  $ assistant skills inspect slack
-  $ assistant skills inspect resend-setup --json
-  $ assistant skills search react
-  $ assistant skills search react --limit 5 --json
-  $ assistant skills install weather
-  $ assistant skills install weather --overwrite
-  $ assistant skills uninstall weather
-  $ assistant skills add vercel-labs/skills@find-skills
-  $ assistant skills add vercel-labs/skills/find-skills --overwrite`,
+Manage skills from the Vellum catalog. Skills are instructions of
+workflows and tools that teach the assistant when dynamically relevant.`,
   subcommands: [
     {
       name: "list",
@@ -70,15 +57,25 @@ Examples:
         { flags: "--json", description: "Machine-readable JSON output" },
       ],
       helpText: `
+When to use:
+  Second stop after plugin search when the user asks to set up, install, or
+  add a capability (for example "Setup <app> for me", "add a skill for file management").
+  Search here before searching the web.
+
 Arguments:
   query    Free-text search term matched against skill names, descriptions,
-           and tags. Searches the Vellum catalog, the skills.sh community
+           and tags. Use the product or capability name, not the full user
+           sentence. Searches the Vellum catalog, the skills.sh community
            registry, and the clawhub registry.
 
 Displays results from all sources with clear labels. When a skill ID
 exists in both the Vellum catalog and a community registry, a conflict
 note is shown with guidance on which install command to use.
 
+If a catalog match is found, install it with 'assistant skills install <id>'.
+If a community match is found, add it with 'assistant skills add <source>'.
+If nothing matches, try web search.
+
 Examples:
   $ assistant skills search react
   $ assistant skills search "file management" --limit 3

package/src/cli/commands/trust.ts

@@ -5,6 +5,7 @@
  * IPC route (`trust_rules_list`).
  */
 
+import type { TrustRulesListIpcResponse } from "@vellumai/gateway-client";
 import type { Command } from "commander";
 
 import { cliIpcCall } from "../../ipc/cli-client.js";
@@ -13,18 +14,6 @@ import { registerCommand } from "../lib/register-command.js";
 import { log } from "../logger.js";
 import { trustHelp } from "./trust.help.js";
 
-// -- Types --------------------------------------------------------------------
-
-interface TrustRule {
-  id: string;
-  tool: string;
-  pattern: string;
-  risk: string;
-  origin: string;
-  userModified: boolean;
-  updatedAt: string;
-}
-
 // -- Registration -------------------------------------------------------------
 
 export function registerTrustCommand(program: Command): void {
@@ -44,7 +33,7 @@ export function registerTrustCommand(program: Command): void {
             ...(opts.tool ? { tool: opts.tool } : {}),
           };
 
-          const result = await cliIpcCall<{ rules: TrustRule[] }>(
+          const result = await cliIpcCall<TrustRulesListIpcResponse>(
             "trust_rules_list",
             { body: params },
           );
@@ -86,7 +75,7 @@ export function registerTrustCommand(program: Command): void {
             "MODIFIED",
           ];
 
-          const rows: string[][] = rules.map((r: TrustRule) => [
+          const rows: string[][] = rules.map((r) => [
             r.id.slice(0, 16),
             r.tool,
             r.pattern,

package/src/cli/lib/__tests__/install-from-github.test.ts

@@ -303,6 +303,44 @@ describe("installPlugin — install lifecycle", () => {
     expect(existsSync(join(target, "README.md"))).toBe(true);
   });
 
+  test("--force keeps live config.json instead of a pin-shipped default", async () => {
+    // GIVEN an install with user config, and a clone that materializes a default
+    const target = join(pluginsDir, "caveman");
+    mkdirSync(target, { recursive: true });
+    writeFileSync(join(target, "package.json"), '{"name":"caveman"}');
+    writeFileSync(
+      join(target, "config.json"),
+      '{"provider":"photon","ingressMode":"live"}\n',
+    );
+    mkdirSync(join(target, "data"), { recursive: true });
+    writeFileSync(join(target, "data", "cursor.json"), '{"n":1}\n');
+
+    const fetch = makeContentsFetch({ tree: {}, manifest: CAVEMAN_MANIFEST });
+    const runGit = fakeGitRunner({
+      tree: {
+        "package.json": '{"name":"caveman"}',
+        "README.md": "# caveman",
+        "config.json": '{"provider":"comms","ingressMode":"webhook"}\n',
+      },
+      commit: CAVEMAN_SHA,
+    });
+
+    // WHEN we reinstall with --force
+    await installPlugin(
+      { name: "caveman", force: true, ref: "main" },
+      { fetch, runGit, workspacePluginsDir: pluginsDir },
+    );
+
+    // THEN user-owned state survives and the pin's default config does not win
+    expect(readFileSync(join(target, "config.json"), "utf-8")).toBe(
+      '{"provider":"photon","ingressMode":"live"}\n',
+    );
+    expect(readFileSync(join(target, "data", "cursor.json"), "utf-8")).toBe(
+      '{"n":1}\n',
+    );
+    expect(existsSync(join(target, "README.md"))).toBe(true);
+  });
+
   test("commitOverride materializes a specific commit, keeping repo from the manifest", async () => {
     // GIVEN a manifest pinning the current SHA, but an override to an older one
     const OLD_SHA = "1".repeat(40);

package/src/cli/lib/__tests__/merge-plugin-tree.test.ts

@@ -440,4 +440,38 @@ describe("mergePluginTree", () => {
     expect(result.conflicts).toEqual(["f.txt"]);
     expect(result.binaryConflicts).toEqual([]);
   });
+
+  test("copies live config.json, data/, and .disabled from ours, ignoring the pin", async () => {
+    // GIVEN user-owned state on the live install and different defaults at the pin
+    writeTree(baseDir, { "package.json": '{"name":"p"}\n' });
+    writeTree(oursDir, {
+      "package.json": '{"name":"p"}\n',
+      "config.json": '{"provider":"photon","ingressMode":"live"}\n',
+      "data/cursor.json": '{"n":1}\n',
+      ".disabled": "",
+    });
+    writeTree(theirsDir, {
+      "package.json": '{"name":"p"}\n',
+      "config.json": '{"provider":"comms","ingressMode":"webhook"}\n',
+      "data/cursor.json": '{"n":0}\n',
+    });
+
+    // WHEN merged with theirs (the strategy that would otherwise take the pin)
+    const result = await mergePluginTree({
+      baseDir,
+      oursDir,
+      theirsDir,
+      destDir,
+      strategy: "theirs",
+    });
+
+    // THEN user-owned state is the live copy, not the pin, and is not counted
+    // as a merged source file
+    expect(read("config.json")).toBe(
+      '{"provider":"photon","ingressMode":"live"}\n',
+    );
+    expect(read("data/cursor.json")).toBe('{"n":1}\n');
+    expect(read(".disabled")).toBe("");
+    expect(result.fileCount).toBe(1);
+  });
 });

package/src/cli/lib/__tests__/plugin-surfaces.test.ts

@@ -17,7 +17,11 @@ import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { afterEach, beforeEach, describe, expect, test } from "bun:test";
 
-import { detectPluginSurfaces } from "../plugin-surfaces.js";
+import {
+  detectPluginSurfaces,
+  findPluginSetupSkill,
+  formatPluginSetupHint,
+} from "../plugin-surfaces.js";
 
 let pluginDir: string;
 
@@ -188,3 +192,30 @@ describe("detectPluginSurfaces", () => {
     expect(detectPluginSurfaces(pluginDir).schedules).toEqual([]);
   });
 });
+
+describe("findPluginSetupSkill", () => {
+  test("prefers the generic setup id when both conventions are present", () => {
+    expect(
+      findPluginSetupSkill("imessage", ["setup", "imessage-setup", "other"]),
+    ).toBe("setup");
+  });
+
+  test("falls back to <plugin-name>-setup", () => {
+    expect(findPluginSetupSkill("imessage", ["imessage-setup"])).toBe(
+      "imessage-setup",
+    );
+  });
+
+  test("returns undefined when the plugin ships no setup skill", () => {
+    expect(findPluginSetupSkill("imessage", ["imessage-send"])).toBeUndefined();
+    expect(findPluginSetupSkill("imessage", [])).toBeUndefined();
+  });
+});
+
+describe("formatPluginSetupHint", () => {
+  test("names the skill in the install-command copy", () => {
+    expect(formatPluginSetupHint("imessage-setup")).toBe(
+      "Load the imessage-setup skill to help set up this plugin",
+    );
+  });
+});

package/src/cli/lib/__tests__/toggle-plugin.test.ts

@@ -5,6 +5,7 @@ import {
   rmSync,
   writeFileSync,
 } from "node:fs";
+import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { afterEach, beforeEach, describe, expect, it } from "bun:test";
 
@@ -18,7 +19,7 @@ import {
 
 // VELLUM_WORKSPACE_DIR is set to this; getWorkspacePluginsDir() returns
 // <workspace>/plugins, so the actual plugins directory is TMP_WS_DIR/plugins.
-const TMP_WS_DIR = join(import.meta.dir, "__tmp__toggle-plugin-test");
+const TMP_WS_DIR = join(tmpdir(), `toggle-plugin-test-${process.pid}`);
 const TMP_PLUGINS_DIR = join(TMP_WS_DIR, "plugins");
 
 beforeEach(() => {

package/src/cli/lib/__tests__/upgrade-plugin.test.ts

@@ -920,6 +920,70 @@ describe("upgradePlugin --strategy", () => {
     expect(installedFile("level-up", "local-only.txt")).toBeNull();
   });
 
+  test("keeps live config.json, data/, and .disabled under overwrite, theirs, and ours", async () => {
+    // GIVEN user-owned state on the live install, and a pin that ships defaults
+    const userConfig = '{"provider":"photon","ingressMode":"live"}\n';
+    const pinConfig = '{"provider":"comms","ingressMode":"webhook"}\n';
+    const liveOurs: Tree = {
+      ...OURS,
+      "config.json": userConfig,
+      "data/cursor.json": '{"n":1}\n',
+      ".disabled": "",
+    };
+    const pinTheirs: Tree = {
+      ...THEIRS,
+      "config.json": pinConfig,
+      "data/cursor.json": '{"n":0}\n',
+    };
+
+    for (const strategy of ["overwrite", "theirs", "ours"] as const) {
+      rmSync(join(pluginsDir, "level-up"), { recursive: true, force: true });
+      installMergeCopy("level-up", liveOurs, SHA_A, BASE);
+      const fetch = makeFetch({ manifest: manifestWith("level-up", SHA_B) });
+      const runGit = treeGitRunner({
+        [SHA_A]: BASE,
+        [SHA_B]: pinTheirs,
+      });
+
+      // WHEN the plugin is upgraded
+      const result = await upgradePlugin(
+        { name: "level-up", strategy },
+        { fetch, runGit, workspacePluginsDir: pluginsDir },
+      );
+
+      // THEN user-owned state is unchanged, including under `theirs` and
+      // overwrite, which would otherwise take the pin wholesale
+      expect(result.outcome).toBe("upgraded");
+      expect(installedFile("level-up", "config.json")).toBe(userConfig);
+      expect(installedFile("level-up", "data/cursor.json")).toBe('{"n":1}\n');
+      expect(existsSync(join(pluginsDir, "level-up", ".disabled"))).toBe(true);
+    }
+  });
+
+  test("does not seed config.json from the pin when the live install has none", async () => {
+    // GIVEN an install with no config.json, and a pin that ships defaults
+    installMergeCopy("level-up", OURS, SHA_A, BASE);
+    const fetch = makeFetch({ manifest: manifestWith("level-up", SHA_B) });
+    const runGit = treeGitRunner({
+      [SHA_A]: BASE,
+      [SHA_B]: {
+        ...THEIRS,
+        "config.json": '{"provider":"comms","ingressMode":"webhook"}\n',
+      },
+    });
+
+    // WHEN upgraded (overwrite materializes the pin tree directly)
+    const result = await upgradePlugin(
+      { name: "level-up", strategy: "overwrite" },
+      { fetch, runGit, workspacePluginsDir: pluginsDir },
+    );
+
+    // THEN the pin's config.json does not land: host-owned config is created
+    // by the plugin at runtime, not by the installer
+    expect(result.outcome).toBe("upgraded");
+    expect(installedFile("level-up", "config.json")).toBeNull();
+  });
+
   test("defaults to overwrite when no strategy is given", async () => {
     // GIVEN an install with a local-only file and an advanced pin
     installMergeCopy("level-up", OURS, SHA_A, BASE);

package/src/cli/lib/bundled-marketplace.json

@@ -292,6 +292,19 @@
       "license": "MIT",
       "icon": "🎬"
     },
+    {
+      "name": "natural",
+      "source": {
+        "source": "github",
+        "repo": "vellum-ai/natural",
+        "ref": "3cfc8c97c14df311fc64992ab8b172c1b2d64eb0"
+      },
+      "description": "Hosted MCP access to Natural payment operations for Vellum assistants.",
+      "category": "finance",
+      "homepage": "https://www.natural.com",
+      "license": "MIT",
+      "icon": "💳"
+    },
     {
       "name": "originkit",
       "source": {

package/src/cli/lib/daemon-credential-client.ts

@@ -31,6 +31,8 @@ export interface DeleteSecureKeyResult {
   result: DeleteResult;
   /** Human-readable error reason when result="error". */
   error?: string;
+  /** Daemon error code when result="error" (e.g. `CREDENTIAL_IN_USE`). */
+  code?: string;
 }
 
 // ---------------------------------------------------------------------------
@@ -104,13 +106,18 @@ export async function setSecureKeyViaDaemon(
 /**
  * Delete a secret via the daemon IPC socket. Falls back to direct
  * `deleteSecureKeyAsync()` when the daemon is not running.
+ *
+ * `force` carries the caller's confirmation past the daemon's in-use guard,
+ * which refuses a delete while an LLM provider connection resolves its auth
+ * through the credential.
  */
 export async function deleteSecureKeyViaDaemon(
   type: string,
   name: string,
+  force = false,
 ): Promise<DeleteSecureKeyResult> {
   const ipc = await cliIpcCall<{ success: boolean }>("secrets_delete", {
-    body: { type, name },
+    body: force ? { type, name, force } : { type, name },
   });
 
   if (ipc.ok && ipc.result?.success) {
@@ -122,7 +129,9 @@ export async function deleteSecureKeyViaDaemon(
     if (ipc.error.includes("not found") || ipc.error.includes("404")) {
       return { result: "not-found" };
     }
-    return { result: "error", error: ipc.error };
+    return ipc.errorCode !== undefined
+      ? { result: "error", error: ipc.error, code: ipc.errorCode }
+      : { result: "error", error: ipc.error };
   }
 
   // Daemon returned success=false
@@ -133,7 +142,20 @@ export async function deleteSecureKeyViaDaemon(
     };
   }
 
-  // Daemon unreachable — fall back to direct delete.
+  // Daemon unreachable. The in-use guard runs daemon-side, next to the
+  // provider connection records it reads, so a direct delete here would
+  // remove a credential an LLM connection depends on with no check at all.
+  // The delete is refused until the caller either starts the assistant or
+  // states the intent explicitly.
+  if (!force) {
+    return {
+      result: "error",
+      code: "IN_USE_CHECK_UNAVAILABLE",
+      error:
+        "Cannot check whether LLM provider connections depend on this credential while the assistant is not running. Start the assistant, or re-run with --force to delete without the check.",
+    };
+  }
+
   if (type === "api_key") {
     // Delete from both locations; during migration overlap both may exist.
     const credResult = await deleteSecureKeyAsync(

package/src/cli/lib/install-from-github.ts

@@ -26,7 +26,6 @@
 import { execFile } from "node:child_process";
 import {
   copyFileSync,
-  cpSync,
   existsSync,
   mkdirSync,
   readdirSync,
@@ -40,7 +39,12 @@ import { tmpdir } from "node:os";
 import { dirname, join, resolve, sep } from "node:path";
 import { promisify } from "node:util";
 
-import { PRESERVED_ENTRIES } from "../../plugins/plugin-tree-walk.js";
+import {
+  copyPreservedUserState,
+  isPluginUserStateEntry,
+  PRESERVED_ENTRIES,
+  stripPreservedUserState,
+} from "../../plugins/plugin-tree-walk.js";
 import { ensureBun } from "../../util/bun-runtime.js";
 import { getWorkspacePluginsDir } from "../../util/platform.js";
 import type { FetchLike } from "./fetch-like.js";
@@ -687,27 +691,13 @@ export async function finalizeStagedInstall(
     }
   }
 
-  // Copy preserved entries (config.json, data/, .disabled) from the existing
-  // install into the staging dir before the swap so user-owned state survives
-  // upgrades and reinstalls. Without this, the rm+rename below would destroy
-  // user config and runtime data.
+  // User-owned state (config.json, data/, .disabled) is not plugin source.
+  // Drop any copy the pin materialized into staging, then copy the live
+  // install's bytes over so an upgrade cannot reset user config to pin
+  // defaults. Without the live copy, the rm+rename below would destroy it.
+  stripPreservedUserState(stagingDir);
   if (existsSync(target)) {
-    for (const entry of PRESERVED_ENTRIES) {
-      if (entry === INSTALL_META_FILENAME) {
-        continue;
-      } // sidecar is rewritten above
-      const src = join(target, entry);
-      if (!existsSync(src)) {
-        continue;
-      }
-      const dest = join(stagingDir, entry);
-      const stat = statSync(src);
-      if (stat.isDirectory()) {
-        cpSync(src, dest, { recursive: true });
-      } else {
-        copyFileSync(src, dest);
-      }
-    }
+    copyPreservedUserState(target, stagingDir);
     rmSync(target, { recursive: true, force: true });
   }
   renameSync(stagingDir, target);
@@ -1287,7 +1277,8 @@ function pluginPostinstallEnv(bun: string): NodeJS.ProcessEnv {
 
 /**
  * Recursively copy regular files from `srcRoot` into `destDir`, skipping the
- * top-level `.git` directory, a top-level `bunfig.toml` (see below), and any
+ * top-level `.git` directory, a top-level `bunfig.toml` (see below), top-level
+ * user-state entries (`config.json`, `data/`, `.disabled`), and any
  * symlinks. Returns the file count.
  */
 function copyTreeSkippingGit(srcRoot: string, destDir: string): number {
@@ -1300,6 +1291,12 @@ function copyTreeSkippingGit(srcRoot: string, destDir: string): number {
       if (relDir === "" && entry.name === ".git") {
         continue;
       }
+      // Runtime-owned state is not plugin source. A pin that ships
+      // `config.json` or `data/` must not land in the staged tree, or an
+      // upgrade would reset user config to whatever the revision committed.
+      if (relDir === "" && isPluginUserStateEntry(entry.name)) {
+        continue;
+      }
       // Drop a top-level `bunfig.toml`. The adapter postinstall runs `bun` with
       // its cwd at the staged root, and Bun auto-loads `$cwd/bunfig.toml` as
       // project config — including a `preload` list it executes before the
@@ -1712,6 +1709,7 @@ async function copyDir(
   ref: string,
   destDir: string,
   fetchFn: FetchLike,
+  atPluginRoot = true,
 ): Promise<number> {
   const entries = await listDir(owner, repo, apiPath, ref, fetchFn);
   if (entries === null) {
@@ -1725,11 +1723,24 @@ async function copyDir(
       continue;
     }
     assertSafeFilename("entry name", entry.name);
+    // Adapter stubs overlay the plugin root. User-state names at that root
+    // are host-owned, not adapter input.
+    if (atPluginRoot && isPluginUserStateEntry(entry.name)) {
+      continue;
+    }
 
     if (entry.type === "dir") {
       const subDest = join(destDir, entry.name);
       mkdirSync(subDest, { recursive: true });
-      count += await copyDir(owner, repo, entry.path, ref, subDest, fetchFn);
+      count += await copyDir(
+        owner,
+        repo,
+        entry.path,
+        ref,
+        subDest,
+        fetchFn,
+        false,
+      );
       continue;
     }
 

package/src/cli/lib/merge-plugin-tree.ts

@@ -23,6 +23,13 @@
  * one side) is resolved here before any line merge, since `git merge-file`
  * operates on three existing blobs.
  *
+ * Runtime-owned user state (`config.json`, `data/`, `.disabled`) is never
+ * merged. Those paths are excluded from the fingerprints and copied from the
+ * live install (`ours`) into the destination after the source merge, so a
+ * pin that ships default config cannot reset user settings. The provenance
+ * sidecar is excluded on every side and rewritten by the caller after the swap.
+ *
+ *
  * Binary files cannot be line-merged, so a binary file that diverged on both
  * sides is resolved whole-file by the strategy (`ours`/`theirs`); under
  * `assistant` the local copy is kept and the path is reported as a binary
@@ -44,7 +51,10 @@ import { tmpdir } from "node:os";
 import { dirname, join } from "node:path";
 import { promisify } from "node:util";
 
-import { PRESERVED_ENTRIES } from "../../plugins/plugin-tree-walk.js";
+import {
+  copyPreservedUserState,
+  PRESERVED_ENTRIES,
+} from "../../plugins/plugin-tree-walk.js";
 import { computeFingerprint } from "./plugin-fingerprint.js";
 
 const execFileAsync = promisify(execFile);
@@ -230,9 +240,11 @@ async function threeWayMergeFile(
 
 /**
  * Three-way merge `oursDir`/`theirsDir` against `baseDir` into `destDir` per
- * `strategy`. The provenance sidecar is excluded on every side — it is
- * rewritten by the caller after the swap and must not be carried through the
- * merge.
+ * `strategy`. User-owned state (`config.json`, `data/`, `.disabled`) is
+ * excluded from the line merge and copied from `oursDir` afterward, so the
+ * pin cannot reset it. The provenance sidecar is excluded on every side — it
+ * is rewritten by the caller after the swap and must not be carried through
+ * the merge.
  *
  * Returns the file count plus, for the `assistant` strategy, the paths left for
  * the assistant to resolve (text files with conflict markers and modify/delete
@@ -332,5 +344,11 @@ export async function mergePluginTree({
     // Present only in the base: removed on both sides, so it stays removed.
   }
 
+  // User-owned state is not plugin source, so it is never a merge input.
+  // Copy it from the live install so `--strategy theirs` cannot take a
+  // pin-shipped `config.json` (or drop an untracked one the fingerprint
+  // walk never saw).
+  copyPreservedUserState(oursDir, destDir);
+
   return { fileCount, conflicts, binaryConflicts };
 }

package/src/cli/lib/plugin-surfaces.ts

@@ -216,6 +216,33 @@ function listSchedules(schedulesDir: string): PluginScheduleSurface[] {
   );
 }
 
+/**
+ * Skill id a plugin ships for first-run setup, if any.
+ *
+ * Convention: `skills/setup/SKILL.md` or `skills/<plugin-name>-setup/SKILL.md`.
+ * The directory name is the id `skill_load` uses. `setup` wins when both exist
+ * so the generic name is the standard; `<plugin-name>-setup` covers plugins
+ * that already ship that form.
+ */
+export function findPluginSetupSkill(
+  pluginName: string,
+  skillIds: readonly string[],
+): string | undefined {
+  if (skillIds.includes("setup")) {
+    return "setup";
+  }
+  const named = `${pluginName}-setup`;
+  if (skillIds.includes(named)) {
+    return named;
+  }
+  return undefined;
+}
+
+/** Install-command copy pointing at a plugin's setup skill. */
+export function formatPluginSetupHint(skillId: string): string {
+  return `Load the ${skillId} skill to help set up this plugin`;
+}
+
 /**
  * Detect the {@link PluginSurfaces} an installed plugin contributes by walking
  * its install tree at `pluginDir`. Surface types with no contributions come

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",
     });
   });