@vellumai/assistant 0.11.2 → 0.11.3-staging.2

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

@@ -10,17 +10,29 @@
  *                                (see the external plugin loader's tool walk).
  * - `skills/<id>/SKILL.md`     → a skill owned by the plugin (see the skills
  *                                catalog's `discoverPluginResidentSkills`).
+ * - `schedules/<name>/`        → a declared schedule (see the daemon's plugin
+ *                                schedule declaration parser).
  *
  * This module re-derives those same sets so `plugins inspect` can report exactly
  * what a plugin contributes. Detection is intentionally a self-contained walk of
- * the install tree — `cli/lib` does not reach into the daemon-internal loader or
- * skills catalog — but it mirrors their conventions so inspect agrees with what
- * the runtime actually loads.
+ * the install tree (`cli/lib` does not reach into the daemon-internal loader,
+ * skills catalog, or schedule parser), but it mirrors their conventions so
+ * inspect agrees with what the runtime actually loads.
  */
 
-import { existsSync, readdirSync, statSync } from "node:fs";
+import { existsSync, readdirSync, readFileSync, statSync } from "node:fs";
 import { join } from "node:path";
 
+/** A schedule the plugin declares under `schedules/`. */
+export interface PluginScheduleSurface {
+  /** Schedule name: the declaration directory's name. */
+  readonly name: string;
+  /** Raw schedule `expression` string from the declaration's config. */
+  readonly cadence: string;
+  /** `execute` for a markdown prompt entrypoint, `script` for `index.sh`. */
+  readonly mode: "execute" | "script";
+}
+
 /** The surfaces an installed plugin contributes, each sorted and de-duplicated. */
 export interface PluginSurfaces {
   /** Skill ids shipped at `skills/<id>/SKILL.md`. */
@@ -36,6 +48,18 @@ export interface PluginSurfaces {
    * importing and executing untrusted plugin code, which inspection avoids.
    */
   readonly tools: readonly string[];
+  /**
+   * Schedules declared under `schedules/`, each a `<name>/` directory with a
+   * `config.json` plus exactly one `index.md` (`execute`) or `index.sh`
+   * (`script`) entrypoint. This is a display surface, not the arming path:
+   * a file directly under `schedules/` and structurally malformed directories
+   * (a bad entrypoint set, an empty prompt body, frontmatter in `index.md`, an
+   * unreadable config, a missing `expression`) are skipped rather than reported
+   * as errors. Detection stays permissive beyond structure: `expression`
+   * validity is not checked CLI-side, so a declaration whose expression the
+   * daemon rejects is still listed here.
+   */
+  readonly schedules: readonly PluginScheduleSurface[];
 }
 
 /**
@@ -97,6 +121,101 @@ function listSkillIds(skillsDir: string): string[] {
   return ids.sort();
 }
 
+/** Matches a `---` delimited YAML frontmatter block at the start of a file. */
+const FRONTMATTER_REGEX = /^---\r?\n([\s\S]*?)\r?\n---(?:\r?\n|$)/;
+
+/** Read the raw `expression` string from a parsed schedule config, or `null`. */
+function readConfigExpression(config: unknown): string | null {
+  if (typeof config !== "object" || config === null || Array.isArray(config)) {
+    return null;
+  }
+  const expression = (config as Record<string, unknown>).expression;
+  return typeof expression === "string" && expression.trim() !== ""
+    ? expression
+    : null;
+}
+
+/**
+ * Read a `<name>/` directory declaration: `config.json` supplies the cadence
+ * and the single `index.md`/`index.sh` entrypoint decides the mode. `null` for
+ * anything the loader would refuse (zero or multiple `index.*` entries, an
+ * unsupported entrypoint, an `index.md` carrying frontmatter or an empty
+ * prompt body, an unreadable config, a missing expression).
+ */
+function readDirectorySchedule(
+  dirPath: string,
+): { cadence: string; mode: PluginScheduleSurface["mode"] } | null {
+  let children;
+  try {
+    children = readdirSync(dirPath, { withFileTypes: true });
+  } catch {
+    return null;
+  }
+  const entrypoints = children
+    .filter((c) => c.isFile() && c.name.startsWith("index."))
+    .map((c) => c.name);
+  if (entrypoints.length !== 1) {
+    return null;
+  }
+  const entrypoint = entrypoints[0]!;
+  if (entrypoint !== "index.md" && entrypoint !== "index.sh") {
+    return null;
+  }
+  if (entrypoint === "index.md") {
+    let body: string;
+    try {
+      body = readFileSync(join(dirPath, "index.md"), "utf8");
+    } catch {
+      return null;
+    }
+    // Directory-form config belongs in config.json, so the loader refuses an
+    // index.md with frontmatter; it also refuses an empty prompt body.
+    if (FRONTMATTER_REGEX.test(body) || !body.trim()) {
+      return null;
+    }
+  }
+  let config: unknown;
+  try {
+    config = JSON.parse(readFileSync(join(dirPath, "config.json"), "utf8"));
+  } catch {
+    return null;
+  }
+  const cadence = readConfigExpression(config);
+  if (cadence === null) {
+    return null;
+  }
+  return { cadence, mode: entrypoint === "index.md" ? "execute" : "script" };
+}
+
+/**
+ * List the schedules a plugin declares under `schedules/`, mirroring the
+ * declaration parser's directory-only form. A file directly under
+ * `schedules/` is not a declaration, so it is skipped. Returns schedules
+ * sorted by name; a missing directory yields `[]`.
+ */
+function listSchedules(schedulesDir: string): PluginScheduleSurface[] {
+  let entries;
+  try {
+    entries = readdirSync(schedulesDir, { withFileTypes: true });
+  } catch {
+    return [];
+  }
+
+  const schedules: PluginScheduleSurface[] = [];
+  for (const entry of entries) {
+    if (entry.name.startsWith(".") || !entry.isDirectory()) {
+      continue;
+    }
+    const parsed = readDirectorySchedule(join(schedulesDir, entry.name));
+    if (parsed) {
+      schedules.push({ name: entry.name, ...parsed });
+    }
+  }
+  return schedules.sort((a, b) =>
+    a.name < b.name ? -1 : a.name > b.name ? 1 : 0,
+  );
+}
+
 /**
  * Detect the {@link PluginSurfaces} an installed plugin contributes by walking
  * its install tree at `pluginDir`. Surface types with no contributions come
@@ -110,5 +229,6 @@ export function detectPluginSurfaces(pluginDir: string): PluginSurfaces {
     skills: listSkillIds(join(pluginDir, "skills")),
     hooks: listModuleBasenames(join(pluginDir, "hooks")),
     tools: [...new Set(toolNames)].sort(),
+    schedules: listSchedules(join(pluginDir, "schedules")),
   };
 }

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

@@ -54,6 +54,8 @@ import {
   type PluginLocalInfo,
 } from "./inspect-plugin.js";
 import {
+  type ConfirmStagedInstall,
+  confirmStagedOrAbort,
   finalizeStagedInstall,
   type GitRunner,
   installPlugin,
@@ -138,6 +140,14 @@ export interface UpgradePluginDeps {
    * for dry runs or no-op upgrades (they never reach the swap).
    */
   readonly beforeSwap?: () => Promise<void>;
+  /**
+   * Consent gate between staging and finalize, identical to the install
+   * path's ({@link ConfirmStagedInstall}): the caller inspects the staged
+   * upgraded tree (e.g. its declared schedules) and may abort before
+   * anything replaces the live install. Applies to overwrite and merge
+   * upgrades alike; dry runs and no-ops never stage, so it is not invoked.
+   */
+  readonly confirmStaged?: ConfirmStagedInstall;
 }
 
 /** Result of an upgrade attempt. */
@@ -401,6 +411,7 @@ export async function upgradePlugin(
       runPostinstall: deps.runPostinstall,
       runInstallDeps: deps.runInstallDeps,
       beforeSwap: deps.beforeSwap,
+      confirmStaged: deps.confirmStaged,
     },
   );
 
@@ -591,6 +602,7 @@ async function directUpgrade(
       runPostinstall: deps.runPostinstall,
       runInstallDeps: deps.runInstallDeps,
       beforeSwap: deps.beforeSwap,
+      confirmStaged: deps.confirmStaged,
     },
   );
 
@@ -747,6 +759,12 @@ async function mergeUpgrade(
       conflictLabels,
     });
 
+    // Same staging-to-finalize consent boundary as `installPlugin`: the
+    // caller sees the fully merged tree before it replaces the live install.
+    // A decline removes the staging dir and throws; the outer catch's rmSync
+    // is then a no-op.
+    await confirmStagedOrAbort(name, stagingDir, deps.confirmStaged);
+
     const toCommit = theirs.commit ?? ctx.toCommit;
     const toTimestamp = theirs.committedAt ?? ctx.toTimestamp;
     await finalizeStagedInstall(stagingDir, {

package/src/cli.ts

@@ -513,7 +513,7 @@ export async function startCli(): Promise<void> {
     if (content === "/conversations") {
       pendingSessionPick = true;
       try {
-        const rows = listConversations(20);
+        const rows = listConversations({ limit: 20 });
         const conversations = rows.map((r) => ({
           id: r.id,
           title: r.title || "Untitled",

package/src/config/__tests__/deployment-context-defaults.test.ts

@@ -1,6 +1,7 @@
 import {
   existsSync,
   mkdirSync,
+  readdirSync,
   readFileSync,
   rmSync,
   writeFileSync,
@@ -267,4 +268,76 @@ describe("deployment-context embedding-provider default (via loadConfig)", () =>
     >;
     expect(v3Raw.live).toBe(true);
   });
+
+  describe("quarantine-reseed carries memory.v3.live forward", () => {
+    function removeQuarantineFiles(): void {
+      for (const name of readdirSync(WORKSPACE_DIR)) {
+        if (name.startsWith("config.json.corrupt-")) {
+          rmSync(join(WORKSPACE_DIR, name), { force: true });
+        }
+      }
+    }
+
+    function readV3Raw(): Record<string, unknown> {
+      const raw = readConfig();
+      return ((raw.memory as Record<string, unknown>).v3 ?? {}) as Record<
+        string,
+        unknown
+      >;
+    }
+
+    afterEach(() => {
+      removeQuarantineFiles();
+    });
+
+    test("a corrupt live-v3 config is quarantined and the reseed keeps live=true", () => {
+      // Truncated mid-write JSON from a memory-v3 assistant: unparseable,
+      // but the live flag survives in the raw text.
+      writeFileSync(
+        CONFIG_PATH,
+        `{\n  "memory": {\n    "v3": {\n      "live": true\n    }\n  },\n  "llm": {\n    "activeProf`,
+      );
+
+      const config = loadConfig();
+
+      // The corrupt file was quarantined and the reseeded config carries the
+      // tier forward, both in-memory (this boot runs v3) and on disk
+      // (migration 105 is checkpointed on existing workspaces and never
+      // re-runs, so the persisted value is the only durable record).
+      expect(config.memory.v3.live).toBe(true);
+      expect(readV3Raw()).toEqual({ live: true });
+    });
+
+    test("a corrupt config without live=true reseeds with the leaf absent", () => {
+      writeFileSync(
+        CONFIG_PATH,
+        `{\n  "memory": {\n    "v3": {\n      "live": false\n    }\n  },\n  "llm": {\n    "activeProf`,
+      );
+
+      const config = loadConfig();
+
+      // Never carry `false` forward: an absent leaf is the only way to keep
+      // "no decision recorded" representable for migration 105.
+      expect(config.memory.v3.live).toBe(false);
+      expect(readV3Raw()).toEqual({});
+    });
+
+    test("a deleted config with an earlier live-v3 quarantine still carries live=true", () => {
+      if (existsSync(CONFIG_PATH)) {
+        rmSync(CONFIG_PATH, { force: true });
+      }
+      writeFileSync(
+        join(
+          WORKSPACE_DIR,
+          "config.json.corrupt-2026-08-01T10-00-00.000Z.json",
+        ),
+        JSON.stringify({ memory: { v3: { live: true } } }, null, 2),
+      );
+
+      const config = loadConfig();
+
+      expect(config.memory.v3.live).toBe(true);
+      expect(readV3Raw()).toEqual({ live: true });
+    });
+  });
 });

package/src/config/__tests__/memory-retrospective-schema.test.ts

@@ -0,0 +1,40 @@
+/**
+ * Guards over the `memory.retrospective` switches that decide whether the
+ * retrospective pass runs at all, and how its fork is materialized.
+ *
+ * The defaults are the contract: a workspace that says nothing must behave
+ * exactly as it did before the fields existed: retrospectives on, forks
+ * cloned.
+ */
+
+import { describe, expect, test } from "bun:test";
+
+import { MemoryRetrospectiveConfigSchema } from "../schemas/memory-retrospective.js";
+
+describe("memory.retrospective config schema", () => {
+  test("an empty block leaves retrospectives on and forks cloning", () => {
+    const parsed = MemoryRetrospectiveConfigSchema.parse({});
+    expect(parsed.enabled).toBe(true);
+    expect(parsed.forkStrategy).toBe("cloning");
+  });
+
+  test("enabled is a boolean-only kill switch", () => {
+    expect(
+      MemoryRetrospectiveConfigSchema.parse({ enabled: false }).enabled,
+    ).toBe(false);
+    expect(
+      MemoryRetrospectiveConfigSchema.safeParse({ enabled: "false" }).success,
+    ).toBe(false);
+  });
+
+  test("forkStrategy accepts only cloning and reference", () => {
+    expect(
+      MemoryRetrospectiveConfigSchema.parse({ forkStrategy: "reference" })
+        .forkStrategy,
+    ).toBe("reference");
+    expect(
+      MemoryRetrospectiveConfigSchema.safeParse({ forkStrategy: "referential" })
+        .success,
+    ).toBe(false);
+  });
+});

package/src/config/__tests__/plugin-updates-schema.test.ts

@@ -0,0 +1,43 @@
+/**
+ * Guards over the `pluginUpdates` config block — the opt-in that decides
+ * whether installed plugins move on their own.
+ *
+ * The defaults are the contract: a workspace that says nothing must behave
+ * exactly as it did before the block existed (manual, nothing upgrades), and
+ * an opted-in workspace must land on the `theirs` merge strategy.
+ */
+
+import { describe, expect, test } from "bun:test";
+
+import { PluginUpdatesConfigSchema } from "../schemas/plugin-updates.js";
+
+describe("pluginUpdates config schema", () => {
+  test("defaults are manual / theirs / hourly", () => {
+    expect(PluginUpdatesConfigSchema.parse({})).toEqual({
+      mode: "manual",
+      strategy: "theirs",
+      checkIntervalMs: 3_600_000,
+    });
+  });
+
+  test("mode accepts only manual and auto", () => {
+    expect(PluginUpdatesConfigSchema.parse({ mode: "auto" }).mode).toBe("auto");
+    expect(
+      PluginUpdatesConfigSchema.safeParse({ mode: "sometimes" }).success,
+    ).toBe(false);
+  });
+
+  test("the unattended-hostile `assistant` strategy is not selectable", () => {
+    // It resolves conflicts by writing git conflict markers into the plugin's
+    // files, which unattended would leave broken code on disk.
+    expect(
+      PluginUpdatesConfigSchema.safeParse({ strategy: "assistant" }).success,
+    ).toBe(false);
+  });
+
+  test("sub-5-minute sweep intervals are rejected", () => {
+    expect(
+      PluginUpdatesConfigSchema.safeParse({ checkIntervalMs: 1_000 }).success,
+    ).toBe(false);
+  });
+});

package/src/config/__tests__/profile-tool-support.test.ts

@@ -0,0 +1,131 @@
+import { describe, expect, test } from "bun:test";
+
+import { profileSupportsTools } from "../profile-tool-support.js";
+import type { AssistantConfig } from "../schema.js";
+import { LLMSchema } from "../schemas/llm.js";
+
+/** An `AssistantConfig` carrying nothing the profile lookup does not read. */
+function configWithProfiles(
+  profiles: Record<string, Record<string, unknown>>,
+): AssistantConfig {
+  return { llm: LLMSchema.parse({ profiles }) } as AssistantConfig;
+}
+
+describe("profileSupportsTools", () => {
+  test("reports false only for a model the catalog declares tool-less", () => {
+    const config = configWithProfiles({
+      "no-tools": { provider: "openrouter", model: "minimax/minimax-01" },
+    });
+    expect(profileSupportsTools("no-tools", config)).toBe(false);
+  });
+
+  test("reports true for a catalog model declared tool-capable", () => {
+    const config = configWithProfiles({
+      "with-tools": { provider: "anthropic", model: "claude-fable-5" },
+    });
+    expect(profileSupportsTools("with-tools", config)).toBe(true);
+  });
+
+  test("fails open for a BYOK model the catalog has never seen", () => {
+    const config = configWithProfiles({
+      byok: { provider: "openrouter", model: "acme/private-llm-9" },
+    });
+    expect(profileSupportsTools("byok", config)).toBeUndefined();
+  });
+
+  test("fails open for a profile key that resolves to nothing", () => {
+    expect(profileSupportsTools("not-a-profile", configWithProfiles({}))).toBe(
+      undefined,
+    );
+  });
+
+  test("verdicts a catalog default profile through the code catalog", () => {
+    // `balanced` carries no workspace body, so the answer has to come from the
+    // resolved default-profile catalog rather than `llm.profiles`.
+    expect(profileSupportsTools("balanced", configWithProfiles({}))).toBe(true);
+  });
+
+  test("reports false for a mix whose every arm is catalog-denied", () => {
+    const config = configWithProfiles({
+      "no-tools": { provider: "openrouter", model: "minimax/minimax-01" },
+      "no-tools-too": {
+        provider: "openrouter",
+        model: "minimax/minimax-m2-her",
+      },
+      blend: {
+        mix: [
+          { profile: "no-tools", weight: 1 },
+          { profile: "no-tools-too", weight: 1 },
+        ],
+      },
+    });
+    // Whichever arm the child's seed lands on, it cannot call tools.
+    expect(profileSupportsTools("blend", config)).toBe(false);
+  });
+
+  test("reports true for a mix whose every arm is catalog-capable", () => {
+    const config = configWithProfiles({
+      "tools-a": { provider: "anthropic", model: "claude-fable-5" },
+      "tools-b": {
+        provider: "anthropic",
+        model: "claude-haiku-4-5-20251001",
+      },
+      blend: {
+        mix: [
+          { profile: "tools-a", weight: 3 },
+          { profile: "tools-b", weight: 1 },
+        ],
+      },
+    });
+    expect(profileSupportsTools("blend", config)).toBe(true);
+  });
+
+  test("fails open for a mix of tool-capable and tool-less arms", () => {
+    const config = configWithProfiles({
+      "no-tools": { provider: "openrouter", model: "minimax/minimax-01" },
+      "with-tools": { provider: "anthropic", model: "claude-fable-5" },
+      blend: {
+        mix: [
+          { profile: "no-tools", weight: 1 },
+          { profile: "with-tools", weight: 1 },
+        ],
+      },
+    });
+    // The arm is picked from the child conversation's seed, and no such
+    // conversation exists yet, so neither arm's verdict speaks for the mix.
+    expect(profileSupportsTools("blend", config)).toBeUndefined();
+  });
+
+  test("fails open for a mix with an arm the catalog has never seen", () => {
+    const config = configWithProfiles({
+      "no-tools": { provider: "openrouter", model: "minimax/minimax-01" },
+      byok: { provider: "openrouter", model: "acme/private-llm-9" },
+      blend: {
+        mix: [
+          { profile: "no-tools", weight: 1 },
+          { profile: "byok", weight: 1 },
+        ],
+      },
+    });
+    expect(profileSupportsTools("blend", config)).toBeUndefined();
+  });
+
+  test("answers a mixed-capability mix identically on every call", () => {
+    const config = configWithProfiles({
+      "no-tools": { provider: "openrouter", model: "minimax/minimax-01" },
+      "with-tools": { provider: "anthropic", model: "claude-fable-5" },
+      blend: {
+        mix: [
+          { profile: "no-tools", weight: 1 },
+          { profile: "with-tools", weight: 1 },
+        ],
+      },
+    });
+    // No seed is available at probe time, so the verdict must not ride on a
+    // random arm pick.
+    const verdicts = new Set(
+      Array.from({ length: 50 }, () => profileSupportsTools("blend", config)),
+    );
+    expect([...verdicts]).toEqual([undefined]);
+  });
+});

package/src/config/bundled-skills/AGENTS.md

@@ -8,7 +8,7 @@ Each executor is a module exporting `run(input, context)` (the `SkillToolScript`
 
 ## Keeping public docs in sync
 
-A bundled skill's `SKILL.md` is its behavioral source of truth. Skills with a public reference page (`https://www.vellum.ai/docs/skills-reference/<slug>`, authored in the **vellum-assistant-platform** repo) are fingerprinted in `scripts/skill-docs-sync.manifest.json` and enforced by `skill-docs-sync-guard.test.ts`.
+A bundled skill's `SKILL.md` is its behavioral source of truth. Skills with a public reference page (`https://www.vellum.ai/docs/skills-reference/<slug>`, authored in this repo at `clients/docs/src/app/docs/_components/skills-reference-<slug>-content.tsx`) are fingerprinted in `scripts/skill-docs-sync.manifest.json` and enforced by `skill-docs-sync-guard.test.ts`.
 
 When you change a documented skill's `SKILL.md`, the guard fails until you reconcile it:
 

package/src/config/bundled-skills/app-builder/SKILL.md

@@ -75,7 +75,7 @@ Every app is multi-file TSX (`formatVersion: 2`). Never write a root-level `inde
 
 ## Responsive & design system
 
-Every app works phone (~360px) to desktop (~1400px+). The `<turn_context>` block carries a `client_os:` field (the OS the user is on): `ios`/`android` → mobile-first (design narrow first, body 17px); `macos`/`web` → desktop-first (multi-column, body 14px); absent → desktop-first unless the request implies phone use ("for my iPhone"). (The sibling `interface:` field is the transport surface — always `web` for the apps — so key layout off `client_os`, not `interface`.)
+Every app works phone (~360px) to desktop (~1400px+). The `<turn_context>` block carries a `client_os:` field (the OS the user is on): `ios`/`android` → mobile-first (design narrow first, body 17px); `macos`/`windows`/`web` → desktop-first (multi-column, body 14px); absent → desktop-first unless the request implies phone use ("for my iPhone"). (The sibling `interface:` field is the transport surface, always `web` for the apps, so key layout off `client_os`, not `interface`.)
 
 **Universal baseline — every build, regardless of client_os:**
 - Viewport meta: `width=device-width, initial-scale=1, viewport-fit=cover`. Never `user-scalable=no` (blocks accessibility zoom).

package/src/config/bundled-skills/app-builder/references/RESPONSIVE.md

@@ -4,12 +4,12 @@ Every app must work across phone (~360px) to desktop (~1400px+).
 
 ## Mode selection
 
-The conversation context's `<turn_context>` block carries a `client_os:` field — the OS the user is on. (Key layout off this, not the sibling `interface:` field: `interface:` is the transport surface and is always `web` for the web/iOS/macOS apps.)
+The conversation context's `<turn_context>` block carries a `client_os:` field, the OS the user is on. (Key layout off this, not the sibling `interface:` field: `interface:` is the transport surface and is always `web` for the web/iOS/macOS/Windows apps.)
 
 **If `client_os: ios` or `android`** (any mobile-web surface):
 → Mobile-first build. Design the narrow viewport first, enhance upward.
 
-**If `client_os: macos` or `web`**:
+**If `client_os: macos`, `windows`, or `web`**:
 → Desktop-first build. Design larger composition first; narrow fallback still meets the universal baseline below.
 
 **If field is absent or ambiguous**:
@@ -52,6 +52,6 @@ The conversation context's `<turn_context>` block carries a `client_os:` field
 - Bottom-anchor the primary action: `position: sticky; bottom: env(safe-area-inset-bottom)`.
 - Replace side modals and popovers with bottom sheets.
 
-## Desktop-first priorities (`client_os: macos` / `web`)
+## Desktop-first priorities (`client_os: macos` / `windows` / `web`)
 
 Multi-column composition, hover-rich affordances, denser information, side modals, inline primary actions. The universal baseline above is still the floor — narrow view must still work.

package/src/config/bundled-skills/schedule/SKILL.md

@@ -121,7 +121,11 @@ assistant conversations wake "$id" --hint "Summarize the new items" --external-c
 
 ## Inference Profile
 
-Execute-mode runs use the default `mainAgent` model selection unless the schedule pins an `inference_profile` (a key from `llm.profiles`). Pin a profile when a recurring task should run on a specific model — e.g. a cost-optimized profile for a high-frequency digest. Pass `inference_profile: null` on update to revert to the default. The pinned profile is shown on the schedule's details page in settings.
+Every schedule carries a pinned `inference_profile` (a key from `llm.profiles`), so a recurring task keeps running on the model (and the price) it was created under even when the user later changes their default profile. Creating a schedule without `inference_profile` pins it to the user's current default; pass one explicitly when a task should run on a specific model, e.g. a cost-optimized profile for a high-frequency digest. Passing `inference_profile: null` on update re-pins the schedule to the user's current default rather than unpinning it. The pinned profile is shown on the schedule's details page in settings.
+
+The pin governs the agent turns a schedule runs: an execute-mode run, a wake-mode run, and the turns a script-mode schedule hands off to the agent loop with `assistant conversations wake`. The schedule id travels in the script's environment, so those handoff turns run on the pinned profile with no flag to pass. Notify-mode schedules run no agent turn at all, so their pin never comes into play.
+
+Workflow-mode schedules are the exception: each step of a workflow resolves its own model, so a workflow schedule's pin does not govern its runs.
 
 ## Conversation Group
 

package/src/config/bundled-skills/schedule/TOOLS.json

@@ -110,7 +110,7 @@
           },
           "inference_profile": {
             "type": "string",
-            "description": "Inference profile (a key from llm.profiles) the schedule's LLM-executed runs should use, e.g. to run a heavy task on a cheaper model. When omitted, runs use the default mainAgent model selection. Must name a configured profile."
+            "description": "Inference profile (a key from llm.profiles) the schedule's LLM-executed runs should use, e.g. to run a heavy task on a cheaper model. When omitted, the schedule is pinned to the user's current default profile so its model does not change later. Must name a configured profile. Workflow-mode schedules resolve a model per workflow step, so the pin does not govern their runs."
           },
           "group": {
             "type": "string",
@@ -233,7 +233,7 @@
           },
           "inference_profile": {
             "type": ["string", "null"],
-            "description": "Inference profile (a key from llm.profiles) the schedule's LLM-executed runs should use. Pass null to clear the override and revert to the default mainAgent model selection. Must name a configured profile."
+            "description": "Inference profile (a key from llm.profiles) the schedule's LLM-executed runs should use. Pass null to re-pin the schedule to the user's current default profile. Must name a configured profile. Workflow-mode schedules resolve a model per workflow step, so the pin does not govern their runs."
           },
           "group": {
             "type": ["string", "null"],

package/src/config/bundled-skills/settings/TOOLS.json

@@ -3,7 +3,7 @@
   "tools": [
     {
       "name": "voice_config_update",
-      "description": "Update a voice configuration setting. Use tts_provider / stt_provider to switch the active TTS / STT provider. Provider \"vellum\" is Vellum-managed speech (billed to your organization; requires a Vellum platform connection via 'assistant platform connect'); any other provider uses the user's own API key. Valid TTS providers come from the provider catalog (vellum, elevenlabs, fish-audio, deepgram, xai); valid STT providers: vellum, deepgram, google-gemini, openai-whisper, xai. Use tts_voice_id to change the voice; it targets whichever TTS provider is currently active. For elevenlabs, pass an ElevenLabs voice ID. For vellum (managed), pass a managed voice model ID: this may be an ElevenLabs voice ID (managed speech serves the same ElevenLabs voices) or a Deepgram Aura model ID (e.g. aura-2-thalia-en); only rate-carded voices synthesize, so prefer models from the managed catalog (daemon route GET tts/managed-voices, also used by the web voice picker). For deepgram, pass a Deepgram Aura model ID. Use fish_audio_reference_id for Fish Audio voice reference. Use stt_language to set the spoken language for speech recognition: one of the 49 base language codes on the verified Deepgram nova-3 monolingual roster (e.g. en, es, hi, ta, zh, ko), or \"multi\" for code-switching mid-sentence across its 10-language roster (English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch; e.g. Hinglish); plain language names like \"tamil\" or \"multilingual\" are accepted and normalized. Accepted values follow the configured STT provider: vellum-managed and deepgram accept the full roster plus \"multi\"; xai accepts only the 10 multilingual-roster codes (en, es, fr, de, hi, ru, pt, ja, it, nl) and rejects \"multi\" and the extended codes (both are verified for Deepgram nova-3 only); google-gemini / openai-whisper auto-detect natively, so the value persists but is ignored while they are active. Deepgram uses the same API key for TTS and STT. Changes persist to services.stt / services.tts config and take effect immediately.",
+      "description": "Update a voice configuration setting. Use tts_provider / stt_provider to switch the active TTS / STT provider. Provider \"vellum\" is Vellum-managed speech (billed to your organization; requires a Vellum platform connection via 'assistant platform connect'); any other provider uses the user's own API key. Valid TTS providers come from the provider catalog (vellum, elevenlabs, fish-audio, deepgram, xai); valid STT providers: vellum, deepgram, google-gemini, openai-whisper, xai. Use tts_voice_id to change the voice; it targets whichever TTS provider is currently active. For elevenlabs, pass an ElevenLabs voice ID. For vellum (managed), pass a managed voice model ID: this may be an ElevenLabs voice ID (managed speech serves the same ElevenLabs voices) or a Deepgram Aura model ID (e.g. aura-2-thalia-en); only rate-carded voices synthesize, so prefer models from the managed catalog (daemon route GET tts/managed-voices, also used by the web voice picker). For deepgram, pass a Deepgram Aura model ID. Use fish_audio_reference_id for Fish Audio voice reference. Use stt_language to set the spoken language for speech recognition: one of the 50 base language codes on the verified Deepgram nova-3 monolingual roster (e.g. en, es, hi, ta, zh, ko), or \"multi\" for code-switching mid-sentence across its 10-language roster (English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch; e.g. Hinglish); plain language names like \"tamil\" or \"multilingual\" are accepted and normalized. Accepted values follow the configured STT provider: vellum-managed and deepgram accept the full roster plus \"multi\"; xai accepts only the 10 multilingual-roster codes (en, es, fr, de, hi, ru, pt, ja, it, nl) and rejects \"multi\" and the extended codes (both are verified for Deepgram nova-3 only); google-gemini / openai-whisper auto-detect natively, so the value persists but is ignored while they are active. Deepgram uses the same API key for TTS and STT. Changes persist to services.stt / services.tts config and take effect immediately.",
       "category": "system",
       "risk": "low",
       "input_schema": {
@@ -23,7 +23,7 @@
             "description": "The voice setting to change. tts_provider / stt_provider select the active provider for each service (\"vellum\" = Vellum-managed speech; anything else = the user's own API key). tts_voice_id sets the voice for the currently active TTS provider (ElevenLabs voice ID for elevenlabs; a managed voice model ID for vellum, meaning an ElevenLabs voice ID or a Deepgram Aura model ID like aura-2-thalia-en; a Deepgram Aura model ID for deepgram; voice ID for xai). fish_audio_reference_id sets the Fish Audio voice reference. stt_language sets the spoken language for speech recognition (vellum-managed/deepgram: the full roster plus \"multi\"; xai: only the 10 multilingual-roster codes, with \"multi\" and the extended codes rejected; google-gemini and openai-whisper auto-detect and ignore it). Deepgram shares one API key across TTS and STT."
           },
           "value": {
-            "description": "The new value for the setting. For tts_provider: one of vellum, elevenlabs, fish-audio, deepgram, xai. For stt_provider: one of vellum, deepgram, google-gemini, openai-whisper, xai. For tts_voice_id: a voice ID for the active TTS provider: an alphanumeric ElevenLabs voice ID (elevenlabs), a managed voice model ID which may be an ElevenLabs voice ID or a Deepgram Aura model ID like aura-2-thalia-en (vellum), or a Deepgram Aura model ID (deepgram). For fish_audio_reference_id: a Fish Audio voice reference ID. For stt_language: one of the 49 base codes on the nova-3 monolingual roster (e.g. en, es, hi, ta, zh, ko), or multi (code-switching across its 10-language roster); language names like \"hindi\", \"tamil\", or \"multilingual\" are also accepted and normalized; when the configured STT provider is xai, only the 10 multilingual-roster codes (en, es, fr, de, hi, ru, pt, ja, it, nl) are accepted. For conversation_timeout: seconds (5, 10, 15, 30, or 60). For activation_key: key identifier string."
+            "description": "The new value for the setting. For tts_provider: one of vellum, elevenlabs, fish-audio, deepgram, xai. For stt_provider: one of vellum, deepgram, google-gemini, openai-whisper, xai. For tts_voice_id: a voice ID for the active TTS provider: an alphanumeric ElevenLabs voice ID (elevenlabs), a managed voice model ID which may be an ElevenLabs voice ID or a Deepgram Aura model ID like aura-2-thalia-en (vellum), or a Deepgram Aura model ID (deepgram). For fish_audio_reference_id: a Fish Audio voice reference ID. For stt_language: one of the 50 base codes on the nova-3 monolingual roster (e.g. en, es, hi, ta, zh, ko), or multi (code-switching across its 10-language roster); language names like \"hindi\", \"tamil\", or \"multilingual\" are also accepted and normalized; when the configured STT provider is xai, only the 10 multilingual-roster codes (en, es, fr, de, hi, ru, pt, ja, it, nl) are accepted. For conversation_timeout: seconds (5, 10, 15, 30, or 60). For activation_key: key identifier string."
           }
         }
       },

package/src/config/bundled-skills/settings/tools/voice-config-update.ts

@@ -179,6 +179,7 @@ const STT_LANGUAGE_ALIASES: Record<
   thai: "th",
   turkish: "tr",
   ukrainian: "uk",
+  urdu: "ur",
   vietnamese: "vi",
   multilingual: "multi",
   auto: "multi",