@vellumai/assistant 0.9.1-staging.1 → 0.10.0-staging.2

package/src/config/bundled-skills/contacts/tools/contact-search.ts

@@ -18,7 +18,6 @@ function formatContactSummary(c: ContactWithChannels): string {
       .map((ch) => {
         let s = `${ch.type}:${ch.address}${ch.isPrimary ? "*" : ""}`;
         const extras: string[] = [];
-        if (ch.externalUserId) extras.push(`userId: ${ch.externalUserId}`);
         if (ch.externalChatId) extras.push(`chatId: ${ch.externalChatId}`);
         if (extras.length > 0) s += ` (${extras.join(", ")})`;
         return s;

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

@@ -28,6 +28,8 @@ When the user mentions "email" - sending, reading, checking, decluttering, draft
 
 Do not offer the assistant's own email as an option unless the user specifically asks. If Gmail and Outlook are not connected, guide them through setup.
 
+Reading, searching, or summarizing the user's inbox is a **messaging task** — use the messaging tools (or the Gmail connection flow below if nothing is connected). Never run `assistant channels` commands for a mailbox request: channels are the assistant's own inbound delivery routes, not the user's mailbox, and inspecting them sends you down the wrong path.
+
 When a platform is connected (auth test succeeds), always use the messaging API tools for that platform. Never fall back to browser automation, shell commands (bash, curl), or any other approach for operations that messaging tools can handle. The messaging tools handle authentication internally - never try to access tokens or call APIs directly. Browser automation is only appropriate for initial credential setup (OAuth consent screens), not for day-to-day messaging operations.
 
 **Exception: Slack.** Slack messaging should use the Slack Web API directly via CLI, not messaging tools. See the **slack** skill for details.
@@ -50,13 +52,13 @@ When the user asks to "connect my email", "set up email", "manage my email", or
 
 ### Gmail
 
-1. **Try connecting directly first.** Run `assistant oauth status google`. This will show whether or not the user had previously connected their google account. If so, they are ready to go.
-2. **If no connections are found:** Call `skill_load` with `skill: "vellum-oauth-integrations"`. The skill will evaluate whether managed or your-own mode is appropriate and guide the user accordingly.
+1. **Check the connection.** Run `assistant oauth status google`. If a connection exists, the user is ready to go — proceed with their request.
+2. **If no connections are found, render the connect button in one step:** call `ui_show` with `surface_type: "oauth_connect"` and `data.providerKey: "google"`. That surface is always available — do **not** run `assistant channels`, `oauth providers get`, or load `vellum-oauth-integrations` just to display the button, and do **not** fall back to `assistant oauth connect`. Only load `vellum-oauth-integrations` when the managed-vs-your-own-credentials decision genuinely needs handling (e.g. a BYOK setup where the managed connect surface isn't available).
 
 ### Outlook
 
-1. **Try connecting directly first.** Run `assistant oauth status outlook`. This will show whether the user has previously connected their Outlook account.
-2. **If no connections are found:** Call `skill_load` with `skill: "vellum-oauth-integrations"`. The skill will evaluate whether managed or your-own mode is appropriate and guide the user accordingly.
+1. **Check the connection.** Run `assistant oauth status outlook`. If a connection exists, the user is ready to go — proceed with their request.
+2. **If no connections are found, render the connect button in one step:** call `ui_show` with `surface_type: "oauth_connect"` and `data.providerKey: "outlook"`. The same rules as Gmail apply — don't probe further or load a setup skill just to show the button; only load `vellum-oauth-integrations` when a managed-vs-your-own-credentials decision genuinely needs handling.
 
 ### Slack
 

package/src/config/bundled-skills/messaging/tools/messaging-archive-by-sender.ts

@@ -1,3 +1,4 @@
+import { isArchiveBySenderAuthorized } from "../../../../runtime/effective-capabilities.js";
 import type {
   ToolContext,
   ToolExecutionResult,
@@ -8,14 +9,14 @@ export async function run(
   input: Record<string, unknown>,
   context: ToolContext,
 ): Promise<ToolExecutionResult> {
-  const userApproved =
-    input.user_approved === true && context.trustClass === "guardian";
-  if (
-    !context.triggeredBySurfaceAction &&
-    !context.batchAuthorizedByTask &&
-    !context.approvedViaPrompt &&
-    !userApproved
-  ) {
+  const authorized = isArchiveBySenderAuthorized({
+    trustClass: context.trustClass,
+    triggeredBySurfaceAction: context.triggeredBySurfaceAction,
+    batchAuthorizedByTask: context.batchAuthorizedByTask,
+    approvedViaPrompt: context.approvedViaPrompt,
+    userApproved: input.user_approved === true,
+  });
+  if (!authorized) {
     return err(
       "This tool requires either a surface action or a scheduled task run with this tool in required_tools. Present results in a selection table with action buttons and wait for the user to click before proceeding.",
     );

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

@@ -1,18 +1,22 @@
 ---
 name: workflows
-description: Author and run autonomous multi-agent workflows that fan work across parallel leaf agents
+description: Delegate a big or high-stakes job to a fleet of parallel subagents, orchestrated deterministically; runs unattended and reports back
 compatibility: "Designed for Vellum personal assistants"
 metadata:
   emoji: "⚙️"
   vellum:
     display-name: "Workflows"
     category: "system"
+    always-candidate: true
     activation-hints:
-      - "A task decomposes into many similar sub-tasks that can run concurrently (score every item, extract a field from each of many documents, draft-then-verify a batch)"
-      - "You want fan-out orchestrated deterministically and the result reported back when the whole run finishes"
+      - "Batch — apply one operation to each of MANY items (score / rate / rank / classify / extract / summarize each of a large set)"
+      - "Comprehensive coverage — exhaustively sweep, audit, or find EVERY instance across a large surface"
+      - "Research & synthesize — gather across many sources or pages and combine into one answer"
+      - "Confidence — generate several independent attempts and judge them, or adversarially verify findings before trusting the result"
+      - "Scale — work too large to finish well in one inline pass"
     avoid-when:
-      - "The task is a single tool call or a quick lookup — do it inline"
-      - "The work needs interactive, conversational back-and-forth rather than unattended fan-out"
+      - "A single inline answer, a quick lookup, or a small one-off"
+      - "Interactive, conversational back-and-forth rather than unattended fan-out"
 ---
 
 A workflow is a short JS/TS script you author that runs in a sandbox and fans work
@@ -21,8 +25,11 @@ one with `run_workflow` (inline `script` OR saved `name`, exactly one). It retur
 `runId` immediately; the run is asynchronous and you are notified in this
 conversation when it completes — **do NOT poll**.
 
-Reach for one when a task decomposes into many similar small sub-tasks that can run
-concurrently. For a single task or a quick lookup, do it inline.
+Reach for one when a job is too big, too parallel, or too important for one inline
+pass. That is more than batch/map-reduce over many items — it also covers exhaustively
+sweeping or auditing a large surface, researching across many sources and synthesizing,
+and generating several independent attempts to judge or adversarially verify before
+trusting the result. For a single task or a quick lookup, do it inline.
 
 ## The script model
 

package/src/config/call-site-defaults.ts

@@ -69,6 +69,9 @@ export const CALL_SITE_DEFAULTS: Record<LLMCallSite, CallSiteDefaultConfig> = {
   meetConsentMonitor: { profile: "cost-optimized" },
   meetChatOpportunity: { profile: "cost-optimized" },
   inference: { profile: "cost-optimized" },
+  // The advisor consults the strongest managed profile by default; a workspace
+  // overrides this via `llm.advisorProfile` (which floats above this).
+  advisor: { profile: "quality-optimized" },
 
   heartbeatAgent: {
     profile: "cost-optimized",

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

@@ -24,7 +24,10 @@
       "label": "Pre-chat Onboarding Experiment 2026-06-06",
       "description": "Pre-chat onboarding experiment with control and variant-a arms. Control shows the full funnel; variant-a shows the condensed pared-down flow.",
       "defaultEnabled": "control",
-      "values": ["control", "variant-a"]
+      "values": [
+        "control",
+        "variant-a"
+      ]
     },
     {
       "id": "experiment-activation-flow-2026-06-03",
@@ -33,7 +36,11 @@
       "label": "Activation Flow Experiment 2026-06-03",
       "description": "Multivariate activation-flow experiment. control = standard flow; variant-a = activation rail; personal-page = new sign-up-page variant (front-end only). Targeted via LaunchDarkly.",
       "defaultEnabled": "control",
-      "values": ["control", "variant-a", "personal-page"]
+      "values": [
+        "control",
+        "variant-a",
+        "personal-page"
+      ]
     },
     {
       "id": "local-docker-enabled",
@@ -51,6 +58,14 @@
       "description": "Enable the A2A (Agent-to-Agent) channel for inter-assistant communication via the open A2A protocol",
       "defaultEnabled": false
     },
+    {
+      "id": "trace-collection",
+      "scope": "assistant",
+      "key": "trace-collection",
+      "label": "Trace Collection",
+      "description": "Gate per-turn conversation trace (user/assistant/tool-call/tool-response) collection. The daemon attaches a trace to its turn telemetry only when this flag AND the owner's share_diagnostics consent are both on. Defaults off (fail-closed) — the live value is delivered from LaunchDarkly.",
+      "defaultEnabled": false
+    },
     {
       "id": "workspace-tools-watcher",
       "scope": "assistant",
@@ -171,14 +186,6 @@
       "description": "Enable Anthropic fast mode for Opus models (4.6, 4.7, 4.8), delivering up to 2.5x higher output tokens per second at premium pricing",
       "defaultEnabled": false
     },
-    {
-      "id": "conversation-groups-ui",
-      "scope": "assistant",
-      "key": "conversation-groups-ui",
-      "label": "Conversation Groups",
-      "description": "Enable custom conversation group creation, move-to-group, and group management in the sidebar",
-      "defaultEnabled": false
-    },
     {
       "id": "teleport",
       "scope": "client",
@@ -387,14 +394,6 @@
       "description": "When on, runs the new memory-v3 topic-tree retrieval alongside v2 in shadow mode: logs selections to memory_v3_selections, does not modify injected context. Off by default.",
       "defaultEnabled": false
     },
-    {
-      "id": "memory-v3-live",
-      "scope": "assistant",
-      "key": "memory-v3-live",
-      "label": "Memory v3 Live",
-      "description": "When on, memory-v3 topic-tree retrieval becomes the live injected memory source (suppressing v2 injection). Off by default; only enable after shadow telemetry validates the design.",
-      "defaultEnabled": false
-    },
     {
       "id": "self-intro-greeting",
       "scope": "both",
@@ -402,6 +401,38 @@
       "label": "Self-intro first message",
       "description": "On the first conversation, send a natural self-introduction (e.g. \"Hi Vela, I'm alex. Nice to meet you.\") on the user's behalf and route it through real LLM inference, instead of serving the canned first greeting. Names come from the onboarding context; falls back to the canned greeting when no name is known. Exposed to clients so pre-hatch onboarding can compute the source-of-truth initial message, and to the assistant so older clients can still be gated server-side. See assistant/src/daemon/first-greeting.ts (buildSelfIntroMessage).",
       "defaultEnabled": false
+    },
+    {
+      "id": "research-onboarding",
+      "scope": "client",
+      "key": "research-onboarding",
+      "label": "Research onboarding (spike)",
+      "description": "Spike: replace pre-chat onboarding with a web-research \"here's what I know about you\" flow, including a Google Calendar \"Let's chat tomorrow\" step shown over the streaming research. Gates the /assistant/onboarding/research route and the research mock harness. Off by default; enable locally via the feature-flags panel.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "channel-trust-floors",
+      "scope": "assistant",
+      "key": "channel-trust-floors",
+      "label": "Channel Trust Floors",
+      "description": "Expose the Channel Trust Floors settings card (per-channel inbound admission policy) on the Privacy settings page. When off, the card is hidden and channels fall back to their default admission floors. Off by default while the feature is in development.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "mcp-settings",
+      "scope": "assistant",
+      "key": "mcp-settings",
+      "label": "MCP Settings",
+      "description": "Show the MCP page in Settings for managing Model Context Protocol server connections: view status, enable/disable, configure, and inspect registered tools per server.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "os-beta",
+      "scope": "assistant",
+      "key": "os-beta",
+      "label": "OS Beta",
+      "description": "Enable the OS Beta model profile (GLM 5.2 / Fireworks) in the assistant's model profile selection.",
+      "defaultEnabled": false
     }
   ]
 }

package/src/config/llm-resolver.ts

@@ -366,6 +366,9 @@ function profileConfigFragment(profile: ProfileEntry): Mergeable {
     // lower-precedence (e.g. active) profile into one that merely inherited it.
     // `RetryProvider` resolves it from the applied profile, not the merge.
     logitBias: _logitBias,
+    // Per-profile advisor toggle is profile identity, not inheritable model
+    // config — strip it so it can't leak into the merged `LLMConfigBase`.
+    advisorEnabled: _advisorEnabled,
     ...config
   } = profile;
   return config as Mergeable;

package/src/config/memory-v3-gate.ts

@@ -0,0 +1,11 @@
+import type { AssistantConfig } from "./schema.js";
+
+/**
+ * Whether memory-v3 is the live injected memory source for this assistant,
+ * suppressing v2 injection. Gated by workspace config (`memory.v3.live`): new
+ * assistants are switched on at creation via a workspace migration, while
+ * existing assistants stay on v2 until the value is set explicitly.
+ */
+export function isMemoryV3Live(config: AssistantConfig): boolean {
+  return config.memory?.v3?.live === true;
+}

package/src/config/schema.ts

@@ -136,16 +136,18 @@ export const AssistantConfigSchema = z
       .describe(
         "Per-plugin configuration keyed by plugin name. Validated downstream by each plugin's manifest.config validator at bootstrap.",
       ),
-    collectUsageData: z
+    legacyTelemetryOptOut: z
       .boolean()
-      .default(true)
+      .optional()
       .describe(
-        "Whether to collect anonymous usage data to help improve the assistant",
+        "Fail-closed telemetry marker: set for a workspace that had an explicit local usage-data opt-out before telemetry moved to platform share_analytics consent. While set, usage telemetry stays disabled regardless of platform consent.",
       ),
-    sendDiagnostics: z
+    legacyDiagnosticsOptOut: z
       .boolean()
-      .default(true)
-      .describe("Whether to send diagnostic/crash reports"),
+      .optional()
+      .describe(
+        "Fail-closed diagnostics marker: set for a workspace that had an explicit local sendDiagnostics opt-out before crash reporting moved to platform share_diagnostics consent. While set, Sentry stays disabled regardless of platform consent.",
+      ),
     maxStepsPerSession: z
       .number({ error: "maxStepsPerSession must be a number" })
       .int("maxStepsPerSession must be an integer")

package/src/config/schemas/__tests__/memory-v3.test.ts

@@ -6,6 +6,7 @@ describe("MemoryV3ConfigSchema", () => {
   test("parses an empty object to documented defaults", () => {
     const parsed = MemoryV3ConfigSchema.parse({});
     expect(parsed).toEqual({
+      live: false,
       prune: { maxResidentBytes: 393216, targetResidentBytes: 262144 },
       hotSet: { k: 40, halfLifeDays: 14 },
       freshSet: { k: 100 },

package/src/config/schemas/call-site-catalog.ts

@@ -300,6 +300,13 @@ const CATALOG_RECORD: CatalogRecord = {
     description: "General-purpose LLM inference call site for skill use.",
     domain: "skills",
   },
+  advisor: {
+    id: "advisor",
+    displayName: "Advisor",
+    description:
+      "Stronger reviewer model consulted mid-task to shape or pressure-test the plan.",
+    domain: "skills",
+  },
   homeGreeting: {
     id: "homeGreeting",
     displayName: "Home Greeting",

package/src/config/schemas/channels.ts

@@ -136,5 +136,16 @@ export const SlackConfigSchema = z
       .string({ error: "slack.botUsername must be a string" })
       .default("")
       .describe("Slack bot display name"),
+    threadMode: z
+      .enum(["mention_only", "mention_then_thread"], {
+        error:
+          "slack.threadMode must be 'mention_only' or 'mention_then_thread'",
+      })
+      .default("mention_then_thread")
+      .describe(
+        "Controls whether the bot follows threads after an initial @mention. " +
+          "'mention_only' requires every message to @-mention the bot. " +
+          "'mention_then_thread' auto-follows the thread after the first mention.",
+      ),
   })
   .describe("Slack channel configuration");

package/src/config/schemas/llm.ts

@@ -23,6 +23,7 @@ export const LLMProvider = z
     "openrouter",
     "openai-compatible",
     "minimax",
+    "atlascloud",
   ])
   .meta({ id: "LLMProvider" });
 type LLMProvider = z.infer<typeof LLMProvider>;
@@ -77,6 +78,7 @@ export const LLMCallSiteEnum = z.enum([
   "meetConsentMonitor",
   "meetChatOpportunity",
   "inference",
+  "advisor",
   "trustRuleSuggestion",
   "homeGreeting",
   "homeSuggestedPrompts",
@@ -126,6 +128,11 @@ const VerbosityEnum = z.enum(["low", "medium", "high"]);
 const ModelSchema = z.string().min(1);
 const MaxTokensSchema = z.number().int().positive();
 const TemperatureSchema = z.number().min(0).max(2).nullable();
+// `top_p` (nucleus sampling). Range 0–1; `null` = "no opinion — let the
+// provider pick its own default" (matches TemperatureSchema's null
+// semantics). `RetryProvider` renames `topP`→`top_p` and only forwards a
+// non-null value, so providers never receive `top_p: null`.
+const TopPSchema = z.number().min(0).max(1).nullable();
 // Named, code-resolved logit-bias preset a profile may opt into. The value is a
 // preset *name*, not an inline token→bias map, so the workspace config stays
 // small. This is profile-identity metadata, not inheritable config: the resolver
@@ -326,6 +333,7 @@ export const LLMConfigBase = z.object({
   speed: SpeedEnum.default("standard"),
   verbosity: VerbosityEnum.default("medium"),
   temperature: TemperatureSchema.default(null),
+  topP: TopPSchema.default(null),
   thinking: ThinkingSchema.default(ThinkingSchema.parse({})),
   contextWindow: ContextWindowSchema.default(ContextWindowSchema.parse({})),
   openrouter: OpenRouterSchema.default(OpenRouterSchema.parse({})),
@@ -361,6 +369,7 @@ export const LLMConfigFragment = z
     speed: SpeedEnum.optional(),
     verbosity: VerbosityEnum.optional(),
     temperature: TemperatureSchema.optional(),
+    topP: TopPSchema.optional(),
     thinking: ThinkingFragmentSchema.optional(),
     contextWindow: ContextWindowDeepPartialSchema.optional(),
     openrouter: OpenRouterDeepPartialSchema.optional(),
@@ -431,6 +440,13 @@ export const ProfileEntry = LLMConfigFragment.extend({
    * #30362 even though the schema didn't accept it until now.
    */
   status: ProfileStatusSchema.nullable().optional(),
+  /**
+   * Whether the advisor is active while this profile is the chat profile.
+   * Absent/null means enabled (default on); only an explicit `false` disables
+   * it. `.nullable()` matches `status`/`label` so the PUT route's "send null
+   * to clear" sentinel resets it back to the default-on state.
+   */
+  advisorEnabled: z.boolean().nullable().optional(),
   /**
    * When present, this profile is a "mix": it carries no model config and
    * instead references a weighted list of standard profiles. The resolver
@@ -474,6 +490,11 @@ export const LLMSchema = z
     // schema level, so `LLMSchema.parse({})` yields an empty map.
     callSites: z.partialRecord(LLMCallSiteEnum, LLMCallSiteConfig).default({}),
     activeProfile: z.string().min(1).optional(),
+    // The profile the advisor consults (chosen under Models & Services). It is
+    // excluded from the chat-profile pickers so it can't be selected as the
+    // assistant's chat model. Absent falls back to the `advisor` call-site
+    // default (`quality-optimized`).
+    advisorProfile: z.string().min(1).optional(),
     // TTL bounds for inference profile sessions. `defaultTtlSeconds` is read by
     // the CLI to apply when `--ttl` is omitted; the daemon handler itself only
     // reads `maxTtlSeconds` (to clamp caller-supplied values).
@@ -507,6 +528,16 @@ export const LLMSchema = z
         message: `Profile "${config.activeProfile}" referenced by llm.activeProfile is not defined in llm.profiles`,
       });
     }
+    if (
+      config.advisorProfile != null &&
+      !profileNames.has(config.advisorProfile)
+    ) {
+      ctx.addIssue({
+        code: "custom",
+        path: ["advisorProfile"],
+        message: `Profile "${config.advisorProfile}" referenced by llm.advisorProfile is not defined in llm.profiles`,
+      });
+    }
 
     // --- Mix profile validation --------------------------------------------
     // Config keys a mix profile must NOT also set (a mix only references other

package/src/config/schemas/memory-lifecycle.ts

@@ -133,13 +133,9 @@ export const MemoryCleanupConfigSchema = z
       .nonnegative(
         "memory.cleanup.llmRequestLogRetentionMs must be non-negative",
       )
-      // Upper bound must match gateway MAX_LLM_REQUEST_LOG_RETENTION_MS in
-      // gateway/src/http/routes/privacy-config.ts. If a manually edited
-      // config.json sets a value larger than this, the gateway GET would
-      // return it and the macOS picker would snap it to its largest known
-      // option, and the next PATCH would silently truncate the value —
-      // causing quiet data loss. Enforcing the same cap here prevents the
-      // daemon from accepting out-of-range values in the first place.
+      // Cap retention at 365 days. Enforced daemon-side only: the cleanup jobs
+      // honor this bound, so a manually edited config.json with a larger value
+      // is rejected here rather than silently retained.
       .max(
         365 * 24 * 60 * 60 * 1000,
         "memory.cleanup.llmRequestLogRetentionMs must be <= 365 days in ms",

package/src/config/schemas/memory-v3.ts

@@ -220,6 +220,12 @@ export const MemoryV3PruneSchema = z
 // so legacy configs keep parsing. Do not make this object `.strict()`.
 export const MemoryV3ConfigSchema = z
   .object({
+    live: z
+      .boolean({ error: "memory.v3.live must be a boolean" })
+      .default(false)
+      .describe(
+        "Whether memory-v3 is the live injected memory source, suppressing v2 injection. Off by default; brand-new assistants are switched on at creation via a workspace migration, while existing assistants stay on v2 until explicitly enabled.",
+      ),
     prune: MemoryV3PruneSchema.default(MemoryV3PruneSchema.parse({})),
     hotSet: MemoryV3HotSetSchema.default(MemoryV3HotSetSchema.parse({})),
     freshSet: MemoryV3FreshSetSchema.default(MemoryV3FreshSetSchema.parse({})),

package/src/config/schemas/services.ts

@@ -1,6 +1,7 @@
 import { z } from "zod";
 
 import { DEFAULT_IMAGE_MODEL } from "../../media/image-models.js";
+import { FETCH_PROVIDER_IDS } from "../../providers/fetch-provider-catalog.js";
 import { SEARCH_PROVIDER_IDS } from "../../providers/search-provider-catalog.js";
 import { SttServiceSchema } from "./stt.js";
 import { TtsServiceSchema } from "./tts.js";
@@ -28,6 +29,13 @@ const VALID_IMAGE_GEN_PROVIDERS = ["gemini", "openai"] as const;
  */
 const VALID_WEB_SEARCH_PROVIDERS = SEARCH_PROVIDER_IDS;
 
+/**
+ * Derived from `FETCH_PROVIDER_CATALOG`. Adding a new web-fetch provider
+ * to the catalog automatically extends the config-schema enum — no edit
+ * here required.
+ */
+const VALID_WEB_FETCH_PROVIDERS = FETCH_PROVIDER_IDS;
+
 const BaseServiceSchema = z.object({
   mode: ServiceModeSchema.default("your-own"),
 });
@@ -58,6 +66,15 @@ const WebSearchServiceSchema = BaseServiceSchema.extend({
     .default("inference-provider-native"),
 });
 
+const WebFetchServiceSchema = BaseServiceSchema.extend({
+  // Provider that backs the `web_fetch` tool. `default` is the daemon's
+  // built-in HTTP fetch + extract path (no key). BYOK providers (e.g.
+  // `firecrawl`) scrape via their hosted API and reuse the same stored key as
+  // their web-search counterpart. The `mode` field is inherited from
+  // `BaseServiceSchema` for symmetry; web-fetch has no managed proxy today.
+  provider: z.enum(VALID_WEB_FETCH_PROVIDERS).default("default"),
+});
+
 const GoogleOAuthServiceSchema = BaseServiceSchema.extend({
   mode: ServiceModeSchema.default("managed"),
 });
@@ -142,6 +159,7 @@ export const ServicesSchema = z.object({
   "web-search": WebSearchServiceSchema.default(
     WebSearchServiceSchema.parse({}),
   ),
+  "web-fetch": WebFetchServiceSchema.default(WebFetchServiceSchema.parse({})),
   stt: SttServiceSchema.default({
     mode: "your-own" as const,
     provider: "deepgram" as const,