@vellumai/assistant 0.8.5 → 0.8.7

package/src/daemon/disk-pressure-guard.ts

@@ -1,5 +1,8 @@
-import { isAssistantFeatureFlagEnabled } from "../config/assistant-feature-flags.js";
-import { getConfig } from "../config/loader.js";
+import {
+  type DiskPressureBlockedCapability,
+  type DiskPressureState,
+  type DiskPressureStatus,
+} from "../api/events/disk-pressure-status-changed.js";
 import { buildAssistantEvent } from "../runtime/assistant-event.js";
 import { assistantEventHub } from "../runtime/assistant-event-hub.js";
 import { cancelBackgroundTools } from "../tools/background-tool-registry.js";
@@ -8,34 +11,32 @@ import { getLogger } from "../util/logger.js";
 
 export const DISK_PRESSURE_WARNING_THRESHOLD_PERCENT = 80;
 export const DISK_PRESSURE_THRESHOLD_PERCENT = 95;
+// Hysteresis lower bound: once locked, the guard stays locked until usage
+// falls below this clear threshold. The deadband between this and the
+// critical threshold stops the lock from flapping when usage hovers near
+// 95% — otherwise clearing the lock immediately resumes background work,
+// which can refill the disk and re-trip the lock on the next sample.
+export const DISK_PRESSURE_CLEAR_THRESHOLD_PERCENT = 90;
+// Warning-side hysteresis lower bound: once in the warning state, usage must
+// fall below this clear threshold before warning resolves. The deadband between
+// this and the warning threshold stops the in-chat disk-pressure banner from
+// flapping when usage hovers near 80% — without it, a brief dip below 80%
+// clears the warning state, which discards the banner's (state-scoped) dismissal
+// so it re-appears the moment usage ticks back up.
+export const DISK_PRESSURE_WARNING_CLEAR_THRESHOLD_PERCENT = 77;
 export const DISK_PRESSURE_CHECK_INTERVAL_MS = 60_000;
 export const DISK_PRESSURE_OVERRIDE_CONFIRMATION = "I understand the risks";
 export const DISK_PRESSURE_BLOCKED_CAPABILITIES = [
   "agent-turns",
   "background-work",
   "remote-ingress",
-] as const;
-
-export type DiskPressureState = "disabled" | "ok" | "warning" | "critical" | "unknown";
-
-export type DiskPressureBlockedCapability =
-  (typeof DISK_PRESSURE_BLOCKED_CAPABILITIES)[number];
-
-export interface DiskPressureStatus {
-  enabled: boolean;
-  state: DiskPressureState;
-  locked: boolean;
-  acknowledged: boolean;
-  overrideActive: boolean;
-  effectivelyLocked: boolean;
-  lockId: string | null;
-  usagePercent: number | null;
-  thresholdPercent: number;
-  path: string | null;
-  lastCheckedAt: string | null;
-  blockedCapabilities: DiskPressureBlockedCapability[];
-  error: string | null;
-}
+] as const satisfies readonly DiskPressureBlockedCapability[];
+
+export {
+  type DiskPressureBlockedCapability,
+  type DiskPressureState,
+  type DiskPressureStatus,
+};
 
 export type DiskPressureTransitionResult =
   | { ok: true; status: DiskPressureStatus }
@@ -119,24 +120,10 @@ function replaceStatus(next: DiskPressureStatus): DiskPressureStatus {
   return cloneStatus(state.status);
 }
 
-function isEnabled(): boolean {
-  return isAssistantFeatureFlagEnabled("safe-storage-limits", getConfig());
-}
-
-function resetToDisabled(): DiskPressureStatus {
-  const previous = cloneStatus(state.status);
-  stopDiskPressureGuard();
-  state.status = cloneStatus(DISABLED_STATUS);
-  publishStatusChangedIfNeeded(previous);
-  return cloneStatus(state.status);
-}
-
-function ensureEnabledStatus(): DiskPressureStatus | null {
-  if (!isEnabled()) return resetToDisabled();
+function ensureEnabledStatus(): void {
   if (!state.status.enabled) {
     state.status = cloneStatus(OPEN_STATUS);
   }
-  return null;
 }
 
 function nextLockId(): string {
@@ -192,8 +179,7 @@ function rejectTransition(
 }
 
 export function startDiskPressureGuard(): DiskPressureStatus {
-  const disabledStatus = ensureEnabledStatus();
-  if (disabledStatus) return disabledStatus;
+  ensureEnabledStatus();
 
   if (!state.timer) {
     state.timer = setInterval(() => {
@@ -212,8 +198,7 @@ export function stopDiskPressureGuard(): void {
 }
 
 export function evaluateDiskPressureNow(): DiskPressureStatus {
-  const disabledStatus = ensureEnabledStatus();
-  if (disabledStatus) return disabledStatus;
+  ensureEnabledStatus();
 
   let usageInfo: ReturnType<typeof getDiskUsageInfo>;
   try {
@@ -229,8 +214,28 @@ export function evaluateDiskPressureNow(): DiskPressureStatus {
   const usagePercent = roundPercent(
     (usageInfo.usedMb / usageInfo.totalMb) * 100,
   );
-  const isCritical = usagePercent >= DISK_PRESSURE_THRESHOLD_PERCENT;
-  const isWarning = !isCritical && usagePercent >= DISK_PRESSURE_WARNING_THRESHOLD_PERCENT;
+  // Hysteresis: while locked, hold until usage drops below the lower clear
+  // threshold; otherwise lock at the critical threshold.
+  const criticalThreshold = state.status.locked
+    ? DISK_PRESSURE_CLEAR_THRESHOLD_PERCENT
+    : DISK_PRESSURE_THRESHOLD_PERCENT;
+  const isCritical = usagePercent >= criticalThreshold;
+  // Mirror the critical deadband for the warning band: once in an active
+  // pressure state (warning or critical), hold warning until usage clears the
+  // lower warning-clear threshold. Treating "critical" as active here matters
+  // for a direct step-down: when cleanup frees a lot of space in one sample,
+  // usage can drop straight from a critical lock to e.g. 78%, which is below
+  // the 80% warning trigger but above the 77% clear threshold. Using the full
+  // 80% threshold in that case would report "ok" and discard the warning/
+  // dismissal, reopening the flapping window on the next tick back up. The
+  // deadband must apply consistently whether we arrived in the warning band
+  // from below (rising past 80%) or from above (falling out of critical).
+  const inActivePressureState =
+    state.status.state === "warning" || state.status.state === "critical";
+  const warningThreshold = inActivePressureState
+    ? DISK_PRESSURE_WARNING_CLEAR_THRESHOLD_PERCENT
+    : DISK_PRESSURE_WARNING_THRESHOLD_PERCENT;
+  const isWarning = !isCritical && usagePercent >= warningThreshold;
   const lastCheckedAt = new Date().toISOString();
 
   if (!isCritical && !isWarning) {
@@ -277,14 +282,13 @@ export function evaluateDiskPressureNow(): DiskPressureStatus {
 }
 
 export function getDiskPressureStatus(): DiskPressureStatus {
-  if (!isEnabled()) return cloneStatus(DISABLED_STATUS);
   if (!state.status.enabled) return cloneStatus(OPEN_STATUS);
   return cloneStatus(state.status);
 }
 
 export function acknowledgeDiskPressureLock(): DiskPressureTransitionResult {
-  const disabledStatus = ensureEnabledStatus();
-  const status = disabledStatus ?? cloneStatus(state.status);
+  ensureEnabledStatus();
+  const status = cloneStatus(state.status);
   if (!status.locked) {
     return rejectTransition(
       "not_locked",
@@ -310,8 +314,8 @@ export function acknowledgeDiskPressureLock(): DiskPressureTransitionResult {
 export function overrideDiskPressureLock(
   confirmation: string,
 ): DiskPressureTransitionResult {
-  const disabledStatus = ensureEnabledStatus();
-  const status = disabledStatus ?? cloneStatus(state.status);
+  ensureEnabledStatus();
+  const status = cloneStatus(state.status);
   if (!status.locked) {
     return rejectTransition(
       "not_locked",

package/src/daemon/external-plugins-bootstrap.ts

@@ -65,6 +65,7 @@ import { getConfig } from "../config/loader.js";
 import type { AssistantConfig } from "../config/schema.js";
 import { HOOKS } from "../plugin-api/constants.js";
 import { registerDefaultPlugins } from "../plugins/defaults/index.js";
+import { installPluginRuntime } from "../plugins/external-api.js";
 import { buildExternalPlugin } from "../plugins/external-plugin-loader.js";
 import {
   registerPluginSkills,
@@ -82,6 +83,7 @@ import {
   type PluginShutdownContext,
   type PluginSkillRegistration,
 } from "../plugins/types.js";
+import { loadUserPlugins } from "../plugins/user-loader.js";
 import {
   registerSkillRoute,
   type SkillRouteHandle,
@@ -99,16 +101,6 @@ import { registerShutdownHook } from "./shutdown-registry.js";
 
 const log = getLogger("plugins-bootstrap");
 
-/**
- * Minimal context required to bootstrap the plugin layer. Kept intentionally
- * small so the call site in `lifecycle.ts` can construct it from whatever
- * state is already available at that point in startup.
- */
-export interface DaemonContext {
-  config: AssistantConfig;
-  assistantVersion: string;
-}
-
 /**
  * Resolve one credential value. Returns the raw secret string or throws a
  * {@link PluginExecutionError} tagged with the plugin name so the caller can
@@ -191,6 +183,37 @@ function getDisabledPluginFlag(
   return undefined;
 }
 
+/**
+ * Bring the plugin layer up during daemon startup. Runs the full sequence in
+ * the one order the rest of the system depends on:
+ *
+ * 1. Install the `globalThis.__vellumPluginRuntime` bridge so plugins can touch
+ *    it from their module body (see `plugins/external-api.ts` — compiled-binary
+ *    module identity).
+ * 2. Register the first-party defaults so their middleware and injectors
+ *    compose innermost, ahead of any user plugins.
+ * 3. Load user plugins from `<workspaceDir>/plugins/*`. A failing user plugin is
+ *    logged and skipped; `loadUserPlugins()` closes the registration window
+ *    when it returns, so the defaults must already be registered by then.
+ * 4. Run every registered plugin's `init()` via {@link bootstrapPlugins}.
+ *
+ * Plugin bootstrap is wrapped so a failing plugin cannot block daemon startup —
+ * the daemon comes up with degraded plugin functionality instead.
+ */
+export async function initializePlugins(): Promise<void> {
+  installPluginRuntime();
+  registerDefaultPlugins();
+  await loadUserPlugins();
+  try {
+    await bootstrapPlugins();
+  } catch (err) {
+    log.warn(
+      { err },
+      "Plugin bootstrap failed — continuing startup with degraded plugin functionality",
+    );
+  }
+}
+
 /**
  * Run every registered plugin's `init()` hook sequentially and install a
  * reverse-order shutdown hook. See the module docstring for full semantics.
@@ -203,7 +226,7 @@ function getDisabledPluginFlag(
  * and before the first conversation is served. First-party defaults are
  * registered inline via {@link registerDefaultPlugins}.
  */
-export async function bootstrapPlugins(ctx: DaemonContext): Promise<void> {
+export async function bootstrapPlugins(): Promise<void> {
   // Register first-party default plugins. Each default wraps one of the
   // assistant's canonical pipelines (`toolExecute`, `llmCall`, ...) with a
   // passthrough so the pipeline shape is explicit at boot even when no
@@ -223,6 +246,8 @@ export async function bootstrapPlugins(ctx: DaemonContext): Promise<void> {
 
   log.info({ count: plugins.length }, "bootstrapPlugins: initializing plugins");
 
+  const assistantConfig = getConfig();
+
   // Plugins that passed `requiresFlag` gating and therefore need the full
   // init → contribute → shutdown lifecycle. Plugins skipped by the flag gate
   // are omitted from this list so the shutdown hook below never tears down
@@ -245,7 +270,7 @@ export async function bootstrapPlugins(ctx: DaemonContext): Promise<void> {
   // shutdown hook below. Only `assistantVersion` is exposed today; future
   // additions live on {@link PluginShutdownContext}.
   const shutdownContext: PluginShutdownContext = {
-    assistantVersion: ctx.assistantVersion,
+    assistantVersion: APP_VERSION,
   };
 
   async function rollbackPlugin(active: ActivePlugin): Promise<void> {
@@ -266,7 +291,7 @@ export async function bootstrapPlugins(ctx: DaemonContext): Promise<void> {
 
   for (const plugin of plugins) {
     const name = plugin.manifest.name;
-    const disabledFlag = getDisabledPluginFlag(plugin, ctx.config);
+    const disabledFlag = getDisabledPluginFlag(plugin, assistantConfig);
     if (disabledFlag !== undefined) {
       log.info(
         { plugin: name, flag: disabledFlag },
@@ -277,7 +302,7 @@ export async function bootstrapPlugins(ctx: DaemonContext): Promise<void> {
     }
 
     try {
-      activePlugins.push(await initializePlugin(plugin, ctx));
+      activePlugins.push(await initializePlugin(plugin, assistantConfig));
     } catch (err) {
       unregisterPlugin(name);
       await teardownPartialInit();
@@ -326,7 +351,7 @@ interface ActivePlugin {
 
 async function initializePlugin(
   plugin: Plugin,
-  ctx: DaemonContext,
+  assistantConfig: AssistantConfig,
 ): Promise<ActivePlugin> {
   const name = plugin.manifest.name;
   const routeHandles: SkillRouteHandle[] = [];
@@ -341,7 +366,7 @@ async function initializePlugin(
     const config = validatePluginConfig(
       name,
       plugin.manifest.config,
-      getPluginConfigRaw(ctx.config, name),
+      getPluginConfigRaw(assistantConfig, name),
     );
 
     const initContext = {
@@ -349,7 +374,7 @@ async function initializePlugin(
       credentials,
       logger: log.child({ plugin: name }),
       pluginStorageDir: ensurePluginStorageDir(name),
-      assistantVersion: ctx.assistantVersion,
+      assistantVersion: APP_VERSION,
     };
 
     if (plugin.tools && plugin.tools.length > 0) {
@@ -417,7 +442,7 @@ async function initializePlugin(
   } catch (err) {
     if (initCompleted) {
       await teardownPlugin({ plugin, routeHandles }, "bootstrap-failed", {
-        assistantVersion: ctx.assistantVersion,
+        assistantVersion: APP_VERSION,
       });
     } else {
       for (const handle of routeHandles) {
@@ -519,11 +544,8 @@ export async function reregisterExternalPlugin(
     return;
   }
 
-  const ctx: DaemonContext = {
-    config: getConfig(),
-    assistantVersion: APP_VERSION,
-  };
-  const disabledFlag = getDisabledPluginFlag(plugin, ctx.config);
+  const assistantConfig = getConfig();
+  const disabledFlag = getDisabledPluginFlag(plugin, assistantConfig);
   if (disabledFlag !== undefined) {
     log.info(
       { plugin: pluginName, flag: disabledFlag },
@@ -535,7 +557,7 @@ export async function reregisterExternalPlugin(
   const existing = getRegisteredPlugin(pluginName);
   if (existing === undefined) {
     try {
-      await initializePlugin(plugin, ctx);
+      await initializePlugin(plugin, assistantConfig);
       setRegisteredPlugin(plugin);
       log.info({ plugin: pluginName }, "external plugin registered post-boot");
     } catch (err) {

package/src/daemon/first-greeting.ts

@@ -14,12 +14,6 @@ export interface OnboardingGreetingContext {
   googleConnected?: boolean;
 }
 
-export const CANNED_FIRST_GREETING = [
-  "Hey,",
-  "",
-  "We can get into whatever you've got, or just talk first — that tends to go better. Up to you.",
-].join("\n");
-
 /**
  * Returns `true` when all of the following are true:
  * - `conversationMessageCount === 0` (no prior messages in this conversation)
@@ -49,6 +43,28 @@ export function getCannedFirstGreeting(
   return CANNED_FIRST_GREETING;
 }
 
+/**
+ * Builds a natural self-introduction to send *on behalf of the user* in place
+ * of the wake-up greeting, so the assistant generates a real response instead
+ * of replaying canned copy. Names come from the onboarding context; missing
+ * names are dropped so the line stays natural:
+ *   - both:           "Hi Vela, I'm alex. Nice to meet you."
+ *   - assistant only: "Hi Vela. Nice to meet you."
+ *   - user only:      "Hi, I'm alex. Nice to meet you."
+ * When neither name is known there is nothing personal to say, so this returns
+ * `undefined` and the caller falls back to the canned greeting.
+ */
+export function buildSelfIntroMessage(
+  onboarding?: OnboardingGreetingContext,
+): string | undefined {
+  const assistant = onboarding?.assistantName?.trim();
+  const user = onboarding?.userName?.trim();
+  if (!assistant && !user) return undefined;
+  const hi = assistant ? `Hi ${assistant}` : "Hi";
+  const intro = user ? `, I'm ${user}` : "";
+  return `${hi}${intro}. Nice to meet you.`;
+}
+
 const TONE_INTRO_CLOSE: Record<Tone, string> = {
   grounded: "",
   warm: "Good to meet you.",
@@ -67,30 +83,54 @@ function buildIntroLine(
   return [greeting, who, close].filter(Boolean).join(" ");
 }
 
-const TONE_INVITE: Record<Tone, string> = {
+// Every greeting variant — the no-onboarding CANNED greeting and all four
+// personalized tones — ends with a migration offer. The opener (task-vs-talk)
+// and the migration offer are kept as separate per-tone maps and composed in
+// `buildInvite`, rather than inlined as one full sentence per variant, so the
+// offer cannot be silently dropped from a variant when the opener copy is
+// edited. The first-greeting test asserts every variant still includes it.
+const TONE_INVITE_OPENER: Record<Tone, string> = {
   grounded:
     "We can get into whatever you've got, or just talk first — that tends to go better. Up to you.",
   warm: "We can start on something specific, or just talk for a bit first — honestly that tends to work out better. Either way, I'm here.",
   energetic:
-    "We can jump straight into whatever you've got, or take a few minutes to just talk first. What sounds right?",
+    "We can jump straight into whatever you've got, or take a few minutes to just talk first.",
   poetic:
     "We can start with whatever's in front of you, or just talk for a bit first. Either way.",
 };
 
+const TONE_MIGRATION_OFFER: Record<Tone, string> = {
+  grounded:
+    "And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll learn from it fast. Best head start you can give me.",
+  warm: "And you don't have to start me from scratch — if there's a ChatGPT or Claude that already knows you, bring it over and I'll get up to speed fast. Honestly, it's the best head start you could give me.",
+  energetic:
+    "And you don't have to start me from scratch — if you've built up a ChatGPT or Claude, bring it over and I'll get up to speed fast. Best head start you can give me — want to start there?",
+  poetic:
+    "And you don't have to start me from nothing — if there's a ChatGPT or Claude that already knows you, bring it over and I'll learn from it. The best head start you could give me.",
+};
+
 const TONE_GOOGLE_SCAN: Record<Tone, string> = {
   grounded:
-    "I can scan your email, calendar, and drive in the background while we talk — just say the word.",
-  warm: "Also — I can scan your email, calendar, and drive in the background while we chat, if you'd like. Just let me know.",
+    "I can scan Gmail in the background while we talk — just say the word.",
+  warm: "Also — I can scan Gmail in the background while we chat, if you'd like. Just let me know.",
   energetic:
-    "Oh, and I can scan your email, calendar, and drive in the background right now — want me to?",
+    "Oh, and I can scan Gmail in the background right now — want me to?",
   poetic:
-    "I can also look through your email, calendar, and drive quietly in the background — say the word.",
+    "I can also look through Gmail quietly in the background — say the word.",
 };
 
 function buildInvite(tone: Tone = "grounded"): string {
-  return TONE_INVITE[tone];
+  return `${TONE_INVITE_OPENER[tone]} ${TONE_MIGRATION_OFFER[tone]}`;
 }
 
+// Composed from the grounded opener + migration offer so the no-onboarding
+// greeting reuses the same source as the personalized grounded greeting rather
+// than duplicating the copy. Defined after the tone maps so they are
+// initialized before this module-level evaluation runs.
+export const CANNED_FIRST_GREETING = ["Hey,", "", buildInvite("grounded")].join(
+  "\n",
+);
+
 const VALID_TONES = new Set<string>([
   "grounded",
   "warm",

package/src/daemon/guardian-action-generators.ts

@@ -44,9 +44,9 @@ export function createGuardianActionCopyGenerator(): GuardianActionCopyGenerator
 
     const response = await provider.sendMessage(
       [{ role: "user", content: [{ type: "text", text: prompt }] }],
-      [],
-      GUARDIAN_ACTION_COPY_SYSTEM_PROMPT,
       {
+        tools: [],
+        systemPrompt: GUARDIAN_ACTION_COPY_SYSTEM_PROMPT,
         config: {
           max_tokens: options.maxTokens ?? GUARDIAN_ACTION_COPY_MAX_TOKENS,
           callSite: "guardianQuestionCopy",

package/src/daemon/handlers/conversations.ts

@@ -3,7 +3,6 @@ import { v4 as uuid } from "uuid";
 import { clearAll, getConversation } from "../../memory/conversation-crud.js";
 import { resolveConversationId } from "../../memory/conversation-key-store.js";
 import { broadcastMessage } from "../../runtime/assistant-event-hub.js";
-import * as pendingInteractions from "../../runtime/pending-interactions.js";
 import { getSubagentManager } from "../../subagent/index.js";
 import { createAbortReason } from "../../util/abort-reasons.js";
 import { truncate } from "../../util/truncate.js";
@@ -28,23 +27,11 @@ export function handleConfirmationResponse(msg: ConfirmationResponse): void {
   for (const [conversationId, conversation] of conversationEntries()) {
     if (conversation.hasPendingConfirmation(msg.requestId)) {
       touchConversation(conversationId);
-      conversation.handleConfirmationResponse(
-        msg.requestId,
-        decision,
-        msg.selectedPattern,
-        msg.selectedScope,
-        undefined,
-        { source: "button" },
-      );
-      // Idempotent cleanup: the prompter's `resolveConfirmation` already
-      // resolved the interaction with the correct approved/rejected state
-      // before we reach here, so the entry is typically gone. This call is
-      // a safety net for paths where the prompter no-ops; map decision to
-      // state to keep the emitted event accurate when it does fire.
-      pendingInteractions.resolve(
-        msg.requestId,
-        decision === "allow" ? "approved" : "rejected",
-      );
+      conversation.handleConfirmationResponse(msg.requestId, decision, {
+        selectedPattern: msg.selectedPattern,
+        selectedScope: msg.selectedScope,
+        emissionContext: { source: "button" },
+      });
       return;
     }
   }
@@ -225,10 +212,7 @@ export function steerToMessage(
   | { steered: true }
   | {
       steered: false;
-      reason:
-        | "conversation_not_found"
-        | "message_not_found"
-        | "not_processing";
+      reason: "conversation_not_found" | "message_not_found" | "not_processing";
     } {
   const conversation = findConversation(conversationId);
   if (!conversation) {

package/src/daemon/handlers/shared.ts

@@ -85,11 +85,18 @@ export interface HistorySurface {
   surfaceType: string;
   title?: string;
   data: Record<string, unknown>;
-  actions?: Array<{ id: string; label: string; style?: string }>;
+  actions?: Array<{
+    id: string;
+    label: string;
+    style?: string;
+    data?: Record<string, unknown>;
+  }>;
   display?: string;
   persistent?: boolean;
   completed?: boolean;
   completionSummary?: string;
+  /** Id of the tool call that produced this surface (the `ui_show` proxy tool). Lets the client gate app previews on the tool result's arrival rather than whole-turn streaming state. */
+  toolCallId?: string;
 }
 
 /**
@@ -371,6 +378,8 @@ export function renderHistoryContent(content: unknown): RenderedHistoryContent {
           typeof block.completionSummary === "string"
             ? block.completionSummary
             : undefined,
+        toolCallId:
+          typeof block.toolCallId === "string" ? block.toolCallId : undefined,
       };
       surfaces.push(surface);
       contentOrder.push(`surface:${surfaces.length - 1}`);

package/src/daemon/handlers/skills.ts

@@ -328,6 +328,7 @@ function toSlimSkillResponse(
   const origin = deriveOrigin(kind, summary.directoryPath, installMeta);
   const status: SlimSkillResponse["status"] = state;
 
+  const category = inferCategory(summary.displayName, summary.description);
   const base = {
     id: summary.id,
     name: summary.displayName,
@@ -335,6 +336,7 @@ function toSlimSkillResponse(
     emoji: summary.emoji,
     kind,
     status,
+    category,
   } as const;
 
   switch (origin) {
@@ -511,16 +513,13 @@ export async function listSkillsFiltered(filter: SkillListFilter): Promise<{
   // Compute category counts BEFORE applying the category filter
   const categoryCounts: Record<string, number> = {};
   for (const s of skills) {
-    const cat = inferCategory(s.name, s.description);
-    categoryCounts[cat] = (categoryCounts[cat] ?? 0) + 1;
+    categoryCounts[s.category] = (categoryCounts[s.category] ?? 0) + 1;
   }
   const totalCount = skills.length;
 
   // Apply category filter
   if (filter.category) {
-    skills = skills.filter(
-      (s) => inferCategory(s.name, s.description) === filter.category,
-    );
+    skills = skills.filter((s) => s.category === filter.category);
   }
 
   // Sort: installed first, community origins before core within installed,
@@ -610,6 +609,7 @@ export async function getSkill(
       kind: slim.kind,
       origin: slim.origin,
       status: slim.status,
+      category: slim.category,
       slug: slim.slug,
       author: slim.author,
       stars: slim.stars,
@@ -649,6 +649,7 @@ export async function getSkill(
       kind: slim.kind,
       origin: slim.origin,
       status: slim.status,
+      category: slim.category,
       slug: slim.slug,
       sourceRepo: slim.sourceRepo,
       installs: slim.installs,
@@ -679,6 +680,7 @@ export async function getSkill(
     kind: slim.kind,
     origin: slim.origin,
     status: slim.status,
+    category: slim.category,
   };
   return { skill: detail };
 }
@@ -1367,6 +1369,7 @@ export async function searchSkills(
         kind: "catalog" as const,
         origin: "vellum" as const,
         status: "available" as const,
+        category: inferCategory(s.displayName, s.description),
       };
     });
 
@@ -1385,6 +1388,7 @@ export async function searchSkills(
         kind: "catalog" as const,
         origin: "clawhub" as const,
         status: "available" as const,
+        category: inferCategory(s.name, s.description),
         slug: s.slug,
         author: s.author,
         stars: s.stars,
@@ -1411,6 +1415,7 @@ export async function searchSkills(
         kind: "catalog" as const,
         origin: "skillssh" as const,
         status: "available" as const,
+        category: inferCategory(r.name, ""),
         slug: r.id,
         sourceRepo: r.source,
         installs: r.installs,
@@ -1562,15 +1567,11 @@ export async function draftSkill(params: {
               body.slice(0, 2000),
             ].join("\n");
 
-            const response = await provider.sendMessage(
-              [userMessage(prompt)],
-              [],
-              undefined,
-              {
-                config: { callSite: "skillCategoryInference", max_tokens: 256 },
-                signal,
-              },
-            );
+            const response = await provider.sendMessage([userMessage(prompt)], {
+              tools: [],
+              config: { callSite: "skillCategoryInference", max_tokens: 256 },
+              signal,
+            });
             cleanup();
 
             const responseText = extractText(response);