@vellumai/assistant 0.4.52 → 0.4.54

package/src/daemon/session-agent-loop.ts

@@ -133,7 +133,7 @@ const log = getLogger("session-agent-loop");
  *
  * Returns the actual token count or null if it cannot be parsed.
  */
-function parseActualTokensFromError(
+export function parseActualTokensFromError(
   errorMessage: string | null,
 ): number | null {
   if (!errorMessage) return null;
@@ -710,10 +710,11 @@ export async function runAgentLoopImpl(
     const preflightBudget = Math.floor(providerMaxTokens * (1 - safetyMargin));
     let reducerState: ReducerState | undefined;
 
+    const toolTokenBudget = ctx.agentLoop.getToolTokenBudget(runMessages);
     const preflightTokens = estimatePromptTokens(
       runMessages,
       ctx.systemPrompt,
-      { providerName: ctx.provider.name },
+      { providerName: ctx.provider.name, toolTokenBudget },
     );
 
     if (overflowRecovery.enabled && preflightTokens > preflightBudget) {
@@ -747,6 +748,7 @@ export async function runAgentLoopImpl(
             systemPrompt: ctx.systemPrompt,
             contextWindow: config.contextWindow,
             targetTokens: preflightBudget,
+            toolTokenBudget,
           },
           reducerState,
           (msgs, signal, opts) =>
@@ -863,7 +865,7 @@ export async function runAgentLoopImpl(
         const estimated = estimatePromptTokens(
           checkpoint.history,
           ctx.systemPrompt,
-          { providerName: ctx.provider.name },
+          { providerName: ctx.provider.name, toolTokenBudget },
         );
         if (estimated > midLoopThreshold) {
           rlog.warn(
@@ -993,6 +995,23 @@ export async function runAgentLoopImpl(
       );
     }
 
+    // If mid-loop compaction exhausted all attempts but the agent loop
+    // still yielded (yieldedForBudget is true), the turn is incomplete.
+    // Escalate to the convergence loop's more aggressive reducer tiers
+    // (tool-result truncation, media stubbing, injection downgrade)
+    // instead of silently treating an incomplete turn as done.
+    if (yieldedForBudget && !abortController.signal.aborted) {
+      rlog.warn(
+        {
+          phase: "mid-loop-compact",
+          midLoopCompactAttempts,
+          maxAttempts: overflowRecovery.maxAttempts,
+        },
+        "Mid-loop compaction exhausted all attempts — escalating to convergence loop",
+      );
+      state.contextTooLargeDetected = true;
+    }
+
     // One-shot ordering error retry
     if (
       state.orderingErrorDetected &&
@@ -1045,6 +1064,7 @@ export async function runAgentLoopImpl(
             ),
         });
         preRepairMessages = updatedHistory;
+        preRunHistoryLength = updatedHistory.length;
       }
       if (!reducerState) {
         reducerState = createInitialReducerState();
@@ -1061,7 +1081,7 @@ export async function runAgentLoopImpl(
       const estimatedTokensAtOverflow = estimatePromptTokens(
         ctx.messages,
         ctx.systemPrompt,
-        { providerName: ctx.provider.name },
+        { providerName: ctx.provider.name, toolTokenBudget },
       );
       let correctedTarget = preflightBudget;
       if (actualTokens && estimatedTokensAtOverflow > 0) {
@@ -1113,6 +1133,7 @@ export async function runAgentLoopImpl(
             systemPrompt: ctx.systemPrompt,
             contextWindow: config.contextWindow,
             targetTokens: correctedTarget,
+            toolTokenBudget,
           },
           reducerState,
           (msgs, signal, opts) =>
@@ -1124,12 +1145,6 @@ export async function runAgentLoopImpl(
         ctx.messages = step.messages;
         currentInjectionMode = step.state.injectionMode;
 
-        // If the reducer is now exhausted without compacting, break out
-        // so the overflow policy path can attempt emergency compaction.
-        if (reducerState.exhausted && !step.compactionResult?.compacted) {
-          break;
-        }
-
         if (step.compactionResult?.compacted) {
           ctx.contextCompactedMessageCount +=
             step.compactionResult.compactedPersistedMessages;
@@ -1183,77 +1198,10 @@ export async function runAgentLoopImpl(
         );
       }
 
-      // When all reducer tiers are exhausted but the context is still too
-      // large, attempt one last emergency compaction before consulting the
-      // overflow policy. This covers the case where progress was made
-      // (messages grew) and the normal tiers couldn't compact enough.
-      if (state.contextTooLargeDetected && reducerState.exhausted) {
-        const emergencyCompact = await ctx.contextWindowManager.maybeCompact(
-          ctx.messages,
-          abortController.signal,
-          {
-            lastCompactedAt: ctx.contextCompactedAt ?? undefined,
-            force: true,
-            minKeepRecentUserTurns: 0,
-            targetInputTokensOverride: correctedTarget,
-          },
-        );
-        if (emergencyCompact.compacted) {
-          ctx.messages = emergencyCompact.messages;
-          ctx.contextCompactedMessageCount +=
-            emergencyCompact.compactedPersistedMessages;
-          ctx.contextCompactedAt = Date.now();
-          updateConversationContextWindow(
-            ctx.conversationId,
-            emergencyCompact.summaryText,
-            ctx.contextCompactedMessageCount,
-          );
-          onEvent({
-            type: "context_compacted",
-            previousEstimatedInputTokens:
-              emergencyCompact.previousEstimatedInputTokens,
-            estimatedInputTokens: emergencyCompact.estimatedInputTokens,
-            maxInputTokens: emergencyCompact.maxInputTokens,
-            thresholdTokens: emergencyCompact.thresholdTokens,
-            compactedMessages: emergencyCompact.compactedMessages,
-            summaryCalls: emergencyCompact.summaryCalls,
-            summaryInputTokens: emergencyCompact.summaryInputTokens,
-            summaryOutputTokens: emergencyCompact.summaryOutputTokens,
-            summaryModel: emergencyCompact.summaryModel,
-          });
-          emitUsage(
-            ctx,
-            emergencyCompact.summaryInputTokens,
-            emergencyCompact.summaryOutputTokens,
-            emergencyCompact.summaryModel,
-            onEvent,
-            "context_compactor",
-            reqId,
-            emergencyCompact.summaryCacheCreationInputTokens ?? 0,
-            emergencyCompact.summaryCacheReadInputTokens ?? 0,
-            collapseRawResponses(emergencyCompact.summaryRawResponses),
-          );
-
-          runMessages = applyRuntimeInjections(ctx.messages, {
-            ...injectionOpts,
-            mode: currentInjectionMode,
-          });
-          preRepairMessages = runMessages;
-          preRunHistoryLength = runMessages.length;
-          state.contextTooLargeDetected = false;
-
-          updatedHistory = await ctx.agentLoop.run(
-            runMessages,
-            eventHandler,
-            abortController.signal,
-            reqId,
-            onCheckpoint,
-          );
-        }
-      }
-
       // All reducer tiers exhausted but provider still rejects —
       // consult the overflow policy for latest-turn compression.
+      // Emergency compaction is deferred to the policy-gated paths below
+      // so that `request_user_approval` sessions collect consent first.
       if (state.contextTooLargeDetected) {
         const action = resolveOverflowAction({
           overflowRecovery,

package/src/daemon/session-error.ts

@@ -218,7 +218,7 @@ function classifyCore(
         return {
           code: "PROVIDER_WEB_SEARCH",
           userMessage:
-            "An internal error occurred with web search. Retrying...",
+            "An internal error occurred with web search. Please try again.",
           retryable: true,
           errorCategory: "web_search_ordering",
         };
@@ -226,7 +226,7 @@ function classifyCore(
       if (isOrderingError(message)) {
         return {
           code: "PROVIDER_ORDERING",
-          userMessage: "An internal error occurred. Retrying...",
+          userMessage: "An internal error occurred. Please try again.",
           retryable: true,
           errorCategory: "tool_ordering",
         };
@@ -308,7 +308,8 @@ function classifyByMessage(
   if (isWebSearchOrderingError(message)) {
     return {
       code: "PROVIDER_WEB_SEARCH",
-      userMessage: "An internal error occurred with web search. Retrying...",
+      userMessage:
+        "An internal error occurred with web search. Please try again.",
       retryable: true,
       errorCategory: "web_search_ordering",
     };
@@ -318,7 +319,7 @@ function classifyByMessage(
   if (isOrderingError(message)) {
     return {
       code: "PROVIDER_ORDERING",
-      userMessage: "An internal error occurred. Retrying...",
+      userMessage: "An internal error occurred. Please try again.",
       retryable: true,
       errorCategory: "tool_ordering",
     };

package/src/daemon/session-process.ts

@@ -15,7 +15,7 @@ import type {
   TurnInterfaceContext,
 } from "../channels/types.js";
 import { parseChannelId, parseInterfaceId } from "../channels/types.js";
-import { getConfig } from "../config/loader.js";
+import { API_KEY_PROVIDERS, getConfig } from "../config/loader.js";
 import { listPendingRequestsByConversationScope } from "../memory/canonical-guardian-store.js";
 import {
   addMessage,
@@ -27,6 +27,7 @@ import { extractPreferences } from "../notifications/preference-extractor.js";
 import { createPreference } from "../notifications/preferences-store.js";
 import type { Message } from "../providers/types.js";
 import { routeGuardianReply } from "../runtime/guardian-reply-router.js";
+import { getSecureKeyAsync } from "../security/secure-keys.js";
 import { getLogger } from "../util/logger.js";
 import type {
   ServerMessage,
@@ -47,12 +48,15 @@ import { resolveVerificationSessionIntent } from "./verification-session-intent.
 const log = getLogger("session-process");
 
 /** Build a model_info event with fresh config data. */
-export function buildModelInfoEvent(): ServerMessage {
+export async function buildModelInfoEvent(): Promise<ServerMessage> {
   const config = getConfig();
-  const configured = Object.keys(config.apiKeys).filter(
-    (k) => !!config.apiKeys[k],
-  );
-  if (!configured.includes("ollama")) configured.push("ollama");
+  const configured: string[] = ["ollama"];
+  for (const p of API_KEY_PROVIDERS) {
+    if (p === "ollama") continue;
+    if (await getSecureKeyAsync(p)) {
+      configured.push(p);
+    }
+  }
   return {
     type: "model_info",
     model: config.model,
@@ -296,7 +300,10 @@ export async function drainQueue(
   }
 
   // Resolve slash commands for queued messages
-  const slashResult = resolveSlash(next.content, buildSlashContext(session));
+  const slashResult = await resolveSlash(
+    next.content,
+    buildSlashContext(session),
+  );
 
   // Unknown slash — persist the exchange and continue draining.
   // Persist each message before pushing to session.messages so that a
@@ -365,7 +372,7 @@ export async function drainQueue(
         isModelSlashCommand(next.content) ||
         isProviderShortcut(next.content)
       ) {
-        next.onEvent(buildModelInfoEvent());
+        next.onEvent(await buildModelInfoEvent());
       }
       next.onEvent({ type: "assistant_text_delta", text: slashResult.message });
       session.traceEmitter.emit(
@@ -651,7 +658,7 @@ export async function processMessage(
   }
 
   // Resolve slash commands before persistence
-  const slashResult = resolveSlash(content, buildSlashContext(session));
+  const slashResult = await resolveSlash(content, buildSlashContext(session));
 
   // Unknown slash command — persist the exchange (user + assistant) so the
   // messageId is real.  Persist each message before pushing to session.messages
@@ -715,7 +722,7 @@ export async function processMessage(
     // Emit fresh model info before the text delta so the client has
     // up-to-date configuredProviders when rendering /model or /models UI.
     if (isModelSlashCommand(content) || isProviderShortcut(content)) {
-      onEvent(buildModelInfoEvent());
+      onEvent(await buildModelInfoEvent());
     }
     onEvent({ type: "assistant_text_delta", text: slashResult.message });
     session.traceEmitter.emit(

package/src/daemon/session-runtime-assembly.ts

@@ -37,6 +37,8 @@ export interface ChannelCapabilities {
   pttActivationKey?: string;
   /** Whether the client has been granted microphone permission by the OS. */
   microphonePermissionGranted?: boolean;
+  /** Chat type from the gateway (e.g. "private", "group", "supergroup", "channel", "im", "mpim"). */
+  chatType?: string;
 }
 
 /**
@@ -296,6 +298,7 @@ export function resolveChannelCapabilities(
   sourceChannel?: string | null,
   sourceInterface?: string | null,
   pttMetadata?: PttMetadata | null,
+  chatType?: string | null,
 ): ChannelCapabilities {
   // Normalise legacy pseudo-channel IDs to canonical ChannelId values.
   let channel: string;
@@ -330,6 +333,8 @@ export function resolveChannelCapabilities(
     }
   }
 
+  const resolvedChatType = chatType ?? undefined;
+
   switch (channel) {
     case "vellum": {
       const supportsDesktopUi = iface === "macos";
@@ -342,6 +347,7 @@ export function resolveChannelCapabilities(
           pttMetadata?.pttActivationKey,
         ),
         microphonePermissionGranted: pttMetadata?.microphonePermissionGranted,
+        chatType: resolvedChatType,
       };
     }
     case "telegram":
@@ -354,6 +360,7 @@ export function resolveChannelCapabilities(
         dashboardCapable: false,
         supportsDynamicUi: false,
         supportsVoiceInput: false,
+        chatType: resolvedChatType,
       };
     default:
       return {
@@ -361,10 +368,28 @@ export function resolveChannelCapabilities(
         dashboardCapable: false,
         supportsDynamicUi: false,
         supportsVoiceInput: false,
+        chatType: resolvedChatType,
       };
   }
 }
 
+/**
+ * Returns true when the chat type indicates a group/multi-party conversation
+ * (Telegram group/supergroup, Slack channel/group/mpim, etc.).
+ */
+export function isGroupChatType(chatType?: string): boolean {
+  if (!chatType) return false;
+  switch (chatType) {
+    case "group":
+    case "supergroup":
+    case "channel":
+    case "mpim":
+      return true;
+    default:
+      return false;
+  }
+}
+
 /** Context about the active workspace surface, passed to applyRuntimeInjections. */
 export interface ActiveSurfaceContext {
   surfaceId: string;
@@ -632,6 +657,31 @@ export function injectChannelCapabilityContext(
     }
   }
 
+  // Inject group chat etiquette only when the chat type indicates a multi-party
+  // conversation, avoiding misconditioned "stay silent" guidance in 1:1 DMs.
+  if (isGroupChatType(caps.chatType)) {
+    lines.push(`chat_type: ${caps.chatType}`);
+    lines.push("");
+    lines.push("GROUP CHAT ETIQUETTE:");
+    lines.push(
+      "- You are a **participant**, not the user's proxy. Think before you speak.",
+    );
+    lines.push(
+      "- **Respond when:** directly mentioned, you can add genuine value, something witty fits naturally, or correcting important misinformation.",
+    );
+    lines.push(
+      '- **Stay silent when:** casual banter between humans, someone already answered, your response would just be "yeah" or "nice", or the conversation flows fine without you.',
+    );
+    lines.push(
+      "- **The human rule:** humans don't respond to every message in a group chat. Neither should you. Quality over quantity.",
+    );
+    if (caps.channel === "slack") {
+      lines.push(
+        "- Use emoji reactions naturally to acknowledge without cluttering.",
+      );
+    }
+  }
+
   lines.push("</channel_capabilities>");
 
   const block = lines.join("\n");

package/src/daemon/session-slash.ts

@@ -5,10 +5,16 @@ import { join } from "node:path";
 import QRCode from "qrcode";
 
 import { getGatewayPort, getIngressPublicBaseUrl } from "../config/env.js";
-import { getConfig, loadRawConfig, saveRawConfig } from "../config/loader.js";
+import {
+  API_KEY_PROVIDERS,
+  getConfig,
+  loadRawConfig,
+  saveRawConfig,
+} from "../config/loader.js";
 import { resolveSkillStates } from "../config/skill-state.js";
 import { loadSkillCatalog } from "../config/skills.js";
 import { initializeProviders } from "../providers/registry.js";
+import { getSecureKeyAsync } from "../security/secure-keys.js";
 import {
   buildInvocableSlashCatalog,
   resolveSlashSkillCommand,
@@ -53,14 +59,12 @@ export interface SlashContext {
 
 const AVAILABLE_MODELS = [
   "claude-opus-4-6",
-  "claude-opus-4-6-fast",
   "claude-sonnet-4-6",
   "claude-haiku-4-5-20251001",
 ] as const;
 
 const MODEL_DISPLAY_NAMES: Record<string, string> = {
   "claude-opus-4-6": "Claude Opus 4.6",
-  "claude-opus-4-6-fast": "Claude Opus 4.6 Fast",
   "claude-sonnet-4-6": "Claude Sonnet 4.6",
   "claude-haiku-4-5-20251001": "Claude Haiku 4.5",
 };
@@ -75,11 +79,6 @@ const PROVIDER_MODEL_SHORTCUTS: Record<
     model: "claude-opus-4-6",
     displayName: "Claude Opus 4.6",
   },
-  "opus-fast": {
-    provider: "anthropic",
-    model: "claude-opus-4-6-fast",
-    displayName: "Claude Opus 4.6 Fast",
-  },
   sonnet: {
     provider: "anthropic",
     model: "claude-sonnet-4-6",
@@ -146,7 +145,9 @@ function matchModel(input: string): string | undefined {
   return AVAILABLE_MODELS.find((m) => m.includes(lower));
 }
 
-function resolveProviderModelCommand(content: string): SlashResolution | null {
+async function resolveProviderModelCommand(
+  content: string,
+): Promise<SlashResolution | null> {
   const trimmed = content.trim();
   if (!trimmed.startsWith("/")) return null;
 
@@ -163,7 +164,7 @@ function resolveProviderModelCommand(content: string): SlashResolution | null {
   const name = getAssistantName();
 
   // Check if API key exists for this provider (Ollama doesn't require an API key)
-  if (provider !== "ollama" && !config.apiKeys[provider]) {
+  if (provider !== "ollama" && !(await getSecureKeyAsync(provider))) {
     return {
       kind: "unknown",
       message: `Cannot switch to ${displayName}. No API key configured for ${provider}.\n\nSet it with: \`keys set ${provider} <your-key>\``,
@@ -189,7 +190,7 @@ function resolveProviderModelCommand(content: string): SlashResolution | null {
 
   // Re-initialize providers with new config
   const newConfig = getConfig();
-  initializeProviders(newConfig);
+  await initializeProviders(newConfig);
 
   const switchedMsg = name
     ? `Switched ${name} to **${displayName}**. New conversations will use this model.`
@@ -201,14 +202,23 @@ function resolveProviderModelCommand(content: string): SlashResolution | null {
   };
 }
 
-function resolveModelList(): SlashResolution {
+async function resolveModelList(): Promise<SlashResolution> {
   const config = getConfig();
+
+  // Build a set of providers that have a configured API key.
+  const configuredProviders = new Set<string>(["ollama"]);
+  for (const p of API_KEY_PROVIDERS) {
+    if (await getSecureKeyAsync(p)) {
+      configuredProviders.add(p);
+    }
+  }
+
   const lines = ["Available models:\n"];
 
   for (const [cmd, { provider, model, displayName }] of Object.entries(
     PROVIDER_MODEL_SHORTCUTS,
   )) {
-    const hasKey = provider === "ollama" || !!config.apiKeys[provider];
+    const hasKey = configuredProviders.has(provider);
     const isCurrent = config.provider === provider && config.model === model;
     const status = hasKey ? "✓" : "✗";
     const current = isCurrent ? " **[current]**" : "";
@@ -224,11 +234,13 @@ function resolveModelList(): SlashResolution {
   };
 }
 
-function resolveModelCommand(content: string): SlashResolution | null {
+async function resolveModelCommand(
+  content: string,
+): Promise<SlashResolution | null> {
   const trimmed = content.trim();
   // Match /models → route to list
   if (trimmed === "/models") {
-    return resolveModelList();
+    return await resolveModelList();
   }
 
   if (!trimmed.startsWith("/model")) return null;
@@ -251,7 +263,7 @@ function resolveModelCommand(content: string): SlashResolution | null {
 
   // Handle /model list
   if (args === "list") {
-    return resolveModelList();
+    return await resolveModelList();
   }
 
   // Try to match the model name
@@ -280,7 +292,7 @@ function resolveModelCommand(content: string): SlashResolution | null {
   }
 
   // Validate that Anthropic provider is available
-  if (!currentConfig.apiKeys.anthropic) {
+  if (!(await getSecureKeyAsync("anthropic"))) {
     const displayName = MODEL_DISPLAY_NAMES[matched] ?? matched;
     return {
       kind: "unknown",
@@ -294,7 +306,7 @@ function resolveModelCommand(content: string): SlashResolution | null {
   raw.model = matched;
   saveRawConfig(raw);
   const config = getConfig();
-  initializeProviders(config);
+  await initializeProviders(config);
 
   const displayName = MODEL_DISPLAY_NAMES[matched] ?? matched;
   const switchedMsg = name
@@ -343,16 +355,16 @@ function resolveStatusCommand(context: SlashContext): SlashResolution {
  * Resolve slash commands against the current skill catalog.
  * Returns `unknown` with a deterministic message, or the (possibly rewritten) content.
  */
-export function resolveSlash(
+export async function resolveSlash(
   content: string,
   context?: SlashContext,
-): SlashResolution {
+): Promise<SlashResolution> {
   // Check provider shortcuts first (/gpt4, /opus, etc.)
-  const providerResult = resolveProviderModelCommand(content);
+  const providerResult = await resolveProviderModelCommand(content);
   if (providerResult) return providerResult;
 
   // Handle /model command
-  const modelResult = resolveModelCommand(content);
+  const modelResult = await resolveModelCommand(content);
   if (modelResult) return modelResult;
 
   // Handle /pair command

package/src/daemon/session.ts

@@ -348,6 +348,7 @@ export class Session {
       provider,
       systemPrompt: () => resolveSystemPromptCallback([]).systemPrompt,
       config: config.contextWindow,
+      toolTokenBudget: this.agentLoop.getToolTokenBudget(),
     });
 
     void getHookManager().trigger("session-start", {

package/src/daemon/shutdown-handlers.ts

@@ -24,6 +24,7 @@ export interface ShutdownDeps {
   memoryWorker: { stop(): void };
   qdrantManager: QdrantManager;
   mcpManager: McpServerManager | null;
+  telemetryReporter: { stop(): Promise<void> } | null;
   cleanupPidFile: () => void;
 }
 
@@ -87,6 +88,20 @@ export function installShutdownHandlers(deps: ShutdownDeps): void {
       log.warn({ err }, "Enrichment service shutdown failed (non-fatal)");
     }
 
+    if (deps.telemetryReporter) {
+      try {
+        const timeout = new Promise<void>((_, reject) =>
+          setTimeout(
+            () => reject(new Error("Telemetry flush timed out")),
+            3_000,
+          ),
+        );
+        await Promise.race([deps.telemetryReporter.stop(), timeout]);
+      } catch (err) {
+        log.warn({ err }, "Telemetry reporter shutdown failed (non-fatal)");
+      }
+    }
+
     if (deps.runtimeHttp) await deps.runtimeHttp.stop();
     await browserManager.closeAllPages();
     deps.scheduler.stop();

package/src/daemon/watch-handler.ts

@@ -107,7 +107,7 @@ export async function handleWatchObservation(
 
 async function generateCommentary(session: WatchSession): Promise<void> {
   try {
-    const provider = getConfiguredProvider();
+    const provider = await getConfiguredProvider();
     if (!provider) {
       log.warn(
         { watchId: session.watchId },
@@ -198,7 +198,7 @@ export async function generateSummary(session: WatchSession): Promise<void> {
       },
       "generateSummary starting — calling LLM",
     );
-    const provider = getConfiguredProvider();
+    const provider = await getConfiguredProvider();
     if (!provider) {
       log.warn(
         { watchId: session.watchId },

package/src/email/guardrails.ts

@@ -104,7 +104,7 @@ export function setDailySendCap(cap: number): void {
   saveState(state);
 }
 
-export function isAddressAllowed(email: string): {
+function isAddressAllowed(email: string): {
   allowed: boolean;
   reason?: string;
   rule?: AddressRule;

package/src/email/service.ts

@@ -382,8 +382,3 @@ export function getEmailService(): EmailService {
   }
   return instance;
 }
-
-/** @internal Test-only: reset singleton. */
-export function _resetEmailService(): void {
-  instance = null;
-}

package/src/events/domain-events.ts

@@ -25,6 +25,7 @@ export interface ToolDomainEvents {
       | "allow_10m"
       | "allow_thread"
       | "always_allow"
+      | "always_allow_high_risk"
       | "deny"
       | "always_deny"
       | "temporary_override";

package/src/hooks/templates.ts

@@ -18,7 +18,7 @@ const log = getLogger("hooks-templates");
 
 /**
  * Install bundled hook templates into the user's hooks directory.
- * Templates are copied from `assistant/hook-templates/` to `~/.vellum/workspace/hooks/`.
+ * Templates are copied from `assistant/hook-templates/` to `~/.vellum/hooks/`.
  * - Never overwrites existing hooks (user modifications are preserved).
  * - Newly installed hooks are disabled by default.
  */

package/src/media/app-icon-generator.ts

@@ -15,6 +15,7 @@ import {
   buildManagedBaseUrl,
   resolveManagedProxyContext,
 } from "../providers/managed-proxy/context.js";
+import { getSecureKeyAsync } from "../security/secure-keys.js";
 import { getLogger } from "../util/logger.js";
 import {
   generateImage,
@@ -36,15 +37,15 @@ export async function generateAppIcon(
   appDescription?: string,
 ): Promise<void> {
   const config = getConfig();
-  const apiKey = config.apiKeys.gemini ?? process.env.GEMINI_API_KEY;
+  const apiKey = await getSecureKeyAsync("gemini");
 
   let credentials: ImageGenCredentials | undefined;
   if (apiKey) {
     credentials = { type: "direct", apiKey };
   } else {
-    const managedBaseUrl = buildManagedBaseUrl("vertex");
+    const managedBaseUrl = await buildManagedBaseUrl("vertex");
     if (managedBaseUrl) {
-      const ctx = resolveManagedProxyContext();
+      const ctx = await resolveManagedProxyContext();
       credentials = {
         type: "managed-proxy",
         assistantApiKey: ctx.assistantApiKey,