@vellumai/assistant 0.4.55 → 0.4.57

package/src/config/bundled-skills/skill-management/TOOLS.json

@@ -3,7 +3,7 @@
   "tools": [
     {
       "name": "scaffold_managed_skill",
-      "description": "Create or update a managed skill in ~/.vellum/workspace/skills. The skill becomes available for skill_load immediately.",
+      "description": "Create or update a managed skill in ~/.vellum/workspace/skills. The skill becomes available for skill_load immediately. Never persist a skill without explicit user consent. Before persisting, test the snippet: write to a temp file with bash and run with `bun run /tmp/vellum-eval/snippet.ts`. Iterate up to 3 attempts, then ask the user. Clean up temp files after. Do not use file_write for temp files outside the working directory. After a skill is written, the next turn may run in a recreated session due to file-watcher eviction - continue normally.",
       "category": "skills",
       "risk": "high",
       "input_schema": {
@@ -23,20 +23,12 @@
           },
           "body_markdown": {
             "type": "string",
-            "description": "The full skill body in markdown — instructions, prompts, templates, etc."
+            "description": "The full skill body in markdown - instructions, prompts, templates, etc."
           },
           "emoji": {
             "type": "string",
             "description": "Optional emoji icon for the skill."
           },
-          "user_invocable": {
-            "type": "boolean",
-            "description": "Whether users can invoke this skill directly (default: true)."
-          },
-          "disable_model_invocation": {
-            "type": "boolean",
-            "description": "Whether to prevent the model from auto-invoking this skill (default: false)."
-          },
           "overwrite": {
             "type": "boolean",
             "description": "Whether to overwrite an existing skill with the same ID (default: false)."
@@ -50,7 +42,7 @@
             "items": { "type": "string" },
             "description": "Optional list of child skill IDs that this skill includes (metadata only, no auto-activation)."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of what you are creating and why, shown to the user as a status update. Use simple language a non-technical person would understand."
           }
@@ -62,7 +54,7 @@
     },
     {
       "name": "delete_managed_skill",
-      "description": "Delete a managed skill from ~/.vellum/workspace/skills and remove it from the SKILLS.md index.",
+      "description": "Delete a managed skill from ~/.vellum/workspace/skills and remove it from the SKILLS.md index. Never delete a skill without explicit user confirmation. After deletion, the next turn may run in a recreated session due to file-watcher eviction - continue normally.",
       "category": "skills",
       "risk": "high",
       "input_schema": {
@@ -76,7 +68,7 @@
             "type": "boolean",
             "description": "Whether to remove the skill from SKILLS.md index (default: true)."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of what you are deleting and why, shown to the user as a status update. Use simple language a non-technical person would understand."
           }

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

@@ -0,0 +1,84 @@
+---
+name: skills-catalog
+description: Discover bundled skills and search/install community skills from the skills.sh registry
+compatibility: "Designed for Vellum personal assistants"
+metadata:
+  emoji: "\U0001F9E9"
+  vellum:
+    display-name: "Skills Catalog"
+    activation-hints:
+      - "what can you do"
+      - "find a skill"
+      - "install a skill"
+      - "community skills"
+      - "skills.sh"
+      - "search for skills"
+    avoid-when:
+      - "User is asking about a specific bundled skill already visible in the catalog"
+---
+
+You can help the user discover what skills are available and find community skills to extend the assistant's capabilities.
+
+## Bundled skills (first-party)
+
+First-party skills are **bundled** with the assistant - they are compiled in and always available. They do not need to be installed or downloaded. To activate a bundled skill, use the `skill_load` tool:
+
+```
+skill_load skill=<skill-id>
+```
+
+The skill catalog shown in the system prompt lists all bundled skills with their IDs. When a user asks about capabilities, refer to this list to find relevant bundled skills and load them as needed.
+
+## Community skills (skills.sh)
+
+Community skills are published on the skills.sh registry and can be searched, inspected, and installed on demand using the `assistant skills` CLI.
+
+### Searching for community skills
+
+```bash
+assistant skills search "<query>"
+```
+
+Returns matching skills with their slug, source, install counts, and security audit badges. Use this when the user asks for a capability not covered by bundled skills.
+
+### Installing a community skill
+
+**Trust model - check the source owner before installing:**
+
+- **Vellum-owned** (`vellum-ai/*`): First-party skills published by the Vellum team. Install these directly without prompting - they are vetted and trusted.
+- **Third-party** (any other owner): Ask the user for permission first. Present the skill name, source, audit results, and install count. Say something like: "I found a community skill that could help, but it's published by a third party - we haven't vetted it. Want to install it anyway?"
+
+```bash
+assistant skills add <owner>/<repo>@<skill-name>
+```
+
+For example:
+```bash
+assistant skills add vercel-labs/skills@find-skills
+```
+
+Once installed, the skill appears in the workspace skills directory and can be loaded with `skill_load` like any other skill.
+
+## Typical flow
+
+1. **User asks about capabilities** - "Can you order food?" or "What can you do?"
+   - Check the bundled skills list in the system prompt
+   - Present relevant skills to the user
+   - Load any that match with `skill_load`
+
+2. **User wants a capability not covered by bundled skills** - "Can you do X?"
+   - Search with `assistant skills search "<query>"`
+   - Present matching results with descriptions, install counts, and audit badges
+   - Check the source owner to determine trust level (see trust model above)
+   - Install with `assistant skills add <owner>/<repo>@<skill-name>`
+   - Load it with `skill_load`
+
+3. **Skill has dependencies** - if `includes` lists other skill IDs, load those first with `skill_load`
+
+## Notes
+
+- Bundled skills are always available and do not need installation
+- Community skills are installed to the workspace skills directory
+- After installing a community skill, it is auto-enabled and immediately loadable
+- Skills can be enabled or disabled via feature flags without uninstalling them
+- Never install third-party community skills without explicit user confirmation

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

@@ -1,12 +1,11 @@
 ---
 name: slack
-description: Scan channels, summarize threads, and manage Slack with privacy guardrails
+description: Scan channels, summarize threads, and manage Slack
 compatibility: "Designed for Vellum personal assistants"
 metadata:
   emoji: "💬"
   vellum:
     display-name: "Slack"
-    user-invocable: true
 ---
 
 You are a Slack assistant that helps users stay on top of their Slack workspace. Use the slack tools for channel scanning, thread summarization, and Slack-specific operations.
@@ -69,7 +68,7 @@ When responding to messages from Slack channels, replies are automatically threa
 When you need to **send** content to Slack proactively (e.g. a scheduled digest, a scan summary, or a report):
 
 - Use `messaging_send` with `platform: "slack"` and the target `conversation_id` (Slack channel or DM ID). This posts directly via `chat.postMessage` and preserves the full message content.
-- Do **NOT** use `send_notification` for rich content like digests — the notification router's decision engine rewrites content into short alerts, stripping the actual digest.
+- Do **NOT** use `send_notification` for rich content like digests - the notification router's decision engine rewrites content into short alerts, stripping the actual digest.
 - `send_notification` is appropriate for short alerts and status updates where you want the router to pick the best channel. `messaging_send` is appropriate when you have specific content to deliver to a specific Slack destination.
 - For scheduled tasks (cron/RRULE), always end with a `messaging_send` call so the results actually reach the user. Without it, the output only lives in the conversation log.
 

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

@@ -33,7 +33,7 @@
             "enum": ["text", "blocks"],
             "description": "Output format: 'text' returns JSON (default), 'blocks' returns Slack Block Kit structured output"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -54,7 +54,7 @@
             "type": "string",
             "description": "The Slack channel ID to get details for"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -84,7 +84,7 @@
             },
             "description": "Channel IDs to add, remove, or set (not needed for 'list')"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -118,7 +118,7 @@
             "type": "number",
             "description": "Confidence score (0-1) for this action"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -148,7 +148,7 @@
             "type": "number",
             "description": "Confidence score (0-1) for this action"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -182,7 +182,7 @@
             "type": "number",
             "description": "Confidence score (0-1) for this action"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -208,7 +208,7 @@
             "type": "number",
             "description": "Confidence score (0-1) for this action"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -258,7 +258,7 @@
             "enum": ["restricted", "standard"],
             "description": "Trust level override: 'restricted' limits tool access, 'standard' uses defaults (for 'set' action)"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }

package/src/config/bundled-skills/slack/tools/slack-scan-digest.ts

@@ -165,7 +165,7 @@ function buildBlockKitOutput(
     type: "header",
     text: {
       type: "plain_text",
-      text: `Slack Digest — ${digests.length} channel${digests.length !== 1 ? "s" : ""} scanned`,
+      text: `Slack Digest - ${digests.length} channel${digests.length !== 1 ? "s" : ""} scanned`,
     },
   });
 
@@ -185,7 +185,7 @@ function buildBlockKitOutput(
         type: "section",
         text: {
           type: "mrkdwn",
-          text: `${digest.isPrivate ? "\ud83d\udd12 " : ""}*#${digest.channelName}* — _Error: ${digest.error}_`,
+          text: `${digest.isPrivate ? "\ud83d\udd12 " : ""}*#${digest.channelName}* - _Error: ${digest.error}_`,
         },
       });
       blocks.push({ type: "divider" });
@@ -198,7 +198,7 @@ function buildBlockKitOutput(
       type: "section",
       text: {
         type: "mrkdwn",
-        text: `${privacyIcon}*#${digest.channelName}* — ${digest.messageCount} message${digest.messageCount !== 1 ? "s" : ""}`,
+        text: `${privacyIcon}*#${digest.channelName}* - ${digest.messageCount} message${digest.messageCount !== 1 ? "s" : ""}`,
       },
     });
 

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

@@ -15,12 +15,12 @@ Subagent orchestration -- spawn background agents to work on tasks in parallel.
 Subagents follow this status flow: `pending` -> `running` -> `completed` / `failed` / `aborted`
 
 - **Spawn**: Use `subagent_spawn` with a label and objective. The subagent runs autonomously.
-- **Auto-notification**: The parent session is automatically notified when a subagent reaches a terminal status. Do NOT poll `subagent_status`.
+- **Auto-notification**: The parent conversation is automatically notified when a subagent reaches a terminal status. Do NOT poll `subagent_status`.
 - **Read output**: Use `subagent_read` only after the subagent reaches a terminal status (completed/failed/aborted).
 
 ## Ownership
 
-Only the parent session that spawned a subagent can interact with it (check status, send messages, abort, or read output).
+Only the parent conversation that spawned a subagent can interact with it (check status, send messages, abort, or read output).
 
 ## Silent Mode
 

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

@@ -25,7 +25,7 @@
             "type": "boolean",
             "description": "Whether to present the subagent's result to the user when it completes. Defaults to true. Set to false for internal/silent processing."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -37,7 +37,7 @@
     },
     {
       "name": "subagent_status",
-      "description": "Get the status of a specific subagent or list all subagents for the current session. Only use this when the user explicitly asks about subagent status \u2014 do NOT poll automatically, as you will be notified when subagents complete.",
+      "description": "Get the status of a specific subagent or list all subagents for the current conversation. Only use this when the user explicitly asks about subagent status \u2014 do NOT poll automatically, as you will be notified when subagents complete.",
       "category": "orchestration",
       "risk": "low",
       "input_schema": {
@@ -45,9 +45,9 @@
         "properties": {
           "subagent_id": {
             "type": "string",
-            "description": "Optional subagent ID to query. If omitted, returns all subagents for this session."
+            "description": "Optional subagent ID to query. If omitted, returns all subagents for this conversation."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -69,7 +69,7 @@
             "type": "string",
             "description": "The ID of the subagent to abort."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -95,7 +95,7 @@
             "type": "string",
             "description": "The message content to send to the subagent."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -117,7 +117,7 @@
             "type": "string",
             "description": "The ID of the subagent whose output to read."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }

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

@@ -29,4 +29,4 @@ Use `task_list_add` to enqueue items (ad-hoc or from a template), `task_list_sho
 - When the user says "add to my tasks" or "add to my queue", use `task_list_add` (NOT schedule_create).
 - Use `task_save` only when the user wants to capture a conversation pattern as a reusable template.
 - `task_list` shows saved templates; `task_list_show` shows the active work queue.
-- **Always specify `required_tools`** when calling `task_list_add`. Think about what tools the task will need at execution time and list them explicitly (e.g. `["host_bash"]` for shell commands, `["host_file_read", "host_file_write"]` for file operations, `["web_search", "web_fetch"]` for web lookups). The user must approve these tools before the task can run — omitting them forces a fallback to all tools, which is noisy and may miss non-standard tools the task actually needs.
+- **Always specify `required_tools`** when calling `task_list_add`. Think about what tools the task will need at execution time and list them explicitly (e.g. `["host_bash"]` for shell commands, `["host_file_read", "host_file_write"]` for file operations, `["web_search", "web_fetch"]` for web lookups). The user must approve these tools before the task can run - omitting them forces a fallback to all tools, which is noisy and may miss non-standard tools the task actually needs.

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

@@ -17,7 +17,7 @@
             "type": "string",
             "description": "Optional override for the auto-generated task title"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -50,7 +50,7 @@
               "type": "string"
             }
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -67,7 +67,7 @@
       "input_schema": {
         "type": "object",
         "properties": {
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -91,7 +91,7 @@
             },
             "description": "One or more task IDs to delete."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -123,7 +123,7 @@
             ],
             "description": "Optional status filter. A single status string (e.g. \"queued\") or an array of statuses to include."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -180,7 +180,7 @@
             },
             "description": "Tools the task will need at execution time. Always specify this \u2014 think about what tools are needed to accomplish the task (e.g. [\"host_bash\"] for shell commands, [\"host_file_read\", \"host_file_write\"] for file operations, [\"web_search\"] for web lookups). The user must approve these before the task runs. Pass [] if no tools are needed."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -259,7 +259,7 @@
             "type": "number",
             "description": "Disambiguation filter: pick the Nth oldest match (1 = oldest, 2 = second oldest, etc.) when multiple items share the same title/task_id."
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -311,7 +311,7 @@
             "type": "number",
             "description": "Disambiguator: 1-indexed creation order among matches (1 = oldest, 2 = second oldest, etc.)"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }
@@ -340,7 +340,7 @@
             "type": "string",
             "description": "Work item title to search for (case-insensitive substring match)"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }

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

@@ -1,6 +1,6 @@
 ---
 name: transcribe
-description: Transcribe audio and video files using Whisper (cloud API or local)
+description: Transcribe audio and video files using Whisper
 compatibility: "Designed for Vellum personal assistants"
 metadata:
   emoji: "🎙️"
@@ -8,14 +8,14 @@ metadata:
     display-name: "Transcribe"
 ---
 
-Transcribe audio and video files using OpenAI's Whisper model — either via the cloud API or locally via whisper.cpp.
+Transcribe audio and video files using OpenAI's Whisper model - either via the cloud API or locally via whisper.cpp.
 
 ## Choosing a Mode
 
 Before transcribing, **ask the user which mode they prefer** if they haven't specified:
 
-1. **`api`** — Uses the OpenAI Whisper API. Fast, accurate, no setup needed. Requires an OpenAI API key (check if one is already configured). Audio is sent to OpenAI's servers. Costs ~$0.006/min.
-2. **`local`** — Uses whisper.cpp installed via Homebrew. Free, private, runs entirely on-device. Requires a one-time `brew install whisper-cpp`. Slightly slower but no data leaves the machine.
+1. **`api`** - Uses the OpenAI Whisper API. Fast, accurate, no setup needed. Requires an OpenAI API key (check if one is already configured). Audio is sent to OpenAI's servers. Costs ~$0.006/min.
+2. **`local`** - Uses whisper.cpp installed via Homebrew. Free, private, runs entirely on-device. Requires a one-time `brew install whisper-cpp`. Slightly slower but no data leaves the machine.
 
 If the user says "cloud", "API", or "online" → use `api`.
 If the user says "local", "offline", "private", or "on-device" → use `local`.
@@ -25,5 +25,5 @@ If the user says "local", "offline", "private", or "on-device" → use `local`.
 - The tool accepts either a `file_path` (absolute path to a local file) or an `attachment_id` (for uploaded attachments). Prefer `file_path` when the user references a file on disk.
 - Supported formats: any video (mp4, mov, etc.) or audio (mp3, wav, m4a, etc.) file.
 - For video files, audio is automatically extracted via ffmpeg before transcription.
-- The API mode has a 25MB per-request limit — large files are automatically split into chunks.
+- The API mode has a 25MB per-request limit - large files are automatically split into chunks.
 - Local mode requires whisper.cpp (`brew install whisper-cpp`). The model is downloaded automatically on first use.

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

@@ -22,7 +22,7 @@
             "enum": ["api", "local"],
             "description": "Transcription backend: 'api' for OpenAI Whisper API (cloud), 'local' for whisper.cpp (on-device)"
           },
-          "reason": {
+          "activity": {
             "type": "string",
             "description": "Brief non-technical explanation of why this tool is being called"
           }

package/src/config/bundled-skills/transcribe/tools/transcribe-media.ts

@@ -10,12 +10,16 @@ import {
 import { tmpdir } from "node:os";
 import { extname, join } from "node:path";
 
-import { getAttachmentsByIds } from "../../../../memory/attachments-store.js";
-import { getSecureKeyAsync } from "../../../../security/secure-keys.js";
+import {
+  getAttachmentsByIds,
+  getFilePathForAttachment,
+} from "../../../../memory/attachments-store.js";
+import { getProviderKeyAsync } from "../../../../security/secure-keys.js";
 import type {
   ToolContext,
   ToolExecutionResult,
 } from "../../../../tools/types.js";
+import { silentlyWithLog } from "../../../../util/silently.js";
 import {
   FFMPEG_TRANSCODE_TIMEOUT_MS,
   FFPROBE_TIMEOUT_MS,
@@ -46,7 +50,7 @@ const AUDIO_EXTENSIONS = new Set([
 /** Max file size for a single OpenAI Whisper API request (25MB). */
 const WHISPER_API_MAX_BYTES = 25 * 1024 * 1024;
 
-/** Duration per chunk when splitting for the API (10 minutes — stays well under 25MB as WAV). */
+/** Duration per chunk when splitting for the API (10 minutes - stays well under 25MB as WAV). */
 const API_CHUNK_DURATION_SECS = 600;
 
 /** Timeout for a single Whisper API request. */
@@ -160,6 +164,26 @@ async function resolveSource(
         isError: true,
       };
     }
+    // Check if this is a file-backed attachment (large files stored on disk)
+    const onDiskPath = getFilePathForAttachment(attachment.id);
+    if (onDiskPath) {
+      // File-backed attachment - use the on-disk file directly
+      try {
+        await access(onDiskPath);
+      } catch {
+        return {
+          content: `Attachment file not found on disk: ${onDiskPath}`,
+          isError: true,
+        };
+      }
+      return {
+        inputPath: onDiskPath,
+        isVideo: mime.startsWith("video/"),
+        tempFile: null,
+      };
+    }
+
+    // Inline attachment - decode base64 to a temp file
     const ext = mime.startsWith("video/") ? ".mp4" : ".m4a";
     const tempPath = join(
       tmpdir(),
@@ -193,7 +217,7 @@ async function toWav(inputPath: string, isVideo: boolean): Promise<string> {
 }
 
 // ---------------------------------------------------------------------------
-// API mode — OpenAI Whisper API
+// API mode - OpenAI Whisper API
 // ---------------------------------------------------------------------------
 
 async function transcribeViaApi(
@@ -220,7 +244,7 @@ async function transcribeViaApi(
     context.onOutput?.(
       `Large file (${Math.round(
         duration / 60,
-      )}min) — splitting into chunks...\n`,
+      )}min) - splitting into chunks...\n`,
     );
     const chunks = await splitAudio(
       audioPath,
@@ -239,7 +263,10 @@ async function transcribeViaApi(
     return parts.join(" ");
   } finally {
     const { rm } = await import("node:fs/promises");
-    await rm(chunkDir, { recursive: true, force: true }).catch(() => {});
+    await silentlyWithLog(
+      rm(chunkDir, { recursive: true, force: true }),
+      "transcribe chunk cleanup",
+    );
   }
 }
 
@@ -278,7 +305,7 @@ async function whisperApiRequest(
 }
 
 // ---------------------------------------------------------------------------
-// Local mode — whisper.cpp
+// Local mode - whisper.cpp
 // ---------------------------------------------------------------------------
 
 async function transcribeViaLocal(
@@ -293,20 +320,20 @@ async function transcribeViaLocal(
     );
   }
 
-  // Resolve model path — use the base model, download if needed
+  // Resolve model path - use the base model, download if needed
   const modelPath = await resolveWhisperModel(context);
 
   const duration = await getAudioDuration(audioPath);
 
   if (duration > 0 && duration <= 1800) {
-    // Under 30 minutes — transcribe directly (whisper.cpp handles long files well)
+    // Under 30 minutes - transcribe directly (whisper.cpp handles long files well)
     context.onOutput?.(
       `Transcribing ${Math.round(duration / 60)}min of audio locally...\n`,
     );
     return await whisperCppRun(audioPath, modelPath);
   }
 
-  // Very long files — split into 10-minute chunks to show progress
+  // Very long files - split into 10-minute chunks to show progress
   const chunkDir = join(
     tmpdir(),
     `vellum-transcribe-local-chunks-${randomUUID()}`,
@@ -317,7 +344,7 @@ async function transcribeViaLocal(
     context.onOutput?.(
       `Large file (${Math.round(
         duration / 60,
-      )}min) — splitting into chunks...\n`,
+      )}min) - splitting into chunks...\n`,
     );
     const chunks = await splitAudio(audioPath, chunkDir, 600);
     const parts: string[] = [];
@@ -332,7 +359,10 @@ async function transcribeViaLocal(
     return parts.join(" ");
   } finally {
     const { rm } = await import("node:fs/promises");
-    await rm(chunkDir, { recursive: true, force: true }).catch(() => {});
+    await silentlyWithLog(
+      rm(chunkDir, { recursive: true, force: true }),
+      "transcribe chunk cleanup",
+    );
   }
 }
 
@@ -419,7 +449,7 @@ export async function run(
   // Validate API key for api mode
   let openaiKey: string | undefined;
   if (mode === "api") {
-    openaiKey = await getSecureKeyAsync("openai");
+    openaiKey = await getProviderKeyAsync("openai");
     if (!openaiKey) {
       return {
         content:

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

@@ -12,10 +12,10 @@ Create and manage watchers that poll external services for events and process th
 
 ## Concepts
 
-- **Provider** — The external service to poll (e.g. "gmail"). Each provider defines how to fetch and parse events.
-- **Action prompt** — LLM instructions for handling detected events. Sent along with event data to a background conversation.
-- **Poll interval** — How often to check for new events (minimum 15 seconds, default 60 seconds).
-- **Digest** — Summary of recent watcher activity, grouped by watcher with time-based filtering.
+- **Provider** - The external service to poll (e.g. "gmail"). Each provider defines how to fetch and parse events.
+- **Action prompt** - LLM instructions for handling detected events. Sent along with event data to a background conversation.
+- **Poll interval** - How often to check for new events (minimum 15 seconds, default 60 seconds).
+- **Digest** - Summary of recent watcher activity, grouped by watcher with time-based filtering.
 
 ## Lifecycle