// How to read files (audio, documents, images) that Sir sent to Alfred on any channel — Slack file uploads, Telegram voice memos, MMS audio, generic URLs. Converts file references into transcripts + vault records so Alfred doesn't say "I don't see your file".
| name | alfred-channel-attachments |
| description | How to read files (audio, documents, images) that Sir sent to Alfred on any channel — Slack file uploads, Telegram voice memos, MMS audio, generic URLs. Converts file references into transcripts + vault records so Alfred doesn't say "I don't see your file". |
| triggers | file, attachment, audio, voice memo, voice note, recording, transcribe, slack file, telegram voice, photo, image, document, pdf, uploaded |
Sir sends files to Alfred across several channels — Slack uploads, Telegram voice memos, SMS/MMS media, email attachments, and direct drops on the dashboard. Every one of those files gets bytes into the vault; your job is to fetch them inline whenever the conversation references an uploaded file, so you can actually respond to what's inside them.
If you ever catch yourself saying "I don't see the file you uploaded" — you forgot this skill. Call the endpoint.
self({
endpoint: "/api/v1/channels/attachment/fetch",
method: "POST",
body: {
source: "slack" | "telegram" | "mms" | "url",
file_ref: "<channel-specific id>", // for slack/telegram
file_url: "<direct url>", // for mms/url
file_name: "<optional hint>", // optional
mime_type: "<optional hint>" // optional
}
})
Returns:
{
"ok": true,
"source": "slack",
"file_name": "voice-memo-2026-04-23.m4a",
"mime_type": "audio/m4a",
"size": 482317,
"vault_path": "inbox/voice-memo-2026-04-23.m4a",
"transcript": "Hey Alfred, I wanted to follow up on the Erste meeting...",
"text": "(same as transcript, or text content for docs)"
}
You get:
transcript — ready-to-use transcript for audio (Whisper via Groq, ~3s round trip).text — inline content for text documents (up to 30KB).vault_path — where it lives. The async MediaIngestionWorkflow will build a full vault record too (so search + surveyor linking pick it up later), but you don't have to wait.file_share event or your session context mentions files[]voice/audio in contextMediaUrl0Always call it before replying. Even if you're not sure — err on the side of fetching. It's cheap.
self({ endpoint: "/api/v1/channels/attachment/fetch", method: "POST",
body: { source: "slack", file_ref: "F0ABC123" } })
file_ref is Slack's file ID (starts with F). If your session context has files[].id, use that. Downloads via the Slack Composio OAuth credential — no extra auth needed from you.
self({ endpoint: "/api/v1/channels/attachment/fetch", method: "POST",
body: { source: "telegram", file_ref: "<file_id>" } })
Telegram's file_id is a long opaque string like AAQB.... Works for voice memos (voice.file_id), audio files (audio.file_id), documents (document.file_id), photos (photo[-1].file_id = largest).
self({ endpoint: "/api/v1/channels/attachment/fetch", method: "POST",
body: { source: "mms", file_url: "<MediaUrl0 from Twilio>" } })
self({ endpoint: "/api/v1/channels/attachment/fetch", method: "POST",
body: { source: "url", file_url: "https://..." } })
Use when Sir pastes a pre-signed URL or a public Dropbox/Drive link.
Email attachments have their own endpoint — alfred-email-channel/SKILL.md already covers that flow using /api/v1/email/attachment/<message_id>/<attachment_id>. Don't route email attachments through this skill.
Every call emits a stream_type: media event to system-media-ingestion.jsonl. The MediaIngestionWorkflow will:
You don't need to wait for this. You already have the content from the endpoint's synchronous response. But it means asking "find the transcript of the voice memo Sir sent yesterday" works later even if you didn't save it explicitly.
If the endpoint returns ok: false:
resolve failed — bad file_ref or the channel's connected account is broken. Tell Sir the file ref wasn't resolvable; don't pretend you read it.download failed — network issue or auth expired. Try once more; if it fails again, tell Sir.GROQ_API_KEY not set — will just skip transcription silently and return the file_name + vault_path. You can still tell Sir the file was saved, but you don't have the content to discuss.Never invent content. If you didn't get a transcript, don't pretend you did.
Drive Paperclip (the AI-agent orchestration plane Sir runs at paperclip.<DOMAIN>) end-to-end through the MCP `paperclip` connector. ~48 first-party tools wrapping Paperclip's REST surface — companies, employees (incl. Alfred-as-an-employee), issues, projects, routines, approvals, comments, governance. Use whenever Sir asks about company state, wants to assign work, check what Alfred-the-employee is doing, file/comment on issues, manage routines, or run governance flows. Use whenever Alfred is acting AS an employee under a Paperclip task assignment (the principal/CEO of the company has tasked Alfred — Alfred answers via comments + issue updates).
Schedule reminders, delegate background work to Alfred-the-text-agent, list scheduled jobs, and inspect Hermes runtime — anything that's about Alfred-the-runtime rather than Sir's vault/calendar/finances. Use whenever Sir asks "remind me…", "research X then tell me…", "what reminders do I have set?", or wants to fire-and-forget a long task on text-mode while you stay on the phone with him.
How to behave on a phone call. Voice persona + tool-usage rules for AgentPhone. The Voice Bridge runs this persona as the OpenAI Realtime session prompt; this file is also the canonical text reference for the openclaw main agent so it stays consistent across channels.
Process raw inbound content (emails, voice memos, notes) into structured Obsidian vault records with proper frontmatter, wikilinks, and file placement.
Read operational vault records and extract latent knowledge into structured learning records with proper frontmatter, wikilinks, and file placement.
Fix vault quality issues — broken frontmatter, invalid values, orphaned records, garbage content.