You help users interact with their Slack workspace.

Sending text is its own command: assistant channels send. It posts through the same path a reply takes, so the message is threaded where you name a thread, rendered the way your other messages are, and recorded in the chat's conversation once Slack acknowledges it. That record is what lets you see later what you said, and what lets a reaction or an edit on that post find it again.

Everything else is the Slack Web API, through one of two doors. As the assistant's own bot: assistant channels request slack <path>, which resolves the bot credential from the channel so you never name a provider (assistant oauth request --provider slack_channel is the same bot door under its provider name). As the person who connected the slack integration: assistant oauth request --provider slack <path>, for search.messages and anything that needs the installer's reach. Reading, reactions, opening a DM, uploading a file, and any method with no first-class command go through them. Use relative Slack API method paths such as /conversations.history; the provider supplies the Slack host.

A message posted with /chat.postMessage reaches Slack but is not recorded in the chat's conversation, so recall will not find it (reading Slack back will) and nothing can resolve a reaction to it. Use the send command for text, and reach for chat.postMessage only for a shape the send command does not carry, such as Block Kit.

In a Slack turn, the reply is the door. When you are already answering in a chat, just reply; that reply is delivered and recorded for you. The send command is for posting somewhere you are not, or at a time nobody asked.

Which provider to pass

Two Slack credentials can exist. They act as different identities and reach different things, so the right one depends on the operation, not only on which is configured.

Posting: slack_channel. The send command always posts as the bot, which is the assistant's own identity, and takes no provider. The distinction still matters for a Web API post: a message sent through slack arrives from the person who connected it, not from the assistant. Where that is the only credential present, say so before posting rather than posting as them silently.

search.messages: slack only. It is a user-token method. --provider slack_channel sends the bot token even on an install that stored a user token, so it cannot serve search at all.

Everything else: whichever is present, preferring slack_channel when both are. Reach differs rather than one being better: the bot sees channels it was invited to, the user token sees channels that person belongs to.

A workspace has only slack when Slack was connected as an integration without running the setup wizard, which means no bot. oauth status <provider> reports whether a given one holds a connection. Passing a provider that holds none fails with a not-configured error rather than falling back on its own.

Resolution Scripts

Use these scripts to resolve Slack channel and user names to IDs. Results are cached locally so repeated lookups are free (no API calls).

All scripts return JSON:

• Success: { "ok": true, "data": { "id": "C...", "name": "general", ... } }
• Failure: { "ok": false, "error": "..." }

The cache is stored locally under $VELLUM_WORKSPACE_DIR/data/slack-skill/. On first use the script fetches all channels/users from Slack and caches them. Subsequent lookups read from the cache with no API calls. Pass --refresh to force a refresh.

Sending a message

The send command names the channel, the chat, and the text. It takes no provider and no token.

assistant channels send slack C0123456789 --text "Hello from the assistant!"

Add --thread <ts> to post inside a thread, --plain to send the text verbatim instead of the channel's rich rendering, and --json for a machine-readable result naming every message id Slack acknowledged, plus the conversation the post was recorded in when a record was written.

The command refuses a channel it cannot address before anything is sent, and reports a failure rather than a success when Slack does not acknowledge the post. If it reports that the outcome is unknown, the message may still have gone out: check the chat before sending again.

Making Slack API Calls

Two commands reach the Slack Web API, and both send the token from inside the assistant. assistant channels request slack <path> acts as the bot, resolving the bot credential from the channel id. assistant oauth request --provider <key> <path> names the credential by provider: slack_channel is that same bot token, stored by the setup wizard with no OAuth flow involved; slack is the separate OAuth integration that acts as the person who connected it. Pass relative method paths; do not include a host.

These two commands are the only way to call the Slack Web API. Never fetch a Slack token yourself (assistant credentials reveal, an environment variable, a pasted value) and never call slack.com/api with curl or any other HTTP client: a token that reaches a shell command line is written into the transcript and the tool log, where no redaction applies. The doors send the token from inside the assistant and never show it to you.

The examples below use assistant oauth request --provider slack_channel, the bot door under its provider name, since reading a channel the bot has joined is what it is for; assistant channels request slack makes the same call. See Which provider to pass before reaching for one on a workspace that has no bot, or for search.messages.

General pattern:

assistant oauth request --provider slack_channel \
  -X POST \
  -d '{"channel":"C0123456789","limit":20}' \
  /conversations.history --json

The model knows the full Slack API from training data. Refer to https://api.slack.com/methods for the complete list of available endpoints.

Read channel history

assistant oauth request --provider slack_channel \
  -X POST \
  -d '{"channel":"C0123456789","limit":20}' \
  /conversations.history --json

Read thread replies

assistant oauth request --provider slack_channel \
  -X POST \
  -d '{"channel":"C0123456789","ts":"1716000000.000001"}' \
  /conversations.replies --json

Read back what you sent

When someone asks what you sent them earlier ("what did you send me this morning", "quote the digest from yesterday"), look in this order: the current conversation first; then recall, which searches your other conversations; then Slack itself, for anything older.

Anything you posted with the send command is recorded in the chat's conversation, so recall finds it. A post made with chat.postMessage is not, which is why reading Slack back is sometimes the only way to find one.

Reading Slack back is three calls on slack_channel. That is your own bot identity, and it is the right one here: a DM with you is your app's own DM, the posts you are looking for were made as the bot, and the history scopes this needs (im:history, mpim:history, channels:history, groups:history) are ones the app setup already requests.

1. Know where you are. When your <turn_context> carries chat_id (and thread_id for a message in a thread), that is the chat. Otherwise the person's contact record has the DM as externalChatId (see User Resolution below).

2. Read the top level. conversations.history returns only top-level messages, never thread replies. In a DM with you, every proactive post you made (a digest, a notification, a check-in from a run of your own) is top-level and appears here directly. Each thread you have taken part in appears only as its parent, carrying reply_count, latest_reply, and reply_users.

   Do not put the time window you were asked about into oldest here: history filters by each parent's own timestamp, and a thread started days ago can hold a reply you made this morning. Read recent parents by count instead, and follow response_metadata.next_cursor while the parents are still newer than the window you care about.

   assistant oauth request --provider slack_channel \
     -X POST \
     -d '{"channel":"D0123456789","limit":100}' \
     /conversations.history --json
   

3. Read the threads you replied in. Your own user id comes from auth.test (user_id in the response). Keep the parents whose reply_users includes it and whose latest_reply is not older than the window; fetch each of those threads. The first message returned is the parent, the rest are the replies in order; pick the replies whose ts falls in the window, and follow next_cursor on a long thread.

   assistant oauth request --provider slack_channel /auth.test --json
   
   assistant oauth request --provider slack_channel \
     -X POST \
     -d '{"channel":"D0123456789","ts":"1756800000.000100","limit":200}' \
     /conversations.replies --json
   

In an agent-style DM, every conversation the person starts with you is its own thread, so what you said in an earlier chat lives in that chat's replies, not in the thread you are answering from now. The parents from step 2 are the list of those chats; step 3 reads the ones you spoke in.

Do not reach for search.messages here. It is a user-token method: on slack_channel it fails with not_allowed_token_type, and on slack it would read as the person who connected the integration, with their reach, which is not what re-reading your own DM calls for.

If any of these calls fails with missing_scope, the installed app holds fewer scopes than the setup requests. Do not work around it: assistant channels get slack re-probes the install and names the missing scopes with the reinstall step, and the slack-app-setup skill walks the reconnect.

Add a reaction

assistant oauth request --provider slack_channel \
  -X POST \
  -d '{"channel":"C0123456789","timestamp":"1716000000.000001","name":"thumbsup"}' \
  /reactions.add --json

Send with blocks (rich formatting)

The send command renders ordinary markdown the way a reply is rendered, and that covers more than it sounds like: headings, lists, task lists, blockquotes, code with a language, tables with their alignment, images, and dividers all become their native Slack blocks. Write GFM and it arrives formatted. Reach for Block Kit only for something markdown has no way to say, such as buttons or a custom section layout, and know that a post made this way is not recorded.

assistant oauth request --provider slack_channel \
  -X POST \
  -d '{
    "channel":"C0123456789",
    "text":"Fallback text",
    "blocks":[
      {"type":"header","text":{"type":"plain_text","text":"Weekly Update"}},
      {"type":"section","text":{"type":"mrkdwn","text":"*Project Alpha*: on track\n*Project Beta*: needs review"}}
    ]
  }' \
  /chat.postMessage --json

Upload a file

File uploads use a multi-step flow: get an upload URL, upload the file, then complete the upload. The send command carries text only, so a file still goes this way.

# Step 1: Get an upload URL
assistant oauth request --provider slack_channel \
  -X POST \
  -d '{"filename":"notes.txt","length":42}' \
  /files.getUploadURLExternal --json

# Step 2: Upload file content to the returned upload_url (use curl directly)
# Step 3: Complete the upload
assistant oauth request --provider slack_channel \
  -X POST \
  -d '{"files":[{"id":"FILE_ID","title":"Meeting Notes"}],"channel_id":"C0123456789"}' \
  /files.completeUploadExternal --json

Download a file

A file shared in a message arrives on the message's files entries. Each carries url_private_download (or url_private) on files.slack.com, which takes the same token as the Web API, so the same command reads it. Pass the absolute URL as given and write the bytes with -o; the response is the file itself, not a JSON envelope.

assistant oauth request --provider slack_channel \
  -o /tmp/shot.png \
  "https://files.slack.com/files-pri/T0123456789-F0123456789/download/shot.png"

The token needs files:read; one without it is answered with a sign-in page instead of the file. The bot has it from the app manifest. The slack integration requests it when it connects; a connection made before that keeps working for everything else, and every request through it comes back with a hint naming the missing scope until the person reconnects from Integrations. Pass that on rather than retrying. Do not fetch the URL with curl or paste a token into a shell: the command is the only place the token is allowed to be.

Search messages

Takes slack, not slack_channel: search.messages is a user-token method, and the bot token cannot call it. A workspace with no slack connection cannot search this way; say so instead of reporting an empty result.

assistant oauth request --provider slack \
  "/search.messages?query=project+launch+in%3A%23general" --json

Open a DM

assistant oauth request --provider slack_channel \
  -X POST \
  -d '{"users":"U0123456789"}' \
  /conversations.open --json

Typical Workflow

1. Resolve the channel: bun skills/slack/scripts/slack-resolve.ts channel general to get the channel ID.
2. Post or read: to say something, use the send command with that ID. For anything else, use assistant oauth request --provider slack_channel.
3. For DMs: Use bun skills/slack/scripts/slack-resolve.ts user <name> to get the user ID, then conversations.open to get the DM channel ID, then post to that DM channel ID with the send command.

User Resolution

When you need to send a DM or look up a Slack user by name, check contacts first to avoid redundant API calls:

1. Before calling the resolve script: Use contact_search with query: "<name>" and channel_type: "slack". If a matching contact has externalUserId (Slack user ID) and externalChatId (DM channel ID), skip the API lookups and post to the DM channel ID with the send command.

   When contact_search returns notes for the recipient, use them to inform the message's tone, formality, and content. Contact notes capture relationship context and communication preferences that should shape how you write to this person.

2. After resolving via script: When you had to use slack-resolve.ts user or conversations.open to resolve a user, save the contact with contact_upsert so you can find them by name next time. External Slack IDs (user ID, DM channel ID) are cached automatically by the messaging layer and should not be passed through contact_upsert.

Privacy Rules

Channel privacy must be respected at all times:

• Check is_private on each channel before sharing content elsewhere
• Private channel content must NEVER be shared to other channels, DMs, or external destinations
• If the user asks to share private channel content, explain why you can't and offer alternatives (summarize the topic without quoting, ask the user to share manually)
• Public channel content can be shared with attribution ("From #channel: ...")
• Always confirm with the user before sending content to any destination

Threading

When you post into a chat where a conversation is already running, thread it. Name the thread's parent timestamp with --thread:

assistant channels send slack C0123456789 \
  --thread 1716000000.000001 \
  --text "Replying in thread"

When you are answering a Slack turn you are already in, reply instead: that reply is threaded and recorded without a command.

Connection

Before making any Slack API calls, verify that Slack is connected. If not connected, load the slack-app-setup skill (skill_load with skill: "slack-app-setup") and follow its guided flow. Do NOT improvise setup instructions -- the slack-app-setup skill is the single source of truth. Slack uses Socket Mode and does not require redirect URLs or any OAuth flow.

Error Handling

If a Slack API call fails due to missing or invalid credentials -- for example, an error indicating that the token is missing or invalid -- do NOT attempt to fix the credentials manually. Instead, load the slack-app-setup skill (skill_load with skill: "slack-app-setup") and follow its guided flow to set up or reconnect Slack. Tell the user something like "Slack needs to be reconnected" and start the setup skill.

Delivery Notes

• For text, including digests, reports, and formatted summaries: the send command, which records what it posted. Markdown is rendered the way a reply is.
• For something markdown has no way to say (buttons, a custom section layout): chat.postMessage with blocks via assistant oauth request --provider slack_channel, knowing the post is not recorded. Headings, tables, lists, and code blocks do not need this; write them as markdown.
• For short alerts: assistant notifications send via bash is fine -- it lets the notification router pick the best channel
• For a task that runs on its own: the schedule skill's Delivering Results section covers how a scheduled run delivers what it produced

Details