Slack Connector (experimental)

Post messages to a Slack workspace from a Caffeine canister.

⚠️ Experimental (slack-client@0.1.0). Verified against live Slack: the request path — a real chat.postMessage posts (the earlier query-vs-form bug is resolved; POST params go in an application/x-www-form-urlencoded body). Implemented in 0.1.0 but not yet exercised live: the success-response decode ({"ok":true,…} → the success schema) and the {"ok":false} error-envelope routing (see Known limitations for what each does). The response side is therefore code-complete but unproven — diagnostics is on, so a decode failure surfaces the raw Slack body. Don't present Slack as a fully supported platform feature until one live run confirms a decoded success reply; that condition, not a version number, is the gate.

Orchestrator routing notes

Load this skill when the user, spec, or a prior task mentions Slack, posting to a channel, or notifying a Slack workspace. The generated slack-client package is the preferred path; raw ic.http_request to https://slack.com/api/* is an anti-pattern that re-implements auth, percent-encoding, and JSON parsing by hand.

Intent → capability mapping:

Before generating code, report the token choice back to the prompting user (bot vs user — see Auth model for the one-line rule and the trade-offs) and tell them where to obtain it. They cannot proceed without pasting a token, so surfacing this early avoids an app that traps on first use.

The token is something a human has to go and create, so the build is not done when the backend compiles. It is done when the app itself tells the admin how to get a token and gives them somewhere to paste it — see Obtaining a bot token for the steps, Scopes to request for what to tick, Addressing a channel for the channel ID, and Frontend for the page that MUST ship. Repeat those steps in the completion message too; the chat outlives neither the app nor the admin's memory.

Scope of this generated drop: the messaging-core families chat, conversations, users, files, reactions, and pins (10 API modules) — see Available API surface for the per-module breakdown. Other Slack methods are out of scope until the spec is regenerated.

Auth model — bot token (xoxb-) or user token (xoxp-)

Both are bearer credentials and the client treats them identically, but they differ in who the workspace sees acting. Ask the prompting user which one they want before writing code, and state the trade-off — the answer changes what their app looks like in Slack, and it cannot be swapped later without a re-install.

If the request is "post notifications/alerts from my app", that is xoxb-. Only choose xoxp- when the user explicitly wants messages to look like they came from a human, or needs a user-only API. Say which you picked and why.

Obtaining a bot token (xoxb-)

1. https://api.slack.com/apps → Create New App → From scratch, pick the workspace.
2. OAuth & Permissions → Scopes → Bot Token Scopes: add the scopes the build needs — chat:write at minimum; see Scopes to request below for the rest, and get the list right now (a later addition forces a reinstall).
3. Install to Workspace → authorize → copy the Bot User OAuth Token, which starts with xoxb-.
4. In Slack, invite the app to each target channel: /invite @YourApp. Skipping this is the most common first failure (not_in_channel).

Scopes to request

Step 2 above decides what the app can do — the admin cannot guess the list, so derive it from the features the build actually uses and show exactly those in the settings UI (see Frontend). Bot-token scopes:

⚠️ Adding a scope later invalidates the token. Slack requires a reinstall after any scope change, and the reinstall issues a new xoxb- — the old one keeps working for existing scopes but never gains the new one, so the app fails with missing_scope until the admin pastes the new token. Get the scope list right the first time, and make the settings page re-pasteable rather than one-shot.

Obtaining a user token (xoxp-)

1. Same app → OAuth & Permissions → Scopes → User Token Scopes (a separate list from bot scopes): add e.g. chat:write, search:read.
2. Install to Workspace (or Reinstall, if the app already exists) and authorize — the token represents whoever clicks Allow.
3. Copy the User OAuth Token, which starts with xoxp-.

A single admin-supplied xoxp- is supported by this recipe: it goes through the same setter and the same config.auth. What is out of scope here is per-user OAuth, i.e. each end-user authorising their own account — that needs a full redirect + code-exchange + refresh flow, for which no Slack helper exists yet. Do not attempt to hand-roll it.

Handing the token to the canister

Whichever flavour, the workspace admin pastes it into the canister through an admin-gated setter — gated on AccessControl.hasPermission(state, caller, #admin). The token is held by the canister only and is never returned to the frontend.

⚠️ Never gate the setter on a first-caller-claims-ownership scheme. On the IC every unauthenticated caller is the same anonymous principal, so if an anonymous call claims ownership first, every anonymous caller passes the caller == owner check and can overwrite the workspace token. Use the authorization component's #admin permission, as the example below does.

The canister then hands that token to the client only through config.auth = ?#bearer(token), which every method turns into an Authorization: Bearer … header. No method takes a token argument and no method puts the credential in the URL, so it cannot leak through a logged query string.

Outcalls are already non-replicated

defaultConfig ships is_replicated = ?false, so anything derived from it with record update is correct as-is — nothing to remember, nothing to add.

Do not override it to ?true or null. A replicated outcall repeats the request from every node in the subnet, which for Slack means the message is posted once per replica (~13 duplicates), the Authorization: Bearer xoxb-…/xoxp-… header leaves every node, and consensus fails anyway because Slack's reply carries a per-request ts.

Backend

Add dependencies

The admin gate in the recipe below needs the authorization component alongside the client:

mops add slack-client@0.1.0
mops add caffeineai-authorization@1.0.1

The generated function is ChatApi.chatPostMessage(config, channel, asUser, attachments, blocks, iconEmoji, iconUrl, linkNames, mrkdwn, parse, replyBroadcast, text, threadTs, unfurlLinks, unfurlMedia, username). Pass empty strings / false for the options you don't use. The token is not an argument — it travels only in config.auth (see Auth model above); this holds for every method in the client.

import AccessControl "mo:caffeineai-authorization/access-control";
import MixinAuthorization "mo:caffeineai-authorization/MixinAuthorization";
import MixinSlackConfig "mixins/slack-config";
import MixinSlackMessaging "mixins/slack-messaging";

actor {
  let accessControlState = AccessControl.initState();
  include MixinAuthorization(accessControlState, null);

  // Admin-held Slack token, `xoxb-…` or `xoxp-…` — never returned to the frontend.
  let slackConfig = { var token : Text = "" };
  include MixinSlackConfig(accessControlState, slackConfig);
  include MixinSlackMessaging(slackConfig);
};

import AccessControl "mo:caffeineai-authorization/access-control";
import Runtime "mo:core/Runtime";

mixin (
  accessControlState : AccessControl.AccessControlState,
  slackConfig : { var token : Text },
) {
  public query func isSlackConfigured() : async Bool {
    slackConfig.token.size() > 0;
  };

  // Admin-only; accepts either token flavour. NOTE: `#admin` — never a
  // first-caller-claims-ownership check,
  // which the shared anonymous principal would defeat.
  public shared ({ caller }) func setSlackToken(token : Text) : async () {
    if (not AccessControl.hasPermission(accessControlState, caller, #admin)) {
      Runtime.trap("Unauthorized: Only admins can set the Slack token");
    };
    slackConfig.token := token;
  };
};

import Principal "mo:core/Principal";
import Runtime "mo:core/Runtime";
import { chatPostMessage } "mo:slack-client/Apis/ChatApi";
import { defaultConfig; type Config } "mo:slack-client/Config";

mixin (slackConfig : { var token : Text }) {
  // Token rides `config.auth`; `defaultConfig` is already non-replicated.
  func slackClientConfig(token : Text) : Config {
    {
      defaultConfig with
      auth = ?#bearer(token);
      max_response_bytes = ?(1_000_000 : Nat64);
    };
  };

  // Post `text` to `channel` (channel ID like "C012AB3CD" or "#general").
  // Returns the posted message timestamp (`ts`).
  public shared ({ caller }) func postSlackMessage(channel : Text, text : Text) : async Text {
    if (caller.isAnonymous()) Runtime.trap("Sign in to post to Slack");
    if (slackConfig.token.size() == 0) {
      Runtime.trap("Slack is not configured (an admin must set the token)");
    };
    let res = await* chatPostMessage(
      slackClientConfig(slackConfig.token), // token rides config.auth — never a URL param
      channel,
      "", "", "", "", "", // asUser, attachments, blocks, iconEmoji, iconUrl
      false, // linkNames
      true, // mrkdwn
      "", // parse
      false, // replyBroadcast
      text, // text
      "", // threadTs
      false, // unfurlLinks
      false, // unfurlMedia
      "", // username
    );
    res.ts;
  };
};

Addressing a channel — an ID, and the bot must be in it

chatPostMessage's first argument is a channel ID: C… for a public or private channel, D… for a DM, G… for a legacy group. A #general-style name still resolves for public channels but is deprecated and never works for private ones — prefer the ID and store it, don't hardcode a name.

Where a human finds an ID (put these words in the UI, not just in the completion message):

• Open the channel → click its name in the header → About tab → the ID sits at the bottom of the panel (C012AB3CD), with a copy button.
• Or right-click the channel in the sidebar → Copy link → the ID is the last path segment of …/archives/C012AB3CD.

Where the canister gets one: ConversationsApi.conversationsList and match on name (needs channels:read). Resolve once and cache the ID in a stable variable — do not resolve on every post; it doubles the outcalls and the cycles.

The bot must be a member, or chatPostMessage answers not_in_channel — the single most common first failure. Three ways out, in order of preference:

1. the admin runs /invite @YourApp in the target channel;
2. grant chat:write.public, which lets the bot post to any public channel with no invite at all;
3. call conversationsJoin (public channels only, needs channels:join).

To DM a person: usersLookupByEmail (needs users:read.email) or usersList → user ID → conversationsOpen → post to the D… channel it returns. Needs im:write.

Available API surface

The drop ships 10 API modules / ~140 operations. Only chatPostMessage is runtime-verified (see Known limitations); the rest are generated from the same spec and typecheck, but treat their response decode as unproven.

Binary upload/download is not usable from this client (multipart bodies are outside the generated JSON/form surface). filesRemoteAdd, which registers a file that lives at an external URL, is the supported alternative.

Frontend

Slack needs no OAuth callback: the credential is a long-lived token the admin pastes, so there is no redirect URI, no /connect/slack route, and no per-user handshake. Do not build one. What a Slack build MUST ship is the page that lets the admin get and enter the token — these are acceptance criteria, not suggestions, and a build missing them is broken, not merely incomplete:

• The token page exists and is reachable. A "post to Slack" feature with no page to enter a token is unusable. A signed-in admin must reach /settings/slack from the nav or from the not-configured prompt.
• The token-generation steps are in the UI, not only in the agent's chat reply. The admin returns to this page weeks later, after the chat is gone.
• The exact scope list is displayed. The admin cannot infer which scopes this app needs; show the ones the build actually uses (Scopes to request).

1. A login flow — required. setSlackToken gates on #admin, so the app needs non-anonymous callers. Take login, useInternetIdentity / useActor plumbing, and the admin-role gate from extension-authorization.

2. An admin settings page — /settings/slack (admin-gated). Required:

   • A "How to get your Slack bot token" panel above the input, framed as a one-time ~5-minute setup, with these numbered steps (the agent's completion message must repeat them verbatim):
     1. open https://api.slack.com/apps → Create New App → From scratch → name it and pick the workspace;
     2. OAuth & Permissions → Scopes → Bot Token Scopes → add the scopes listed on this page;
     3. Install to Workspace → Allow → copy the Bot User OAuth Token (it starts with xoxb-);
     4. paste it below and save;
     5. in Slack, run /invite @YourApp in every channel the app posts to. Include a convenience link that opens https://api.slack.com/apps.
   • The scope list rendered as copyable text — exactly the scopes this build uses, not the whole table.
   • One password input bound to setSlackToken(token). Submit on enter, clear on success, and keep it re-pasteable: a scope change forces a reinstall and a new token, so this is not a one-shot form.
   • Status driven by isSlackConfigured() (Bool) — "Configured" / "Not configured". Never render the token back, not even masked-with-suffix.
   • If the app posts to a fixed channel, its ID belongs on this page too: extend the config mixin to { var token : Text; var channel : Text } with an admin-gated setSlackChannel, and put the "where to find a channel ID" hint (see Addressing a channel) inline next to the field. Do not make users type a raw ID with no explanation of where it comes from.
   • Make the page reachable. The shared Layout nav MUST link here when isCallerAdmin is true and hide it otherwise. Add the link where the nav is defined, not inside this page.

3. Empty-state nudges. When isSlackConfigured() is false, never render a dead "Send to Slack" button: admins get a "Set up Slack" link to /settings/slack; non-admins get an explanation — e.g. "Slack isn't set up yet — an administrator needs to add the workspace token in Settings."

4. Translate Slack's errors. Logical failures arrive as rejected calls whose message carries Slack's own error string (see Known limitations). Map at least these three to an action instead of showing the raw reject:

   • not_in_channel → "Invite the app to the channel: /invite @YourApp"
   • invalid_auth / not_authed → "The Slack token is invalid — paste a new one in Settings" (admins get the link)
   • channel_not_found → "Check the channel ID" (with the how-to-find hint)
   • missing_scope → "The app needs another Slack scope — add it, reinstall, and paste the new token"

Suggested route layout:

/                →  Main UI (any signed-in user; empty-state when unconfigured)
/settings/slack  →  Admin token + default channel (admin-only)
# No /connect/slack: Slack uses a pasted long-lived token, not a redirect flow.

What the composer must tell the Caffeine user

The app cannot work until a human creates a Slack app and pastes a token, so the completion message in the composer is part of the deliverable, not a summary of it. It MUST contain, in this order:

1. That a token is required, and who enters it — an admin, on /settings/slack, reachable from the nav once signed in.
2. The five numbered steps verbatim from Frontend item 2: create the app at https://api.slack.com/apps, add the bot scopes, Install to Workspace, copy the xoxb-… token, paste it, then /invite @YourApp in each target channel.
3. The exact scope list this build needs, from Scopes to request — as text the user can copy straight into Slack's scope picker, not a description of it.
4. Where to find a channel ID — the About-tab / copy-link recipe from Addressing a channel — whenever the user has to name a channel.
5. The failure map, one line each: not_in_channel → invite the app; invalid_auth → re-paste the token; channel_not_found → check the ID; missing_scope → add the scope, reinstall, paste the new token.

Do not compress this to "configure Slack in Settings", and do not substitute a link to Slack's documentation. The user is mid-build, has very likely never seen the Slack app dashboard, and the composer is where they are looking. Use the same wording here as in the settings-page panel so the two cannot drift.

Known limitations (experimental)

• {"ok": false} is reported, as a rejected call. Slack signals logical failures (bad token, missing scope, channel not found) as {"ok": false, "error": "…"} over HTTP 200. Since 0.1.0 the client detects that and routes it through the error path, so the reject message names Slack's own error string — e.g. not_authed, invalid_auth, channel_not_found — instead of a decode failure. Handle it as a rejected call: try { … } catch (e) { Error.message(e) }. Do not write if (res.ok) …: a returned value has already decoded, so ok is always true there and the check is dead code. Turning these into #ok/#err return values would change every method's signature, so that waits for a major.
• Reachability (IPv4) — no proxy needed. slack.com is IPv4-only, which used to put it out of reach of IC HTTPS outcalls. Since 2025-08-04 the IC tries a direct (IPv6) connection and automatically retries through an IC-managed SOCKS proxy when that fails, so IPv4-only hosts work: leave config.baseUrl at the default https://slack.com/api. The TLS session is end-to-end between node and Slack, so the proxy sees only ciphertext. Expect some added latency on the fallback path (non-replicated outcalls are also the slower path — see above).
• Auth is header-based, for every method. The token rides the Authorization: Bearer header only — never in the URL, so it cannot land in a logged query string, and it is never a method argument.
• Partial runtime verification. The request shape is proven against live Slack (a real post lands); the success-response decode is not yet runtime-confirmed — diagnostics is on, so any decode failure surfaces the raw Slack body. The schema is also generated from an archived (~2020) spec revision.

Related

• mops add slack-client@0.1.0 — the generated Slack Web API bindings.
• Slack Web API reference — every method, its scopes, and its error strings.
• chat.postMessage — the one runtime-verified path; its channel/text/blocks semantics.
• Slack token types — bot (xoxb-) vs user (xoxp-), and what a reinstall does to them.
• Slack app management — where the admin creates the app, sets scopes, and installs to the workspace.
• extension-authorization — required prerequisite. Internet Identity login, useInternetIdentity / useActor plumbing, and the #admin role gate the token setter needs.