feat(teams): the slash-command seam, and the first command through it

Phase 7 of TEAMS.md. `api.registerSlashCommands` stops throwing: a module
registers a command's DEFINITION and its HANDLER together, the bot pulls the
definitions over the internal listener and runs none of our code, and the
handler executes here — forced by the bot container having no `modules` volume,
and the right boundary anyway.

Registration validates what Discord would reject as a batch (names, description
lengths, the four option types, required-before-optional), because the bot
registers the whole set in one PUT and a single bad entry costs every command
including the bot's own. Commands are not namespaced under their owner — there
is no dot in Discord's name grammar — so collisions are first-come with the
holder named.

The dispatcher is the access boundary: `linked` has no Discord equivalent, so
the platform-side permission default can only ever be advertising. It resolves
the actor by `auth_providers.kind` rather than the id slug, treats a banned
account as unlinked, bounds a handler under the bot's own timeout, and keeps
`ok` outside the envelope so a handler cannot forge it.

Liveness is asked at both the pull and the dispatch. The registries have no
removal path, so a module an operator disables at runtime would otherwise keep
a live handler behind a command Discord still advertises.

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
2026-08-18 18:53:34 -05:00
parent b1d3b87cd6
commit cecd72915f
21 changed files with 1459 additions and 27 deletions

View File

@@ -5,6 +5,7 @@ const { Client, GatewayIntentBits, REST, Routes } = require('discord.js')
const createLogger = require('../utils/logger')
const commands = require('./commands')
const dynamicCommands = require('./dynamicCommands')
const messageFilter = require('./messageFilter')
const scheduler = require('../scheduler/scheduler')
const roleMenuHandler = require('./roleMenuHandler')
@@ -22,12 +23,46 @@ let status = 'disconnected' // disconnected | connecting | connected | error
let statusDetail = null
let lastConnectedAt = null
// One whole-set PUT of the bot's own commands plus whatever the app has
// registered (TEAMS.md §7.1). Because it replaces the set rather than adding to
// it, DEREGISTRATION is free: a module that is gone is simply absent from the
// next pull, and nobody has to remember to take its command back.
async function registerCommands(applicationId, targetGuildId) {
const dynamic = dynamicCommands.definitions()
const rest = new REST({ version: '10' }).setToken(client.token)
await rest.put(Routes.applicationGuildCommands(applicationId, targetGuildId), {
body: commands.all.map((c) => c.data),
body: [...commands.all.map((c) => c.data), ...dynamic],
})
log.info('registered guild slash commands', { guildId: targetGuildId, count: commands.all.length })
log.info('registered guild slash commands', {
guildId: targetGuildId,
builtIn: commands.all.length,
fromApp: dynamic.length,
})
}
/**
* Re-pull the app's commands and re-register the set if it moved.
*
* Called on `ready` and again whenever the app nudges
* (`POST /internal/refresh-commands`). A no-op when nothing changed, so a nudge
* per module state change costs one cheap GET rather than a REST.put per
* install — and a disconnected bot does nothing at all, since there is no
* application to register against until it logs in.
*/
async function refreshCommands() {
const result = await dynamicCommands.pull()
if (!result.ok || !result.changed) return result
if (!client || !client.isReady()) return result
try {
await registerCommands(client.application.id, guildId)
} catch (err) {
// The PUT is all-or-nothing: a definition Discord rejects costs every
// command, the built-ins included. Loud, and never fatal to the process.
log.error('re-registering slash commands failed — the previous set is still live', {
message: err.message,
})
}
return result
}
async function stop() {
@@ -54,6 +89,11 @@ async function stop() {
// failure here leaves the client connected but flags an error status.
async function onReady() {
try {
// Pull BEFORE the single PUT, so the app's commands are in the very first
// registration rather than appearing a beat later. The pull never throws —
// an unreachable app costs the module commands and nothing else, and the
// bot's own set registers exactly as it always did.
await dynamicCommands.pull()
await registerCommands(client.application.id, guildId)
await scheduler.start(client)
tempRoleSweeper.start(client)
@@ -70,14 +110,19 @@ async function onReady() {
}
}
// Route an interaction: role-menu handler first, then chat-input slash commands.
// Route an interaction: role-menu handler first, then chat-input slash commands
// — the bot's own, then the app's. Built-ins are consulted FIRST and the pull
// already drops any module name that collides with one, so the two orderings
// agree; checking here as well means a name that somehow reached Discord twice
// still runs the bot's version rather than whichever registry answered first.
async function onInteractionCreate(interaction) {
if (await roleMenuHandler.handleInteraction(interaction)) return
if (!interaction.isChatInputCommand()) return
const command = commands.get(interaction.commandName)
if (!command) return
if (!command && !dynamicCommands.has(interaction.commandName)) return
try {
await command.execute(interaction)
if (command) await command.execute(interaction)
else await dynamicCommands.execute(interaction)
} catch (err) {
log.error('command execution failed', { command: interaction.commandName, message: err.message })
const payload = { content: 'Something went wrong running that command.', ephemeral: true }
@@ -146,4 +191,4 @@ function getConnection() {
return { client, guildId }
}
module.exports = { start, stop, getStatus, getConnection }
module.exports = { start, stop, getStatus, getConnection, refreshCommands }

View File

@@ -0,0 +1,210 @@
// Slash commands whose DEFINITION and HANDLER live in the website process
// (TEAMS.md §7.1). The bot pulls the definitions, registers them alongside its
// own, and executes one by deferring, asking the app, and editing the reply in.
//
// Everything Discord-specific is here and nothing else is: the app's dispatcher
// resolves the actor, enforces access and produces a platform-neutral envelope,
// and this file turns that envelope into an interaction reply. A module never
// touches an interaction, which is what makes the registration API something a
// second platform could implement.
const { PermissionFlagsBits } = require('discord.js')
const appInternal = require('../site/appInternalClient')
const staticCommands = require('./commands')
const createLogger = require('../utils/logger')
const log = createLogger('dynamic-commands')
// §7.1.1's four types, and the only four. The app rejects anything else at
// registration; this map is the second half of that agreement.
const OPTION_TYPE = { string: 3, integer: 4, boolean: 5, user: 6 }
// The pulled set, and the app's module-state counter it came from. `null`
// version means "never successfully pulled", which is distinct from 0 ("pulled
// while the app had no modules loaded") — the first should retry, the second is
// a true answer.
let pulled = []
let version = null
/**
* Ask the app for the current definitions.
*
* **A failed pull KEEPS the previous set.** The app being briefly unreachable is
* not the same as it having no commands, and treating it as such would
* deregister every module command from Discord on a restart blip — then
* re-register them a minute later, with members watching commands appear and
* disappear. Nothing changes until the app actually answers.
*
* @returns {Promise<{ok: boolean, changed: boolean, count: number}>}
*/
async function pull() {
const res = await appInternal.fetchCommands()
if (!res.ok) {
log.warn('command pull failed — keeping the set already registered', {
error: res.error,
holding: pulled.length,
})
return { ok: false, changed: false, count: pulled.length }
}
const { version: pulledVersion, commands } = res.data || {}
const next = Array.isArray(commands) ? commands.filter(usable) : []
const changed = version === null || pulledVersion !== version || next.length !== pulled.length
pulled = next
version = typeof pulledVersion === 'number' ? pulledVersion : 0
return { ok: true, changed, count: pulled.length }
}
/**
* Drop a pulled definition the bot cannot honour.
*
* **The name collision the app cannot see.** The app validates a command against
* everything IT has registered; it does not know the bot's own static array
* exists. A module registering `ping` would produce two `ping` entries in one
* `REST.put`, which Discord rejects as a batch — taking down every command
* including the bot's own. The bot's built-ins win, because they are the ones a
* module cannot be asked to change.
*/
function usable(definition) {
if (!definition || typeof definition.name !== 'string') return false
if (staticCommands.get(definition.name)) {
log.warn('module slash command collides with a built-in and is ignored', {
command: definition.name,
owner: definition.owner,
})
return false
}
return true
}
/**
* The pulled definitions as Discord command data, for the whole-set PUT.
*
* `access: 'staff'` becomes a Discord-side permission default; `linked` cannot
* be expressed in Discord's permission model at all — there is no "has a website
* account" predicate — so it is simply not advertised and the app's dispatcher
* refuses it. That asymmetry is the reason §7.1 says access is enforced twice
* and that only the server half is the gate.
*/
function definitions() {
return pulled.map((c) => {
const data = {
name: c.name,
description: c.description,
options: (c.options || []).map((o) => ({
name: o.name,
description: o.description,
type: OPTION_TYPE[o.type],
required: Boolean(o.required),
...(o.choices ? { choices: o.choices } : {}),
})),
}
if (c.access === 'staff') data.default_member_permissions = PermissionFlagsBits.ModerateMembers.toString()
return data
})
}
/** Is this a command the app owns? Asked before the static registry is consulted. */
const has = (name) => pulled.some((c) => c.name === name)
// Read the options the member actually supplied, by the names the definition
// declared. A `user` option is passed on as the Discord user id and nothing else
// — a handler receives platform ids, never a platform object.
function collectOptions(interaction, definition) {
const out = {}
for (const option of definition.options || []) {
const supplied = interaction.options.get(option.name)
if (supplied === null || supplied === undefined) continue
out[option.name] = option.type === 'user' ? String(supplied.value) : supplied.value
}
return out
}
// What the caller sees when the app declined. The COPY lives here rather than in
// the app on purpose: the app answers with a machine reason, and how a refusal is
// phrased to a member is the platform's own voice.
function refusal({ reason, access }) {
if (reason === 'forbidden' && access === 'linked') {
return 'Link your Discord account on the site to use this command.'
}
if (reason === 'forbidden') return 'You do not have access to that command.'
if (reason === 'unknown') return 'That command is no longer available.'
return 'Something went wrong running that command.'
}
// Envelope → interaction payload. A response with fields or a title is an embed;
// a bare `text` is plain content, which reads better for a one-line answer.
function render(envelope) {
const { text, title, fields, url } = envelope
if (!title && !fields) return { content: text || '' }
const embed = {}
if (title) embed.title = title
if (text) embed.description = text
if (url) embed.url = url
if (fields) embed.fields = fields
return { embeds: [embed] }
}
/**
* Defer, dispatch, edit.
*
* **The deferral comes first, always.** Discord gives three seconds to acknowledge
* an interaction; the app is given four to answer. Deferring before the dispatch
* is what keeps the website out of that critical path entirely — a wedged handler
* costs its own reply and never an "application did not respond".
*
* A failure at any point after the defer is an edit, not a reply: the interaction
* has already been acknowledged, and `reply()` on a deferred interaction throws.
*/
async function execute(interaction) {
const definition = pulled.find((c) => c.name === interaction.commandName)
if (!definition) return false
// Ephemerality has to be decided BEFORE the answer exists, because it is a
// property of the deferral. A command declared for linked users only is
// answered privately by default — its answer is about the caller's own
// account — and everything else defers publicly; a handler that wants the
// opposite says so, and the follow-up carries it.
const ephemeral = definition.access === 'linked'
await interaction.deferReply({ ephemeral })
const res = await appInternal.dispatchCommand({
command: definition.name,
options: collectOptions(interaction, definition),
platformUserId: interaction.user.id,
guildId: interaction.guildId,
})
// A transport failure and a handler failure are the same sentence to the
// member and different lines in the log: one is the app being unreachable,
// the other is a module's code.
if (!res.ok) {
log.warn('command dispatch failed', { command: definition.name, error: res.error })
await interaction.editReply({ content: refusal({ reason: 'error' }) })
return true
}
if (!res.data || !res.data.ok) {
await interaction.editReply({ content: refusal(res.data || {}) })
return true
}
const envelope = res.data.response || {}
await interaction.editReply(render(envelope))
// The private aside beside a public answer (§9 answer 5). Skipped when the
// reply was already private — the member would just be told the same thing
// twice, in the same place.
if (envelope.notice && !ephemeral && !envelope.ephemeral) {
await interaction.followUp({ content: envelope.notice, ephemeral: true })
}
return true
}
// Test-only: the pulled set is process-global, so a test that pulls has to be
// able to hand the process back.
function _reset() {
pulled = []
version = null
}
module.exports = { pull, definitions, has, execute, _reset }

View File

@@ -99,4 +99,26 @@ async function reverseModAction(req, res) {
}
}
module.exports = { setConfig, getStatus: getStatusHandler, announce, reverseModAction }
// POST /internal/refresh-commands — the app's nudge that its registered
// slash-command set has moved (TEAMS.md §7.1). No body: the bot re-pulls
// `/internal/commands` and re-registers only if the set actually changed, so the
// nudge stays a cheap thing the app can send on every module state change.
//
// Deliberately its OWN endpoint rather than riding on /internal/config, which
// carries the decrypted bot token: saying "commands changed" should not require
// the app to read a secret out of the database.
//
// Answers 200 even when disconnected — there is no application to register
// against until the bot logs in, and `ready` pulls again anyway. A 5xx here
// would make an ordinary module install look like a failure in the admin panel.
async function refreshCommands(req, res) {
try {
const result = await discordManager.refreshCommands()
return res.json({ ok: true, ...result })
} catch (err) {
log.error('refresh-commands failed', { message: err.message })
return res.json({ ok: false, error: err.message })
}
}
module.exports = { setConfig, getStatus: getStatusHandler, announce, reverseModAction, refreshCommands }

View File

@@ -11,5 +11,6 @@ router.post('/config', ctrl.setConfig)
router.get('/status', ctrl.getStatus)
router.post('/announce', ctrl.announce)
router.post('/mod-reverse', ctrl.reverseModAction)
router.post('/refresh-commands', ctrl.refreshCommands)
module.exports = router

View File

@@ -0,0 +1,76 @@
// Shared-secret client for the APP's internal listener (port 3001) — the
// bot→app direction of the channel `botInternalClient.js` runs app→bot.
//
// Two callers, both slash-command plumbing (TEAMS.md §7.1): pull the registered
// command definitions, and dispatch one that a member has just run. Distinct
// from siteApiClient.js, which reads the site's PUBLIC API with no secret at all.
//
// **The base URL is derived from `SITE_INTERNAL_URL`'s origin, not configured
// separately.** That variable already points at the app's internal listener —
// `http://app:3001/internal/bot-config` — and adding a second variable naming the
// same host would be one more thing an operator can get half-right. Deriving it
// means every existing deployment gains these endpoints with no compose change.
const createLogger = require('../utils/logger')
const log = createLogger('app-internal')
const KEY = process.env.BOT_INTERNAL_KEY || ''
// §7.1's budget, and the same 4s `botInternalClient` uses in the other
// direction. The app bounds its own handlers UNDER this (3s), so a timeout here
// normally means the app itself is unreachable rather than a module being slow.
const TIMEOUT_MS = 4000
function baseUrl() {
const configured = process.env.SITE_INTERNAL_URL
if (!configured) return null
try {
return new URL(configured).origin
} catch {
log.error('SITE_INTERNAL_URL is not a URL — slash-command registration is off', { configured })
return null
}
}
async function call(path, { method = 'GET', body } = {}) {
const base = baseUrl()
if (!base || !KEY) return { ok: false, error: 'SITE_INTERNAL_URL or BOT_INTERNAL_KEY not set' }
const controller = new AbortController()
const timeout = setTimeout(() => controller.abort(), TIMEOUT_MS)
try {
const res = await fetch(`${base}${path}`, {
method,
headers: { 'Content-Type': 'application/json', 'X-Internal-Key': KEY },
body: body ? JSON.stringify(body) : undefined,
signal: controller.signal,
})
if (!res.ok) return { ok: false, status: res.status, error: `app responded ${res.status}` }
return { ok: true, status: res.status, data: await res.json() }
} catch (err) {
log.warn('app internal call failed', { path, message: err.message })
return { ok: false, status: 0, error: err.message }
} finally {
clearTimeout(timeout)
}
}
/** The registered slash-command definitions, plus the version they belong to. */
function fetchCommands() {
return call('/internal/commands')
}
/**
* Run one command in the app and get the response envelope back.
*
* The bot has already deferred by the time this is called, so the only deadline
* that matters is Discord's 15-minute follow-up window — TIMEOUT_MS is about not
* holding an interaction open on a wedged app, not about the 3-second ack.
*/
function dispatchCommand({ command, options, platformUserId, guildId }) {
return call('/internal/commands/dispatch', {
method: 'POST',
body: { command, options, platform: 'discord', platformUserId, guildId },
})
}
module.exports = { fetchCommands, dispatchCommand }