website/server/src/model/teams/teamSync.model.js

// ── The reconciler ─────────────────────────────────────────────────────────
//
// Core's projection of the module's Teams, kept in step (TEAMS.md §2.4). This is
// the only thing that writes `team_members`, and one of only two things that
// write `teams.status`.
//
// **The four places it refuses to act** are the point of the file, and they are
// all one rule stated four ways: *a result derived from an answer core does not
// trust is never applied.* Anything less specific tends to collapse, under
// maintenance, into "a failed call means no teams" — which is invariant 1's
// failure mode and would empty every roster on the site the first time a sidecar
// restarted.
//
//   1. `getTeams()` not ok            → write sync state, touch NOTHING, return.
//   2. ok but empty, core holds ≥1    → quarantine; apply only if the NEXT
//                                       authoritative answer agrees.
//   3. `getTeamMembers()` not ok      → that Team's roster untouched and stale;
//                                       the other Teams carry on.
//   4. ok but zero members, had some  → the same two-strikes quarantine, per Team.
//
// Gates 2 and 4 exist because an authoritative-looking empty answer during a cold
// start is the one failure indistinguishable from a real wipe. "Every Team on the
// shard disbanded at once" costs one interval of delay to confirm; getting it
// wrong costs every roster on the site.
//
// Events (§2.3) are an OPTIMISATION, never the source of truth. They make the
// common case immediate; reconciliation is what makes it correct. Nothing
// destructive at Team level is ever driven by one — §2.2 scopes archival to an
// authoritative full list, so a `team.disbanded` event schedules a run rather
// than archiving, and a spurious event costs a reconcile instead of a Team.

const teamsDb = require('./teams.db')
const teamProvider = require('./teamProvider')
const moderation = require('./teamModeration.model')
const { slugify, uniqueSlug } = require('./teamSlug')
const settings = require('../settings/settings.model')
const log = require('../../utils/logger')('teams')

// At most one run per 30s (§2.4), so a sidecar flapping cannot become a
// reconciliation storm — every flap publishes events, and every event asks for a
// run.
const DEBOUNCE_MS = 30_000
const DEFAULT_INTERVAL_S = 900
const MIN_INTERVAL_S = 60
const INTERVAL_KEY = 'teams_reconcile_interval_s'

// The six kinds a module may publish (§2.3). Six rather than the four a
// membership-shaped reading suggests, because leadership is its own authority
// path and a leadership change must be expressible without pretending someone
// joined or left.
const EVENT_KINDS = new Set([
  'team.created', 'team.disbanded',
  'team.member.added', 'team.member.removed',
  'team.leader.added', 'team.leader.removed',
])

// Kinds that can only be answered by a full list. `team.created` cannot be
// applied from a delta — a Team built from one has no name, no roster and no
// leaders — and `team.disbanded` must not be, per §2.2.
const RECONCILE_ONLY = new Set(['team.created', 'team.disbanded'])

// ── Scheduling state (in-process; one provider per deployment) ─────────────

let running = false
let rerunReason = null
let lastRunAt = 0
let debounceTimer = null
let pollTimer = null
let started = false

/** Resolve the poll interval, floored so a bad setting cannot become a hot loop. */
async function intervalSeconds() {
  let raw
  try {
    raw = await settings.get(INTERVAL_KEY)
  } catch {
    return DEFAULT_INTERVAL_S
  }
  const n = Number.parseInt(raw, 10)
  if (!Number.isFinite(n) || n < MIN_INTERVAL_S) return DEFAULT_INTERVAL_S
  return n
}

/**
 * Backoff, capped at the poll interval (§2.4).
 *
 * The cap is what keeps this a backoff rather than an outage: a module down for a
 * day would otherwise reach a delay measured in weeks and stay stale long after
 * it recovered.
 */
function backoffSeconds(consecutiveFailures, intervalS) {
  if (!consecutiveFailures) return intervalS
  return Math.min(intervalS, 2 ** Math.min(consecutiveFailures, 16) * 15)
}

// ── Applying one Team ──────────────────────────────────────────────────────

/**
 * Create the row for a Team core has not seen, deriving its slug and screening
 * its name against the reserved list (§2.8).
 *
 * The row is created whatever the screening says, and hidden if it matched. Core
 * cannot refuse a name: the guild already exists in the game and core is a mirror
 * of it, not an authority over it. A hidden Team is absent from public surfaces
 * and completely functional for its own members — the people in it are not being
 * punished for a name their leader chose.
 */
async function createTeam(moduleId, team) {
  const taken = await teamsDb.slugsLike(slugify(team.name) || 'team')
  const slug = uniqueSlug(team.name, taken)
  const screened = await moderation.screenForCreate(team.name)
  const id = await teamsDb.insertTeam({ moduleId, slug, ...team, ...screened })
  log.info('team created', {
    moduleId, externalId: team.externalId, name: team.name, slug, hidden: Boolean(screened.hidden),
  })
  return id
}

/**
 * The §2.2 rename rule: same id and a different name is an archive plus a create.
 *
 * The old row keeps its forum, its activity and its grants, all read-only, and
 * points at its successor so the old slug can explain itself instead of 404ing.
 * Core never decides whether this is "really" the same team — that judgement is
 * the module's, expressed in whether it reuses the external id (§10.5).
 */
async function applyRename(moduleId, existing, team) {
  const successorId = await createTeam(moduleId, team)
  await teamsDb.archiveTeam(existing.id, 'renamed', successorId)
  log.info('team renamed; previous row archived', {
    externalId: team.externalId, from: existing.name, to: team.name, archivedId: existing.id, successorId,
  })
  return successorId
}

/**
 * Sync one Team's roster and leadership. Gates 3 and 4 live here.
 *
 * Returns whether the roster was applied, so the caller can tell "synced" from
 * "left alone", which is the difference between fresh and stale on that Team's
 * page.
 */
async function syncRoster(team) {
  const answer = await teamProvider.getTeamMembers(team.external_id)

  // Gate 3. One Team's unanswerable roster is not the other Teams' problem, and
  // it is certainly not an empty roster.
  if (!answer.ok) {
    log.warn('roster left untouched; provider could not answer', {
      externalId: team.external_id, reason: answer.reason,
    })
    return false
  }

  const known = await teamsDb.memberKeys(team.id)

  // Gate 4, the per-Team twin of gate 2.
  if (answer.complete && answer.members.length === 0 && known.length > 0) {
    if (!team.members_empty_since) {
      await teamsDb.setMembersEmptySince(team.id, new Date())
      log.warn('empty roster quarantined; awaiting a second answer', {
        externalId: team.external_id, had: known.length,
      })
      return false
    }
    log.warn('empty roster confirmed by a second answer; departing every member', {
      externalId: team.external_id, had: known.length,
    })
  } else if (team.members_empty_since) {
    // Any non-empty answer clears the quarantine.
    await teamsDb.setMembersEmptySince(team.id, null)
  }

  for (const member of answer.members) {
    // eslint-disable-next-line no-await-in-loop
    await teamsDb.upsertMember({
      teamId: team.id,
      memberKey: member.memberKey,
      displayName: member.displayName,
      userId: member.userId,
      isLeader: member.leader,
      rankLabel: member.rankLabel,
      online: member.online,
    })
  }

  // Removals only from a COMPLETE answer. `complete: false` means "valid but
  // partial", so additions and updates apply and nothing is taken away.
  if (answer.complete) {
    const seen = new Set(answer.members.map((m) => m.memberKey))
    await teamsDb.markDeparted(team.id, known.filter((key) => !seen.has(key)))
  }

  // Leadership is a separate question with a separate answer, and a provider that
  // cannot answer it leaves the synced value alone rather than demoting everyone.
  const leaders = await teamProvider.getTeamLeaders(team.external_id)
  if (leaders.ok) {
    await teamsDb.setLeaders(team.id, leaders.leaders)
  } else {
    log.warn('leadership left untouched; provider could not answer', {
      externalId: team.external_id, reason: leaders.reason,
    })
  }

  await teamsDb.recount(team.id)
  await teamsDb.markRosterSynced(team.id)
  return true
}

// ── The run ────────────────────────────────────────────────────────────────

/**
 * One full reconciliation. Callers use `request()`; this is the body it guards.
 *
 * Never throws: a reconcile is a background job, and a rejection here would
 * surface as an unhandled rejection in the poll timer rather than as anything an
 * operator could act on. The failure is recorded where it can be read — in
 * `team_sync_state`, which Admin → Teams shows verbatim.
 */
async function runOnce(reason) {
  const moduleId = teamProvider.providerModuleId()
  if (!moduleId) return { ok: false, reason: 'no team provider is registered' }

  await teamsDb.recordAttempt(moduleId)
  const answer = await teamProvider.getTeams()

  // Gate 1.
  if (!answer.ok) {
    await teamsDb.recordFailure(moduleId, answer.reason)
    log.warn('reconcile refused; provider could not answer', { reason: answer.reason, trigger: reason })
    return { ok: false, reason: answer.reason }
  }

  const existing = await teamsDb.activeByModule(moduleId)

  // Gate 2. Only a COMPLETE answer can mean "there are no teams" — a partial one
  // removes nothing by definition.
  if (answer.complete && answer.teams.length === 0 && existing.length > 0) {
    const state = await teamsDb.syncState(moduleId)
    if (!state || !state.pending_empty_since) {
      await teamsDb.setPendingEmpty(moduleId, new Date())
      await teamsDb.recordSuccess(moduleId)
      log.warn('empty team list quarantined; awaiting a second answer', { held: existing.length })
      return { ok: true, quarantined: true, applied: 0 }
    }
    const intervalS = await intervalSeconds()
    const waited = (Date.now() - new Date(state.pending_empty_since).getTime()) / 1000
    if (waited < intervalS) {
      await teamsDb.recordSuccess(moduleId)
      log.warn('empty team list still quarantined', { waitedSeconds: Math.round(waited), intervalS })
      return { ok: true, quarantined: true, applied: 0 }
    }
    log.warn('empty team list confirmed; archiving every active team', { count: existing.length })
  } else if (answer.teams.length) {
    // Any non-empty answer clears the quarantine.
    await teamsDb.setPendingEmpty(moduleId, null)
  }

  const byExternalId = new Map(existing.map((t) => [t.external_id, t]))
  const seen = new Set()
  let created = 0
  let renamed = 0
  let rosters = 0

  for (const team of answer.teams) {
    seen.add(team.externalId)
    const current = byExternalId.get(team.externalId)
    let id
    if (!current) {
      id = await createTeam(moduleId, team)
      created += 1
    } else if (current.name !== team.name) {
      id = await applyRename(moduleId, current, team)
      renamed += 1
    } else {
      id = current.id
      await teamsDb.updateTeam(id, { abbr: team.abbr, meta: team.meta })
    }

    // Re-read rather than reusing `current`: a create or a rename has just made a
    // row this loop has never seen, and syncRoster reads the quarantine stamp off
    // it. Passing a stale object would drop the second strike of gate 4.
    const row = await teamsDb.findById(id)
    if (row && await syncRoster(row)) rosters += 1
  }

  // Archive what the module no longer lists — the §2.2 disband path, and the only
  // one. Guarded by `complete` for the same reason removals are.
  let archived = 0
  if (answer.complete) {
    for (const team of existing) {
      if (seen.has(team.external_id)) continue
      await teamsDb.archiveTeam(team.id, 'disbanded')
      archived += 1
      log.info('team archived; absent from an authoritative list', {
        externalId: team.external_id, name: team.name,
      })
    }
  }

  // Re-screen the names no human has ruled on. Names are immutable per row, so
  // this only changes an outcome when the reserved TERMS changed — an operator
  // adding one, or the deployment being renamed — which is exactly the case a
  // create-time-only check would miss forever.
  const rehidden = await moderation.rescreen(moduleId)

  await teamsDb.recordSuccess(moduleId)
  log.info('reconcile complete', {
    trigger: reason, created, renamed, archived, rosters, rehidden, total: answer.teams.length,
  })
  return { ok: true, created, renamed, archived, rosters, rehidden }
}

// ── The public entry points ────────────────────────────────────────────────

/**
 * Run now, awaited, with the per-module lock held. Admin → Resync uses this,
 * because an operator pressing a button is owed the outcome rather than a
 * promise that something will happen soon.
 *
 * A run already in progress is JOINED rather than queued: the caller wants "the
 * projection is now current", and a run that started a moment ago delivers that.
 */
async function reconcileNow(reason = 'manual') {
  if (running) {
    rerunReason = reason
    return { ok: true, joined: true }
  }
  running = true
  try {
    const result = await runOnce(reason)
    lastRunAt = Date.now()
    return result
  } catch (err) {
    log.error('reconcile threw', { message: err.message, trigger: reason })
    return { ok: false, reason: err.message }
  } finally {
    running = false
    const queued = rerunReason
    rerunReason = null
    // Something asked while this run was in flight, so it saw state this run may
    // have been too early to include. Ask again, through the debounce.
    if (queued) request({ reason: queued })
  }
}

/**
 * Ask for a reconciliation. Returns immediately and never rejects — this is what
 * `ctx.teams.reconcile()` is (§2.3), and a module must not be able to make its
 * own call site slow or its own errors someone else's.
 */
function request({ reason = 'module' } = {}) {
  if (debounceTimer) return
  const since = Date.now() - lastRunAt
  if (running) {
    rerunReason = reason
    return
  }
  if (since >= DEBOUNCE_MS) {
    reconcileNow(reason).catch(() => {})
    return
  }
  debounceTimer = setTimeout(() => {
    debounceTimer = null
    reconcileNow(reason).catch(() => {})
  }, DEBOUNCE_MS - since)
  // Unreffed for the same reason the provider's deadline is: a pending debounce
  // must not hold a shutdown open waiting to do background work.
  if (typeof debounceTimer.unref === 'function') debounceTimer.unref()
}

/**
 * Apply a module-published event (§2.3).
 *
 * Deltas are applied only for a Team core already knows, and only for the four
 * kinds a delta can express. Everything else — an unknown Team, a create, a
 * disband — asks for a reconciliation instead, because a Team invented from a
 * delta has no name, no roster and no leaders, and an archive driven by one is
 * destruction on the strength of a message that may simply have been repeated.
 */
async function publish(event) {
  const { kind, externalId } = event || {}
  if (!EVENT_KINDS.has(kind)) throw new Error(`teams.publish: unknown event kind "${kind}"`)
  const id = typeof externalId === 'string' ? externalId.trim() : ''
  if (!id) throw new Error(`teams.publish: ${kind} has no externalId`)

  const moduleId = teamProvider.providerModuleId()
  if (!moduleId) return

  if (RECONCILE_ONLY.has(kind)) {
    request({ reason: kind })
    return
  }

  const team = await teamsDb.findActive(moduleId, id)
  if (!team) {
    request({ reason: `${kind} for an unknown team` })
    return
  }

  const memberKey = typeof event.memberKey === 'string' ? event.memberKey.trim() : ''
  if (!memberKey) throw new Error(`teams.publish: ${kind} has no memberKey`)

  switch (kind) {
    case 'team.member.added':
      await teamsDb.upsertMember({
        teamId: team.id,
        memberKey,
        displayName: typeof event.displayName === 'string' ? event.displayName.trim() : null,
        userId: Number.isInteger(event.userId) && event.userId > 0 ? event.userId : null,
        isLeader: Boolean(event.leader),
        rankLabel: typeof event.rankLabel === 'string' ? event.rankLabel.trim() : null,
        online: Boolean(event.online),
      })
      break
    case 'team.member.removed':
      await teamsDb.markDeparted(team.id, [memberKey])
      break
    case 'team.leader.added':
    case 'team.leader.removed':
      // A no-op when the member is unknown: the row is created by the roster, not
      // by a leadership delta, and inventing one here would put a member on the
      // roster whose only evidence is that someone promoted them.
      await teamsDb.setMemberLeader(team.id, memberKey, kind === 'team.leader.added')
      break
    default:
      break
  }

  await teamsDb.recount(team.id)
  // A delta is a hint that something changed, not a claim to have applied all of
  // it, so every one still asks for the run that makes it correct.
  request({ reason: kind })
}

// ── The poll ───────────────────────────────────────────────────────────────

async function scheduleNextPoll() {
  const intervalS = await intervalSeconds()
  const moduleId = teamProvider.providerModuleId()
  let delayS = intervalS
  if (moduleId) {
    const state = await teamsDb.syncState(moduleId).catch(() => null)
    if (state) delayS = backoffSeconds(state.consecutive_failures, intervalS)
  }
  pollTimer = setTimeout(() => {
    reconcileNow('poll').catch(() => {}).then(() => { if (started) scheduleNextPoll().catch(() => {}) })
  }, delayS * 1000)
  if (typeof pollTimer.unref === 'function') pollTimer.unref()
}

/**
 * Start the boot reconcile and the poll. Called from the module lifecycle, after
 * every module has started — the website may have been down across a whole guild
 * war, so the first thing it does on the way up is ask.
 */
async function start() {
  if (started) return
  started = true
  if (!teamProvider.providerModuleId()) {
    log.info('no team provider registered; the reconciler stays idle')
    return
  }
  await reconcileNow('boot')
  await scheduleNextPoll()
}

function stop() {
  started = false
  if (pollTimer) clearTimeout(pollTimer)
  if (debounceTimer) clearTimeout(debounceTimer)
  pollTimer = null
  debounceTimer = null
}

// Test-only: the scheduler is module-level state, so a test that triggers a run
// has to be able to put it back.
function _reset() {
  stop()
  running = false
  rerunReason = null
  lastRunAt = 0
}

module.exports = {
  reconcileNow,
  request,
  publish,
  start,
  stop,
  intervalSeconds,
  backoffSeconds,
  EVENT_KINDS,
  DEBOUNCE_MS,
  DEFAULT_INTERVAL_S,
  _reset,
}