Skip to main content

Audit Trail & Call Logs (Admin)

Audience: Support escalations · Engineering · Ops · Where in app: Admin → Logs & Events → Audit Trail (/admin/audit-logs) and Call Logs (/admin/call-logs) · Access: Super-admin only — both controllers return 403 unless request.user.user_type === 'admin'.

Two read-only forensic surfaces. The Audit Trail answers "who changed this setting/record, when, from what to what". Call Logs answers "what happened, step by step, on this call" — across server, sockets, the widget and the dashboard.

Audit Trail — /admin/audit-logs​

How it works​

backend/app/lib/auditLog.js is a Mongoose schema plugin (auditPlugin) attached to 39 models. On every create/update/delete it reads the document before and after, diffs the leaves, and writes one row to audit_logs (AuditLogModel). Shipped 2026-10-06 (378faa826/e4f72967c, frontend 6be764ec).

  • Actor. app.middlewares.passport wraps the request in runWithActor (an AsyncLocalStorage context), so every write the request causes is stamped with actor.{user_id, email, user_type} and actor.source = "<METHOD> <path>". Writes outside a passport request (cron jobs, webhooks, api-key /migrations/* routes) log as source: 'system'.
  • Best-effort. Audit writes never fail or delay the audited write.
  • Diff shape. changes is { 'path.to.field': { from, to } }; arrays compare whole; values over 2,000 chars are truncated. select:false fields are read too, so a secret rotation shows up (as [redacted] → [redacted]).
  • Redaction. Any path matching pass(word)|token|secret|api_key|private|hash|otp|credential|invite_key|*_enc|authorization|signing|webhook_url|webhook_header is [redacted], including keys nested inside arrays (e.g. an agent's webhook_secret). Models can add extra redact paths.
  • Ignored paths. Per-model ignore lists machine-written fields (counters, heartbeats, online status, token refreshes, billing cursors, knox.cumulative_*, recap last_sent_at…). A write that touches only ignored paths skips auditing and its before/after reads entirely — this is what keeps the plugin cheap on hot paths.
  • Bulk actions. updateMany, deleteMany, multi-doc insertMany and bulkWrite write one entry with bulk: true, record_ids (max 1,000; truncated: true beyond), record_count, and the redacted filter/update in operation — not a per-document diff.
  • Identity records. people / person_profiles (resolver-owned): deletes always (ids only, no snapshot), edits only when the actor is a super-admin repairing by hand, not via POST /admin/identity/maintenance/* sweeps (adminEditsAndDeletes). Resolver traffic is never logged.
  • Settings. settings skips its two job-cursor keys. Upsert-creates are logged as create.
  • Tenant stamp. company_id / agency_id are filled from the record so a company filter finds its rows.
  • Retention. TTL index on created_at — 180 days.

Audited collections: companies, agencies, users, teams, branches, subscriptions, agency_subscription, user_invoices, agency_client_billings, packages, agency_packages, addons, agency_addons, agency_invites, agency_domains, webhooks, platform_webhooks, automation_rules, bot_flows, chatbots, page_rules, call_auto_trigger, company_faqs, company_websites, ai_conversations, blocked, do_not_contacts, knox_visitor_directives, identity_graph_defaults, people, person_profiles, settings, certification_applicants, certification_cohorts, certification_tests, partner_certifications, partner_invite_codes, defaultvideos, default_call_background. To audit another model add Schema.plugin(auditPlugin, { ignore: [...] }) — and list its machine-written fields in ignore, or every heartbeat becomes an entry.

Screen​

Filters: record type (grouped Accounts / Billing / Integrations / Tenant config / Identity / …), action (create / update / delete), record ID, company ID, actor email (case-insensitive substring), date range. Rows expand to the field-by-field changes, or for bulk entries to the operation and the affected ids. ?record_id= and ?company_id= deep-link a record's or tenant's history. Record/company IDs only enter the query once they are a complete ObjectId; search inputs debounce 500 ms like every admin list.

Endpoint​

GET /admin/audit-logs (AuditLogController@index, routes/auditLogRoutes.js) — query collection_name, action, record_id, company_id, actor, start_date, end_date, skip, limit (max 100). A record_id filter also matches bulk entries whose record_ids contain it. Queries cap at maxTimeMS 15 s; an unfiltered count uses estimatedDocumentCount. Indexes: {collection_name, record_id, created_at}, {record_ids, created_at} (sparse), {company_id, created_at}, {actor.user_id, created_at}.

Not to be confused with the Nox Funnel Audit (/admin/nox-funnel/audit-logs, funnel signal trail — see Nox admin), the person-level identity audit on Person detail, or AI Credit Logs (credit ledger).

Call Logs — /admin/call-logs​

How it works​

Every call — AI or human, web or phone — has a service_rooms document, so its id is the call id. Since 2026-10-06 (backend 0e183b2e1/60c81f408, frontend 460ff2a7, ms-communication 6f8798e) each room carries call_logs, an array of step entries { at, source, event, message, level, data } written by ms-communication's app/helpers/CallLogHelper.js:

sourceWhere it comes from
serverevery LokiService.log(service_room_id, …) in ms-communication (call lifecycle)
socketcall socket events that name a room (SocketService middleware)
widget / dashboardclient steps sent on the call-log socket event — the dashboard's src/utils/callLog.js batches per room and flushes every second (e.g. ringing, room joined, hang-up)
  • Writes are batched (one $push per room per second) and capped at the last 1,000 entries ($slice). Logging never throws or blocks the call path.
  • data drops keys matching token|secret|password|authorization|api_key (Twilio access tokens never land here) and is truncated at 4,000 chars; messages at 500.
  • Client timestamps are trusted only within 5 minutes of server time, else the arrival time is used.
  • The field is select: false and every CallController aggregate projects it out — read it only with select('+call_logs').

Screen & endpoints​

The list shows calls that have at least one log entry, newest first, with log/error/warn counts and the last step time; filters: company ID, search by service-room id or session id, kind (AI / human), errors only, date range, page size. A row opens the step timeline (CallLogTimeline.vue).

  • GET /admin/calls/logs (AdminCallsController@listCallLogs) — company_id, search (must be an ObjectId), kind=ai|human, errors_only=true, start_date/end_date (applied as an _id range), skip, limit (default 5, max 100).
  • GET /admin/calls/:id/logs (@showCallLogs) — the room's call metadata + its logs sorted by at.