Build or Review an Audit System with evlog

For application developers who either need to add an audit trail to their product, or who already have one and want it reviewed. Walks through the design calls, the end-to-end implementation, and a review checklist for an existing setup.

This skill assumes the audit lives in your app. To extend the evlog package itself (new audit helper, new drain wrapper), see the contributor skills under .agents/skills/.

Quick reference — call-site cheat sheet

When you already know the system is wired and just need to remember the API:

AuditFields schema (always provide action, actor, outcome; target strongly recommended; the rest is filled in for you):

interface AuditFields {
  action: string                                  // 'invoice.refund'
  actor: { type: 'user' | 'system' | 'api' | 'agent', id: string, email?, displayName?, model?, tools?, reason?, promptId? }
  outcome: 'success' | 'failure' | 'denied'
  target?: { type: string, id: string, [k: string]: unknown }
  reason?: string
  changes?: { before?: unknown, after?: unknown, patch?: AuditPatchOp[] }
  causationId?: string
  correlationId?: string
  version?: number                                // defaults to AUDIT_SCHEMA_VERSION
  idempotencyKey?: string                         // auto-derived from action+actor+target+timestamp
  context?: { requestId?, traceId?, ip?, userAgent?, tenantId?, ... }   // filled by auditEnricher
  signature?: string                              // added by signed(drain, { strategy: 'hmac' })
  prevHash?: string                               // added by signed(drain, { strategy: 'hash-chain' })
  hash?: string                                   // added by signed(drain, { strategy: 'hash-chain' })
}

What "audit logging" actually means

An audit log answers a forensic question: who did what, on which resource, when, from where, with which outcome. That's a different shape from observability logs, which is why the operational rules differ:

evlog ships the audit layer as a thin extension of its wide-event pipeline (a typed audit field on BaseWideEvent plus a few helpers and drain wrappers). The point is that you compose with the primitives the app already uses — same drains, same enrichers, same redact, same framework integration. There is no parallel system to maintain.

Mental model

log.audit(...) ──► sets event.audit ──► force-keep ──► auditEnricher ──► redact ──► every drain
                                                                                  └─► auditOnly(signed(fsDrain))

Design calls before writing code

Make these explicit and write them down somewhere a security reviewer can find. Without a written rule, the system can't be audited — auditors look for the policy first, then the enforcement.

1. Where do audits live?

Rule of thumb. Pick at least two: a queryable one (Axiom / Postgres) for day-to-day forensics + an append-only one (FS journal with hash-chain, or S3 Object Lock) as the compliance artefact. The two-drain pattern protects against vendor outages and admin mistakes on the queryable side.

2. Do you need integrity (signed)?

Yes if any of:

• A compliance framework requires tamper-evidence (SOC2 CC7, HIPAA §164.312(c)(1), PCI 10.5).
• The sink is mutable by engineers / admins.
• You may need to prove to a regulator that no events were modified after the fact.

Skip if:

• Sink is already WORM (S3 Object Lock, BigQuery append-only, Postgres with row-level immutability + monitored DDL).
• You're prototyping.

Strategies:

• 'hmac' — per-event signature; quick to verify; rotate secret annually and embed a key id (extend AuditFields).
• 'hash-chain' — sequence integrity; deleting a row breaks the chain forward; persist state.{load,save} if you run multiple processes (Redis is the typical store).

3. Multi-tenancy?

If the app is multi-tenant, tenant isolation on every audit event is non-negotiable — a query that mixes tenants is a privacy incident. Wire it once in the enricher:

auditEnricher({ tenantId: ctx => resolveTenant(ctx) })

Then either (a) partition the audit dataset by audit.context.tenantId, or (b) one sink per tenant if hard isolation is required. Never query audits without a tenant filter.

4. Retention

Pick a window per sink and document it. Enforce at the sink layer, not in app code — the sink already has audited mechanisms for it (lifecycle policies, DELETE jobs, dataset retention).

How to enforce per sink:

• FS: createFsDrain({ maxFiles }) + a daily compactor.
• Postgres: DELETE FROM audit_events WHERE timestamp < now() - interval '7 years' on a cron.
• Axiom / Datadog / Loki: dataset-level retention policy.

5. GDPR vs append-only

The right to be forgotten collides with audit immutability. Recommended pattern:

1. Keep audit rows immutable and chain-verified.
2. Encrypt PII fields (email, name, IP) with a per-actor key held outside the audit store.
3. To "forget" a user, delete their key. The audit row stays, the chain stays valid, the PII becomes unreadable (crypto-shredding).

A built-in cryptoShredding helper is on the roadmap; until then, encrypt in a custom enricher.

Step-by-step buildout

Step 1 — Wire the pipeline (one-time)

The wiring shape is the same in every framework: register auditEnricher() so event.audit.context gets requestId, traceId, ip, userAgent, and (if configured) tenantId automatically, then add a main drain plus an audit-only sink.

The minimal Nuxt / Nitro setup looks like this:

// server/plugins/evlog.ts
import { auditEnricher, auditOnly, signed } from 'evlog'
import { createAxiomDrain } from 'evlog/axiom'
import { createFsDrain } from 'evlog/fs'

export default defineNitroPlugin((nitroApp) => {
  const auditSink = auditOnly(
    signed(createFsDrain({ dir: '.audit/' }), { strategy: 'hash-chain' }),
    { await: true },
  )
  const main = createAxiomDrain({ dataset: 'logs' })

  nitroApp.hooks.hook('evlog:enrich', auditEnricher({
    tenantId: ctx => ctx.headers?.['x-tenant-id'],
  }))
  nitroApp.hooks.hook('evlog:drain', async (ctx) => {
    await Promise.all([main(ctx), auditSink(ctx)])
  })
})

For Hono, Express, Next.js, or standalone scripts / workers, see references/framework-wiring.md. The pattern is identical — only the framework integration helper changes.

Step 2 — Define the action vocabulary

Audits get queried and alerted on by audit.action. A typo is a missing alert, so centralise the list:

// app/audit/actions.ts
import { defineAuditAction } from 'evlog'

export const InvoiceRefund = defineAuditAction('invoice.refund', { target: 'invoice' })
export const UserUpdate    = defineAuditAction('user.update',    { target: 'user' })
export const ApiKeyRevoke  = defineAuditAction('apiKey.revoke',  { target: 'apiKey' })
export const RolePromote   = defineAuditAction('role.promote',   { target: 'user' })

Naming conventions:

• noun.verb (invoice.refund, not refundInvoice).
• Past tense if the audit is logged after the fact (invoice.refunded); present tense when wrapped by withAudit() (which resolves the outcome itself).
• Lowercase, dot-delimited, no spaces.

Step 3 — Instrument call sites

Three patterns, in order of preference:

A. Wrap the action with withAudit() — pure audit-worthy actions (refund, delete, role change, password reset). Outcome resolution is automatic, so you can't forget to log a denial or failure:

import { withAudit, AuditDeniedError } from 'evlog'

export const refundInvoice = withAudit({
  action: 'invoice.refund',
  target: ({ id }: { id: string }) => ({ type: 'invoice', id }),
})(async ({ id }, ctx) => {
  if (!ctx.actor) throw new AuditDeniedError('Anonymous refund denied')
  return db.invoices.refund(id)
})

Outcome resolution:

• fn resolves → outcome: 'success'.
• fn throws AuditDeniedError (or any error with status === 403) → outcome: 'denied', error message becomes reason.
• Any other thrown error → outcome: 'failure', then re-thrown.

B. Manual log.audit() inside a handler — when the audit is one of several decisions in a larger handler, or when you need to emit before the action completes:

const log = useLogger(event)

if (!user.canRefund(invoice)) {
  log.audit.deny('Insufficient permissions', {
    action: 'invoice.refund',
    actor: { type: 'user', id: user.id },
    target: { type: 'invoice', id: invoice.id },
  })
  throw createError({ status: 403 })
}

const after = await db.invoices.refund(invoice.id)

log.audit({
  action: 'invoice.refund',
  actor: { type: 'user', id: user.id, email: user.email },
  target: { type: 'invoice', id: after.id },
  outcome: 'success',
  changes: auditDiff(invoice, after),
})

C. Standalone audit() for jobs / scripts — no request, no logger. Same shape, no context auto-fill:

import { audit } from 'evlog'

audit({
  action: 'cron.cleanup',
  actor: { type: 'system', id: 'cron' },
  target: { type: 'job', id: 'cleanup-stale-sessions' },
  outcome: 'success',
})

Step 4 — Add denial coverage

Auditors care most about denials — they're how you prove the policy is actually being enforced. Every authorisation check should have a paired log.audit.deny(). Pulling the deny into a single helper guarantees coverage parity with successes:

function authorize(actor, action, resource) {
  const allowed = policy.check(actor, action, resource)
  if (!allowed) {
    useLogger().audit.deny(`Policy denied ${action}`, {
      action,
      actor,
      target: { type: resource.type, id: resource.id },
    })
    throw createError({ status: 403 })
  }
}

Step 5 — Redact

Apply auditRedactPreset (or merge it into the existing RedactConfig). It drops Authorization / Cookie headers and common credential field names (password, token, apiKey, cardNumber, cvv, ssn) wherever they appear inside audit.changes.before / audit.changes.after:

import { initLogger, auditRedactPreset } from 'evlog'

initLogger({
  redact: {
    paths: [...(auditRedactPreset.paths ?? []), 'user.password', 'user.token'],
  },
})

Step 6 — Test it

mockAudit() captures audit events for assertions without going through any drain. Make the denial test mandatory in code review — untested denial paths are the most common cause of audit gaps:

import { mockAudit } from 'evlog'

it('refunds the invoice and records an audit', async () => {
  const captured = mockAudit()

  await refundInvoice({ id: 'inv_889' }, { actor: { type: 'user', id: 'u1' } })

  expect(captured.events).toHaveLength(1)
  expect(captured.toIncludeAuditOf({
    action: 'invoice.refund',
    target: { type: 'invoice', id: 'inv_889' },
    outcome: 'success',
  })).toBe(true)

  captured.restore()
})

it('denies refund for non-owners and records the denial', async () => {
  const captured = mockAudit()

  await expect(refundInvoice({ id: 'inv_889' }, { actor: null })).rejects.toThrow()

  expect(captured.toIncludeAuditOf({
    action: 'invoice.refund',
    outcome: 'denied',
  })).toBe(true)

  captured.restore()
})

Step 7 — Production readiness checklist

Walk through this with a security stakeholder before declaring the system production-ready (the same checklist powers the review mode below):

• auditEnricher is registered on every framework integration.
• Every authorisation check has a paired log.audit.deny() (greppable).
• Every mutating endpoint either uses withAudit() or calls log.audit() explicitly.
• At least two sinks: one queryable (Axiom / Postgres) and one tamper-evident (FS + signed hash-chain, or S3 Object Lock).
• auditRedactPreset (or stricter) is in the global RedactConfig.
• Retention policy is documented per sink and enforced at the sink layer.
• Multi-tenant apps: tenantId is set on every audit event; queries always filter by tenant.
• Hash-chain state.{load,save} persists across process restarts (file / Redis / Postgres).
• HMAC secret rotation procedure is documented; keyId is embedded in AuditFields (extend via declare module).
• Tests include a denial path for every privileged action.
• Audit dataset access is itself logged — meta-auditing matters.

Review an existing audit setup

When the user already has an audit system and wants it reviewed, work through the four passes below in order. Each pass tells you exactly what to grep, what to look for, and what to flag.

Pass 1 — Pipeline wiring

Find where the logger is initialised and where drains / enrichers are registered:

rg -n "initLogger|defineNitroPlugin|createLogger|evlog:enrich|evlog:drain" --type ts
rg -n "auditEnricher|auditOnly|signed\(" --type ts

Flag if:

• auditEnricher() is missing → event.audit.context is empty, no requestId / IP / tenant correlation.
• An audit-only sink exists but is not wrapped in auditOnly(...) → main events leak into the audit dataset (privacy & cost incident).
• Only one drain → no tamper-evident copy. Acceptable only if the single sink is WORM (S3 Object Lock, BigQuery append-only, Postgres immutable).
• signed() is used without a persisted state while running multiple processes → hash-chain breaks across restarts / instances.
• await: true is missing on the audit-only sink → events may be lost on crash.

Pass 2 — Coverage (call sites)

Inventory every mutating action and every authorisation check:

rg -n "log\.audit\(|log\.audit\.deny\(|withAudit\(|^.*audit\(" --type ts
rg -n "createError\(.*403|throw .*Forbidden|status:\s*403|statusCode:\s*403" --type ts
rg -n "(?i)\b(delete|update|create|refund|grant|revoke|promote|demote|reset|impersonate)\b.*async\s+function|defineEventHandler" --type ts

For each match, check:

• Mutating endpoint without a log.audit() or withAudit() → coverage gap.
• 403 / Forbidden thrown without a paired log.audit.deny() → silent denial. This is the single most common gap.
• actor: { type: 'user', id: 'cron' } or hard-coded actors in cron / queue handlers → wrong actor.type. Should be 'system', 'api', or 'agent'.
• actor.id set to a session id or token instead of the stable user id → forensic ambiguity.
• log.set({ audit: ... }) without using the helpers → bypasses force-keep, may be dropped by tail-sampling.
• withAudit() action name in present tense (invoice.refund) is fine; manual log.audit() after the fact should use past tense (invoice.refunded).

Pass 3 — Redaction & integrity

rg -n "auditRedactPreset|RedactConfig|paths:\s*\[" --type ts
rg -n "auditDiff\(" --type ts
rg -n "strategy:\s*['\"](?:hmac|hash-chain)" --type ts

Flag if:

• auditRedactPreset is not merged into the global redact config → Authorization, Cookie, password, token, apiKey, cardNumber, cvv, ssn may leak through audit.changes.
• auditDiff() is called on objects containing PII fields not listed in redactPaths → leak in the patch payload.
• HMAC secret is hard-coded or read from process.env.SECRET without a rotation plan / keyId → events become unverifiable after rotation.
• Hash-chain state is in-memory only → chain restarts each process boot, breaking continuity.

Pass 4 — Tests

rg -n "mockAudit\(|toIncludeAuditOf\(" --type ts

Flag if:

• No tests use mockAudit() → audit pipeline silently drifts unnoticed.
• Tests only assert success outcomes → denial paths can rot. Every privileged action should have at least one denied-outcome test.
• Tests assert against RegExp actions broadly → typos in audit.action slip through (an action typo is a missing alert in production).

Reporting the findings

Group findings by severity for the user:

• P0 (blocker): missing auditOnly wrap on an audit sink, missing auditRedactPreset, denials not logged, no tamper-evident sink in a regulated context.
• P1 (compliance gap): missing tenant isolation, hash-chain state not persisted, no HMAC rotation, no denial test coverage.
• P2 (hygiene): action naming inconsistency, in-line actor objects (should use defineAuditAction), missing causationId / correlationId on chained operations.

Then map each finding to the relevant step in the buildout above (e.g. P0 → Step 5 redact, P1 → Step 7 checklist) so the fix is unambiguous.

Common pitfalls

• Logging only successes. Auditors care most about denials. Pair log.audit() with log.audit.deny() on every negative branch of every check.
• Leaking PII through changes. auditDiff() runs through RedactConfig, but only if the field paths are listed. Add password, token, apiKey once globally so it's never a per-call-site decision.
• Treating audits as observability. Don't sample, downsample, or summarise audit events. Force-keep is on by default — don't disable it.
• Conflating actor.id with the session id. actor.id is the stable user id (or system identity); correlate sessions via context.requestId / context.traceId.
• Forgetting standalone jobs. Cron, queue workers, CLIs trigger audit-worthy actions too — use audit() or withAudit().
• Faking the actor type. actor.type: 'user' for cron jobs gets flagged in audits. Use 'system', 'api', or 'agent' accurately.
• Single global secret with no rotation. HMAC keys must rotate; without a keyId, old events become unverifiable after rotation.
• One drain that fails takes audits down. Sinks must fail in isolation. The default drain: [...] array does this; if you wrap in Promise.all, don't throw on a single rejection — log it.

Glossary

• Action — audit.action, the verb-on-noun identifier (invoice.refund).
• Actor — who/what performed the action (user, system, api, agent).
• Target — the resource the action was performed on.
• Outcome — success, failure, or denied.
• Idempotency key — auto-derived hash of action + actor + target + timestamp; safe retries across drains.
• Causation id — id of the action that caused this one (admin action → system reactions).
• Correlation id — shared by every action in one logical operation.
• Hash-chain — each event's prevHash matches the previous event's hash, forming a verifiable sequence.

Reference

• Per-framework wiring (Hono, Express, Next.js, standalone): references/framework-wiring.md
• Docs: Audit logs overview — source at apps/docs/content/4.use-cases/4.audit/
• Source: packages/evlog/src/audit.ts
• Tests: packages/evlog/test/core/audit.test.ts