feat(perf): measure the auth proxy per request (Server-Timing + proxy completed log) (#1922)

The proxy in front of every page, RSC, prefetch and /api request makes
several sequential network calls (getUser, session state, the
resolve_active_company RPC, MFA factor lookups) and nothing measured them,
while the route wrapper has logged authMs/companyMs/handlerMs per API call
for months. This is the first PR of the responsiveness plan (customer
report: "it takes time before all fields load when clicking around"): the
baseline every later change is measured against.

- lib/supabase/proxy-timing.ts: pure helpers (request classification from
  the app-router headers, route template that collapses ids and tokens,
  Server-Timing formatting, a timed() accumulator).
- lib/supabase/middleware.ts: updateSession wraps updateSessionInner, times
  each phase, sets Server-Timing on page/RSC/prefetch responses and
  X-Proxy-Timing on /api responses (withRouteContext owns Server-Timing
  there), and emits one "proxy completed" log line per request.
- scripts/perf/log-percentiles.ts: p50/p90/p99 per group over
  `vercel logs --json` output, for both "op completed" and
  "proxy completed"; scripts/perf/README.md documents the protocol and
  targets.

Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

lib/supabase/__tests__/middleware.test.ts

@@ -72,6 +72,19 @@ vi.mock('@supabase/ssr', () => ({
   })),
 }))
 
+const logState = vi.hoisted(() => ({ info: vi.fn() }))
+
+vi.mock('@/lib/logger', () => {
+  const logger = {
+    info: (...args: unknown[]) => logState.info(...args),
+    warn: vi.fn(),
+    error: vi.fn(),
+    debug: vi.fn(),
+    child: () => logger,
+  }
+  return { createLogger: () => logger }
+})
+
 import { updateSession } from '../middleware'
 import {
   createSessionTimeoutState,
@@ -103,6 +116,7 @@ describe('updateSession redirect destinations', () => {
 
   beforeEach(() => {
     vi.clearAllMocks()
+    logState.info.mockClear()
     state.user = null
     state.sessionId = 'session-1'
     state.authError = null
@@ -535,6 +549,67 @@ describe('updateSession redirect destinations', () => {
 
   // ── MFA semantics that must not change ────────────────────────────────
 
+  describe('per-request timing header and log line', () => {
+    const TIMING_RE =
+      /^mw-auth;dur=\d+, mw-session;dur=\d+, mw-company;dur=\d+, mw-mfa;dur=\d+, mw-total;dur=\d+$/
+
+    function lastLog() {
+      expect(logState.info).toHaveBeenCalledTimes(1)
+      const [msg, ctx] = logState.info.mock.calls[0] as [string, Record<string, unknown>]
+      expect(msg).toBe('proxy completed')
+      return ctx
+    }
+
+    it('page responses carry Server-Timing and log kind=page with the route', async () => {
+      state.user = SIGNED_IN
+      const res = await run('/invoices')
+      expect(res.status).toBe(200)
+      expect(res.headers.get('server-timing')).toMatch(TIMING_RE)
+      expect(res.headers.get('x-proxy-timing')).toBeNull()
+      const ctx = lastLog()
+      expect(ctx.kind).toBe('page')
+      expect(ctx.route).toBe('/invoices')
+      expect(ctx.status).toBe(200)
+      expect(typeof ctx.totalMs).toBe('number')
+      expect(typeof ctx.authMs).toBe('number')
+      expect(typeof ctx.companyMs).toBe('number')
+    })
+
+    it('classifies prefetch and RSC requests from the app-router headers', async () => {
+      state.user = SIGNED_IN
+      await run('/invoices', { headers: { 'next-router-prefetch': '1', rsc: '1' } })
+      expect(lastLog().kind).toBe('prefetch')
+      logState.info.mockClear()
+      await run('/invoices', { headers: { rsc: '1' } })
+      expect(lastLog().kind).toBe('rsc')
+    })
+
+    it('/api responses use X-Proxy-Timing and leave Server-Timing to the route wrapper', async () => {
+      state.user = SIGNED_IN
+      const res = await run('/api/settings')
+      expect(res.headers.get('x-proxy-timing')).toMatch(TIMING_RE)
+      expect(res.headers.get('server-timing')).toBeNull()
+      expect(lastLog().kind).toBe('api')
+    })
+
+    it('redirect responses also carry the header and log their status', async () => {
+      const res = await run('/invoices')
+      expect(res.status).toBe(307)
+      expect(res.headers.get('server-timing')).toMatch(TIMING_RE)
+      expect(lastLog().status).toBe(307)
+    })
+
+    it('never logs a token-carrying path or a raw entity id', async () => {
+      const res = await run('/invite/9f8e7d6c5b4a3928171605f4e3d2c1b0')
+      expect(res.status).toBe(200)
+      expect(lastLog().route).toBe('/invite/*')
+      logState.info.mockClear()
+      state.user = SIGNED_IN
+      await run('/invoices/6f1c2a3e-1234-4bcd-9abc-0123456789ab')
+      expect(lastLog().route).toBe('/invoices/:id')
+    })
+  })
+
   describe('MFA-disabled and self-hosted paths are unchanged', () => {
     it('does not redirect when NEXT_PUBLIC_REQUIRE_MFA is unset', async () => {
       state.user = SIGNED_IN

lib/supabase/__tests__/proxy-timing.test.ts

@@ -0,0 +1,83 @@
+import { describe, it, expect } from 'vitest'
+import {
+  classifyProxyRequest,
+  createProxyTimings,
+  formatProxyServerTiming,
+  proxyRouteTemplate,
+  timed,
+} from '../proxy-timing'
+
+describe('classifyProxyRequest', () => {
+  it('treats /api paths as api regardless of headers', () => {
+    const headers = new Headers({ 'next-router-prefetch': '1', rsc: '1' })
+    expect(classifyProxyRequest('/api/settings', headers)).toBe('api')
+  })
+
+  it('recognises app-router prefetch and RSC requests by header', () => {
+    expect(
+      classifyProxyRequest('/invoices', new Headers({ 'Next-Router-Prefetch': '1', RSC: '1' })),
+    ).toBe('prefetch')
+    expect(classifyProxyRequest('/invoices', new Headers({ RSC: '1' }))).toBe('rsc')
+  })
+
+  it('falls back to page for a plain document request', () => {
+    expect(classifyProxyRequest('/invoices', new Headers())).toBe('page')
+  })
+})
+
+describe('proxyRouteTemplate', () => {
+  it('replaces UUID and numeric segments with placeholders', () => {
+    expect(proxyRouteTemplate('/invoices/6f1c2a3e-1234-4bcd-9abc-0123456789ab/edit')).toBe(
+      '/invoices/:id/edit',
+    )
+    expect(proxyRouteTemplate('/salary/runs/42')).toBe('/salary/runs/:n')
+  })
+
+  it('collapses token-carrying prefixes so secrets never reach the log', () => {
+    expect(proxyRouteTemplate('/invite/9f8e7d6c5b4a3928171605f4e3d2c1b0')).toBe('/invite/*')
+    expect(proxyRouteTemplate('/payslip/abc')).toBe('/payslip/*')
+    expect(proxyRouteTemplate('/auth/callback')).toBe('/auth/*')
+    expect(proxyRouteTemplate('/auth')).toBe('/auth/*')
+  })
+
+  it('masks long opaque segments outside the known prefixes', () => {
+    expect(proxyRouteTemplate('/e/sector/aVeryLongOpaqueSlugThatLooksLikeAToken')).toBe(
+      '/e/sector/:token',
+    )
+  })
+
+  it('keeps ordinary routes and the root untouched', () => {
+    expect(proxyRouteTemplate('/settings/company')).toBe('/settings/company')
+    expect(proxyRouteTemplate('/')).toBe('/')
+  })
+})
+
+describe('formatProxyServerTiming', () => {
+  it('emits one mw-* metric per phase plus the total', () => {
+    const timing = { authMs: 12, sessionMs: 3, companyMs: 40, mfaMs: 0 }
+    expect(formatProxyServerTiming(timing, 61)).toBe(
+      'mw-auth;dur=12, mw-session;dur=3, mw-company;dur=40, mw-mfa;dur=0, mw-total;dur=61',
+    )
+  })
+})
+
+describe('timed', () => {
+  it('returns the wrapped value and accumulates elapsed time on the key', async () => {
+    const timing = createProxyTimings()
+    const value = await timed(timing, 'sessionMs', async () => 'ok')
+    expect(value).toBe('ok')
+    await timed(timing, 'sessionMs', async () => undefined)
+    expect(timing.sessionMs).toBeGreaterThanOrEqual(0)
+    expect(timing.authMs).toBe(0)
+  })
+
+  it('still records time when the wrapped call throws', async () => {
+    const timing = createProxyTimings()
+    await expect(
+      timed(timing, 'authMs', async () => {
+        throw new Error('boom')
+      }),
+    ).rejects.toThrow('boom')
+    expect(timing.authMs).toBeGreaterThanOrEqual(0)
+  })
+})

lib/supabase/middleware.ts

@@ -1,5 +1,15 @@
 import { createServerClient } from '@supabase/ssr'
 import { NextResponse, type NextRequest } from 'next/server'
+import { createLogger } from '@/lib/logger'
+import {
+  PROXY_TIMING_HEADER,
+  classifyProxyRequest,
+  createProxyTimings,
+  formatProxyServerTiming,
+  proxyRouteTemplate,
+  timed,
+  type ProxyTimings,
+} from '@/lib/supabase/proxy-timing'
 import { shouldEnforceMfa } from '@/lib/auth/mfa'
 import { apiPathSkipsMfaGate } from '@/lib/auth/api-mfa-gate'
 import { DEFAULT_LOCALE, LOCALE_COOKIE, isLocale } from '@/i18n/config'
@@ -27,7 +37,43 @@ import {
   type SessionTimeoutReason,
 } from '@/lib/auth/session-timeout-shared'
 
+const log = createLogger('proxy')
+
+/**
+ * Auth proxy entry point. Wraps the real work so every response carries a
+ * per-phase timing header and emits one structured log line, mirroring what
+ * withRouteContext does for API routes: without it the proxy's sequential
+ * network calls (getUser, session state, company RPC, MFA lookups) were the
+ * one part of a request nobody could measure. Page/RSC/prefetch responses
+ * get `Server-Timing` (visible in the browser Timing tab); /api responses
+ * get `X-Proxy-Timing` so the route wrapper's own Server-Timing is left
+ * alone. Token-carrying paths are collapsed before logging.
+ */
 export async function updateSession(request: NextRequest) {
+  const start = Date.now()
+  const timing = createProxyTimings()
+  const response = await updateSessionInner(request, timing)
+  const totalMs = Date.now() - start
+  const pathname = request.nextUrl.pathname
+  const kind = classifyProxyRequest(pathname, request.headers)
+  response.headers.set(
+    kind === 'api' ? PROXY_TIMING_HEADER : 'Server-Timing',
+    formatProxyServerTiming(timing, totalMs),
+  )
+  log.info('proxy completed', {
+    kind,
+    route: proxyRouteTemplate(pathname),
+    status: response.status,
+    ...timing,
+    totalMs,
+  })
+  return response
+}
+
+async function updateSessionInner(
+  request: NextRequest,
+  timing: ProxyTimings,
+): Promise<NextResponse> {
   let supabaseResponse = NextResponse.next({
     request,
   })
@@ -62,7 +108,7 @@ export async function updateSession(request: NextRequest) {
   const {
     data: { user },
     error: authError,
-  } = await supabase.auth.getUser()
+  } = await timed(timing, 'authMs', () => supabase.auth.getUser())
 
   // Get the pathname
   const pathname = request.nextUrl.pathname
@@ -93,8 +139,12 @@ export async function updateSession(request: NextRequest) {
     !apiRequestSkipsSessionTimeout(pathname, hasAuthorizationHeader)
   ) {
     const encodedState = request.cookies.get(SESSION_TIMEOUT_COOKIE)?.value
-    const sessionId = await getSupabaseSessionId(supabase)
-    const verifiedState = await verifySessionTimeoutState(encodedState)
+    const sessionId = await timed(timing, 'sessionMs', () =>
+      getSupabaseSessionId(supabase),
+    )
+    const verifiedState = await timed(timing, 'sessionMs', () =>
+      verifySessionTimeoutState(encodedState),
+    )
 
     if (encodedState && !verifiedState) {
       await signOutTimedOutSession(supabase)
@@ -121,7 +171,9 @@ export async function updateSession(request: NextRequest) {
       const method = isSessionAuthMethod(hintedMethod)
         ? hintedMethod
         : 'password'
-      const autoLogout = await fetchAutoLogoutPreference(supabase, user.id)
+      const autoLogout = await timed(timing, 'sessionMs', () =>
+        fetchAutoLogoutPreference(supabase, user.id),
+      )
 
       // Unknown preference (failed read): mint nothing, so no fail-open
       // snapshot gets persisted; the next request retries the read.
@@ -184,8 +236,9 @@ export async function updateSession(request: NextRequest) {
       hasAuthorizationHeader,
     )
     if (!skipMfaGate && user && shouldEnforceMfa(user)) {
-      const { data: aal } =
-        await supabase.auth.mfa.getAuthenticatorAssuranceLevel()
+      const { data: aal } = await timed(timing, 'mfaMs', () =>
+        supabase.auth.mfa.getAuthenticatorAssuranceLevel(),
+      )
       if (aal?.nextLevel === 'aal2' && aal?.currentLevel === 'aal1') {
         return NextResponse.json({ error: 'MFA-verifiering krävs.' }, { status: 403 })
       }
@@ -310,11 +363,15 @@ export async function updateSession(request: NextRequest) {
     degraded: boolean
   } | null = null
   const resolveCompanyOnce = async () =>
-    (resolvedCompany ??= await resolveCompanyForMiddleware(supabase, user.id, request))
+    (resolvedCompany ??= await timed(timing, 'companyMs', () =>
+      resolveCompanyForMiddleware(supabase, user.id, request),
+    ))
 
   // MFA enforcement (application-side only, not RLS)
   if (shouldEnforceMfa(user)) {
-    const { data: aal } = await supabase.auth.mfa.getAuthenticatorAssuranceLevel()
+    const { data: aal } = await timed(timing, 'mfaMs', () =>
+      supabase.auth.mfa.getAuthenticatorAssuranceLevel(),
+    )
 
     // User has MFA enrolled but hasn't verified this session → redirect to verify
     if (aal?.nextLevel === 'aal2' && aal?.currentLevel === 'aal1') {
@@ -325,7 +382,9 @@ export async function updateSession(request: NextRequest) {
     // Skip for users with no companies (still setting up)
     const { companyId: companyIdForMfa } = await resolveCompanyOnce()
     if (companyIdForMfa) {
-      const { data: factors } = await supabase.auth.mfa.listFactors()
+      const { data: factors } = await timed(timing, 'mfaMs', () =>
+        supabase.auth.mfa.listFactors(),
+      )
       const hasVerifiedFactor = factors?.totp?.some(f => f.status === 'verified')
 
       if (!hasVerifiedFactor) {

lib/supabase/proxy-timing.ts

@@ -0,0 +1,115 @@
+/**
+ * Per-request timing for the auth proxy (lib/supabase/middleware.ts).
+ *
+ * The proxy runs in front of every page, RSC, prefetch and /api request and
+ * makes several sequential network calls (Supabase Auth, the active-company
+ * RPC, MFA factor lookups). Nothing measured that cost until now, while the
+ * route wrapper (lib/api/with-route-context.ts) has logged authMs/companyMs/
+ * handlerMs per API call for months. These helpers give the proxy the same
+ * Server-Timing header and one structured "proxy completed" log line per
+ * request so the fixed per-request cost can be read off Vercel logs and the
+ * browser Timing tab instead of guessed.
+ *
+ * Pure functions only: the middleware test suite mocks Supabase heavily, so
+ * classification and formatting live here where they can be unit-tested
+ * without that harness.
+ */
+
+export type ProxyRequestKind = 'api' | 'prefetch' | 'rsc' | 'page'
+
+export interface ProxyTimings {
+  /** supabase.auth.getUser(): a network round trip to Supabase Auth. */
+  authMs: number
+  /** Session-timeout state: getClaims + HMAC verify + auto_logout read. */
+  sessionMs: number
+  /** resolve_active_company RPC (or the query fallback) incl. write-back. */
+  companyMs: number
+  /** MFA assurance level + listFactors (the latter is a network call). */
+  mfaMs: number
+}
+
+export type ProxyTimingKey = keyof ProxyTimings
+
+/**
+ * Header used on /api responses. withRouteContext owns `Server-Timing` on
+ * API routes and only sets it when absent, so the proxy's numbers travel on
+ * a separate header there; on page/RSC responses nothing else sets
+ * Server-Timing and the proxy uses the standard header directly.
+ */
+export const PROXY_TIMING_HEADER = 'X-Proxy-Timing'
+
+export function createProxyTimings(): ProxyTimings {
+  return { authMs: 0, sessionMs: 0, companyMs: 0, mfaMs: 0 }
+}
+
+/** Run `fn` and add its wall time to `timing[key]` (accumulates on repeats). */
+export async function timed<T>(
+  timing: ProxyTimings,
+  key: ProxyTimingKey,
+  fn: () => Promise<T>,
+): Promise<T> {
+  const start = Date.now()
+  try {
+    return await fn()
+  } finally {
+    timing[key] += Date.now() - start
+  }
+}
+
+/**
+ * Which kind of request the proxy is fronting. Prefetch and RSC requests
+ * are recognised by the headers the Next.js app router sends
+ * (`Next-Router-Prefetch: 1`, `RSC: 1`); header names are case-insensitive
+ * on the Fetch `Headers` interface, so lower-case lookups are fine.
+ */
+export function classifyProxyRequest(
+  pathname: string,
+  headers: Headers,
+): ProxyRequestKind {
+  if (pathname.startsWith('/api')) return 'api'
+  if (headers.get('next-router-prefetch') === '1') return 'prefetch'
+  if (headers.get('rsc') === '1') return 'rsc'
+  return 'page'
+}
+
+const UUID_RE =
+  /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i
+const NUMERIC_RE = /^\d+$/
+/** Prefixes whose tail is a secret (invite tokens, payslip links, PKCE). */
+const TOKEN_PREFIXES = ['/invite', '/payslip', '/auth']
+
+/**
+ * Collapse a concrete pathname to a loggable route template: UUIDs become
+ * `:id`, numbers `:n`, long opaque segments `:token`, and paths under the
+ * token-carrying prefixes are cut to the prefix so a secret never lands in
+ * a log line.
+ */
+export function proxyRouteTemplate(pathname: string): string {
+  for (const prefix of TOKEN_PREFIXES) {
+    if (pathname === prefix || pathname.startsWith(`${prefix}/`)) {
+      return `${prefix}/*`
+    }
+  }
+  const segments = pathname.split('/').map((segment) => {
+    if (segment === '') return segment
+    if (UUID_RE.test(segment)) return ':id'
+    if (NUMERIC_RE.test(segment)) return ':n'
+    if (segment.length >= 24) return ':token'
+    return segment
+  })
+  return segments.join('/') || '/'
+}
+
+/** `Server-Timing` value: one `mw-*` metric per phase plus the total. */
+export function formatProxyServerTiming(
+  timing: ProxyTimings,
+  totalMs: number,
+): string {
+  return [
+    `mw-auth;dur=${timing.authMs}`,
+    `mw-session;dur=${timing.sessionMs}`,
+    `mw-company;dur=${timing.companyMs}`,
+    `mw-mfa;dur=${timing.mfaMs}`,
+    `mw-total;dur=${totalMs}`,
+  ].join(', ')
+}

scripts/__tests__/log-percentiles.test.ts

@@ -0,0 +1,123 @@
+import { describe, it, expect } from 'vitest'
+import {
+  extractRecord,
+  parseArgs,
+  percentile,
+  renderMarkdown,
+  summarize,
+} from '../perf/log-percentiles'
+
+const opLine = (operation: string, durationMs: number, authMs = 2) =>
+  JSON.stringify({
+    level: 'info',
+    module: `api/${operation}`,
+    msg: 'op completed',
+    operation,
+    durationMs,
+    authMs,
+  })
+
+describe('extractRecord', () => {
+  it('parses a bare logger line', () => {
+    expect(extractRecord(opLine('period.list', 77))).toMatchObject({
+      operation: 'period.list',
+      durationMs: 77,
+    })
+  })
+
+  it('unwraps the JSON embedded in a vercel logs --json envelope', () => {
+    const envelope = JSON.stringify({
+      timestamp: 1,
+      source: 'serverless',
+      requestPath: '/api/bookkeeping/fiscal-periods',
+      message: `    ${opLine('period.list', 77)}`,
+    })
+    expect(extractRecord(envelope)).toMatchObject({
+      source: 'serverless',
+      operation: 'period.list',
+      durationMs: 77,
+    })
+  })
+
+  it('keeps a non-JSON message as the envelope only', () => {
+    expect(extractRecord(JSON.stringify({ message: 'plain text' }))).toEqual({
+      message: 'plain text',
+    })
+  })
+
+  it('ignores blank and unparseable lines', () => {
+    expect(extractRecord('')).toBeNull()
+    expect(extractRecord('not json')).toBeNull()
+    expect(extractRecord('{broken')).toBeNull()
+  })
+})
+
+describe('percentile', () => {
+  it('uses nearest rank', () => {
+    const sorted = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+    expect(percentile(sorted, 50)).toBe(5)
+    expect(percentile(sorted, 90)).toBe(9)
+    expect(percentile(sorted, 99)).toBe(10)
+    expect(percentile([42], 50)).toBe(42)
+    expect(Number.isNaN(percentile([], 50))).toBe(true)
+  })
+})
+
+describe('summarize', () => {
+  const records = [
+    { operation: 'a', durationMs: 10, authMs: 1 },
+    { operation: 'a', durationMs: 30, authMs: 1 },
+    { operation: 'b', durationMs: 100, authMs: 50 },
+    { operation: 'b', durationMs: 'oops', authMs: 5 },
+  ]
+
+  it('groups by the requested keys and ranks slowest first', () => {
+    const rows = summarize(records, { fields: ['durationMs', 'authMs'], groupBy: ['operation'] })
+    expect(rows.map((r) => r.group)).toEqual(['b', 'a'])
+    expect(rows[1].count).toBe(2)
+    expect(rows[1].fields.durationMs).toEqual({ p50: 10, p90: 30, p99: 30, max: 30 })
+    // The non-numeric durationMs is skipped for that field but the row still counts the sample.
+    expect(rows[0].count).toBe(2)
+    expect(rows[0].fields.durationMs.max).toBe(100)
+  })
+
+  it('applies exact-match filters and min-count', () => {
+    const rows = summarize(records, {
+      fields: ['durationMs'],
+      groupBy: ['operation'],
+      filters: [{ key: 'operation', value: 'a' }],
+    })
+    expect(rows).toHaveLength(1)
+    expect(rows[0].group).toBe('a')
+    expect(summarize(records, { fields: ['durationMs'], groupBy: ['operation'], minCount: 3 })).toEqual([])
+  })
+
+  it('produces a single "all" row without grouping', () => {
+    const rows = summarize(records, { fields: ['durationMs'] })
+    expect(rows).toHaveLength(1)
+    expect(rows[0].group).toBe('all')
+    expect(rows[0].count).toBe(4)
+  })
+})
+
+describe('renderMarkdown + parseArgs', () => {
+  it('renders a markdown table with one column set per field', () => {
+    const rows = summarize([{ k: 'x', v: 5 }], { fields: ['v'], groupBy: ['k'] })
+    const md = renderMarkdown(rows, ['v'])
+    expect(md.split('\n')[0]).toBe('| group | n | v p50 | v p90 | v p99 | v max |')
+    expect(md).toContain('| x | 1 | 5 | 5 | 5 | 5 |')
+  })
+
+  it('parses the documented CLI flags', () => {
+    expect(
+      parseArgs(['--field', 'a,b', '--group', 'kind,route', '--filter', 'msg=op completed', '--min-count', '5']),
+    ).toEqual({
+      fields: ['a', 'b'],
+      groupBy: ['kind', 'route'],
+      filters: [{ key: 'msg', value: 'op completed' }],
+      minCount: 5,
+    })
+    expect(() => parseArgs([])).toThrow('--field is required')
+    expect(() => parseArgs(['--bogus'])).toThrow('unknown option')
+  })
+})

scripts/perf/README.md

@@ -0,0 +1,69 @@
+# Request latency measurement (auth proxy + route wrapper)
+
+Why this exists: a customer reported that "it takes time before all fields load when clicking around" (2026-08-26). The API handlers themselves are fast (p50 38 ms); the cost is the number of sequential calls a page makes and the fixed per-request tax in front of each one (auth proxy + route wrapper). This page is the protocol for measuring that tax before and after every change in the responsiveness plan, so no PR claims a win without a number.
+
+## What is instrumented
+
+| Surface | Where | Header | Log line |
+|---|---|---|---|
+| Auth proxy (every page, RSC, prefetch, `/api` request) | `lib/supabase/middleware.ts` via `lib/supabase/proxy-timing.ts` | `Server-Timing: mw-auth, mw-session, mw-company, mw-mfa, mw-total` on page/RSC/prefetch responses; `X-Proxy-Timing` (same value) on `/api` responses | `proxy completed` with `kind` (`page`, `rsc`, `prefetch`, `api`), `route` (ids and tokens collapsed), `status`, `authMs`, `sessionMs`, `companyMs`, `mfaMs`, `totalMs` |
+| Route wrapper (`withRouteContext`, 367 of 529 routes) | `lib/api/with-route-context.ts` | `Server-Timing: auth, company, handler` | `op completed` with `operation`, `status`, `durationMs`, `authMs`, `companyMs`, `handlerMs` |
+| Browser | `@vercel/speed-insights` mounted in `app/layout.tsx` | n/a | Vercel dashboard > Speed Insights > Routes (p75 TTFB, FCP, LCP, INP, CLS per route) |
+
+Phase meanings for the proxy: `authMs` = `supabase.auth.getUser()` (a network call to Supabase Auth); `sessionMs` = session-timeout cookie state (`getClaims`, HMAC verify, `auto_logout` read when re-minting); `companyMs` = `resolve_active_company` RPC including the write-back on fallback; `mfaMs` = assurance-level check plus `listFactors()` (a second network call) on the enforced-MFA path.
+
+## Reading the numbers
+
+Percentiles per route from production logs (the `vercel` CLI is linked to the project; `--limit` defaults to 100, raise it):
+
+```bash
+# route wrapper, grouped by operation
+vercel logs --environment production --since 24h --limit 1000 --json --query "op completed" \
+  | npx tsx scripts/perf/log-percentiles.ts --field durationMs,authMs,companyMs,handlerMs --group operation
+
+# auth proxy, grouped by request kind and route
+vercel logs --environment production --since 24h --limit 1000 --json --query "proxy completed" \
+  | npx tsx scripts/perf/log-percentiles.ts --field totalMs,authMs,sessionMs,companyMs,mfaMs --group kind,route
+
+# auth proxy, one row per kind (the headline fixed cost)
+vercel logs --environment production --since 24h --limit 1000 --json --query "proxy completed" \
+  | npx tsx scripts/perf/log-percentiles.ts --field totalMs,authMs,companyMs,mfaMs --group kind
+```
+
+If `--limit` above 100 is not honoured by the installed CLI, loop `--since`/`--until` windows and concatenate before piping. The Vercel MCP `get_runtime_logs` tool (`group_by: route`) gives request counts per route, not percentiles; use it for prefetch volume (`kind=prefetch` per page load).
+
+In the browser: DevTools > Network, pick a document or `_rsc` request, Timing tab, the `mw-*` metrics show the proxy phases; `/api` responses show `auth`/`company`/`handler` from the route wrapper and the proxy numbers in the `X-Proxy-Timing` response header.
+
+Request count per interaction (the number that maps to "fields load late"), in the console after each action on a production build (`next build && next start`):
+
+```js
+performance.getEntriesByType('resource')
+  .filter((r) => /\/api\/|\/rest\/v1\//.test(r.name))
+  .map((r) => `${Math.round(r.startTime)} ${Math.round(r.duration)}ms ${r.name}`)
+```
+
+Interactions to record every time: `/transactions` soft navigation; open Bokför on a row; `/bookkeeping` then Nytt verifikat; `/invoices` then Ny faktura; `/reports`; `/supplier-invoices/new`; a list row to its detail page on customers and invoices. Note the request count and how many dependent rounds (start times that wait on an earlier response).
+
+## Targets (p75 unless stated)
+
+| Metric | Target | Why |
+|---|---|---|
+| Proxy `page`/`rsc` totalMs | p50 < 40 ms, p90 < 100 ms | Two parallel DB waves at most, no auth network call |
+| Proxy `api` totalMs | p50 < 5 ms | Local JWT verification only |
+| Route wrapper auth + company | p50 < 45 ms (read), same for write | One company resolution per call, never two |
+| TTFB (hard load) | < 400 ms | Layout on two DB waves |
+| LCP | < 1.5 s | |
+| INP | < 200 ms | The metric "clicking around" maps to |
+| Reference-data requests per form open | 0 blocking | Seeded and cached client-side |
+
+## Protocol per PR
+
+1. Before merging: record the current numbers for the interactions above on a production build, and the last 24 h of `proxy completed` / `op completed` percentiles, in the PR description.
+2. 24 h after the production deploy: re-run the same three commands and the same interactions; append one row per PR to the table below.
+3. A PR that claims a latency win without a before/after row is not done.
+
+## Baseline and results
+
+| Date | Change | Proxy page p50/p90 | Proxy api p50/p90 | Route auth+company p50 | Notes |
+|---|---|---|---|---|---|
+| 2026-08-26 | Route wrapper only (proxy unmeasured) | n/a | n/a | 44 ms (auth 3, company 41) | 100-call sample; handler p50 38 ms, total p50 96 ms, p90 274 ms |

scripts/perf/log-percentiles.ts

@@ -0,0 +1,205 @@
+/**
+ * Percentiles over structured log lines.
+ *
+ * Reads JSON Lines on stdin (the shape `vercel logs --json` emits, or raw
+ * logger output) and prints a markdown table of count / p50 / p90 / p99 /
+ * max per group for the numeric fields asked for. Used for the
+ * "op completed" lines from lib/api/with-route-context.ts and the
+ * "proxy completed" lines from lib/supabase/middleware.ts.
+ *
+ *   vercel logs --environment production --since 24h --limit 1000 --json \
+ *     --query "op completed" \
+ *     | npx tsx scripts/perf/log-percentiles.ts \
+ *         --field durationMs,authMs,companyMs,handlerMs --group operation
+ *
+ *   vercel logs --environment production --since 24h --limit 1000 --json \
+ *     --query "proxy completed" \
+ *     | npx tsx scripts/perf/log-percentiles.ts \
+ *         --field totalMs,authMs,companyMs,mfaMs --group kind,route
+ *
+ * Options: --field a,b (required), --group x,y (default: none, one row),
+ * --filter key=value (repeatable; exact match on the parsed record),
+ * --min-count N (drop groups with fewer samples, default 1).
+ *
+ * No dependencies on purpose: this must run from a clean checkout.
+ */
+
+export type LogRecord = Record<string, unknown>
+
+/**
+ * Turn one input line into a flat record. `vercel logs --json` wraps the
+ * application line in a `message` (or `text`) string, so an embedded JSON
+ * object inside that string is parsed and merged over the envelope; a bare
+ * JSON logger line is used as-is. Unparseable lines yield null.
+ */
+export function extractRecord(line: string): LogRecord | null {
+  const trimmed = line.trim()
+  if (!trimmed.startsWith('{')) return null
+  let envelope: LogRecord
+  try {
+    envelope = JSON.parse(trimmed) as LogRecord
+  } catch {
+    return null
+  }
+  const message = envelope.message ?? envelope.text
+  if (typeof message === 'string') {
+    const start = message.indexOf('{')
+    if (start >= 0) {
+      try {
+        const embedded = JSON.parse(message.slice(start)) as LogRecord
+        return { ...envelope, ...embedded }
+      } catch {
+        // Not a JSON payload: fall through and use the envelope alone.
+      }
+    }
+  }
+  return envelope
+}
+
+/** Nearest-rank percentile on an ascending-sorted array. */
+export function percentile(sorted: number[], p: number): number {
+  if (sorted.length === 0) return Number.NaN
+  const rank = Math.ceil((p / 100) * sorted.length)
+  return sorted[Math.min(sorted.length, Math.max(1, rank)) - 1]
+}
+
+export interface FieldStats {
+  p50: number
+  p90: number
+  p99: number
+  max: number
+}
+
+export interface GroupRow {
+  group: string
+  count: number
+  fields: Record<string, FieldStats>
+}
+
+export interface SummarizeOptions {
+  fields: string[]
+  groupBy?: string[]
+  filters?: Array<{ key: string; value: string }>
+  minCount?: number
+}
+
+function matchesFilters(record: LogRecord, filters: SummarizeOptions['filters']): boolean {
+  if (!filters || filters.length === 0) return true
+  return filters.every(({ key, value }) => String(record[key]) === value)
+}
+
+export function summarize(records: LogRecord[], options: SummarizeOptions): GroupRow[] {
+  const groupBy = options.groupBy ?? []
+  const minCount = options.minCount ?? 1
+  const buckets = new Map<string, { count: number; values: Record<string, number[]> }>()
+
+  for (const record of records) {
+    if (!matchesFilters(record, options.filters)) continue
+    const groupKey = groupBy.length
+      ? groupBy.map((key) => String(record[key] ?? '')).join(' / ')
+      : 'all'
+    let bucket = buckets.get(groupKey)
+    if (!bucket) {
+      bucket = {
+        count: 0,
+        values: Object.fromEntries(options.fields.map((f) => [f, [] as number[]])),
+      }
+      buckets.set(groupKey, bucket)
+    }
+    bucket.count += 1
+    for (const field of options.fields) {
+      const value = record[field]
+      if (typeof value === 'number' && Number.isFinite(value)) bucket.values[field].push(value)
+    }
+  }
+
+  const rows: GroupRow[] = []
+  for (const [group, bucket] of buckets) {
+    const count = bucket.count
+    if (count < minCount) continue
+    const fields: Record<string, FieldStats> = {}
+    for (const field of options.fields) {
+      const sorted = [...bucket.values[field]].sort((a, b) => a - b)
+      fields[field] = {
+        p50: percentile(sorted, 50),
+        p90: percentile(sorted, 90),
+        p99: percentile(sorted, 99),
+        max: sorted.length ? sorted[sorted.length - 1] : Number.NaN,
+      }
+    }
+    rows.push({ group, count, fields })
+  }
+
+  // Slowest first by the first field's p50 so the table reads as a ranking.
+  const primary = options.fields[0]
+  rows.sort((a, b) => (b.fields[primary]?.p50 ?? 0) - (a.fields[primary]?.p50 ?? 0))
+  return rows
+}
+
+function fmt(n: number): string {
+  return Number.isNaN(n) ? '-' : String(Math.round(n))
+}
+
+export function renderMarkdown(rows: GroupRow[], fields: string[]): string {
+  const header = ['group', 'n', ...fields.flatMap((f) => [`${f} p50`, `${f} p90`, `${f} p99`, `${f} max`])]
+  const lines = [
+    `| ${header.join(' | ')} |`,
+    `|${header.map(() => '---').join('|')}|`,
+  ]
+  for (const row of rows) {
+    const cells = [row.group, String(row.count)]
+    for (const field of fields) {
+      const s = row.fields[field]
+      cells.push(fmt(s.p50), fmt(s.p90), fmt(s.p99), fmt(s.max))
+    }
+    lines.push(`| ${cells.join(' | ')} |`)
+  }
+  return lines.join('\n')
+}
+
+export function parseArgs(argv: string[]): SummarizeOptions {
+  const options: SummarizeOptions = { fields: [], groupBy: [], filters: [], minCount: 1 }
+  for (let i = 0; i < argv.length; i += 1) {
+    const arg = argv[i]
+    const next = () => {
+      i += 1
+      const value = argv[i]
+      if (value === undefined) throw new Error(`${arg} needs a value`)
+      return value
+    }
+    if (arg === '--field') options.fields = next().split(',').filter(Boolean)
+    else if (arg === '--group') options.groupBy = next().split(',').filter(Boolean)
+    else if (arg === '--filter') {
+      const [key, ...rest] = next().split('=')
+      options.filters!.push({ key, value: rest.join('=') })
+    } else if (arg === '--min-count') options.minCount = Number(next())
+    else throw new Error(`unknown option ${arg}`)
+  }
+  if (options.fields.length === 0) throw new Error('--field is required')
+  return options
+}
+
+async function readStdin(): Promise<string> {
+  const chunks: Buffer[] = []
+  for await (const chunk of process.stdin) chunks.push(chunk as Buffer)
+  return Buffer.concat(chunks).toString('utf8')
+}
+
+async function main(): Promise<void> {
+  const options = parseArgs(process.argv.slice(2))
+  const input = await readStdin()
+  const records = input
+    .split('\n')
+    .map(extractRecord)
+    .filter((r): r is LogRecord => r !== null)
+  const rows = summarize(records, options)
+  process.stdout.write(`${records.length} records parsed\n\n`)
+  process.stdout.write(`${renderMarkdown(rows, options.fields)}\n`)
+}
+
+if (process.argv[1] && /log-percentiles\.(?:ts|mts|js|mjs)$/.test(process.argv[1])) {
+  main().catch((err) => {
+    console.error(err instanceof Error ? err.message : err)
+    process.exit(1)
+  })
+}