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>

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)
+  })
+}