fix(onboarding): atomic DB append for step completion (deferred bug L3)

markOnboardingStepDone read space_members.onboarding_progress, appended the step
id in JS, and wrote the whole jsonb object back — a lost-update race where two
near-simultaneous completions (double-submit, or two steps ticked quickly) both
read the same base array and the second write clobbered the first, silently
dropping a step.

Replace it with mark_onboarding_step_done (migration 055): a single atomic
UPDATE that dedup-appends the step id into onboarding_progress.completed_step_ids
under the row lock — no update can be lost (READ COMMITTED re-evaluates the jsonb
expression against the latest tuple via EvalPlanQual), other keys are preserved,
and re-appending an existing id is a no-op. Ownership-guarded by auth.uid() so a
caller can only complete steps on their own membership row; the no-JWT service
path is trusted and scoped by p_member_id (mirrors class_signup_tx et al.).

integration/onboarding-progress.test.ts covers dedup/idempotency, key
preservation, the auth.uid ownership gate, and the point of the fix: 12
concurrent completions all survive.

Migration 055 folded into schema.sql; types/database.ts regenerated.
typecheck 0 · lint 0 errors · 645 unit · 58 integration · build ok

integration/onboarding-progress.test.ts

@@ -0,0 +1,110 @@
+// Validates migration 055 (mark_onboarding_step_done) against a real Postgres:
+// the atomic dedup-append into space_members.onboarding_progress that replaces
+// the old read-modify-write in markOnboardingStepDone (deferred bug L3). The
+// central case is CONCURRENCY: many simultaneous completions must all survive —
+// the previous SELECT-then-UPDATE could silently drop steps under a lost update.
+import { describe, it, expect, beforeAll, afterAll } from 'vitest'
+import { promisify } from 'node:util'
+import { execFile } from 'node:child_process'
+import {
+  dbReady,
+  DB_URL,
+  rows,
+  execService,
+  execAsUser,
+  seedSpace,
+  seedMember,
+  dropSpace,
+  uuid,
+} from './db'
+
+const execFileP = promisify(execFile)
+const d = describe.skipIf(!dbReady)
+
+d('mark_onboarding_step_done (migration 055)', () => {
+  let spaceId: string
+
+  beforeAll(() => {
+    spaceId = seedSpace().spaceId
+  })
+  afterAll(() => dropSpace(spaceId))
+
+  // Completed step ids as the app stores/reads them: onboarding_progress
+  // -> completed_step_ids (a jsonb string array), or [] when unset.
+  function completedSteps(memberId: string): string[] {
+    const r = rows<{ ids: string[] | null }>(
+      `select (onboarding_progress -> 'completed_step_ids') as ids
+         from space_members where id = '${memberId}'`,
+    )
+    return (r[0]?.ids as string[] | null) ?? []
+  }
+
+  const markService = (memberId: string, stepId: string) =>
+    execService(`select mark_onboarding_step_done('${memberId}', '${stepId}')`)
+
+  it('dedup-appends and is idempotent', () => {
+    const { memberId } = seedMember(spaceId)
+    const a = uuid()
+    const b = uuid()
+    expect(markService(memberId, a).ok).toBe(true)
+    expect(markService(memberId, b).ok).toBe(true)
+    expect(markService(memberId, a).ok).toBe(true) // repeat -> no-op
+
+    const ids = completedSteps(memberId)
+    expect([...ids].sort()).toEqual([a, b].sort())
+    expect(ids).toHaveLength(2) // 'a' not duplicated
+  })
+
+  it('preserves other keys already in onboarding_progress', () => {
+    const { memberId } = seedMember(spaceId)
+    execService(
+      `update space_members set onboarding_progress = '{"foo":"bar"}'::jsonb where id = '${memberId}'`,
+    )
+    const a = uuid()
+    markService(memberId, a)
+
+    const r = rows<{ p: { foo?: string; completed_step_ids?: string[] } }>(
+      `select onboarding_progress as p from space_members where id = '${memberId}'`,
+    )
+    expect(r[0].p.foo).toBe('bar')
+    expect(r[0].p.completed_step_ids).toEqual([a])
+  })
+
+  it('concurrent completions never lose an update (the L3 fix)', async () => {
+    const { memberId } = seedMember(spaceId)
+    const steps = Array.from({ length: 12 }, () => uuid())
+
+    // Fire every append at once through separate psql connections. The old
+    // read-modify-write would have interleaved SELECTs and dropped steps; the
+    // single-statement atomic UPDATE serializes on the row lock so all survive.
+    await Promise.all(
+      steps.map(s =>
+        execFileP('psql', [
+          DB_URL,
+          '-v',
+          'ON_ERROR_STOP=1',
+          '-tAc',
+          `select mark_onboarding_step_done('${memberId}', '${s}')`,
+        ]),
+      ),
+    )
+
+    const ids = completedSteps(memberId)
+    expect([...ids].sort()).toEqual([...steps].sort()) // all 12 present, none lost
+  })
+
+  it('ownership: with a JWT a member can only complete steps on their OWN row', () => {
+    const owner = seedMember(spaceId)
+    const other = seedMember(spaceId)
+    const step = uuid()
+
+    // `other`, acting as a logged-in user, aims the RPC at `owner`'s row.
+    // auth.uid() != owner.user_id -> the WHERE matches nothing -> no-op.
+    execAsUser(other.userId, `select mark_onboarding_step_done('${owner.memberId}', '${step}')`)
+    expect(completedSteps(owner.memberId)).toEqual([])
+
+    // The owner completing their own step works.
+    execAsUser(owner.userId, `select mark_onboarding_step_done('${owner.memberId}', '${step}')`)
+    expect(completedSteps(owner.memberId)).toEqual([step])
+  })
+})

lib/actions/onboarding.ts

@@ -130,20 +130,15 @@ export async function markOnboardingStepDone(stepId: string) {
   if (!auth.ok) return { error: auth.error }
   const { member } = auth
 
-  const { data: row } = await supabase
+  // Atomic dedup-append in the DB (migration 055). The old read-modify-write on
-    .from('space_members')
+  // onboarding_progress had a lost-update race: two near-simultaneous
-    .select('onboarding_progress')
+  // completions both read the same base array and the second write clobbered
-    .eq('id', member.id)
+  // the first, silently dropping a step. The RPC pins the write to the caller's
-    .single()
+  // own row via auth.uid().
-
+  const { error } = await supabase.rpc('mark_onboarding_step_done', {
-  const progress = (row?.onboarding_progress ?? {}) as { completed_step_ids?: string[] }
+    p_member_id: member.id,
-  const done = new Set(progress.completed_step_ids ?? [])
+    p_step_id: v.data,
-  done.add(v.data)
+  })
-
-  const { error } = await supabase
-    .from('space_members')
-    .update({ onboarding_progress: { ...progress, completed_step_ids: Array.from(done) } })
-    .eq('id', member.id)
 
   if (error) return { error: error.message }
   return { success: true as const }

scripts/055_onboarding_step_atomic_append.sql

@@ -0,0 +1,63 @@
+-- =============================================================================
+-- 055: atomic, concurrency-safe onboarding step completion
+-- =============================================================================
+-- markOnboardingStepDone previously did a read-modify-write on the
+-- space_members.onboarding_progress jsonb column: SELECT the object, add the
+-- step id to completed_step_ids in JS, then UPDATE the whole object back. Two
+-- near-simultaneous completions (a double-submit, or two steps ticked in quick
+-- succession) both read the same base array and the second write clobbers the
+-- first -- a lost update that silently drops a completed step (deferred bug L3).
+--
+-- mark_onboarding_step_done replaces that with a single atomic UPDATE that
+-- dedup-appends the step id into onboarding_progress.completed_step_ids. Because
+-- it is one statement, the row lock serializes concurrent callers and Postgres
+-- re-evaluates the jsonb expression against the latest committed tuple
+-- (EvalPlanQual) under READ COMMITTED -- so no update is ever lost, and other
+-- keys in the object are preserved. Appending an already-present id is a no-op
+-- (idempotent), matching the old Set semantics.
+--
+-- Ownership: the id-append is harmless, but a caller must still only be able to
+-- touch their OWN membership row. When invoked with a JWT (auth.uid() set, the
+-- normal RLS/user client path) the WHERE clause pins the row to
+-- user_id = auth.uid(), so passing another member's id updates nothing. The
+-- service/no-JWT path (auth.uid() IS NULL) is trusted and scoped by p_member_id,
+-- mirroring the existing atomic RPCs (class_signup_tx et al.).
+--
+-- Idempotent migration: CREATE OR REPLACE. Apply as-is (Supabase SQL editor /
+-- psql); re-runs are no-ops.
+-- =============================================================================
+
+CREATE OR REPLACE FUNCTION public.mark_onboarding_step_done(
+  p_member_id uuid, p_step_id uuid
+) RETURNS jsonb
+  LANGUAGE plpgsql SECURITY DEFINER SET search_path = public AS $$
+DECLARE
+  v_completed jsonb;
+BEGIN
+  UPDATE public.space_members m
+  SET onboarding_progress = jsonb_set(
+        coalesce(m.onboarding_progress, '{}'::jsonb),
+        '{completed_step_ids}',
+        CASE
+          WHEN coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
+                 @> to_jsonb(p_step_id::text)
+            THEN m.onboarding_progress -> 'completed_step_ids'
+          ELSE coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
+                 || to_jsonb(p_step_id::text)
+        END
+      )
+  WHERE m.id = p_member_id
+    AND (auth.uid() IS NULL OR m.user_id = auth.uid())
+  RETURNING m.onboarding_progress -> 'completed_step_ids' INTO v_completed;
+
+  -- NULL when no row matched (not the caller's membership); the resulting
+  -- completed_step_ids array otherwise.
+  RETURN v_completed;
+END;
+$$;
+
+-- Lock down default PUBLIC execute. The user (RLS) client calls this as the
+-- authenticated role; the ownership guard above keeps it to the caller's own
+-- row. service_role (the admin client) may also call it for the trusted path.
+REVOKE EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) FROM PUBLIC;
+GRANT  EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) TO authenticated, service_role;

scripts/schema.sql

@@ -1655,6 +1655,41 @@ CREATE INDEX IF NOT EXISTS idx_onboarding_steps_space ON public.space_onboarding
 ALTER TABLE public.space_members ADD COLUMN IF NOT EXISTS onboarding_completed_at timestamptz;
 ALTER TABLE public.space_members ADD COLUMN IF NOT EXISTS onboarding_progress jsonb NOT NULL DEFAULT '{}';
 
+-- Atomic, concurrency-safe step completion (migration 055). Single-statement
+-- dedup-append into onboarding_progress.completed_step_ids so two concurrent
+-- completions can't lose an update (the old read-modify-write in
+-- markOnboardingStepDone could). Ownership-guarded: with a JWT it only touches
+-- the caller's own row; the no-JWT service path is trusted and scoped by
+-- p_member_id.
+CREATE OR REPLACE FUNCTION public.mark_onboarding_step_done(
+  p_member_id uuid, p_step_id uuid
+) RETURNS jsonb
+  LANGUAGE plpgsql SECURITY DEFINER SET search_path = public AS $$
+DECLARE
+  v_completed jsonb;
+BEGIN
+  UPDATE public.space_members m
+  SET onboarding_progress = jsonb_set(
+        coalesce(m.onboarding_progress, '{}'::jsonb),
+        '{completed_step_ids}',
+        CASE
+          WHEN coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
+                 @> to_jsonb(p_step_id::text)
+            THEN m.onboarding_progress -> 'completed_step_ids'
+          ELSE coalesce(m.onboarding_progress -> 'completed_step_ids', '[]'::jsonb)
+                 || to_jsonb(p_step_id::text)
+        END
+      )
+  WHERE m.id = p_member_id
+    AND (auth.uid() IS NULL OR m.user_id = auth.uid())
+  RETURNING m.onboarding_progress -> 'completed_step_ids' INTO v_completed;
+
+  RETURN v_completed;
+END;
+$$;
+REVOKE EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) FROM PUBLIC;
+GRANT  EXECUTE ON FUNCTION public.mark_onboarding_step_done(uuid, uuid) TO authenticated, service_role;
+
 CREATE OR REPLACE FUNCTION public.seed_default_onboarding_steps()
   RETURNS trigger LANGUAGE plpgsql SECURITY DEFINER SET search_path = public
 AS $$

types/database.ts

@@ -3462,6 +3462,10 @@ export type Database = {
       }
       expire_proposals: { Args: never; Returns: number }
       get_user_space_ids: { Args: { uid: string }; Returns: string[] }
+      mark_onboarding_step_done: {
+        Args: { p_member_id: string; p_step_id: string }
+        Returns: Json
+      }
       members_with_permission: {
         Args: { perm: string; sid: string }
         Returns: {