@luanpdd/kit-mcp 1.20.0 → 1.22.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (51) hide show
  1. package/README.md +1 -1
  2. package/gates/dept-cycle-prevention.md +179 -0
  3. package/gates/multi-tenant-rls-coverage.md +102 -0
  4. package/gates/service-role-not-in-user-facing.md +113 -0
  5. package/kit/README.md +24 -0
  6. package/kit/agents/audit-log-implementer.md +175 -0
  7. package/kit/agents/auditor-consistencia-isolamento.md +380 -0
  8. package/kit/agents/b2b-saas-architect.md +156 -0
  9. package/kit/agents/crm-pipeline-implementer.md +167 -0
  10. package/kit/agents/detector-tenant-quente.md +337 -0
  11. package/kit/agents/evolution-go-integrator.md +179 -0
  12. package/kit/agents/invite-flow-implementer.md +137 -0
  13. package/kit/agents/lgpd-compliance-auditor.md +206 -0
  14. package/kit/agents/multi-tenant-isolation-auditor.md +253 -0
  15. package/kit/agents/multi-tenant-rls-writer.md +262 -0
  16. package/kit/agents/org-onboarding-implementer.md +202 -0
  17. package/kit/agents/supabase-architect.md +10 -0
  18. package/kit/agents/supabase-migration-writer.md +12 -0
  19. package/kit/agents/super-admin-implementer.md +182 -0
  20. package/kit/agents/validador-evolucao-schema.md +335 -0
  21. package/kit/commands/dados-distribuidos.md +188 -0
  22. package/kit/commands/multi-tenant.md +163 -0
  23. package/kit/file-manifest.json +48 -9
  24. package/kit/skills/_shared-dados-distribuidos/glossary.md +224 -0
  25. package/kit/skills/_shared-multi-tenant/glossary.md +186 -0
  26. package/kit/skills/armadilhas-sistemas-distribuidos/SKILL.md +447 -0
  27. package/kit/skills/audit-log-multi-tenant/SKILL.md +340 -0
  28. package/kit/skills/b2b-saas-architecture/SKILL.md +300 -0
  29. package/kit/skills/cascading-failures/SKILL.md +4 -0
  30. package/kit/skills/consistencia-leitura-replica/SKILL.md +385 -0
  31. package/kit/skills/crm-lead-pipeline-patterns/SKILL.md +343 -0
  32. package/kit/skills/escolha-modelo-consistencia/SKILL.md +495 -0
  33. package/kit/skills/evolucao-schema-compativel/SKILL.md +448 -0
  34. package/kit/skills/evolution-go-whatsapp-integration/SKILL.md +322 -0
  35. package/kit/skills/lgpd-multi-tenant-compliance/SKILL.md +340 -0
  36. package/kit/skills/member-invite-flow/SKILL.md +305 -0
  37. package/kit/skills/member-management-react-shadcn/SKILL.md +328 -0
  38. package/kit/skills/multi-tenant-performance-scaling/SKILL.md +316 -0
  39. package/kit/skills/multi-tenant-rls-hierarchy/SKILL.md +342 -0
  40. package/kit/skills/org-onboarding-flow/SKILL.md +257 -0
  41. package/kit/skills/org-switcher-react-pattern/SKILL.md +349 -0
  42. package/kit/skills/permission-gate-react-pattern/SKILL.md +271 -0
  43. package/kit/skills/postgres-isolamento-concorrencia/SKILL.md +552 -0
  44. package/kit/skills/rbac-permissions-matrix-supabase/SKILL.md +301 -0
  45. package/kit/skills/streams-eventos-cdc/SKILL.md +712 -0
  46. package/kit/skills/supabase-cron-queues/SKILL.md +9 -0
  47. package/kit/skills/supabase-migrations/SKILL.md +10 -0
  48. package/kit/skills/super-admin-platform-pattern/SKILL.md +326 -0
  49. package/kit/skills/tenant-quente-mitigacao/SKILL.md +605 -0
  50. package/kit/skills/whatsapp-conversation-state-machine/SKILL.md +287 -0
  51. package/package.json +1 -1
@@ -0,0 +1,316 @@
1
+ ---
2
+ name: multi-tenant-performance-scaling
3
+ description: Use ao escalar Postgres multi-tenant em Supabase — Supavisor transaction mode (porta 6543), partial indexes obrigatórios em colunas de RLS, helper functions STABLE, partitioning por org_id quando >50k rows/tenant, MVs per-tenant para query caching.
4
+ ---
5
+
6
+ # Multi-Tenant Performance & Scaling — Postgres + Supabase
7
+
8
+ ## Quando usar
9
+
10
+ LLM carrega esta skill ao escalar app B2B multi-tenant em Supabase para alta carga (>1k req/s, >50k rows/tenant, >100 tenants ativos). Trigger phrases:
11
+
12
+ - "Supavisor connection pooling", "transaction mode 6543"
13
+ - "RLS performance multi-tenant", "helper function STABLE"
14
+ - "partitioning por org_id", "particionamento Postgres tenant"
15
+ - "materialized view per tenant", "MV per-tenant"
16
+ - "scaling multi-tenant Postgres", "connection pool exhaustion"
17
+ - "queries lentas multi-tenant"
18
+
19
+ Esta skill é consumida por `multi-tenant-rls-writer` (Phase 108) ao gerar policies (índices acompanham), por `b2b-saas-architecture` (skill irmã) ao definir schema, e por `multi-tenant-isolation-auditor` ao auditar performance gaps.
20
+
21
+ ## Regras absolutas
22
+
23
+ **REGRA #1 (connection pooling Vercel):** **SEMPRE** porta **6543** (Supavisor transaction mode) para Vercel Edge/Serverless. Porta 5432 (direct) só para long-running Node.js processes (workers, schedulers). Session mode na porta 6543 foi **deprecado em 2025-02-28** — não usar.
24
+
25
+ **REGRA #2 (helper functions STABLE):** Funções PG usadas em RLS policies **DEVEM** ser marcadas `STABLE` (não `VOLATILE` que é o default). VOLATILE re-executa por linha — em tabela 100k rows = 100k chamadas extras.
26
+
27
+ **REGRA #3 (partial indexes obrigatórios):** Cada coluna referenciada por RLS policy precisa de índice. Para multi-tenant, usar **partial indexes** em status='active' — exclui suspended/left que não são consultados em hot path.
28
+
29
+ **REGRA #4 (partitioning threshold):** Particionar tabela por `org_id` (LIST partitioning) quando **>50k rows/tenant** OU **>100 tenants × 1k rows**. Abaixo disso, índices regulares são mais simples e performáticos.
30
+
31
+ **REGRA #5 (MV refresh strategy):** Materialized Views per-tenant devem ter `REFRESH MATERIALIZED VIEW CONCURRENTLY` (não bloqueante) + `pg_cron` schedule. Refresh inline em request é anti-pattern.
32
+
33
+ ## Patterns canônicos
34
+
35
+ ### Supavisor — connection string Vercel
36
+
37
+ ```typescript
38
+ // .env.local (Vercel) — porta 6543 SEMPRE
39
+ DATABASE_URL="postgres://postgres.{project_ref}:{password}@aws-0-{region}.pooler.supabase.com:6543/postgres"
40
+
41
+ // Para Prisma adicionar pgbouncer flag (Supavisor é compatível com pgbouncer protocol)
42
+ DATABASE_URL="postgres://...:6543/postgres?pgbouncer=true&connection_limit=1"
43
+
44
+ // Para long-running workers (cron jobs, BullMQ) usar 5432 direct connection
45
+ WORKER_DATABASE_URL="postgres://postgres:{password}@db.{project_ref}.supabase.co:5432/postgres"
46
+ ```
47
+
48
+ **Por que `connection_limit=1` com Prisma:** Supavisor é transaction-pooled — cada connection serverless deve ter `connection_limit=1` para não esgotar o pool no servidor.
49
+
50
+ ### Helper function STABLE — exemplo correto
51
+
52
+ ```sql
53
+ -- STABLE — re-executa apenas 1× por query (Postgres pode cachear)
54
+ create or replace function private.is_member_of(p_org_id uuid)
55
+ returns boolean
56
+ language sql
57
+ stable -- ⭐ CRÍTICO — não VOLATILE
58
+ security invoker
59
+ set search_path = ''
60
+ as $$
61
+ select exists (
62
+ select 1 from public.organization_members
63
+ where org_id = p_org_id
64
+ and user_id = (select auth.uid())
65
+ and status = 'active'
66
+ );
67
+ $$;
68
+ ```
69
+
70
+ **Diferença em produção:**
71
+ - VOLATILE em policy SELECT em tabela 100k rows = 100k chamadas a `is_member_of` = ~10s query
72
+ - STABLE em mesma policy = 1 chamada cacheada = ~50ms query
73
+ - Speedup 200×
74
+
75
+ ### Partial indexes obrigatórios
76
+
77
+ ```sql
78
+ -- Index parcial em organization_members (status='active' é hot path)
79
+ create index organization_members_user_org_active_idx
80
+ on public.organization_members (user_id, org_id)
81
+ where status = 'active';
82
+
83
+ -- Composite index em roles (lookup por nome dentro da org)
84
+ create index roles_org_name_idx
85
+ on public.roles (org_id, name);
86
+
87
+ -- Composite index em permissions (lookup por action+resource)
88
+ create index permissions_action_resource_idx
89
+ on public.permissions (action, resource);
90
+
91
+ -- Composite index em role_permissions (lookup por role)
92
+ create index role_permissions_role_idx
93
+ on public.role_permissions (role_id);
94
+
95
+ -- Index em departments (lookup por org)
96
+ create index departments_org_idx
97
+ on public.departments (org_id);
98
+
99
+ -- Index em audit_logs (sempre filtrar por tenant_id primeiro)
100
+ create index audit_logs_tenant_created_idx
101
+ on public.audit_logs (tenant_id, created_at desc);
102
+ ```
103
+
104
+ **Sem esses indexes, RLS força sequential scan a cada query** — degradação cresce linear com tamanho da tabela.
105
+
106
+ ### Partitioning por org_id (LIST partitioning)
107
+
108
+ Aplicar quando **uma tabela** atinge >50k rows/tenant ou >5M total. Exemplo `audit_logs`:
109
+
110
+ ```sql
111
+ -- Tabela particionada por LIST de org_id
112
+ create table public.audit_logs (
113
+ id uuid not null,
114
+ tenant_id uuid not null,
115
+ event_type text not null,
116
+ actor_id uuid,
117
+ payload jsonb,
118
+ created_at timestamptz not null default now(),
119
+ primary key (id, tenant_id)
120
+ ) partition by list (tenant_id);
121
+
122
+ -- Função que cria partição automaticamente para nova org
123
+ create or replace function private.create_audit_partition(p_org_id uuid)
124
+ returns void
125
+ language plpgsql
126
+ security definer
127
+ set search_path = ''
128
+ as $$
129
+ declare
130
+ partition_name text;
131
+ begin
132
+ partition_name := 'audit_logs_' || replace(p_org_id::text, '-', '_');
133
+ execute format(
134
+ 'create table if not exists public.%I partition of public.audit_logs for values in (%L)',
135
+ partition_name, p_org_id
136
+ );
137
+ end;
138
+ $$;
139
+
140
+ -- Trigger em organizations: cria partição ao criar org
141
+ create or replace function private.on_org_created()
142
+ returns trigger
143
+ language plpgsql
144
+ security definer
145
+ set search_path = ''
146
+ as $$
147
+ begin
148
+ perform private.create_audit_partition(new.id);
149
+ return new;
150
+ end;
151
+ $$;
152
+
153
+ create trigger create_audit_partition_on_org_create
154
+ after insert on public.organizations
155
+ for each row execute function private.on_org_created();
156
+ ```
157
+
158
+ **Quando NÃO particionar:**
159
+ - <50k rows/tenant (overhead > benefit)
160
+ - Queries cross-tenant frequentes (super-admin views) — partitioning torna isso lento
161
+ - <100 tenants total (overhead > benefit)
162
+
163
+ ### Materialized Views per-tenant
164
+
165
+ ```sql
166
+ -- MV agregando métricas por org (lead count, member count, etc.)
167
+ create materialized view public.org_metrics as
168
+ select
169
+ o.id as org_id,
170
+ o.name,
171
+ count(distinct om.user_id) filter (where om.status = 'active') as active_members,
172
+ count(distinct d.id) as departments_count,
173
+ max(om.joined_at) as last_member_joined
174
+ from public.organizations o
175
+ left join public.organization_members om on om.org_id = o.id
176
+ left join public.departments d on d.org_id = o.id
177
+ group by o.id, o.name;
178
+
179
+ create unique index org_metrics_org_id_idx on public.org_metrics (org_id);
180
+
181
+ -- Refresh CONCURRENT via pg_cron (não bloqueante)
182
+ select cron.schedule(
183
+ 'refresh-org-metrics',
184
+ '*/15 * * * *', -- a cada 15min
185
+ $$ refresh materialized view concurrently public.org_metrics; $$
186
+ );
187
+
188
+ -- RLS sobre a MV (mesma policy de organizations)
189
+ alter materialized view public.org_metrics enable row level security;
190
+
191
+ create policy "org_metrics_select_member"
192
+ on public.org_metrics
193
+ for select
194
+ to authenticated
195
+ using (private.is_member_of(org_id));
196
+ ```
197
+
198
+ **REFRESH CONCURRENTLY exige unique index** na MV (linha `unique index ... on public.org_metrics (org_id)` acima).
199
+
200
+ ### Diagnóstico de performance — queries canônicas
201
+
202
+ ```sql
203
+ -- Top 10 queries lentas com RLS
204
+ select
205
+ query,
206
+ calls,
207
+ mean_exec_time,
208
+ total_exec_time
209
+ from pg_stat_statements
210
+ where query ilike '%organization_members%'
211
+ order by mean_exec_time desc
212
+ limit 10;
213
+
214
+ -- Tabelas sem index nas colunas de RLS
215
+ select
216
+ c.relname as table_name,
217
+ pg_size_pretty(pg_total_relation_size(c.oid)) as size
218
+ from pg_class c
219
+ where c.relrowsecurity = true
220
+ and c.relkind = 'r'
221
+ and not exists (
222
+ select 1 from pg_index i
223
+ where i.indrelid = c.oid
224
+ );
225
+
226
+ -- Connection pool usage (Supavisor)
227
+ -- Verificar via Supabase Dashboard → Settings → Database → Connection Pooling
228
+ ```
229
+
230
+ ## Anti-patterns
231
+
232
+ ### Anti-pattern 1: Helper function VOLATILE (default)
233
+
234
+ **Errado:**
235
+ ```sql
236
+ create function private.is_member_of(p_org_id uuid)
237
+ returns boolean
238
+ language sql
239
+ -- sem STABLE — default é VOLATILE
240
+ as $$ select exists(...); $$;
241
+ ```
242
+
243
+ **Por quê:** PG re-executa a função para CADA linha avaliada pela policy. Em tabela 100k rows, isso é 100k chamadas — cada uma com query interna ao banco. Latência cresce linearmente com tamanho da tabela.
244
+
245
+ **Certo:** marcar `STABLE` (acima). PG executa 1× por query e cacheia o resultado para o restante das linhas.
246
+
247
+ ### Anti-pattern 2: Conexão direta porta 5432 em Vercel
248
+
249
+ **Errado:**
250
+ ```typescript
251
+ // Vercel Edge Function usando porta 5432 direct
252
+ DATABASE_URL="postgres://...:5432/postgres"
253
+ ```
254
+
255
+ **Por quê:** cada invocação Edge cria conexão direta ao Postgres. 100 req/s × 5 segundos = 500 conexões abertas simultaneamente. Postgres tem limite ~100 conns padrão — esgota → "too many connections" → 500 errors em produção.
256
+
257
+ **Certo:** porta 6543 Supavisor transaction mode. Pool gerenciado por Supabase, conns recicladas após cada transaction.
258
+
259
+ ### Anti-pattern 3: Particionar tabela com poucos rows
260
+
261
+ **Errado:**
262
+ ```sql
263
+ -- App ainda em pre-launch, 5 tenants, 200 rows total
264
+ create table public.events (...) partition by list (org_id);
265
+ ```
266
+
267
+ **Por quê:** overhead de partition pruning + management > benefit. Cada query passa por partition routing, manutenção é complexa, dump/restore mais lento. Premature optimization clássica.
268
+
269
+ **Certo:** começar com tabela regular + indexes. Particionar quando atingir threshold real (>50k rows/tenant).
270
+
271
+ ### Anti-pattern 4: MV refresh inline em request
272
+
273
+ **Errado:**
274
+ ```typescript
275
+ // Edge Function ao servir dashboard
276
+ await supabase.rpc('refresh_org_metrics') // 30s+ blocking!
277
+ const metrics = await supabase.from('org_metrics').select()
278
+ ```
279
+
280
+ **Por quê:** REFRESH MATERIALIZED VIEW (sem CONCURRENTLY) bloqueia a tabela. CONCURRENTLY não bloqueia mas leva tempo. Em request síncrono, user espera 30s.
281
+
282
+ **Certo:** REFRESH CONCURRENTLY agendado por pg_cron. Request lê snapshot atual da MV (instantâneo).
283
+
284
+ ### Anti-pattern 5: Index full table quando partial cobre 90%
285
+
286
+ **Errado:**
287
+ ```sql
288
+ -- Index full sobre uma tabela onde 90% rows são status='left'
289
+ create index members_user_idx on organization_members (user_id);
290
+ ```
291
+
292
+ **Por quê:** index inclui rows inativas (status in ('suspended', 'left')) que não são consultadas em hot path. Tamanho do index 10× maior que necessário, refresh 10× mais lento.
293
+
294
+ **Certo:**
295
+ ```sql
296
+ create index members_user_active_idx
297
+ on organization_members (user_id, org_id)
298
+ where status = 'active';
299
+ ```
300
+
301
+ Apenas rows ativas no index. Tamanho 10× menor, query plan idêntico em hot path.
302
+
303
+ ## Detecção e Mitigação de Tenant Quente (v1.22+)
304
+
305
+ > Para detectar e mitigar o "tenant Justin Bieber" (1 tenant >>> outros), ver skill [`tenant-quente-mitigacao`](../tenant-quente-mitigacao/SKILL.md) (v1.22 — DDIA Ch 6). Cobre 3 métricas canônicas (queries/min, storage GB, conexões), 5 estratégias de mitigação (rate limit, pool isolado, replica dedicada, MV, request shaping), particionamento range vs hash para tenant_id, e rebalanceamento sem downtime.
306
+
307
+ ## Ver também
308
+
309
+ - [b2b-saas-architecture](../b2b-saas-architecture/SKILL.md) — schema canônico que esta skill performance-otimiza (skill irmã)
310
+ - [multi-tenant-rls-hierarchy](../multi-tenant-rls-hierarchy/SKILL.md) — helper functions PG (Phase 108) consumem REGRA #2 (STABLE)
311
+ - [supabase-rls-policies](../supabase-rls-policies/SKILL.md) — `(select auth.uid())` wrapper já cobre 1 dimensão de performance (esta skill cobre as outras)
312
+ - [supabase-database-functions](../supabase-database-functions/SKILL.md) — padrões PG functions (security invoker, search_path)
313
+ - [supabase-cron-queues](../supabase-cron-queues/SKILL.md) — pg_cron usado em MV refresh schedule
314
+ - [Supabase Supavisor 1M Connections](https://supabase.com/blog/supavisor-1-million)
315
+ - [Supabase RLS Performance Best Practices](https://supabase.com/docs/guides/troubleshooting/rls-performance-and-best-practices-Z5Jjwv)
316
+ - [Postgres LIST Partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html#DDL-PARTITIONING-DECLARATIVE)
@@ -0,0 +1,342 @@
1
+ ---
2
+ name: multi-tenant-rls-hierarchy
3
+ description: Use ao escrever RLS hierárquica multi-tenant (org→dept→role→permission→super-admin bypass) em Supabase. 4 helper functions PG canônicas em schema private (STABLE), policies compostas com PERMISSIVE para super_admin, herança dept→org via coalesce.
4
+ ---
5
+
6
+ # Multi-Tenant RLS Hierarchy — Helper Functions + Policies
7
+
8
+ ## Quando usar
9
+
10
+ LLM carrega esta skill ao escrever RLS para tabelas em app B2B multi-tenant com hierarquia firm→department→leader→collaborator. Trigger phrases:
11
+
12
+ - "RLS multi-tenant hierárquica", "RLS org dept role"
13
+ - "helper function private", "is_member_of", "has_role", "has_permission", "is_super_admin"
14
+ - "policy composta com super_admin bypass"
15
+ - "department member herda role"
16
+ - "PERMISSIVE policy super admin"
17
+
18
+ Esta skill **estende** [`supabase-rls-policies`](../supabase-rls-policies/SKILL.md) (v1.8) — herda anti-pitfalls básicos (`(select auth.uid())` wrapper, no `user_metadata`, granular policies, indexes) e adiciona hierarquia + super_admin bypass.
19
+
20
+ ## Regras absolutas
21
+
22
+ **REGRA #1 (helper functions em schema `private`):** Funções PG de RLS **DEVEM** estar em schema `private` — NÃO em `public`. PostgREST não expõe schema `private` automaticamente, então funções não viram endpoints REST acidentalmente.
23
+
24
+ **REGRA #2 (STABLE marker obrigatório):** Helper functions usadas em policy **DEVEM** ser marcadas `STABLE` (não default `VOLATILE`). VOLATILE re-executa por linha — degradação de até 1000× em tabelas grandes.
25
+
26
+ **REGRA #3 (security invoker + search_path = ''):** Helper functions **DEVEM** ter `security invoker` (default seguro) + `set search_path = ''` (previne search path injection).
27
+
28
+ **REGRA #4 (super_admin via PERMISSIVE separada):** Bypass de super_admin **NÃO** é OR dentro da policy normal. É uma policy `as permissive` separada. PostgreSQL combina policies PERMISSIVE com OR — admin policy concedendo acesso = bypass total preservando granularidade.
29
+
30
+ **REGRA #5 (herança dept→org via coalesce):** `department_members.role_id` NULL = herda do `organization_members.role_id` da mesma `(org, user)`. Resolução via função `private.effective_role_in_dept(p_dept_id, p_user_id)` que retorna `coalesce(dm.role_id, om.role_id)`.
31
+
32
+ **REGRA #6 (todas anti-pitfalls v1.8 herdadas):** Aplicam-se SEMPRE — `(select auth.uid())` wrapper, NUNCA `user_metadata` em authz, 4 policies granulares (não `for all`), `to authenticated`/`to anon` explícito, indexes nas colunas das policies. Ver [`supabase-rls-policies`](../supabase-rls-policies/SKILL.md).
33
+
34
+ ## Patterns canônicos
35
+
36
+ ### 4 helper functions canônicas — DDL completo
37
+
38
+ ```sql
39
+ -- Schema private (não exposto via PostgREST)
40
+ create schema if not exists private;
41
+
42
+ -- 1. is_member_of — checa se user é member ativo de uma org
43
+ create or replace function private.is_member_of(p_org_id uuid)
44
+ returns boolean
45
+ language sql
46
+ stable -- REGRA #2 — re-execução cacheada
47
+ security invoker -- REGRA #3 — usa permissões do caller
48
+ set search_path = '' -- REGRA #3 — previne injection
49
+ as $$
50
+ select exists (
51
+ select 1 from public.organization_members
52
+ where org_id = p_org_id
53
+ and user_id = (select auth.uid())
54
+ and status = 'active'
55
+ );
56
+ $$;
57
+
58
+ -- 2. has_role — checa se user tem role específica numa org
59
+ create or replace function private.has_role(p_org_id uuid, p_role_name text)
60
+ returns boolean
61
+ language sql
62
+ stable
63
+ security invoker
64
+ set search_path = ''
65
+ as $$
66
+ select exists (
67
+ select 1 from public.organization_members om
68
+ join public.roles r on r.id = om.role_id
69
+ where om.org_id = p_org_id
70
+ and om.user_id = (select auth.uid())
71
+ and om.status = 'active'
72
+ and r.name = p_role_name
73
+ );
74
+ $$;
75
+
76
+ -- 3. has_permission — checa se user tem permission resource:action numa org
77
+ create or replace function private.has_permission(p_action text, p_resource text, p_org_id uuid)
78
+ returns boolean
79
+ language sql
80
+ stable
81
+ security invoker
82
+ set search_path = ''
83
+ as $$
84
+ select exists (
85
+ select 1
86
+ from public.organization_members om
87
+ join public.role_permissions rp on rp.role_id = om.role_id
88
+ join public.permissions p on p.id = rp.permission_id
89
+ where om.org_id = p_org_id
90
+ and om.user_id = (select auth.uid())
91
+ and om.status = 'active'
92
+ and p.action = p_action
93
+ and p.resource = p_resource
94
+ );
95
+ $$;
96
+
97
+ -- 4. is_super_admin — checa flag em JWT app_metadata
98
+ create or replace function private.is_super_admin()
99
+ returns boolean
100
+ language sql
101
+ stable
102
+ security invoker
103
+ set search_path = ''
104
+ as $$
105
+ select coalesce(
106
+ ((select auth.jwt())->'app_metadata'->>'super_admin')::boolean,
107
+ false
108
+ );
109
+ $$;
110
+ ```
111
+
112
+ **Indexes obrigatórios** (de [`multi-tenant-performance-scaling`](../multi-tenant-performance-scaling/SKILL.md)):
113
+
114
+ ```sql
115
+ -- Partial index — REGRA #3 da skill performance
116
+ create index if not exists organization_members_user_org_active_idx
117
+ on public.organization_members (user_id, org_id)
118
+ where status = 'active';
119
+
120
+ create index if not exists role_permissions_role_idx
121
+ on public.role_permissions (role_id);
122
+
123
+ create index if not exists permissions_action_resource_idx
124
+ on public.permissions (action, resource);
125
+ ```
126
+
127
+ ### Policy hierárquica composta — exemplo `leads`
128
+
129
+ ```sql
130
+ -- Tabela leads multi-tenant
131
+ create table public.leads (
132
+ id uuid primary key default gen_random_uuid(),
133
+ org_id uuid not null references public.organizations(id) on delete cascade,
134
+ dept_id uuid references public.departments(id) on delete set null,
135
+ contact_name text not null,
136
+ contact_email text,
137
+ contact_phone text,
138
+ stage text not null default 'lead' check (stage in ('lead','qualified','proposal','negotiation','won','lost')),
139
+ owner_id uuid references auth.users(id) on delete set null,
140
+ created_at timestamptz not null default now(),
141
+ unique (org_id, contact_email),
142
+ unique (org_id, contact_phone)
143
+ );
144
+
145
+ alter table public.leads enable row level security;
146
+
147
+ -- POLICY 1: SELECT — member da org pode ler todos leads da org
148
+ create policy "leads_select_member"
149
+ on public.leads
150
+ for select
151
+ to authenticated
152
+ using (private.is_member_of(org_id));
153
+
154
+ -- POLICY 2: INSERT — member com permission leads:create
155
+ create policy "leads_insert_with_permission"
156
+ on public.leads
157
+ for insert
158
+ to authenticated
159
+ with check (
160
+ private.has_permission('create', 'leads', org_id)
161
+ );
162
+
163
+ -- POLICY 3: UPDATE — member com permission leads:update OU é owner do lead
164
+ create policy "leads_update_with_permission_or_owner"
165
+ on public.leads
166
+ for update
167
+ to authenticated
168
+ using (
169
+ private.has_permission('update', 'leads', org_id)
170
+ or owner_id = (select auth.uid())
171
+ )
172
+ with check (
173
+ private.has_permission('update', 'leads', org_id)
174
+ or owner_id = (select auth.uid())
175
+ );
176
+
177
+ -- POLICY 4: DELETE — apenas admin/owner role
178
+ create policy "leads_delete_admin_owner"
179
+ on public.leads
180
+ for delete
181
+ to authenticated
182
+ using (
183
+ private.has_role(org_id, 'admin') or private.has_role(org_id, 'owner')
184
+ );
185
+
186
+ -- POLICY 5 (PERMISSIVE — REGRA #4): super_admin bypass para todas operações
187
+ create policy "leads_super_admin_bypass"
188
+ on public.leads
189
+ as permissive -- combinação OR com policies normais
190
+ for all -- super_admin pode tudo
191
+ to authenticated
192
+ using (private.is_super_admin())
193
+ with check (private.is_super_admin());
194
+
195
+ -- Index obrigatório nas colunas filtradas
196
+ create index leads_org_dept_idx on public.leads (org_id, dept_id);
197
+ create index leads_owner_idx on public.leads (owner_id) where owner_id is not null;
198
+ ```
199
+
200
+ ### Herança dept→org — função `effective_role_in_dept`
201
+
202
+ ```sql
203
+ -- 5. effective_role_in_dept — retorna role do user no contexto do dept
204
+ -- (NULL em department_members.role_id = herda do organization_members)
205
+ create or replace function private.effective_role_in_dept(p_dept_id uuid, p_user_id uuid)
206
+ returns uuid
207
+ language sql
208
+ stable
209
+ security invoker
210
+ set search_path = ''
211
+ as $$
212
+ select coalesce(dm.role_id, om.role_id)
213
+ from public.departments d
214
+ join public.organization_members om on om.org_id = d.org_id and om.user_id = p_user_id
215
+ left join public.department_members dm on dm.dept_id = p_dept_id and dm.user_id = p_user_id
216
+ where d.id = p_dept_id;
217
+ $$;
218
+
219
+ -- Helper: has_role no contexto de um dept (resolve herança)
220
+ create or replace function private.has_role_in_dept(p_dept_id uuid, p_role_name text)
221
+ returns boolean
222
+ language sql
223
+ stable
224
+ security invoker
225
+ set search_path = ''
226
+ as $$
227
+ select exists (
228
+ select 1 from public.roles r
229
+ where r.id = private.effective_role_in_dept(p_dept_id, (select auth.uid()))
230
+ and r.name = p_role_name
231
+ );
232
+ $$;
233
+ ```
234
+
235
+ ### Validar isolation via query — useful em testing
236
+
237
+ ```sql
238
+ -- Listar tabelas com `org_id` mas SEM RLS habilitada (red flag)
239
+ select c.relname
240
+ from pg_class c
241
+ join pg_attribute a on a.attrelid = c.oid
242
+ where a.attname = 'org_id'
243
+ and c.relkind = 'r'
244
+ and c.relrowsecurity = false;
245
+
246
+ -- Listar policies que referenciam helper functions canônicas
247
+ select policyname, tablename, qual
248
+ from pg_policies
249
+ where qual like '%private.is_member_of%'
250
+ or qual like '%private.has_permission%'
251
+ or qual like '%private.is_super_admin%';
252
+ ```
253
+
254
+ ## Anti-patterns
255
+
256
+ ### Anti-pattern 1: Helper functions em `public` (expostos via PostgREST)
257
+
258
+ **Errado:**
259
+ ```sql
260
+ create function public.is_member_of(p_org_id uuid) returns boolean ...
261
+ ```
262
+
263
+ **Por quê:** PostgREST expõe automaticamente `/rpc/is_member_of?p_org_id=...`. Endpoint vira público acessível, atacante pode probe quem é member de qual org.
264
+
265
+ **Certo:** schema `private` (PostgREST ignora por default).
266
+
267
+ ### Anti-pattern 2: super_admin bypass via OR na policy normal
268
+
269
+ **Errado:**
270
+ ```sql
271
+ create policy "leads_select" on public.leads
272
+ for select
273
+ to authenticated
274
+ using (
275
+ private.is_member_of(org_id)
276
+ or private.is_super_admin() -- bypass embutido
277
+ );
278
+ ```
279
+
280
+ **Por quê:** funciona mas mistura semânticas. Mais difícil de auditar (qual é a "policy normal" vs "bypass"?). Mistura severidade — quando você desativar super_admin para teste, precisa editar todas as policies.
281
+
282
+ **Certo:** policy `as permissive` separada para super_admin (REGRA #4). PostgreSQL faz OR entre policies PERMISSIVE.
283
+
284
+ ### Anti-pattern 3: Helper function VOLATILE (default)
285
+
286
+ **Errado:**
287
+ ```sql
288
+ create function private.is_member_of(p_org_id uuid)
289
+ returns boolean
290
+ language sql
291
+ -- sem STABLE — default VOLATILE
292
+ as $$ ... $$;
293
+ ```
294
+
295
+ **Por quê:** ver [`multi-tenant-performance-scaling`](../multi-tenant-performance-scaling/SKILL.md) Anti-pattern 1. Re-execução por linha = degradação 200×.
296
+
297
+ **Certo:** marcar `STABLE` (REGRA #2).
298
+
299
+ ### Anti-pattern 4: department_members sem coalesce — herança quebrada
300
+
301
+ **Errado:**
302
+ ```sql
303
+ -- Policy lê role direto de department_members, ignorando NULL
304
+ using (
305
+ exists (
306
+ select 1 from public.department_members dm
307
+ join public.roles r on r.id = dm.role_id
308
+ where dm.user_id = (select auth.uid()) and r.name = 'admin'
309
+ )
310
+ )
311
+ ```
312
+
313
+ **Por quê:** se `dm.role_id IS NULL`, JOIN não casa, role efetiva não é resolvida → user adicionado ao dept sem role explícita não tem permissão (deveria herdar do org_members).
314
+
315
+ **Certo:** usar `private.effective_role_in_dept` que faz coalesce.
316
+
317
+ ### Anti-pattern 5: super_admin sem audit log
318
+
319
+ **Errado:**
320
+ ```sql
321
+ -- super_admin policy permite ler/modificar tudo sem registrar quem foi
322
+ create policy "super_admin_bypass" on public.leads as permissive for all to authenticated using (private.is_super_admin()) with check (private.is_super_admin());
323
+ -- Mas... onde está o audit?
324
+ ```
325
+
326
+ **Por quê:** super_admin sem audit = ninguém consegue investigar incident "quem deletou todos os leads da org X em 03/04?". Compliance LGPD exige audit de acesso a dados.
327
+
328
+ **Certo:** policy super_admin OK, mas **toda operação super_admin** deve emitir evento `super_admin_action` em `audit_log` (Phase 109). Trigger AFTER INSERT/UPDATE/DELETE em tabelas críticas que checa `private.is_super_admin()` e registra.
329
+
330
+ ## Invariantes Linearizáveis Cross-Tenant (v1.22+)
331
+
332
+ > Para uniqueness constraints cross-org (slug global, license key) e padrões `SELECT FOR UPDATE` em writes cross-tenant, ver skill [`escolha-modelo-consistencia`](../escolha-modelo-consistencia/SKILL.md) (v1.22 — DDIA Ch 9). Resumo: deixe `UNIQUE` constraint Postgres disparar via `INSERT ... ON CONFLICT DO NOTHING RETURNING` em vez de UPDATE+SELECT em nível de app (race window).
333
+
334
+ ## Ver também
335
+
336
+ - [supabase-rls-policies](../supabase-rls-policies/SKILL.md) — anti-patterns base v1.8 herdados (REGRA #6)
337
+ - [b2b-saas-architecture](../b2b-saas-architecture/SKILL.md) — schema canônico que esta skill cobre com RLS
338
+ - [multi-tenant-performance-scaling](../multi-tenant-performance-scaling/SKILL.md) — STABLE marker + partial indexes (REGRA #2)
339
+ - [rbac-permissions-matrix-supabase](../rbac-permissions-matrix-supabase/SKILL.md) — modelagem permissions consumed por `private.has_permission`
340
+ - [super-admin-platform-pattern](../super-admin-platform-pattern/SKILL.md) — Phase 111, super_admin operations
341
+ - [audit-log-multi-tenant](../audit-log-multi-tenant/SKILL.md) — Phase 109, audit `super_admin_action` (Anti-pattern 5)
342
+ - [_shared-multi-tenant/glossary.md](../_shared-multi-tenant/glossary.md) — termos `RBAC`, `permission matrix`, `role escalation rule`