@luanpdd/kit-mcp 1.34.0 → 1.36.0

package/kit/skills/distributed-tracing/SKILL.md

@@ -1,362 +1,362 @@
----
-name: distributed-tracing
-description: Use ao instrumentar tracing — trace_id/span_id/parent_id, propagar W3C TraceContext via header traceparent, stitching além de RPCs (batch, lambda, queue).
----
-
-# Observabilidade — Distributed Tracing
-
-## Quando usar
-
-LLM carrega esta skill ao instrumentar tracing distribuído ou stitching de spans. Trigger phrases:
-
-- "distributed tracing", "traces", "spans"
-- "propagar contexto entre serviços", "trace cross-service"
-- "W3C TraceContext", "traceparent header"
-- "trace_id span_id parent_span_id"
-- "ligar lambda batch job ao trace"
-- "stitching de eventos"
-
-## Regras absolutas
-
-- **trace_id é compartilhado** entre todos os spans de um único request distribuído. **NÃO** mude por hop.
-- **span_id é único por span** — gere novo a cada `startSpan()`. 16 hex chars (8 bytes).
-- **parent_span_id aponta para span pai** — null no root span. Define a árvore.
-- **W3C TraceContext é o padrão** — header HTTP `traceparent: 00-{trace_id}-{span_id}-{flags}`. Adote sempre. B3 é fallback para legacy.
-- **Propague ANTES de fazer call cross-service** — extrair contexto do request inbound, propagar no request outbound. Sem isso, trace quebra.
-- **Stitching ≠ apenas RPC** — também batch jobs, queue messages, lambda invocations, S3 uploads. Carregue `traceparent` em metadata da queue, env var do lambda, header da Step Function.
-- **Sample decision propaga** — bit `01` em flags de `traceparent` significa "sample=true". Decisão tomada no head propaga downstream.
-- **Não invente trace_id** — sempre derive do contexto inbound ou gere via SDK (não `crypto.randomUUID()`).
-- **Spans devem ter `kind`** — `SERVER` (handler de inbound), `CLIENT` (call outbound), `PRODUCER`/`CONSUMER` (queue), `INTERNAL` (subspan dentro do mesmo process).
-
-## Patterns canônicos
-
-### Pattern: extrair contexto inbound + propagar outbound (Node)
-
-```ts
-// PT-BR: handler HTTP — extrai traceparent do request inbound, propaga em call outbound
-import { trace, context, propagation } from '@opentelemetry/api'
-
-const tracer = trace.getTracer('orders-service')
-
-export async function placeOrder(req: Request) {
-  // PT-BR: 1 — extrair contexto inbound do header traceparent
-  const inboundContext = propagation.extract(context.active(), req.headers)
-
-  return tracer.startActiveSpan(
-    'place_order',
-    { kind: SpanKind.SERVER },
-    inboundContext,
-    async (span) => {
-      span.setAttribute('user.id', req.user.id)
-
-      // PT-BR: 2 — fazer call outbound — propagation injeta traceparent automaticamente
-      //         se você usar fetch/grpc instrumentados (ver skill opentelemetry-standard)
-      const outboundHeaders: Record<string, string> = {}
-      propagation.inject(context.active(), outboundHeaders)
-
-      const inventoryRes = await fetch('http://inventory/check', {
-        headers: outboundHeaders,  // PT-BR: traceparent injetado aqui
-        body: JSON.stringify({ items: req.items })
-      })
-
-      span.end()
-      return inventoryRes.json()
-    }
-  )
-}
-```
-
-### Pattern: traceparent format
-
-```text
-traceparent: 00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01
-             ^  ^                                ^                 ^
-             |  |                                |                 |
-             version                             |                 flags (sampled bit)
-                trace_id (32 hex / 16 bytes)     |
-                                                 span_id (16 hex / 8 bytes)
-```
-
-```text
-flags:
-  01 = sampled    (decisão upstream: capture este trace)
-  00 = not sampled (decisão upstream: skip)
-```
-
-### Pattern: trace cross-service via Supabase Edge Function
-
-```ts
-// PT-BR: Edge Function recebe request → propaga para outro service
-import { trace, context, propagation } from 'npm:@opentelemetry/api@1.9.0'
-import { W3CTraceContextPropagator } from 'npm:@opentelemetry/core@1.27.0'
-
-propagation.setGlobalPropagator(new W3CTraceContextPropagator())
-
-const tracer = trace.getTracer('edge-orders')
-
-Deno.serve(async (req) => {
-  // PT-BR: extrair traceparent inbound
-  const inboundCtx = propagation.extract(context.active(), {
-    traceparent: req.headers.get('traceparent') ?? '',
-  })
-
-  return tracer.startActiveSpan(
-    'edge_handler',
-    { kind: 1 /* SERVER */ },
-    inboundCtx,
-    async (span) => {
-      span.setAttribute('endpoint', new URL(req.url).pathname)
-
-      // PT-BR: call outbound para Postgres via PostgREST — injeta traceparent
-      const outHeaders: Record<string, string> = {}
-      propagation.inject(context.active(), outHeaders)
-
-      const dbRes = await fetch(Deno.env.get('SUPABASE_URL') + '/rest/v1/orders', {
-        method: 'POST',
-        headers: {
-          ...outHeaders,
-          'apikey': Deno.env.get('SUPABASE_ANON_KEY')!,
-          'content-type': 'application/json',
-        },
-        body: await req.text(),
-      })
-
-      span.setAttribute('db.status_code', dbRes.status)
-      span.end()
-      return dbRes
-    }
-  )
-})
-```
-
-### Pattern: stitching além de RPC — queue message (não-RPC)
-
-```ts
-// PT-BR: producer — anexa traceparent ao payload da queue (pgmq, SQS, RabbitMQ)
-import { trace, context, propagation } from '@opentelemetry/api'
-
-const tracer = trace.getTracer('producer')
-
-export async function enqueueEmail(emailJob: EmailJob) {
-  return tracer.startActiveSpan(
-    'enqueue_email',
-    { kind: SpanKind.PRODUCER },
-    async (span) => {
-      span.setAttribute('queue.name', 'emails')
-      span.setAttribute('email.recipient', emailJob.to)
-
-      // PT-BR: serializar contexto no payload da mensagem
-      const carrier: Record<string, string> = {}
-      propagation.inject(context.active(), carrier)
-
-      await pgmqEnqueue('emails', {
-        ...emailJob,
-        _trace_context: carrier,  // PT-BR: viaja com o job
-      })
-
-      span.end()
-    }
-  )
-}
-
-// PT-BR: consumer — extrai traceparent do payload, continua o trace
-export async function processEmailJob(job: EmailJobWithContext) {
-  const inboundCtx = propagation.extract(
-    context.active(),
-    job._trace_context ?? {}  // PT-BR: se vazio, novo trace
-  )
-
-  return tracer.startActiveSpan(
-    'process_email',
-    { kind: SpanKind.CONSUMER },
-    inboundCtx,
-    async (span) => {
-      span.setAttribute('email.recipient', job.to)
-      // PT-BR: agora o span do worker faz parte do mesmo trace do producer
-      await sendEmail(job)
-      span.end()
-    }
-  )
-}
-```
-
-### Pattern: stitching de batch job (não-RPC)
-
-```ts
-// PT-BR: cron job processa N items — 1 span por item, todos com mesmo trace_id
-const tracer = trace.getTracer('billing-cron')
-
-export async function dailyBillingJob() {
-  return tracer.startActiveSpan('daily_billing', async (rootSpan) => {
-    rootSpan.setAttribute('job.type', 'cron')
-    rootSpan.setAttribute('build_id', BUILD_ID)
-
-    const customers = await db.getCustomersDueForBilling()
-    rootSpan.setAttribute('customers.count', customers.length)
-
-    // PT-BR: cada customer vira span filho com mesmo trace_id
-    for (const customer of customers) {
-      await tracer.startActiveSpan(
-        'bill_customer',
-        { kind: SpanKind.INTERNAL },
-        async (span) => {
-          span.setAttribute('customer.id', customer.id)
-          span.setAttribute('customer.tier', customer.tier)
-          try {
-            await chargeCustomer(customer)
-            span.setAttribute('result.success', true)
-          } catch (e) {
-            span.setAttribute('result.success', false)
-            span.setAttribute('error.type', classify(e))
-          } finally {
-            span.end()
-          }
-        }
-      )
-    }
-
-    rootSpan.end()
-  })
-}
-```
-
-### Pattern: span kinds
-
-| Kind | Quando usar | Exemplo |
-|---|---|---|
-| `SERVER` | Recebendo request inbound | Handler HTTP, gRPC server method |
-| `CLIENT` | Fazendo call outbound | `fetch()`, gRPC client call, DB query |
-| `PRODUCER` | Enviando msg para queue | `pgmq.enqueue()`, SQS publish |
-| `CONSUMER` | Processando msg de queue | Worker recebendo job |
-| `INTERNAL` | Subdivisão dentro do mesmo process | "json_parse", "validation_step" |
-
-### Pattern: query traces — montar waterfall
-
-```sql
--- PT-BR: pegar todos os spans de um trace em ordem cronológica
-select
-  span_id,
-  parent_span_id,
-  span_name,
-  span_kind,
-  service_name,
-  duration_ms,
-  start_time
-from observability.spans
-where trace_id = '4bf92f3577b34da6a3ce929d0e0e4736'
-order by start_time asc;
-
--- PT-BR: encontrar root span — parent_span_id IS NULL ou span sem parent no mesmo trace
-select * 
-from observability.spans
-where trace_id = '4bf92f3577b34da6a3ce929d0e0e4736'
-  and parent_span_id is null;
-
--- PT-BR: spans mais lentos cross-trace, último 1h
-select 
-  service_name, 
-  span_name,
-  percentile_cont(0.99) within group (order by duration_ms) as p99,
-  count(*) as samples
-from observability.spans
-where start_time > now() - interval '1 hour'
-group by service_name, span_name
-having count(*) > 100
-order by p99 desc
-limit 20;
-```
-
-## Anti-patterns
-
-### ANTI: gerar trace_id por hop
-
-```ts
-// PT-BR: BAD — quebra a cadeia, cada service vê trace diferente
-const traceId = crypto.randomUUID().replace(/-/g, '').slice(0, 32)
-
-// PT-BR: GOOD — extrair do header inbound; deixar SDK gerar root
-const inboundCtx = propagation.extract(context.active(), req.headers)
-tracer.startActiveSpan('handler', {}, inboundCtx, ...)
-```
-
-### ANTI: esquecer de propagar em call outbound
-
-```ts
-// PT-BR: BAD — outbound call sem traceparent — trace quebra no service B
-await fetch('http://service-b/api', { body: ... })
-
-// PT-BR: GOOD — injetar traceparent
-const headers: Record<string, string> = {}
-propagation.inject(context.active(), headers)
-await fetch('http://service-b/api', { headers, body: ... })
-```
-
-### ANTI: trace só de RPCs, não de batch/queue
-
-```ts
-// PT-BR: BAD — producer/consumer não compartilham trace, debug fica fragmentado
-await pgmqEnqueue('emails', payload)  // sem trace context
-// ... depois worker processa sem saber que veio do request X
-
-// PT-BR: GOOD — propagar contexto via metadata da queue
-const carrier = {}
-propagation.inject(context.active(), carrier)
-await pgmqEnqueue('emails', { ...payload, _trace_context: carrier })
-```
-
-### ANTI: span sem `end()`
-
-```ts
-// PT-BR: BAD — span fica aberto forever, duration_ms não calculado, memory leak
-const span = tracer.startSpan('handler')
-// ... handler logic
-return result  // PT-BR: ESQUECEU span.end()
-
-// PT-BR: GOOD — sempre `try/finally`
-const span = tracer.startSpan('handler')
-try {
-  // ... logic
-} finally {
-  span.end()
-}
-```
-
-### ANTI: span hierarchy errada
-
-```ts
-// PT-BR: BAD — usar startSpan sem startActiveSpan, parent não é settado automático
-const parent = tracer.startSpan('parent')
-const child = tracer.startSpan('child')  // PT-BR: parent_span_id ficou null
-parent.end()
-child.end()
-
-// PT-BR: GOOD — startActiveSpan empurra contexto, child herda parent
-tracer.startActiveSpan('parent', (parent) => {
-  tracer.startActiveSpan('child', (child) => {
-    // PT-BR: child.parent_span_id === parent.span_id
-    child.end()
-  })
-  parent.end()
-})
-```
-
-## Verificação
-
-1. **1 trace_id por request** — enviar 1 request, queryar `SELECT DISTINCT trace_id FROM spans WHERE request_id = X` → 1 resultado.
-2. **Cross-service stitching** — request HTTP service A → service B → DB. Queryar `SELECT count(distinct service_name) FROM spans WHERE trace_id = X` → ≥ 3.
-3. **Root span identificável** — `SELECT * FROM spans WHERE trace_id = X AND parent_span_id IS NULL` → 1 row (o root).
-4. **Span hierarchy correta** — graficar via tool (Jaeger UI, Honeycomb, etc.) ou recursivo SQL — deve formar árvore válida (sem ciclos).
-5. **Duration não-zero** — `SELECT min(duration_ms), max(duration_ms) FROM spans` — min ≥ 0, max razoável.
-6. **Sampled flag respeitado** — verificar que se traceparent inbound = `01`, downstream também sample=true.
-7. **Queue stitching funciona** — enqueue + consume → mesmo `trace_id` em ambos os spans.
-
----
-
-## Ver também
-
-- `kit/skills/_shared-observability/glossary.md` — W3C TraceContext, B3, span kinds
-- `kit/skills/structured-events/SKILL.md` — atributos canônicos por span
-- `kit/skills/opentelemetry-standard/SKILL.md` — SDK que faz extract/inject
-- `kit/skills/telemetry-sampling/SKILL.md` *(Phase 34)* — head vs tail sampling decisão
-
-*Material-fonte: Observability Engineering (O'Reilly, 2022) — Cap 6: "Stitching Events into Traces".*
+---
+name: distributed-tracing
+description: Use ao instrumentar tracing — trace_id/span_id/parent_id, propagar W3C TraceContext via header traceparent, stitching além de RPCs (batch, lambda, queue).
+---
+
+# Observabilidade — Distributed Tracing
+
+## Quando usar
+
+LLM carrega esta skill ao instrumentar tracing distribuído ou stitching de spans. Trigger phrases:
+
+- "distributed tracing", "traces", "spans"
+- "propagar contexto entre serviços", "trace cross-service"
+- "W3C TraceContext", "traceparent header"
+- "trace_id span_id parent_span_id"
+- "ligar lambda batch job ao trace"
+- "stitching de eventos"
+
+## Regras absolutas
+
+- **trace_id é compartilhado** entre todos os spans de um único request distribuído. **NÃO** mude por hop.
+- **span_id é único por span** — gere novo a cada `startSpan()`. 16 hex chars (8 bytes).
+- **parent_span_id aponta para span pai** — null no root span. Define a árvore.
+- **W3C TraceContext é o padrão** — header HTTP `traceparent: 00-{trace_id}-{span_id}-{flags}`. Adote sempre. B3 é fallback para legacy.
+- **Propague ANTES de fazer call cross-service** — extrair contexto do request inbound, propagar no request outbound. Sem isso, trace quebra.
+- **Stitching ≠ apenas RPC** — também batch jobs, queue messages, lambda invocations, S3 uploads. Carregue `traceparent` em metadata da queue, env var do lambda, header da Step Function.
+- **Sample decision propaga** — bit `01` em flags de `traceparent` significa "sample=true". Decisão tomada no head propaga downstream.
+- **Não invente trace_id** — sempre derive do contexto inbound ou gere via SDK (não `crypto.randomUUID()`).
+- **Spans devem ter `kind`** — `SERVER` (handler de inbound), `CLIENT` (call outbound), `PRODUCER`/`CONSUMER` (queue), `INTERNAL` (subspan dentro do mesmo process).
+
+## Patterns canônicos
+
+### Pattern: extrair contexto inbound + propagar outbound (Node)
+
+```ts
+// PT-BR: handler HTTP — extrai traceparent do request inbound, propaga em call outbound
+import { trace, context, propagation } from '@opentelemetry/api'
+
+const tracer = trace.getTracer('orders-service')
+
+export async function placeOrder(req: Request) {
+  // PT-BR: 1 — extrair contexto inbound do header traceparent
+  const inboundContext = propagation.extract(context.active(), req.headers)
+
+  return tracer.startActiveSpan(
+    'place_order',
+    { kind: SpanKind.SERVER },
+    inboundContext,
+    async (span) => {
+      span.setAttribute('user.id', req.user.id)
+
+      // PT-BR: 2 — fazer call outbound — propagation injeta traceparent automaticamente
+      //         se você usar fetch/grpc instrumentados (ver skill opentelemetry-standard)
+      const outboundHeaders: Record<string, string> = {}
+      propagation.inject(context.active(), outboundHeaders)
+
+      const inventoryRes = await fetch('http://inventory/check', {
+        headers: outboundHeaders,  // PT-BR: traceparent injetado aqui
+        body: JSON.stringify({ items: req.items })
+      })
+
+      span.end()
+      return inventoryRes.json()
+    }
+  )
+}
+```
+
+### Pattern: traceparent format
+
+```text
+traceparent: 00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01
+             ^  ^                                ^                 ^
+             |  |                                |                 |
+             version                             |                 flags (sampled bit)
+                trace_id (32 hex / 16 bytes)     |
+                                                 span_id (16 hex / 8 bytes)
+```
+
+```text
+flags:
+  01 = sampled    (decisão upstream: capture este trace)
+  00 = not sampled (decisão upstream: skip)
+```
+
+### Pattern: trace cross-service via Supabase Edge Function
+
+```ts
+// PT-BR: Edge Function recebe request → propaga para outro service
+import { trace, context, propagation } from 'npm:@opentelemetry/api@1.9.0'
+import { W3CTraceContextPropagator } from 'npm:@opentelemetry/core@1.27.0'
+
+propagation.setGlobalPropagator(new W3CTraceContextPropagator())
+
+const tracer = trace.getTracer('edge-orders')
+
+Deno.serve(async (req) => {
+  // PT-BR: extrair traceparent inbound
+  const inboundCtx = propagation.extract(context.active(), {
+    traceparent: req.headers.get('traceparent') ?? '',
+  })
+
+  return tracer.startActiveSpan(
+    'edge_handler',
+    { kind: 1 /* SERVER */ },
+    inboundCtx,
+    async (span) => {
+      span.setAttribute('endpoint', new URL(req.url).pathname)
+
+      // PT-BR: call outbound para Postgres via PostgREST — injeta traceparent
+      const outHeaders: Record<string, string> = {}
+      propagation.inject(context.active(), outHeaders)
+
+      const dbRes = await fetch(Deno.env.get('SUPABASE_URL') + '/rest/v1/orders', {
+        method: 'POST',
+        headers: {
+          ...outHeaders,
+          'apikey': Deno.env.get('SUPABASE_ANON_KEY')!,
+          'content-type': 'application/json',
+        },
+        body: await req.text(),
+      })
+
+      span.setAttribute('db.status_code', dbRes.status)
+      span.end()
+      return dbRes
+    }
+  )
+})
+```
+
+### Pattern: stitching além de RPC — queue message (não-RPC)
+
+```ts
+// PT-BR: producer — anexa traceparent ao payload da queue (pgmq, SQS, RabbitMQ)
+import { trace, context, propagation } from '@opentelemetry/api'
+
+const tracer = trace.getTracer('producer')
+
+export async function enqueueEmail(emailJob: EmailJob) {
+  return tracer.startActiveSpan(
+    'enqueue_email',
+    { kind: SpanKind.PRODUCER },
+    async (span) => {
+      span.setAttribute('queue.name', 'emails')
+      span.setAttribute('email.recipient', emailJob.to)
+
+      // PT-BR: serializar contexto no payload da mensagem
+      const carrier: Record<string, string> = {}
+      propagation.inject(context.active(), carrier)
+
+      await pgmqEnqueue('emails', {
+        ...emailJob,
+        _trace_context: carrier,  // PT-BR: viaja com o job
+      })
+
+      span.end()
+    }
+  )
+}
+
+// PT-BR: consumer — extrai traceparent do payload, continua o trace
+export async function processEmailJob(job: EmailJobWithContext) {
+  const inboundCtx = propagation.extract(
+    context.active(),
+    job._trace_context ?? {}  // PT-BR: se vazio, novo trace
+  )
+
+  return tracer.startActiveSpan(
+    'process_email',
+    { kind: SpanKind.CONSUMER },
+    inboundCtx,
+    async (span) => {
+      span.setAttribute('email.recipient', job.to)
+      // PT-BR: agora o span do worker faz parte do mesmo trace do producer
+      await sendEmail(job)
+      span.end()
+    }
+  )
+}
+```
+
+### Pattern: stitching de batch job (não-RPC)
+
+```ts
+// PT-BR: cron job processa N items — 1 span por item, todos com mesmo trace_id
+const tracer = trace.getTracer('billing-cron')
+
+export async function dailyBillingJob() {
+  return tracer.startActiveSpan('daily_billing', async (rootSpan) => {
+    rootSpan.setAttribute('job.type', 'cron')
+    rootSpan.setAttribute('build_id', BUILD_ID)
+
+    const customers = await db.getCustomersDueForBilling()
+    rootSpan.setAttribute('customers.count', customers.length)
+
+    // PT-BR: cada customer vira span filho com mesmo trace_id
+    for (const customer of customers) {
+      await tracer.startActiveSpan(
+        'bill_customer',
+        { kind: SpanKind.INTERNAL },
+        async (span) => {
+          span.setAttribute('customer.id', customer.id)
+          span.setAttribute('customer.tier', customer.tier)
+          try {
+            await chargeCustomer(customer)
+            span.setAttribute('result.success', true)
+          } catch (e) {
+            span.setAttribute('result.success', false)
+            span.setAttribute('error.type', classify(e))
+          } finally {
+            span.end()
+          }
+        }
+      )
+    }
+
+    rootSpan.end()
+  })
+}
+```
+
+### Pattern: span kinds
+
+| Kind | Quando usar | Exemplo |
+|---|---|---|
+| `SERVER` | Recebendo request inbound | Handler HTTP, gRPC server method |
+| `CLIENT` | Fazendo call outbound | `fetch()`, gRPC client call, DB query |
+| `PRODUCER` | Enviando msg para queue | `pgmq.enqueue()`, SQS publish |
+| `CONSUMER` | Processando msg de queue | Worker recebendo job |
+| `INTERNAL` | Subdivisão dentro do mesmo process | "json_parse", "validation_step" |
+
+### Pattern: query traces — montar waterfall
+
+```sql
+-- PT-BR: pegar todos os spans de um trace em ordem cronológica
+select
+  span_id,
+  parent_span_id,
+  span_name,
+  span_kind,
+  service_name,
+  duration_ms,
+  start_time
+from observability.spans
+where trace_id = '4bf92f3577b34da6a3ce929d0e0e4736'
+order by start_time asc;
+
+-- PT-BR: encontrar root span — parent_span_id IS NULL ou span sem parent no mesmo trace
+select * 
+from observability.spans
+where trace_id = '4bf92f3577b34da6a3ce929d0e0e4736'
+  and parent_span_id is null;
+
+-- PT-BR: spans mais lentos cross-trace, último 1h
+select 
+  service_name, 
+  span_name,
+  percentile_cont(0.99) within group (order by duration_ms) as p99,
+  count(*) as samples
+from observability.spans
+where start_time > now() - interval '1 hour'
+group by service_name, span_name
+having count(*) > 100
+order by p99 desc
+limit 20;
+```
+
+## Anti-patterns
+
+### ANTI: gerar trace_id por hop
+
+```ts
+// PT-BR: BAD — quebra a cadeia, cada service vê trace diferente
+const traceId = crypto.randomUUID().replace(/-/g, '').slice(0, 32)
+
+// PT-BR: GOOD — extrair do header inbound; deixar SDK gerar root
+const inboundCtx = propagation.extract(context.active(), req.headers)
+tracer.startActiveSpan('handler', {}, inboundCtx, ...)
+```
+
+### ANTI: esquecer de propagar em call outbound
+
+```ts
+// PT-BR: BAD — outbound call sem traceparent — trace quebra no service B
+await fetch('http://service-b/api', { body: ... })
+
+// PT-BR: GOOD — injetar traceparent
+const headers: Record<string, string> = {}
+propagation.inject(context.active(), headers)
+await fetch('http://service-b/api', { headers, body: ... })
+```
+
+### ANTI: trace só de RPCs, não de batch/queue
+
+```ts
+// PT-BR: BAD — producer/consumer não compartilham trace, debug fica fragmentado
+await pgmqEnqueue('emails', payload)  // sem trace context
+// ... depois worker processa sem saber que veio do request X
+
+// PT-BR: GOOD — propagar contexto via metadata da queue
+const carrier = {}
+propagation.inject(context.active(), carrier)
+await pgmqEnqueue('emails', { ...payload, _trace_context: carrier })
+```
+
+### ANTI: span sem `end()`
+
+```ts
+// PT-BR: BAD — span fica aberto forever, duration_ms não calculado, memory leak
+const span = tracer.startSpan('handler')
+// ... handler logic
+return result  // PT-BR: ESQUECEU span.end()
+
+// PT-BR: GOOD — sempre `try/finally`
+const span = tracer.startSpan('handler')
+try {
+  // ... logic
+} finally {
+  span.end()
+}
+```
+
+### ANTI: span hierarchy errada
+
+```ts
+// PT-BR: BAD — usar startSpan sem startActiveSpan, parent não é settado automático
+const parent = tracer.startSpan('parent')
+const child = tracer.startSpan('child')  // PT-BR: parent_span_id ficou null
+parent.end()
+child.end()
+
+// PT-BR: GOOD — startActiveSpan empurra contexto, child herda parent
+tracer.startActiveSpan('parent', (parent) => {
+  tracer.startActiveSpan('child', (child) => {
+    // PT-BR: child.parent_span_id === parent.span_id
+    child.end()
+  })
+  parent.end()
+})
+```
+
+## Verificação
+
+1. **1 trace_id por request** — enviar 1 request, queryar `SELECT DISTINCT trace_id FROM spans WHERE request_id = X` → 1 resultado.
+2. **Cross-service stitching** — request HTTP service A → service B → DB. Queryar `SELECT count(distinct service_name) FROM spans WHERE trace_id = X` → ≥ 3.
+3. **Root span identificável** — `SELECT * FROM spans WHERE trace_id = X AND parent_span_id IS NULL` → 1 row (o root).
+4. **Span hierarchy correta** — graficar via tool (Jaeger UI, Honeycomb, etc.) ou recursivo SQL — deve formar árvore válida (sem ciclos).
+5. **Duration não-zero** — `SELECT min(duration_ms), max(duration_ms) FROM spans` — min ≥ 0, max razoável.
+6. **Sampled flag respeitado** — verificar que se traceparent inbound = `01`, downstream também sample=true.
+7. **Queue stitching funciona** — enqueue + consume → mesmo `trace_id` em ambos os spans.
+
+---
+
+## Ver também
+
+- `kit/skills/_shared-observability/glossary.md` — W3C TraceContext, B3, span kinds
+- `kit/skills/structured-events/SKILL.md` — atributos canônicos por span
+- `kit/skills/opentelemetry-standard/SKILL.md` — SDK que faz extract/inject
+- `kit/skills/telemetry-sampling/SKILL.md` *(Phase 34)* — head vs tail sampling decisão
+
+*Material-fonte: Observability Engineering (O'Reilly, 2022) — Cap 6: "Stitching Events into Traces".*