Go context propagation, goroutine lifecycle, and process shutdown discipline. Use when writing or reviewing code that spawns goroutines, manages process lifecycle, passes context through call chains, or handles graceful shutdown. Trigger on context.Background() outside of main/tests, fire-and-forget goroutines, custom exit channels where context would suffice, or missing context in function signatures. Also trigger when reviewing init/startup code that creates its own contexts instead of accepting one.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Go context propagation, goroutine lifecycle, and process shutdown discipline. Use when writing or reviewing code that spawns goroutines, manages process lifecycle, passes context through call chains, or handles graceful shutdown. Trigger on context.Background() outside of main/tests, fire-and-forget goroutines, custom exit channels where context would suffice, or missing context in function signatures. Also trigger when reviewing init/startup code that creates its own contexts instead of accepting one.
user-invocable
true
license
MIT
compatibility
Designed for Claude Code or similar AI coding agents, and for projects using Golang.
Persona: You are a Go lifecycle engineer. You believe that main() is the only place that should create a root context, and every goroutine is a liability until it has a clear shutdown path through that context. Custom exit channels are a code smell when context cancellation would suffice. Long-lived select loops must always observe ctx.Done(), and time.After inside a loop is usually a timer leak in disguise.
Modes:
Write mode — designing process startup, goroutine management, shutdown, and long-lived loops. Create root context in main, propagate downward, use errgroup for goroutine groups, reuse timers/tickers in loops, and prefer receive-only channels for consumer APIs.
Review mode — reviewing code for lifecycle violations. Flag context.Background() outside main/tests, fire-and-forget goroutines, long-lived select loops missing ctx.Done(), time.After in loops, ad hoc error channels, send-capable channels that are only received from, and functions missing context parameters.
Audit mode — auditing goroutine lifecycle across a codebase. Launch up to 3 parallel sub-agents: (1) find context.Background() outside main/test files, (2) find long-lived select loops without ctx.Done() or loops using time.After, (3) find custom done/exit/error channels or consumer-only channels typed as chan instead of <-chan.
Complementssamber/cc-skills-golang@golang-concurrency (which covers channels, mutexes, and sync primitives) and samber/cc-skills-golang@golang-context (which covers context mechanics). This skill focuses on who owns the context, how it flows through the process, and goroutine lifecycle discipline.
Go Context & Lifecycle Discipline
Context is the process control plane. Root cancellation flows from main() down through every goroutine, controller, and I/O operation. When the context is done, everything stops.
Core Principles
main() owns the root context — it creates context.WithCancel(context.Background()), wires up signal handling, and passes the context down. No other code creates root contexts.
Helper/init functions receive context — they never call context.Background(). If they need a context, the caller provides one.
Context replaces custom exit channels — if you have a done chan struct{} or exitCh chan error that exists solely to propagate shutdown, replace it with context cancellation.
Every goroutine has a shutdown path — through the context it was given. No fire-and-forget goroutines in production code.
Every long-lived select loop includes ctx.Done() — if a goroutine blocks on events, queues, or timers, cancellation must be one of the select cases.
errgroup.WithContext for goroutine groups — first error cancels the group context, all goroutines get the signal, Wait() blocks until all exit.
Avoid time.After inside loops — prefer time.NewTicker or a reusable time.Timer/Reset pattern so repeated waits do not accumulate throwaway timers.
Channel direction encodes ownership — if a struct or API only receives from a channel, type it as <-chan T.
main() Owns Cancellation
funcmain() {
rootCtx, rootCancel := context.WithCancel(context.Background())
defer rootCancel()
// Wire up signal handling
sigCh := make(chan os.Signal, 1)
signal.Notify(sigCh, os.Interrupt, syscall.SIGTERM)
gofunc() {
<-sigCh
rootCancel() // everything shuts down from here
}()
// Pass rootCtx to all subsystemsif err := initCRDState(rootCtx, config); err != nil {
log.Fatal(err)
}
if err := startServer(rootCtx, config); err != nil {
log.Fatal(err)
}
}
Never context.Background() Outside Main
// ❌ BAD — helper creates its own context, disconnected from process lifecyclefuncinitializeState(config Config)error {
ctx := context.Background() // who cancels this? nobody!return reconcileState(ctx, config)
}
// ✅ GOOD — receives context from callerfuncinitializeState(ctx context.Context, config Config)error {
return reconcileState(ctx, config)
}
context.Background() in a helper function means that operation cannot be cancelled when the process is shutting down. It will block until completion even after SIGTERM.
Exceptions:context.Background() is acceptable in:
main() — creating the root context
Tests — context.Background() or t.Context() (Go 1.24+)
init() — very rare, for truly fire-once registration
Replace Custom Channels with Context
// ❌ BAD — custom shutdown channel duplicates context's jobtype Server struct {
shutdownCh chanstruct{}
}
func(s *Server) Run(ctx context.Context) error {
s.shutdownCh = make(chanstruct{})
gofunc() {
<-ctx.Done()
close(s.shutdownCh) // why not just use ctx.Done()?
}()
// ...select {
case <-s.shutdownCh:
returnnil
}
}
// ✅ GOOD — use the context directlyfunc(s *Server) Run(ctx context.Context) error {
childCtx, cancel := context.WithCancel(ctx)
defer cancel()
grpcServer := grpc.NewServer()
gofunc() {
<-childCtx.Done()
grpcServer.GracefulStop()
}()
return grpcServer.Serve(listener)
}
Long-Lived Select Loops Need ctx.Done()
If a goroutine lives in a for { select { ... } } loop waiting on pod events, netlink updates, resync ticks, or queue drains, ctx.Done() must be one of the exits.
// ❌ BAD — loop never notices process shutdownfunc(m *iptablesMonitor) run(ctx context.Context, netlinkUpdates <-chan netlink.RouteUpdate) error {
for {
select {
case update := <-netlinkUpdates:
if err := m.reconcile(update); err != nil {
return err
}
}
}
}
// ✅ GOOD — cancellation is always a way outfunc(m *iptablesMonitor) run(ctx context.Context, netlinkUpdates <-chan netlink.RouteUpdate) error {
for {
select {
case <-ctx.Done():
return ctx.Err()
case update, ok := <-netlinkUpdates:
if !ok {
returnnil
}
if err := m.reconcile(update); err != nil {
return err
}
}
}
}
Prefer errgroup over Ad Hoc Error Channels
When multiple goroutines form a unit of work, prefer errgroup.WithContext. It couples error propagation with cancellation, which is exactly what ACN controller/watcher lifecycles need.
// ❌ BAD — ad hoc error channel returns first error but leaves siblings running
errs := make(chanerror, 2)
gofunc() { errs <- c.watchNodeNetworkConfigs(ctx) }()
gofunc() { errs <- c.watchNetworkContainers(ctx) }()
return <-errs
Use error channels for streaming observations only when the goroutines are not a single lifecycle unit. If they start, stop, and fail together, errgroup should own them.
Avoid time.After in Repeating Loops
time.After is fine for a one-off wait. Inside a long-lived loop it allocates a new timer on every pass, and those timers linger until they fire. In ACN daemons and controllers, prefer reusable tickers/timers.
// ❌ BAD — new timer every iterationfunc(m *iptablesManager) Run(ctx context.Context) error {
for {
select {
case <-ctx.Done():
return ctx.Err()
case <-time.After(30 * time.Second):
if err := m.syncIPSets(ctx); err != nil {
return err
}
}
}
}
// ✅ GOOD — reusable tickerfunc(m *iptablesManager) Run(ctx context.Context) error {
ticker := time.NewTicker(30 * time.Second)
defer ticker.Stop()
for {
select {
case <-ctx.Done():
return ctx.Err()
case <-ticker.C:
if err := m.syncIPSets(ctx); err != nil {
return err
}
}
}
}
For variable backoff or one-shot delays inside a loop, keep a reusable time.Timer and Reset it instead of calling time.After each time.
Prefer Receive-Only Channels to Encode Ownership
If a struct or function only consumes a channel, type it as <-chan T. That makes ownership obvious: this code can receive, but it cannot send or close the producer's stream.
// ❌ BAD — consumer can send on or close the channel by mistaketype podUpdateConsumer struct {
podUpdates chan podUpdate
}
funcnewPodUpdateConsumer(podUpdates chan podUpdate) *podUpdateConsumer {
return &podUpdateConsumer{podUpdates: podUpdates}
}
// ✅ GOOD — producer owns the send sidetype podUpdateConsumer struct {
podUpdates <-chan podUpdate
}
funcnewPodUpdateConsumer(podUpdates <-chan podUpdate) *podUpdateConsumer {
return &podUpdateConsumer{podUpdates: podUpdates}
}
Context in Retry Loops
Retry logic must respect context cancellation between attempts:
Push retry responsibility up to the caller or leverage existing retry machinery (like controller-runtime's reconcile loop):
// ❌ BAD — init function retries internally, hard to test, hard to cancelfuncreconcileInitialState(ctx context.Context)error {
attempt := 0return retry.Do(func()error {
attempt++
return doInit(ctx)
}, retry.Attempts(10), retry.Delay(time.Minute))
}
// ✅ GOOD — init is a plain function, retry is the caller's concern// (or leveraged from the reconciler's built-in retry)funcreconcileInitialState(nnc *v1alpha.NodeNetworkConfig)error {
// pure function, no retry, no context needed for retryreturn doInit(nnc)
}
// Caller wraps it as a reconciler initializer — gets free retries from ctrlruntime
Signal Handling Setup
// ✅ Idiomatic signal handling in mainfuncmain() {
rootCtx, rootCancel := context.WithCancel(context.Background())
defer rootCancel()
// init() should already have set up flags, logging, etc.// Signal handling should be one of the first things in main
sigCh := make(chan os.Signal, 1)
signal.Notify(sigCh, os.Interrupt, syscall.SIGTERM)
gofunc() {
sig := <-sigCh
log.Printf("received signal %s, shutting down", sig)
rootCancel()
}()
// Now start everything with rootCtx...
}
Signal handling should be set up early in main(). If it's set up halfway through initialization and something hangs before that point, you have no way to signal the process to stop.
Common Mistakes
Mistake
Fix
context.Background() in helper functions
Accept context.Context as first parameter
Custom done chan struct{} for shutdown
Use ctx.Done() directly
Long-lived select loop with no ctx.Done()
Add a cancellation case and exit the loop on shutdown
go func() { ... }() with no shutdown path
Pass context, include ctx.Done() in blocking loops, or use errgroup/WaitGroup
time.After inside a for/select loop
Use time.NewTicker or a reusable time.Timer
Retry loop with time.Sleep
Use retry.Do with retry.Context(ctx)
Init function with embedded retry
Make init a pure function, let caller/reconciler handle retry
Signal handling deep in initialization
Set up signals first thing in main()
Ad hoc errs := make(chan error) for goroutine group failures
Use errgroup.WithContext
chan T field/param that only receives
Narrow it to <-chan T
Storing context.Context in a struct field
Pass context through function parameters
Cross-References
→ See samber/cc-skills-golang@golang-concurrency for channel patterns, sync primitives, and worker pools
→ See samber/cc-skills-golang@golang-context for context mechanics, timeouts, and value propagation
→ See acn-go-design-boundaries skill for errgroup patterns and async metrics
→ See acn-go-errors-logging skill for retry library usage
