Imported from speakeasy-api/gram (.agents/skills/gram-pubsub/SKILL.md). Install upstream with npx skills add speakeasy-api/gram --skill gram-pubsub. Copyright stays with the author.

Gram Pub/Sub

Gram declares its GCP Pub/Sub topology as data on protobuf messages, not as hand-written Terraform or Config Connector YAML. You annotate a "marker" message with a topic or subscription option; a generator walks the compiled descriptors and emits a Config Connector Helm values document (infra/gen/kcc.yaml); the deployment tooling consumes that document to provision real topics and subscriptions. The same proto options drive a type-safe Go publisher/subscriber library, so the infrastructure and the application code can never disagree about a topic's name or a subscription's wiring — both read from one source of truth.

The whole point of the design is that an engineer adding a topic touches only a .proto file under infra/proto/. Everything downstream — the Config Connector specs, the per-environment rollout, and the runtime handle — is derived from it. Everything in this skill lives in this repo (infra/); the deployment side is referred to only abstractly as "the deployment tooling."

Mental model: marker messages

A topic or a subscription is declared by attaching a message option to a protobuf message. The message itself can be the event schema (for a topic) or an empty placeholder (for a subscription) — what matters is the option, not the fields. One message must not carry both a topic and a subscription option; declare them separately. This keeps a topic's payload schema and a consumer's config as distinct, independently evolvable things.

The option definitions live in infra/proto/gcp/pubsub/v1/options.proto:

• (gcp.pubsub.v1.topic) — TopicOptions: optional name, retention_hint, labels.
• (gcp.pubsub.v1.subscription) — SubscriptionOptions: optional name, required topic (the proto full name of the topic-declaring message, e.g. "gram.outbox.v1.Event"), plus retention, ack_deadline, retry_policy, filter, dead_letter, expiration_ttl, retain_acked_messages, labels.

Authoring a topic

Add the topic option to the message that is the event payload, so the schema and the topic travel together. See infra/proto/gram/outbox/v1/event.proto:

message Event {
  string id = 1;
  string type = 2;
  google.protobuf.Timestamp created_at = 3;
  bytes payload = 4;

  option (gcp.pubsub.v1.topic) = {
    retention_hint: { seconds: 604800 /* 7 days */ }
  };
}

With no explicit name, the topic ID is the kebab-cased proto full name: gram.outbox.v1.Event → gram-outbox-v1-event. Set name only when you need to diverge from that.

Authoring a subscription

Declare a subscription on its own marker message (no payload fields needed) and point topic at the topic message's full name. See infra/proto/gram/outbox/v1/processor.proto:

message Processor {
  option (gcp.pubsub.v1.subscription) = {
    topic: "gram.outbox.v1.Event"
    ack_deadline: { seconds: 30 }
    retry_policy: {
      minimum_backoff: { seconds: 10 }
      maximum_backoff: { seconds: 600 }
    }
    dead_letter: { max_delivery_attempts: 5 }
  };
}

With no explicit name, the subscription ID is the kebab-cased proto full name (same rule as topics): gram.outbox.v1.Processor → gram-outbox-v1-processor. Set name only when you need to diverge from that. The topic reference is validated against the discovered topic set during generation, so a typo fails the build rather than producing a dangling subscription.

Dead-letter queues are synthesized

When a subscription sets dead_letter, the generator auto-creates a DLQ topic — you do not declare it. The default name is <subscription>-dlq (override with dead_letter.name). The DLQ topic carries the same message schema as the source and is labeled dlq_for: <subscription>. This is why subscription IDs are length-capped below the topic limit: room must be left for the -dlq suffix.

Regenerating after a proto change

Always run the generator after editing any .proto under infra/proto/ and commit the result — infra/gen/kcc.yaml is the committed artifact the deployment tooling consumes:

mise run gen:infra

This task (.mise-tasks/gen/infra.sh) does three things: buf generate (proto → Go in infra/gen/), buf build (compiled FileDescriptorSet → infra/cmd/infra/descriptors.pb), then go run ./infra/main.go gen-cc to write infra/gen/kcc.yaml. The generated Go and the descriptor blob are gitignored (**/descriptors.pb) and excluded from formatting; infra/gen/kcc.yaml is committed.

How generation works (infra/internal/gcp/)

Key behaviors worth knowing:

• Validation (validateTopicID / validateSubscriptionID): GCP naming rules — must start with a letter, 3–255 chars, no goog prefix, no full resource paths. Duplicate topic or subscription names, an unknown topic reference, or a DLQ name colliding with a declared topic all fail generation.
• Labels: every generated resource gets managed_by: proto_pubsub_orchestrator plus a proto_message label carrying the source message's full name; subscriptions also carry topic_proto_message. These make resources traceable back to their declaration.
• Stable output: topics and subscriptions are sorted by name so the generated file diffs cleanly across runs. Durations render as integer-second strings (e.g. 604800s) as Config Connector expects.
• Separation of concerns: the generator emits only the topology under a pubsub key (names, labels, specs). Per-resource deployment metadata (project, namespace, deletion/prune policy) is applied downstream by the deployment tooling, never here — so don't expect this generator to emit it.

Runtime: publishing and subscribing (infra/pkg/gcp/)

Application code never hard-codes a topic or subscription name. It hands a proto message to a broker and gets back a generic, type-safe handle. The broker resolves names from the same proto options used for generation.

Two brokers implement both PublisherBroker and SubscriberBroker:

• PubSubBroker (pubsub_gcp.go) — talks to real GCP; assumes topics/subs already exist (Config Connector created them).
• EmulatedPubSubBroker (pubsub_local.go) — for local dev against the emulator; reconciles topics and subscriptions on demand since the emulator has no Config Connector.

Both take the embedded descriptors blob. Usage (condensed from infra/cmd/infra/demo.go, the runnable reference):

broker := gcppub.NewEmulatedPubSub(logger, projectID, client, descriptors)

// Publisher for the topic declared by *outboxv1.Event.
pub, _ := gcppub.PubSubPublisherForMessage(ctx, broker, &outboxv1.Event{})

// Subscriber for the *outboxv1.Processor subscription, receiving *outboxv1.Event.
// Read as: "a handle on the Processor subscription delivering Event messages."
sub, _ := gcppub.PubSubSubscriberForMessage(ctx, broker, &outboxv1.Event{}, &outboxv1.Processor{})

pub.Publish(ctx, &msg).Get(ctx)        // proto-marshaled, with content-type + schema attributes

// The callback receives the unmarshaled message plus delivery metadata.
// Return nil to ack; return a non-nil error to nack (and trigger redelivery /
// dead-lettering if enabled for the topic and subscription).
sub.Receive(ctx, func(ctx context.Context, msg *outboxv1.Event, meta gcppub.MessageMetadata) error {
    _ = msg            // already unmarshaled to *outboxv1.Event
    _ = meta.ID        // broker-assigned message ID
    _ = meta.Attributes
    _ = meta.DeliveryAttempt // set when dead-lettering is enabled, else nil
    return nil
})

publisher.go / subscriber.go define the generic Publisher[M] and Subscriber[M] over cloud.google.com/go/pubsub/v2. Messages are proto-marshaled and tagged with content-type: application/x-protobuf and a schema attribute (the message full name); the subscriber unmarshals back into a fresh M and hands it to your callback along with a MessageMetadata. The callback's return value drives ack/nack — nil acks, non-nil nacks — so you no longer call Ack/Nack yourself. Tune behavior with WithPubSubPublishSettings / WithPubSubReceiveSettings.

This is the low-level library. Inside the Gram server you almost never call sub.Receive directly — you write a streams.Handler and register it in the gram streams process, which wraps the receive loop with tracing, panic recovery, and ack/nack plumbing for you. See the next section.

Implementing a subscriber in the server

Subscribers run in a dedicated long-running process, gram streams (server/cmd/gram/streams.go). It is its own service — separate from the API server and the Temporal worker — so message consumers scale and fail independently. Locally it runs as the streams process under pitchfork; start it with pitchfork start streams. Adding a consumer is two steps: write a handler, then register it.

Step 1 — write a handler

A consumer is a streams.Handler[M] (server/internal/streams/handlers.go), where M is the proto message the topic carries:

type Handler[M any] interface {
    Handle(context.Context, M, gcp.MessageMetadata) error
}

Implement it in the domain package that owns the behavior (not in streams/ and not in cmd/gram/). server/internal/ping/handler.go is the canonical reference — a struct holding its dependencies, a NewHandler constructor, and a Handle method:

type Handler struct {
    logger *slog.Logger
}

func NewHandler(logger *slog.Logger) *Handler {
    return &Handler{logger: logger}
}

func (h *Handler) Handle(ctx context.Context, m *pingv2.Message, _ gcp.MessageMetadata) error {
    // ... do the work ...
    return nil
}

The handler's job is narrow: process one message and return. The return value is the ack/nack signal — nil acks the message; a non-nil error nacks it, triggering redelivery and eventual dead-lettering if the subscription declares a dead_letter policy (see "Authoring a subscription"). So return an error only when you genuinely want the message retried; for a poison message you can never process, log it and return nil to drop it. Do not call Ack/Nack, start your own receive goroutine, or open a span for the receive loop — the runner does all of that (next step).

HandlerFunc[M] adapts a bare function to the interface when a struct is overkill.

Step 2 — register it in the streams runner

In streams.go, the Action builds the shared dependencies (db, redis, the gcp.NewPubSubBroker) and a receiverGroup. Register each handler inside the marked block with mustReceive:

// Start subscription receivers in this block
{
    mustReceive(rg, &pingv2.Message{}, &pingv2.Processor{}, ping.NewHandler(logger))
}

The three positional arguments mirror the proto declarations:

• &pingv2.Message{} — the topic message; its type fixes M, the payload your handler receives.
• &pingv2.Processor{} — the subscription marker message (the one carrying (gcp.pubsub.v1.subscription)).
• ping.NewHandler(...) — your handler, with its dependencies injected from the ones the Action already constructed. If a handler needs the db or redis, pass them here; add new shared dependencies to the Action body.

Trailing gcp.SubscriberOptions (e.g. WithPubSubReceiveSettings) can follow. Use mustReceive (it panics on a misconfigured subscription, which is a programmer error caught at startup); receive returns the error if you need to handle it.

What the runner does for you

receive/mustReceive exist so handlers stay tiny. For each registration the runner: resolves the subscription via gcp.PubSubSubscriberForMessage, launches the receive loop in the shared errgroup (so any subscriber's fatal error tears the whole process down for a clean restart), starts a stream.handleMessage span per message, stamps pub/sub subscriber context values for telemetry, recovers panics in your handler (converting them to a nacking error instead of crashing the goroutine), and treats context.Canceled as a clean shutdown rather than a failure. Because all of this is centralized, you get consistent observability and failure semantics across every consumer without writing any of it.

Driving the flow end to end (optional publisher)

ping also ships a StartPublisher (server/internal/ping/publisher.go) launched with group.Go alongside the receivers — a heartbeat that publishes a message every few seconds so you can confirm the publish→subscribe loop is alive in logs. It is a sanity-check harness, not a pattern to copy for real publishers: production code publishes from wherever the event originates (an API handler, a workflow activity) using gcp.PubSubPublisherForMessage, not from a loop in the streams process.

Local development

mise.toml sets PUBSUB_EMULATOR_HOST and compose.yml runs a pubsub-emulator service (the google/cloud-sdk emulators image on port 8085). With the emulator host set, EmulatedPubSubBroker creates topics/subscriptions lazily on first publish/subscribe, so you don't need Config Connector locally. The infra demo command (go run ./infra/main.go demo) is an end-to-end publish/subscribe loop you can run to sanity-check the framework.

How a declaration reaches a real environment

This is the part that most often confuses people, so it's worth being precise about what this repo does and does not do.

• Declaring a topic and committing infra/gen/kcc.yaml does not create anything in GCP. The proto is the source; infra/gen/kcc.yaml is the committed artifact. Provisioning happens later, in separate deployment tooling, not as a side effect of merging.
• Rollout is decoupled and version-pinned per environment. Each environment runs the topology from a specific committed revision, not from main directly. A topology change therefore reaches one environment at a time as that environment is rolled forward — so a freshly declared topic legitimately may not exist yet in a given environment even though it is on main.
• Local working ≠ deployed. The emulator reconciles topics/subscriptions on the fly, so "works locally but missing in GCP" is the expected signature when the rollout simply hasn't reached that environment yet.
• PR previews never create real topics, by design.

So when something seems missing in a real environment, the productive question is not "is the proto correct?" but "did infra/gen/kcc.yaml get regenerated and committed, and has the environment in question been rolled forward to a revision that includes it?" The rollout mechanics themselves live with the deployment tooling, outside this repo.

Gotchas and conventions

• Run mise run gen:infra and commit infra/gen/kcc.yaml after any proto change. A stale kcc.yaml is what actually ships — the proto is only the source. Regenerating without committing the result is the most common reason a declaration silently does nothing downstream.
• topic references use the proto full name, not the resolved topic ID (gram.outbox.v1.Event, not gram-outbox-v1-event).
• Don't declare DLQ topics — they're synthesized from dead_letter.
• One option per message: a message carrying both a topic and a subscription option fails generation by design.
• Retiring a topic is a deliberate, decoupled act. Removing a declaration stops the topology from managing it, but the deployment tooling is configured so existing GCP topics are not destroyed by a topology change. Treat removals as a separate, intentional step rather than assuming a deleted declaration tears down the resource.