Bloodbank Integration

Bloodbank Integration

Route here when a service or harness needs to emit or consume events on the 33GOD bloodbank bus. The bus is the only sanctioned inter-service channel — never bypass it with direct calls.

Operating Principles

• Bus is canon. All inter-service traffic flows through bloodbank. Direct service-to-service calls are an anti-pattern enforced repo-wide.
• Schema first. Every event has a JSON Schema under bloodbank/schemas/. Build envelopes through the canonical builder/validator path, and keep manual envelope examples derivable from the schema and naming contract.
• NATS is the current bus. v3 (Dapr + NATS JetStream + CloudEvents 1.0) is the live target. v2 (RabbitMQ topic exchange) still runs but is migration-only territory.
• Subject convention is load-bearing. CloudEvents type is bloodbank.v1.<domain>.<entity>.<action>. NATS subjects are bloodbank.evt.v1.<domain>.<entity>.<action> for events, bloodbank.cmd.v1.<domain>.<entity>.<action> for commands, and bloodbank.rpy.v1.<domain>.<entity>.<action> for replies. The catch-all event-toaster listens on bloodbank.evt.v1.>.
• Agent hooks use one publisher. All CLI lifecycle hooks call ~/.agents/hooks/bloodbank/publish.py --client <agent> --hook <native-event>. Client-specific prep lives in services/agent-hooks/clients/<agent>.py; per-client publish.py files are wrappers.
• Fail open at the boundary. Hooks must never block the host agent. Producer libs should swallow publish failures by default.

Triage Table

Match the user's intent against the signals on the left; load the cited file first.

Decision Tree: Which Producer Path?

Are you in a 33GOD service container with a Dapr sidecar?
├─ Yes → Dapr pub/sub. Subject = "bloodbank.evt.v1.<domain>.<entity>.<action>", pubsub component = bloodbank-pubsub.
│        See references/producers/methods.md → "Dapr publish".
└─ No
   ├─ One-shot from a shell hook (Claude Code, Copilot, etc.)?
   │  → Canonical agent-hook publisher (stdlib raw TCP, no nats-py). See bloodbank/services/agent-hooks/.
   ├─ Long-running Python service on the host?
   │  → nats-py direct, subject "bloodbank.evt.v1.<...>". See references/producers/methods.md.
   ├─ External webhook (Plane, GitHub, etc.) with HTTP only?
   │  → POST to bloodbank's /event (typed webhook) or /publish (generic). RabbitMQ path.
   └─ HTTP client that needs to issue a COMMAND envelope (not an event)?
      → POST to hookd_bridge :18790/hooks/agent. See bloodbank/hookd_bridge/; command subject is `bloodbank.cmd.v1.agent.invocation.start`.

Decision Tree: Which Consumer Path?

Do you own a 33GOD service container with a Dapr sidecar?
├─ Yes → Dapr subscriber. Declare /dapr/subscribe routes. Reference: services/claude-events-recorder/main.py.
└─ No
   ├─ Need wildcard fan-in across many subjects (observability, audit, notify)?
   │  → NATS core subscribe on "bloodbank.evt.v1.>" (no JetStream consumer, no durability).
   │    Reference: services/event-toaster/main.py.
   ├─ Need durable, replay-capable consumption on a specific subject?
   │  → NATS JetStream durable consumer. Subjects defined in compose/v3/nats/streams.json.
   ├─ Legacy v2 consumer or RabbitMQ-only environment?
   │  → FastStream RabbitMQ consumer bound to exchange bloodbank.events.v1. Avoid for new work.
   └─ Just want desktop notifications for everything?
      → Subscribe to https://ntfy.delo.sh/bloodbank (event-toaster already publishes there).

Cross-Cutting Rules

These apply regardless of producer/consumer path or language:

• Envelope shape is fixed. CloudEvents 1.0 + 33GOD extension fields (producer, service, domain, schemaref, correlationid, causationid). The canonical base lives at bloodbank/schemas/_common/cloudevent_base.v1.json; every event schema allOf-extends it.
• type and NATS subject are bound. The Dapr topic / NATS subject for an event is always the envelope's subject, derived from type by inserting evt after bloodbank (bloodbank.v1.agent.tool.completed → bloodbank.evt.v1.agent.tool.completed). Never publish to a subject that doesn't match the envelope.
• Schema versioning is in the filename. schemas/bloodbank/v1/agent/session.started.v1.json is v1; a breaking change becomes .v2.json with a new dataschema URI.
• Do not make schemas optional. Edit the JSON Schema first, then validate with mise run smoketest:schemas.
• Use Hindsight memory bank bloodbank for integration notes — broker-level decisions, subject-naming surprises, consumer wiring gotchas live there, not in the code.
• Test producers with the toaster. bloodbank-event-toaster subscribes to bloodbank.evt.v1.> and forwards every envelope to https://ntfy.delo.sh/bloodbank. If you don't see your event there, it didn't make it to NATS.

Reading Order

For the most common entry points:

Out of Scope

This skill does NOT cover:

• Generic RabbitMQ / NATS / Kafka setup or tuning unrelated to bloodbank's topology. Use the broker vendor's documentation; this skill assumes the v3 stack (compose/v3/docker-compose.yml) is already running.
• n8n workflow authoring or routing decisions. Use workflow-router to choose between n8n, bloodbank, and other automation tools.
• Hindsight memory recall/retain. Use the hindsight skill for memory-bank operations even when wiring bloodbank events that carry memory references.
• Non-event-bus parts of 33GOD (Candystore persistence internals, Candybar UI work, Bloodbank validator implementation). Use the 33god-ecosystem hub for routing, or the project's own AGENTS.md.
• Schema validator internals. This skill points at the schema workflow; change scripts/validate_schemas.sh and naming smoke tests in the Bloodbank repo when the validator itself changes.
• Claude Code / Copilot CLI hook semantics themselves (timeout flags, OS-specific behavior). Use the vendor docs; this skill covers the wiring layer between those hooks and bloodbank.