Message Types¶

Messages carry type, source, target, payload, and metadata (e.g. trace correlation). PRD §12 defines categories and payloads in full.

Event bus vs streams¶

Durable audit for many lifecycle events lives in the database event log. Redis streams carry operational traffic: dispatch, worker signals, async usage recording. Encoding (JSON vs binary) is an implementation detail.

Partitioning: dispatch traffic is sharded so no single stream becomes the only scalability bottleneck.

Multi-tenancy: all streams carry org_id in the payload when available.

Event types¶

Task lifecycle¶

Type	Description
`TaskScheduled`	Task added to ready queue
`TaskStarted`	Worker claimed and began execution
`TaskCompleted`	Worker returned success result
`TaskFailed`	Worker returned failure or timed out

Phase lifecycle¶

Type	Description
`PhaseStarted`	Execution phase began (e.g. planning, execution, review)
`PhaseCompleted`	Phase finished successfully
`PhaseFailed`	Phase ended in failure
`PhaseSummary`	Aggregate summary emitted at phase end

LLM audit¶

Type	Description
`LLMUsage`	Per-request LLM audit record: `model`, `tokens_in`, `tokens_out`, `latency_ms`, `cost_dollars`

Worker¶

Type	Description
`WorkerHeartbeat`	Liveness signal from worker
`WorkerRegistered`	Worker joined the pool

Goal¶

Type	Description
`GoalCreated`	New goal submitted
`GoalCompleted`	Goal finished; emitted to `astra:goals:completed` for cascade tracking

Redis streams¶

Stream	Purpose
`astra:tasks:<shard>`	Sharded task dispatch to workers
`astra:workers:heartbeat`	Worker liveness signals
`astra:dead_letter`	Tasks that exhausted all retries
`astra:usage`	Async LLM usage persistence fan-out (`request_id`, `agent_id`, `task_id`, `org_id`, `user_id`, `model`, `tokens_in`, `tokens_out`, `latency_ms`, `cost_dollars`, `timestamp`)
`astra:goals:completed`	`GoalCompleted` events for cascade / dependency tracking
`astra:slack:incoming`	Normalized Slack events from slack-adapter
`olympus:triggers:raw`	Normalized trigger events from webhook ingest

Consumer groups process messages with ack semantics. Messages that exceed retry limits are moved to astra:dead_letter.