11 pages
Docs
Operational reference: outputs, sinks, metrics, traces, the effector, CI.
- The errors outputStable
How element-level failures are routed to .errors, the reason taxonomy, and how to configure a sink.
- Runtime metricsStable
The counters and distributions RunAgent publishes, the two identities worth alerting on, and why there are no percentiles.
- TracesStable
The TraceEvent stream, OTel GenAI attribute conventions, and the BigQuery and OTLP sinks.
- The effectorStable
The external service that executes intents — deployment preconditions, lease and TTL budgets, and exactly what is and is not guaranteed.
- Testing tiers and CIStable
Four test tiers, what each gates, the closed marker registry, and the adapter conformance matrix.
- Human in the loopStable
Approval channels, timeout routing with Deny/Drop/Escalate, and the two layers that make timeouts fail closed.
- RunnersPartial
Which runners are supported, at what level, and the test evidence in this repository behind each claim.
- Repository reference docsPartial
The reference documents under docs/ that have no dedicated page here yet, each with what it covers and where it is published.
- The consoleExperimental
A local viewer for what the runtime records — activations, span trees, errors by reason, and the approval queue — over one SQLite file, with no broker and no cloud project.
- Long-term memory storesPartial
The durable MemoryStore tier behind working memory — Bigtable, Redis, Firestore, and SQL backends, seq-guarded upserts, and what retention still is not.
Declare a RunAgent pipeline from a Beam YAML document — one constructor, module:object references, and the same four outputs as the Python surface.