Architecture at a glance
Three nested views — context, services, and inside the backend.
Services & responsibilities
Every process in the topology, what it owns, and what it deliberately does not.
Anatomy of a search
One non-generative round trip, from query expansion to ranked cards.
Anatomy of a chat turn
The backend-runtime callback loop, step by step, with the guard at each hop.
Search vs Chat
Two surfaces over one authorized matching pipeline, and when to use each.
Retrieval & ranking
Lexical and vector branches, the trusted SQL boundary, fusion, and reranking.
Knowledge sources & connectors
Runtime discovery, per-member identity, sync, and freshness.
Agent runtime & the turn engine
The conversational loop, its authority bundle, and how it survives failure.
Tools & the tool boundary
The registry, capability ceilings, parallel calls, and the write guard.
Model gateway
Routing to Kamiwaza, policy, budgets, effort, retries, and response normalization.
Sub-agents & orchestration internals
Delegation as a bounded tool call, and the limits it inherits.
Code sandbox
Deployment-selected isolation, the lease pool, controlled egress, and per-participant workspaces.
Conversation journal & durability
The append-only journal, the turns cache, reconnect, replay, and terminal outcomes.
Durable tasks: scheduler & worker
Definition, occurrence, run, and trace — and how work is claimed exactly once.
Context management & compaction
Always-on system compaction versus the opt-in member compression seams.
Data model & storage
What lives in Postgres, Valkey, and the object store, and who owns each record.
A2A & remote agents
The agent-to-agent proof of concept and why an Agent Card grants nothing.