sequenceDiagram
autonumber
actor U as User
participant D as Dashboard
participant A as Auth / session
participant B as One brain · API
participant P as Planner / LLM router
participant G as Governance gate
participant T as Tool
participant X as LLM observability proxy
participant L as LLM · DeepSeek
participant M as Memlink · memory
participant R as Redis · stream
U->>D: type a message
D->>A: validate session
A-->>D: owner identity (per-owner scope)
D->>B: POST /v1/chat/send
B->>M: ingest message (events_raw)
B->>P: plan — answer or use a tool?
alt needs a tool
P->>G: propose tool action
Note over G: capability · risk · policy · DLP · sign · audit
alt allowed
G->>T: run tool
T-->>B: result
else needs approval
G-->>D: approval chip
D-->>G: operator approves
G->>T: run tool
T-->>B: result
else denied
G-->>B: blocked (policy / DLP)
end
else just answer
P-->>B: answer directly
end
B->>M: recall relevant memory
M-->>B: support bundle
B->>X: generate reply (logged)
X->>L: provider API
L-->>X: streamed tokens
X-->>B: streamed reply
B->>R: publish stream events
R-->>D: fan out (SSE)
D-->>U: reply appears, token by token
B->>M: persist final reply
