Human-in-the-Loop Approval

The clinical review gate (Day 2b) — a LongRunningFunctionTool pauses the pipeline until a clinician decides. This is the closest process to true BPMN: two lanes, one message flow, one exclusive gateway.

flowchart TD
    subgraph LANE_AGENT["Lane: Agent (capstone_agent)"]
        DONE["Extraction pipeline completes"]
        CONF{confidence >= threshold?}
        AUTO["Auto-approve path"]
        PAUSE["Pause: state = needs_review
(LongRunningFunctionTool)"]
        PERSIST["persist_extraction_relational
persist_extraction_vector
(require approved receipt)"]
        DISCARD["Discard + reason logged"]
    end

    subgraph LANE_CLIN["Lane: Clinician (Clinical Inbox UI)"]
        INBOX["Review results in Clinical Inbox"]
        DECIDE{Approve or Reject?}
    end

    DONE --> CONF
    CONF -->|yes| AUTO --> PERSIST
    CONF -->|no| PAUSE --> INBOX --> DECIDE
    DECIDE -->|approve
transition: needs_review -> approved| PERSIST
    DECIDE -->|reject
transition: needs_review -> rejected| DISCARD
    PERSIST --> END1([Stored to relational + vector stores])
    DISCARD --> END2([Audit event with rejection reason])

Key facts:

transition_extraction_review enforces the state machine: only needs_review → approved or needs_review → rejected are valid transitions.
Persistence tools require an approved review receipt — there is no code path that stores unreviewed low-confidence extractions.
The pause/resume mechanism is ADK's LongRunningFunctionTool (human_in_the_loop.py).
Review decisions are clinical audit events (Observability).