ADK Integration

Import Google ADK agent hierarchies — from Agent objects or YAML/JSON configs.

From Agent objects

from google.adk import Agent
from minimal_oversight import analyze_pipeline

root = Agent(
    name="support",
    model="gemini-2.0-flash",
    description="Routes customer queries",
    sub_agents=[
        Agent(name="billing", model="gemini-2.0-flash", description="Billing questions"),
        Agent(name="tech", model="gemini-2.0-flash", description="Technical support"),
    ],
)

# Auto-detected
report = analyze_pipeline(root, p_min=0.80)

From config dicts

from minimal_oversight import analyze_pipeline

config = {
    "name": "support",
    "model": "gemini-2.0-flash",
    "sub_agents": [
        {"name": "billing", "description": "Billing questions"},
        {"name": "tech", "description": "Technical support"},
    ],
}

# Also auto-detected (requires "sub_agents" key)
report = analyze_pipeline(config, p_min=0.80)

Hierarchical IDs

ADK agents form trees. To prevent name collisions (e.g., two sub-agents named "helper" under different parents), the connector generates hierarchical IDs:

support                        → "support"
support/billing                → "support/billing"
support/tech                   → "support/tech"
support/tech/diagnostics       → "support/tech/diagnostics"

Use these IDs for parameter overrides:

from minimal_oversight.connectors.adk import from_adk_config

pipeline = from_adk_config(config, parameter_overrides={
    "support/billing": {"sigma_skill": 0.70},
})

Role inference

The connector infers roles from agent names and descriptions:

Keyword pattern	Inferred role	Default σ_skill
"generate", "write", "draft"	Generator	0.55
"review", "check", "validate"	Reviewer	0.60
"route", "triage", "classify"	Router	0.70
"merge", "gate", "aggregate"	Gate	0.55
"search", "api", "tool"	Tool	0.80
"human", "escalate"	Human	0.85

The root agent defaults to Router if it has sub-agents, Generator if it doesn't.

Calibrating from session logs

from minimal_oversight.connectors.traces import from_adk_session_logs, to_workflow_traces

sessions = [...]  # ADK session dicts with "id" and "events"
traces = to_workflow_traces(from_adk_session_logs(sessions))
report = analyze_pipeline(pipeline, p_min=0.80, traces=traces)

ID matching

ADK session logs use flat agent names ("billing"), but the pipeline uses hierarchical IDs ("support/billing"). For trace calibration, either use a flat pipeline or map trace IDs to hierarchical IDs.