A2A: offline framework-adapter simulation

Twin: examples/sdk_framework_adapter_cert_a2a.py · emits the framework-adapter probe evidence · offline, no credentials. A coding agent can complete this page from the frontmatter alone.

1. What you are testing

Agent2Agent (A2A) ships a preset in FRAMEWORK_PRESETS: method send_message, input mode dict. This page certifies that preset is real — the twin, examples/sdk_framework_adapter_cert_a2a.py, builds a local LocalA2ASession exposing exactly send_message (a message kwarg plus a session side-kwarg) and wraps it through the same wrap_framework / run_framework_adapter_probe path a manifest uses, so the probe exercises the real adapter resolution. The shim returns contract-shaped synthetic evidence — a framework_trace event and a framework_trace_status tool call — and never imports a real a2a-sdk and never touches the network.

The failure class this catches is preset drift: if the send_message/dict shape no longer resolves through the adapter, the probe fails and the preset is corrected. The IO surface this preset binds to is side_kwargs (the message kwarg + session metadata pattern, like pipecat's frame kwarg) — one of the eight existing framework-adapter IO contracts, classified by adapter shape.

A2A is deliberately doubly-covered: this certification probe keeps the closed required set homogeneous, while the deeper A2A protocol surfaces live in the live lane and the protocol-trace example (see section 5).

2. Run it

CLI — the twin is executable and writes the probe artifact:

python examples/sdk_framework_adapter_cert_a2a.py artifacts/framework-cert-a2a.json

SDK, same operation:

from sdk_framework_adapter_cert_a2a import run  # examples/ on sys.path

result = run("artifacts/framework-cert-a2a.json")
assert result["status"] == "passed"
assert result["method"] == "send_message" and result["input_mode"] == "dict"

3. What you built

Postcondition (machine-checkable):

python -c "import json; p=json.load(open('artifacts/framework-cert-a2a.json')); assert p['status']=='passed', p['status']; assert p['method']=='send_message' and p['input_mode']=='dict'; print('ok')"

The artifact carries the resolved method/input mode, the framework runtime trace, and the message round-trip evidence the adapter extracted — a replayable record the release gate re-executes.

4. When it fails

Symptom	First-mile class	Doctor check
vendored import failed	infra	agent-learn doctor -> summary.missing_engine_modules
probe status is failed (method did not resolve)	behavior regression	confirm send_message/dict against the current A2A SDK; if drifted, add a V1_FRAMEWORK_PRESET_CORRECTIONS row and fix the preset
public-boundary mismatch	config fault	agent-learn doctor -> summary.public_boundary_passed

5. Prove it / keep it

The twin is admitted by the framework_adapter_preset_certification_readiness release gate, so every agent-learn release-check re-executes this exact probe — the page stays true or the release fails.

For the deeper A2A protocol surfaces, two existing artifacts go beyond preset certification: the A2A live lane (src/fi/alk/live/a2a_lane.py) and the protocol-trace example (examples/sdk_framework_adapter_a2a_protocol_trace.py, admitted by protocol_adapter_readiness), which export the agent card, the A2A event stream, and the per-task lifecycle. To keep your own A2A endpoint honest, promote the run artifact into a regression baseline with the baseline / promote-to-regression / compare command family.