feat: add working A2A server sample

[Tsuyoshi Ushio (TsuyoshiUshio)]

Summary

Implements the experimental P3 simple non-SSE A2A server above #208. It adds opt-in typed A2A configuration, per-agent Agent Card and JSON-RPC routes, public Microsoft Agent Framework hosting-a2a conversion, and reuse of the existing authenticated non-streaming runtime runner.

The runnable incident-triage sample now uses a real MAF A2AAgent as its primary caller. That client runs in an isolated MAF 1.15+ environment across HTTP, so the server/runtime remains on its verified core 1.13 tuple.

Experimental P3 profile

• Accepts exact A2A wire version 1.0; an absent version is treated as 0.3 and rejected.
• Supports only non-streaming SendMessage requests containing user text parts and returns one direct agent Message.
• Accepts both returnImmediately values; the flag has no effect for a direct Message.
• Explicitly rejects Tasks, task continuation, streaming, subscribe/cancel, push notifications, REST, non-text parts, and outbound client tools.
• Agent Card URLs are explicitly configured per agent; cards advertise streaming: false and pushNotifications: false.
• Preserves the existing HTTP auth boundary, scoped runtime sessions, security/policy behavior, builtin routes, and shared DFApp workflow bindings.
• Enforces bounded request, part, text, response, and per-agent in-flight execution limits; runtime metadata is allowlisted to response text only.

P4 Task persistence and P5 streaming are intentionally not included. Durable dependency optionalization is also out of scope. Any runtime-wide MAF core/OpenAI/Foundry upgrade is a separate standalone compatibility-validation PR (PV), not bundled with P3/P4/P5 or optional-Durable work.

Runnable incident-triage story

From the repository root, create independent server and client environments:

# Server: editable runtime + [a2a], preserving core 1.13
python -m venv .venv-server
.\.venv-server\Scripts\Activate.ps1
python -m pip install -e ".[a2a]"

# Client: only agent-framework-a2a and its core 1.15+ closure
python -m venv samples\a2a-incident-triage\.venv-client
.\samples\a2a-incident-triage\.venv-client\Scripts\Activate.ps1
python -m pip install -r samples\a2a-incident-triage\requirements-client.txt

Copy samples\a2a-incident-triage\src\local.settings.template.json to local.settings.json, configure the repository's normal model credentials, start Azurite separately, and run the host from the server environment:

.\.venv-server\Scripts\Activate.ps1
cd samples\a2a-incident-triage\src
func start

In a second terminal, use the MAF client. It resolves the explicit /agents/main/.well-known/agent-card.json path, selects JSONRPC 1.0, and calls A2AAgent.run():

.\samples\a2a-incident-triage\.venv-client\Scripts\Activate.ps1
cd samples\a2a-incident-triage
python client.py

Use python raw_client.py (and --return-immediately) to inspect and modify the raw Agent Card/JSON-RPC wire contract. Both clients accept --function-key through public HTTP client headers.

Dependencies

The server/runtime [a2a] extra remains:

• agent-framework-hosting-a2a==1.0.0a260730
• agent-framework-hosting==1.0.0a260730
• a2a-sdk[http-server]==1.1.2
• MAF core 1.13.0

The client-only sample requirements pin agent-framework-a2a==1.0.0b260821, which resolves a MAF core 1.15+ client environment. It does not install the editable runtime and is not added to root runtime/dev extras. Existing OpenAI/Foundry pins and the mandatory Durable dependency remain unchanged.

Validation

• python -m ruff check src tests
• python -m mypy src
• python eng/scripts/generate_config_reference.py --check
• python -m pytest --cache-clear --cov=./src/azure_functions_agents --cov-report=xml --cov-branch tests — 1,230 passed, 58 deselected
• Real Functions-host E2E through isolated A2AAgent.run() plus raw JSON-RPC passed
• Client-only closure: agent-framework-a2a 1.0.0b260821, core 1.16.0, SDK 1.1.2, no runtime package
• Python 3.13 and 3.14 clean server-wheel closure remains core 1.13.0 with the original hosting tuple
• Independent follow-up review found no significant issue

Stack

• Base: tsuyoshiushio-a2a-frd-and-design (f435721f6eadbb3d6795b3d2f8f012f9edd7733d)
• This layer: P3 implementation and client-interoperability evidence only
• Bottom PR: docs: propose A2A server architecture #208