A2A on a2a-sdk 1.x: build the routes through ADK's compat layer

[antiv]

Fixes #223.

What was wrong

requirements.lock pins a2a-sdk==1.2.1, but adk_main.py used the 0.3 API (a2a.server.apps.A2AStarletteApplication, a pydantic AgentCard). On 1.x the import fails. --a2a, which the dashboard always passes, printed "A2A dependencies not available" and mounted no /a2a/* routes. A2A calls failed, and so did agents exposed over MCP: the bridge calls /a2a/{agent}.

It went unnoticed locally because a venv that wasn't reinstalled from the lock still had a2a-sdk 0.3.22.

The fix

• shared/utils/a2a_routes.py builds each agent's card and routes through google.adk.a2a._compat (parse_agent_card, attach_a2a_routes_to_app), the same way ADK's own --a2a does. It works on 0.3.x and 1.x. adk_main.py keeps its agent discovery (files, database, runners) and calls it. A missing dependency still gets the "A2A dependencies not available" message.
• Methods: on 1.x the JSON-RPC route serves the 1.x methods and, through the v0.3 compat layer, the 0.3 ones. Existing callers and the MCP bridge (message/send) keep working.
• The default card:
  • url is now http://localhost:8000/a2a/{agent}, the agent's endpoint, instead of the server root. A client resolving the card would have posted to the wrong URL. The base address is still fixed; MATE has no public-URL setting.
  • The modes are text/plain, what the executor reads and writes.
  • The metadata block is dropped; 1.x cards have no such field.
  • agents/<name>/agent.json is still merged over the default, and 0.3-form files are converted by parse_agent_card.
• The MCP bridge (agent_mcp_protocol_handler.py) now takes the reply text from the task's artifacts, or a failed task's status message. It used to return the whole task as a Python dict string, since ADK's executor answers with a task rather than a message.

Checks

• Venv installed from requirements.lock (a2a-sdk 1.2.1, google-adk 2.11.0). The server started through auth_server.py, which starts ADK with --a2a. Through the auth server:

  • GET /a2a/creative_agent/.well-known/agent-card.json → 200. The card is in 1.x form, with supportedInterfaces[0].url = http://localhost:8000/a2a/creative_agent.
  • message/send (0.3) to pisac_root → task completed, reply "Zdravo".
  • SendMessage (1.x, A2A-Version: 1.0) to pisac_root → TASK_STATE_COMPLETED, with a reply.
  • MCP tools/call on chess_mate_root (/agents/chess_mate_root/mcp) → [{"type": "text", "text": "chess"}]. Before the bridge change, the same call returned the whole task as a dict string.

• shared/test/test_a2a_routes.py, 11 tests. It mounts the routes on a FastAPI app with an echo executor and checks:

  • the routes and the card for each agent;
  • that the card names /a2a/{agent};
  • the agent.json merge;
  • message/send and SendMessage both answered;
  • that a failing agent doesn't stop the others;
  • the bridge's text extraction.

  A dependency bump that breaks this now fails CI.

• Full suite: 1467 OK. gen_docs.py --check: OK.

Docs

gen_docs.py drift flags docs/dev/a2a.md and docs/dev/architecture.md.

• architecture.md only says the agent server starts with --a2a, which still holds.
• a2a.md is being rewritten in docs(dev): verify a2a guide against code and remove migrated status (#206) #217, so I left it out of this PR to avoid a conflict. What it needs to say after this:
  • Methods: both the 1.x methods (SendMessage, SendStreamingMessage, GetTask, CancelTask, GetExtendedAgentCard) and the 0.3 ones (message/send, message/stream, tasks/get, ...) are accepted.
  • Card: it is served in the 1.x form (supportedInterfaces, not url). agent.json may be in either form.
  • Card address: the default card points at http://localhost:8000/a2a/{agent}.