hivemind-a2a-agent-plugin

An agent-protocol plugin for hivemind-core. It answers the questions a hive sends to its master with an external agent that speaks the Agent2Agent (A2A) protocol.

• What you can do
• Answer questions with an A2A agent
• Answer an agent that asks a question back
• Handle a remote agent that is down, slow or fails
• Reference
• Deployment
• Troubleshooting

What you can do

A2A is an open protocol that lets one agent call another, whatever framework each one uses inside. An A2A agent publishes an agent card at /.well-known/agent-card.json. The card names the agent, its skills, and the URL that accepts messages.

With this plugin as its agent protocol, hivemind-core answers each QUERY from a satellite or a child node. It sends the question to one A2A agent and returns the answer as speak chunks. The plugin uses the a2a-sdk client. It reads the card, picks the transport the card offers, and speaks A2A 1.0 or 0.3, whichever the agent supports.

The plugin answers QUERY and CASCADE messages, through the natural_language_query seam of the HiveMind agent protocol. A plain BUS recognizer_loop:utterance gets no answer. That is the message that hivemind-client terminal and a voice satellite send. See Troubleshooting.

Answer questions with an A2A agent

Use this setup when you run hivemind-core and have an A2A agent: a LangGraph or CrewAI service, a hosted assistant, or an OVOS box running ovos-persona-server with --a2a-base-url. The hive then answers with that agent.

1. Install the plugin next to hivemind-core:

   pip install --pre hivemind-a2a-agent-plugin
   

2. Select it in ~/.config/hivemind-core/server.json. agent_protocol.module names the plugin, and the key with the plugin name holds its settings:

   {
     "agent_protocol": {
       "module": "hivemind-a2a-agent-plugin",
       "hivemind-a2a-agent-plugin": {
         "agent_url": "http://127.0.0.1:8337/a2a",
         "timeout": 30
       }
     }
   }
   

3. Add a client, allow the message type it will ask with, and start the hub:

   hivemind-core add-client --name kitchen-sat --access-key kitchen-key \
       --password 'Correct-Horse-Battery-77' --allow recognizer_loop:utterance
   hivemind-core listen
   

On start, the log names the plugin:

hivemind_core.service:run:441 - INFO - Agent protocol: A2AAgentProtocol
hivemind_a2a_agent_plugin:__post_init__:99 - INFO - A2AAgentProtocol: configured for A2A agent at http://127.0.0.1:8337/a2a

4. Ask a question with a QUERY. This client script uses hivemind-bus-client:

   import threading
   import uuid
   
   from hivemind_bus_client.client import HiveMessageBusClient
   from hivemind_bus_client.identity import NodeIdentity
   from hivemind_bus_client.message import HiveMessage, HiveMessageType
   from ovos_bus_client.message import Message
   
   client = HiveMessageBusClient(key="kitchen-key", password="Correct-Horse-Battery-77",
                                 host="ws://127.0.0.1", port=5678,
                                 identity=NodeIdentity(app_name="a2a-query-example",
                                                       shared_fallback=False))
   done = threading.Event()
   
   
   def on_query(message: HiveMessage):
       payload = message.payload
       while isinstance(payload, HiveMessage):
           payload = payload.payload
       if payload.msg_type == "speak":
           print(payload.data["utterance"])
       elif payload.msg_type in ("hive.query.complete", "hive.query.timeout"):
           done.set()
   
   
   client.on(HiveMessageType.QUERY, on_query)
   client.connect()
   client.wait_for_handshake()
   question = Message("recognizer_loop:utterance",
                      {"utterances": ["what is the capital of portugal"], "lang": "en-US"})
   client.emit(HiveMessage(HiveMessageType.QUERY,
                           payload=HiveMessage(HiveMessageType.BUS, payload=question),
                           metadata={"query_id": str(uuid.uuid4())}))
   done.wait(30)
   client.close()
   

On success, the agent's answer is printed. With a test agent that echoes its input and its A2A contextId, the hive answered:

echo[8a086dca-691d-499b-8def-a67bc59f3ddd]: what is the capital of portugal

NodeIdentity(app_name="a2a-query-example", shared_fallback=False) gives the script its own identity under ~/.config/hivemind/a2a-query-example/. Without it, the client uses the shared ~/.config/hivemind/_identity.json. That fails when the shared identity already pinned another hub at the same address, and it changes key files that other HiveMind clients of the same user read.

Answer an agent that asks a question back

An agent can stop a task in TASK_STATE_INPUT_REQUIRED or TASK_STATE_AUTH_REQUIRED and ask for more. The satellite hears the question. With a test agent that asks for a city, the hive answered:

which city?

The next QUERY starts a new A2A conversation, because hivemind-core passes no session id. The agent receives the answer without the question that came before it.

Handle a remote agent that is down, slow or fails

The plugin never leaves a query without an answer. When the agent cannot answer, the satellite receives an error sentence, and then the end of the query.

A slow agent within timeout still answers. An agent that took 12 seconds, with timeout: 30, was answered after 13.8 seconds.

Reference

Settings

The plugin also reads the same keys from the OVOS configuration at hivemind.a2a_agent. Keys in server.json win.

What is sent and received

hivemind-core calls natural_language_query(utterance, lang) without a session id. The plugin then sends no contextId, the agent assigns a new one, and every query starts a new A2A conversation. A2A 1.0.1 section 3.4.1 says a client SHOULD NOT send a contextId it generated itself. A caller that passes session_id directly keeps the agent-assigned contextId of that session for its later queries.

Deployment

The plugin runs inside hivemind-core and opens outbound HTTP connections only. A2A has no built-in encryption. Use an https:// agent_url when the agent is on another machine, and auth_header when the agent asks for a token.

Troubleshooting

Development

git clone https://github.com/JarbasHiveMind/hivemind-a2a-agent-plugin
cd hivemind-a2a-agent-plugin
uv venv && uv pip install --prerelease=allow -e ".[test]"
.venv/bin/pytest tests/

The tests start a real a2a-sdk agent (tests/a2a_agent.py) in a subprocess. Run it alone to try the plugin by hand: python tests/a2a_agent.py 9999.

Related projects

• JarbasHiveMind/HiveMind-core: the HiveMind master this plugin connects to.
• OpenVoiceOS/ovos-a2a-agent-plugin: the same A2A client as an ovos-persona solver.

Credits

Funded by NGI0 Commons Fund / NLnet under grant agreement No 101135429, through the European Commission's Next Generation Internet programme.