A2A converse extension, version 1

The Radioso converse extension adds two things to an agent’s A2A endpoint. On the Agent Card it points to the agent’s MCP endpoint and lists each exposed routine with the JSON Schema of its input, because an A2A skill carries no input schema. On replies it carries the same machine-readable envelope an MCP caller gets from ask_agent: answer coverage, ownership, routine status, citations, and who wrote the message.

The extension is identified by this URI:

https://docs.radioso.ai/reference/a2a-converse-extension/v1

The URI does three jobs. The card lists it under capabilities.extensions, a client sends it in the A2A-Extensions header to switch the extension on, and it is the key under which reply metadata travels. The version is part of the path, and this page describes version 1.

Read it in the card

An agent’s card lists the extension with required: false, so a client that ignores it still works. Its params have two members:

{
  "uri": "https://docs.radioso.ai/reference/a2a-converse-extension/v1",
  "description": "Radioso conversation extension. params.mcp names this agent's MCP endpoint and server card; params.tools lists each exposed routine with the JSON Schema of its input.",
  "required": false,
  "params": {
    "mcp": {
      "url": "https://mcp.radioso.ai/mcp/a/ag_7Qb3nT1xK9wZs2Pv0Lm4Rd",
      "serverCardUrl": "https://mcp.radioso.ai/mcp/a/ag_7Qb3nT1xK9wZs2Pv0Lm4Rd/server-card"
    },
    "tools": [
      {
        "toolName": "book_table",
        "description": "Book a table for a date and party size.",
        "inputSchema": {
          "type": "object",
          "properties": {
            "date": { "type": "string" },
            "partySize": { "type": "integer" }
          },
          "required": ["date", "partySize"]
        }
      }
    ]
  }
}

The tools entries are the same descriptors GET /api/v1/mcp/converse/tools returns, so one schema describes a routine whichever door a caller uses.

Activate it

Send the extension’s URI in the A2A-Extensions request header. Several URIs go in one header, separated by commas:

curl https://api.radioso.ai/a2a/ag_7Qb3nT1xK9wZs2Pv0Lm4Rd \
  -H 'Content-Type: application/json' \
  -H 'A2A-Version: 1.0' \
  -H 'A2A-Extensions: https://docs.radioso.ai/reference/a2a-converse-extension/v1' \
  -d '{ "jsonrpc": "2.0", "id": 1, "method": "SendMessage", "params": { "message": { "messageId": "4f0c1a52-9e0b-4b8e-a7a5-0d6f0f6f2a10", "role": "ROLE_USER", "parts": [{ "text": "How long do I have to return an order?" }] } } }'

The response carries an A2A-Extensions header listing the extensions the server activated for that request. A request without the header, or without this URI in it, gets replies with no extension metadata.

Read the reply metadata

When the extension is active, every Message, every task status.message, and every artifact carries a metadata object keyed by the extension’s URI, and lists that URI in its extensions field:

{
  "messageId": "b81d6c0e-2f4a-4c77-9f0e-3f2d5a7c1e44",
  "contextId": "Zk3xQm…",
  "role": "ROLE_AGENT",
  "parts": [{ "text": "You can return an order within 30 days of delivery." }],
  "extensions": ["https://docs.radioso.ai/reference/a2a-converse-extension/v1"],
  "metadata": {
    "https://docs.radioso.ai/reference/a2a-converse-extension/v1": {
      "author": "agent",
      "answerCoverage": {
        "availability": "assessed",
        "coverage": "answered",
        "reason": "sufficient_evidence",
        "originatingTurnId": "c3…",
        "originatingRequestId": "c3…"
      },
      "ownership": { "state": "ai_owned", "suppressed": false },
      "citations": [
        { "documentId": "0c8e…", "chunkId": "5d21…", "title": "Returns policy" }
      ]
    }
  }
}

routine.status is active, waiting_for_input, waiting_for_approval, completed, or abandoned. routine.pendingInput lists every required slot still open plus the current step’s optional ones, each with key, type, required, and description, so a client can supply them all in one message. While a task is TASK_STATE_INPUT_REQUIRED for a routine, this list is what the agent is waiting for.

answerCoverage, ownership, routine, invocation, and citations are the fields of the agent reply envelope; what ask_agent returns explains how a calling model uses them. author is specific to this extension, because A2A has no way to tell a person’s reply from the agent’s.

Common failure modes

• No metadata on replies. The request did not activate the extension. Check that A2A-Extensions holds the exact URI, including the v1 at the end, and that the response header echoes it.
• tools is empty. The operator has not exposed any routine as a tool. A plain message through the ask_agent skill always works.
• A pendingInput slot the client has no value for. Ask the user, then send the answer as text in the next message, with the taskId or the contextId.

Read next

• Connect an A2A client — send a message, keep the contextId, and follow a task from input-required to completed.
• Publish and open an agent — the whole agent card and the walk-in access it describes.
• MCP server — the MCP endpoint that params.mcp points to.