feat: A2A (Agent2Agent) protocol support — Agent Card server + client tool

[cardinalconseils]

Summary

Implements A2A (Agent2Agent) protocol support for Hermes Agent, enabling inter-agent collaboration with any A2A-compliant system (Google ADK, LangGraph, CrewAI, etc.).

MCP connects agents to tools. A2A connects agents to other agents.

Tier 1 — Agent Card Server (plugin)

• GET /.well-known/agent-card.json — serves A2A-compliant Agent Card, auto-generated from live Hermes state (SOUL.md identity, installed skills, version, auth config). Public, no auth required.
• POST /a2a — JSON-RPC 2.0 endpoint accepting:
  • message/send — create task, run agent, return Task with Artifacts
  • task/get — get task status by ID
  • task/list — list recent tasks
  • task/cancel — cancel a running task
• Routes through existing APIServerAdapter._run_agent() — same code path as /v1/chat/completions. Supports blocking and non-blocking modes.
• Plugin hooks into API server via try/except import in connect() — never breaks startup if plugin is missing.
• A2A endpoints advertised in /v1/capabilities.

Tier 2 — A2A Client Tool

• a2a_send tool (delegation toolset) lets Hermes act as an A2A client:
  • discover — fetch and parse Agent Card from remote agent
  • send — send message/task to remote agent, receive structured artifacts
  • status — poll task status
  • list — list recent tasks on remote agent
  • cancel — cancel a running task
• Uses stdlib urllib only — no external dependencies.
• Enables the failed payment scenario: Support Agent discovers a Fraud Agent card, sends a task, receives a risk score artifact.

Governance (config + existing features, no new code)

Governance Concern	Hermes Feature
Agent identity	SOUL.md
Permissions	Toolset restrictions + DELEGATE_BLOCKED_TOOLS
Auth	Bearer token reuse from API_SERVER_KEY
Auditability	Session transcripts in state.db
Cost tracking	hermes insights + delegation cost rollup
Failure handling	Subagent timeout + auto-deny + zombie cleanup

Files

New files (7):

• plugins/a2a/__init__.py — Plugin entry point
• plugins/a2a/agent_card.py — Agent Card generation from SOUL.md + skills + config
• plugins/a2a/models.py — A2A data model (Task, Message, Part, Artifact) with lifecycle
• plugins/a2a/jsonrpc.py — JSON-RPC endpoint handlers
• tools/a2a_send_tool.py — A2A client tool (discover/send/status/list/cancel)
• tests/plugins/test_a2a.py — 20 unit tests
• tests/tools/test_a2a_send.py — 18 unit tests

Modified files (2):

• gateway/platforms/api_server.py — A2A route registration + capabilities advertisement
• toolsets.py — a2a_send added to delegation toolset

Test Results

38 tests pass (20 Agent Card/models, 18 client tool). Security gate clean.

Security

• Agent Card endpoint: no auth required (public discovery, like /.well-known/)
• JSON-RPC endpoint: requires Bearer auth (same API_SERVER_KEY check)
• No secrets in any output (tested with 12 forbidden patterns)
• No SQL injection, no auth bypass, no XSS patterns
• Uses hmac.compare_digest for auth token comparison (same as existing API server)

Spec Reference

https://github.com/a2aproject/A2A/blob/main/docs/specification.md

[alt-glitch]

This was generated by AI during triage.

Related: #514 (the A2A feature request), #46241 (A2A server adapter approach), #41711 (consolidated single-plugin A2A approach). This PR takes a third approach — a plugin Agent Card server (plugins/a2a/) plus a stdlib-only a2a_send client tool. It's not a duplicate of those, but the A2A cluster is now saturated with competing open PRs; a maintainer should pick one direction.