feat: add A2A task delegation over mTLS (#804)

agent/a2a_mtls.py

@@ -29,6 +29,8 @@ import logging
 import os
 import ssl
 import threading
+import time
+import uuid
 from http.server import BaseHTTPRequestHandler, HTTPServer
 from pathlib import Path
 from typing import Any, Callable, Dict, Optional
@@ -441,3 +443,244 @@ class A2AMTLSClient:
     def post(self, url: str, json: Optional[Dict[str, Any]] = None, **kwargs: Any) -> Dict[str, Any]:
         data = (__import__("json").dumps(json).encode() if json is not None else None)
         return self._request("POST", url, data=data, **kwargs)
+
+
+# ---------------------------------------------------------------------------
+# Structured A2A task delegation over mTLS
+# ---------------------------------------------------------------------------
+
+_TERMINAL_TASK_STATES = {"completed", "failed", "canceled", "rejected"}
+
+
+def _iso_now() -> str:
+    return time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
+
+
+def _task_status(state: str, message: str) -> Dict[str, Any]:
+    return {
+        "state": state,
+        "message": message,
+        "timestamp": _iso_now(),
+    }
+
+
+def _coerce_artifact(result: Any) -> Dict[str, Any]:
+    if isinstance(result, dict):
+        if "text" in result:
+            return result
+        if "artifact" in result and isinstance(result["artifact"], dict):
+            return result["artifact"]
+    return {"text": str(result)}
+
+
+def _build_task_record(task_id: str, task: str, requester: Optional[str], metadata: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
+    return {
+        "taskId": task_id,
+        "task": task,
+        "requester": requester,
+        "metadata": metadata or {},
+        "artifacts": [],
+        "status": _task_status("submitted", "Task submitted"),
+    }
+
+
+def _default_agent_card(host: str, port: int) -> Dict[str, Any]:
+    base_url = f"https://{host}:{port}"
+    try:
+        from agent.agent_card import build_agent_card
+        from dataclasses import asdict
+
+        card = asdict(build_agent_card())
+    except Exception as exc:  # pragma: no cover - fallback only exercised when card build breaks
+        logger.warning("Falling back to minimal agent card: %s", exc)
+        card = {
+            "name": os.environ.get("HERMES_AGENT_NAME", "hermes"),
+            "description": "Hermes A2A task server",
+            "version": "unknown",
+        }
+    card["url"] = base_url
+    card["a2aTaskEndpoint"] = f"{base_url}/a2a/rpc"
+    return card
+
+
+def _default_local_hermes_executor(task_payload: Dict[str, Any]) -> Dict[str, Any]:
+    task_text = str(task_payload.get("task", "")).strip()
+    if not task_text:
+        return {"text": ""}
+    from run_agent import AIAgent
+
+    agent = AIAgent(quiet_mode=True)
+    result = agent.chat(task_text)
+    return {
+        "text": result,
+        "metadata": {"executor": "local-hermes"},
+    }
+
+
+class A2ATaskServer:
+    """JSON-RPC A2A task server running over the routing mTLS server."""
+
+    def __init__(
+        self,
+        cert: str | Path,
+        key: str | Path,
+        ca: str | Path,
+        host: str = "127.0.0.1",
+        port: int = 9443,
+        executor: Optional[Callable[[Dict[str, Any]], Dict[str, Any]]] = None,
+        card_factory: Optional[Callable[[], Dict[str, Any]]] = None,
+    ) -> None:
+        self.host = host
+        self.port = port
+        self._server = A2AMTLSServer(cert=cert, key=key, ca=ca, host=host, port=port)
+        self._executor = executor or _default_local_hermes_executor
+        self._card_factory = card_factory or (lambda: _default_agent_card(self.host, self.port))
+        self._tasks: Dict[str, Dict[str, Any]] = {}
+        self._lock = threading.Lock()
+        self._server.add_route("/.well-known/agent-card.json", self._handle_agent_card)
+        self._server.add_route("/agent-card.json", self._handle_agent_card)
+        self._server.add_route("/a2a/rpc", self._handle_rpc)
+
+    def __enter__(self) -> "A2ATaskServer":
+        self.start()
+        return self
+
+    def __exit__(self, *_: Any) -> None:
+        self.stop()
+
+    def start(self) -> None:
+        self._server.start()
+
+    def stop(self) -> None:
+        self._server.stop()
+
+    def _handle_agent_card(self, payload: Dict[str, Any], *, peer_cn: str | None = None) -> Dict[str, Any]:
+        return self._card_factory()
+
+    def _handle_rpc(self, payload: Dict[str, Any], *, peer_cn: str | None = None) -> Dict[str, Any]:
+        req_id = payload.get("id")
+        if payload.get("jsonrpc") != "2.0":
+            return {"jsonrpc": "2.0", "id": req_id, "error": {"code": -32600, "message": "invalid jsonrpc version"}}
+
+        method = payload.get("method")
+        params = payload.get("params") or {}
+        try:
+            if method == "tasks/send":
+                result = self._rpc_send_task(params, peer_cn=peer_cn)
+            elif method == "tasks/get":
+                result = self._rpc_get_task(params)
+            else:
+                return {"jsonrpc": "2.0", "id": req_id, "error": {"code": -32601, "message": f"unknown method: {method}"}}
+        except Exception as exc:
+            logger.exception("A2A task RPC failed: %s", exc)
+            return {"jsonrpc": "2.0", "id": req_id, "error": {"code": -32000, "message": str(exc)}}
+        return {"jsonrpc": "2.0", "id": req_id, "result": result}
+
+    def _rpc_send_task(self, params: Dict[str, Any], *, peer_cn: str | None = None) -> Dict[str, Any]:
+        task_text = str(params.get("task", "")).strip()
+        if not task_text:
+            raise ValueError("task is required")
+        task_id = params.get("taskId") or uuid.uuid4().hex
+        requester = params.get("requester") or peer_cn
+        metadata = dict(params.get("metadata") or {})
+        if peer_cn:
+            metadata.setdefault("peer_cn", peer_cn)
+        record = _build_task_record(task_id, task_text, requester, metadata)
+        with self._lock:
+            self._tasks[task_id] = record
+        worker = threading.Thread(target=self._run_task, args=(task_id,), daemon=True, name=f"a2a-task-{task_id[:8]}")
+        worker.start()
+        return self._copy_task(task_id)
+
+    def _rpc_get_task(self, params: Dict[str, Any]) -> Dict[str, Any]:
+        task_id = str(params.get("taskId", "")).strip()
+        if not task_id:
+            raise ValueError("taskId is required")
+        return self._copy_task(task_id)
+
+    def _copy_task(self, task_id: str) -> Dict[str, Any]:
+        with self._lock:
+            if task_id not in self._tasks:
+                raise KeyError(f"unknown taskId: {task_id}")
+            return json.loads(json.dumps(self._tasks[task_id]))
+
+    def _run_task(self, task_id: str) -> None:
+        with self._lock:
+            task = self._tasks[task_id]
+            task["status"] = _task_status("working", "Task is running")
+            task_payload = {
+                "taskId": task["taskId"],
+                "task": task["task"],
+                "requester": task.get("requester"),
+                "metadata": dict(task.get("metadata") or {}),
+            }
+        try:
+            result = self._executor(task_payload)
+            artifact = _coerce_artifact(result)
+            with self._lock:
+                task = self._tasks[task_id]
+                task["artifacts"] = [artifact]
+                task["status"] = _task_status("completed", "Task completed")
+        except Exception as exc:
+            with self._lock:
+                task = self._tasks[task_id]
+                task["status"] = _task_status("failed", f"Task failed: {exc}")
+
+
+class A2ATaskClient(A2AMTLSClient):
+    """Client helper for A2A JSON-RPC task send/get flows."""
+
+    def discover_card(self, base_url: str) -> Dict[str, Any]:
+        return self.get(f"{base_url.rstrip('/')}/.well-known/agent-card.json")
+
+    def _rpc_call(self, base_url: str, method: str, params: Dict[str, Any]) -> Dict[str, Any]:
+        payload = {
+            "jsonrpc": "2.0",
+            "id": uuid.uuid4().hex,
+            "method": method,
+            "params": params,
+        }
+        response = self.post(f"{base_url.rstrip('/')}/a2a/rpc", json=payload)
+        if "error" in response:
+            error = response["error"]
+            raise RuntimeError(error.get("message") or str(error))
+        return response.get("result", {})
+
+    def send_task(
+        self,
+        base_url: str,
+        *,
+        task: str,
+        requester: str | None = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> Dict[str, Any]:
+        return self._rpc_call(
+            base_url,
+            "tasks/send",
+            {
+                "task": task,
+                "requester": requester,
+                "metadata": metadata or {},
+            },
+        )
+
+    def get_task(self, base_url: str, task_id: str) -> Dict[str, Any]:
+        return self._rpc_call(base_url, "tasks/get", {"taskId": task_id})
+
+    def wait_for_task(
+        self,
+        base_url: str,
+        task_id: str,
+        *,
+        timeout: float = 30.0,
+        poll_interval: float = 0.5,
+    ) -> Dict[str, Any]:
+        deadline = time.monotonic() + timeout
+        while True:
+            task = self.get_task(base_url, task_id)
+            state = str(((task.get("status") or {}).get("state") or "")).lower()
+            if state in _TERMINAL_TASK_STATES:
+                return task
+            if time.monotonic() >= deadline:
+                raise TimeoutError(f"Timed out waiting for task {task_id}")
+            time.sleep(poll_interval)

hermes_cli/a2a_cmd.py

@@ -0,0 +1,132 @@
+"""CLI helpers for A2A task delegation."""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import sys
+import time
+from pathlib import Path
+from typing import Any
+
+from agent.a2a_mtls import A2ATaskClient, A2ATaskServer
+from hermes_cli.config import get_hermes_home
+
+
+def _registry_path() -> Path:
+    return get_hermes_home() / "a2a_agents.json"
+
+
+def _default_identity_paths() -> tuple[str, str, str]:
+    hermes_home = get_hermes_home()
+    agent_name = os.environ.get("HERMES_AGENT_NAME", "hermes").lower()
+    cert = os.environ.get(
+        "HERMES_A2A_CERT",
+        str(hermes_home / "pki" / "agents" / agent_name / f"{agent_name}.crt"),
+    )
+    key = os.environ.get(
+        "HERMES_A2A_KEY",
+        str(hermes_home / "pki" / "agents" / agent_name / f"{agent_name}.key"),
+    )
+    ca = os.environ.get(
+        "HERMES_A2A_CA",
+        str(hermes_home / "pki" / "ca" / "fleet-ca.crt"),
+    )
+    return cert, key, ca
+
+
+def load_agent_registry(path: Path | None = None) -> dict[str, Any]:
+    registry_path = path or _registry_path()
+    if not registry_path.exists():
+        return {}
+    return json.loads(registry_path.read_text(encoding="utf-8"))
+
+
+def resolve_agent_url(agent: str, *, registry_path: Path | None = None) -> str:
+    key = re.sub(r"[^A-Za-z0-9]+", "_", agent).upper()
+    env_value = os.getenv(f"HERMES_A2A_{key}_URL")
+    if env_value:
+        return env_value
+
+    registry = load_agent_registry(registry_path)
+    entry = registry.get(agent)
+    if isinstance(entry, str) and entry:
+        return entry
+    if isinstance(entry, dict):
+        url = entry.get("url") or entry.get("base_url") or entry.get("card_url")
+        if url:
+            return str(url)
+    if agent.startswith("https://") or agent.startswith("http://"):
+        return agent
+    raise SystemExit(f"Unknown A2A agent '{agent}'. Set HERMES_A2A_{key}_URL or add it to {_registry_path()}.")
+
+
+def _print(data: dict[str, Any]) -> None:
+    print(json.dumps(data, indent=2, ensure_ascii=False))
+
+
+def cmd_send(args) -> None:
+    base_url = args.url or resolve_agent_url(args.agent)
+    cert, key, ca = args.cert, args.key, args.ca
+    if not (cert and key and ca):
+        cert, key, ca = _default_identity_paths()
+    client = A2ATaskClient(cert=cert, key=key, ca=ca)
+    card = client.discover_card(base_url)
+    task = client.send_task(
+        base_url,
+        task=args.task,
+        requester=args.requester,
+        metadata={"agent": args.agent},
+    )
+    if args.wait:
+        task = client.wait_for_task(
+            base_url,
+            task["taskId"],
+            timeout=args.timeout,
+            poll_interval=args.poll_interval,
+        )
+    _print({
+        "agent": args.agent,
+        "url": base_url,
+        "card": card,
+        "task": task,
+    })
+
+
+def cmd_status(args) -> None:
+    base_url = args.url or resolve_agent_url(args.agent)
+    cert, key, ca = args.cert, args.key, args.ca
+    if not (cert and key and ca):
+        cert, key, ca = _default_identity_paths()
+    client = A2ATaskClient(cert=cert, key=key, ca=ca)
+    task = client.get_task(base_url, args.task_id)
+    _print({"agent": args.agent, "url": base_url, "task": task})
+
+
+def cmd_serve(args) -> None:
+    cert, key, ca = args.cert, args.key, args.ca
+    if not (cert and key and ca):
+        cert, key, ca = _default_identity_paths()
+    server = A2ATaskServer(cert=cert, key=key, ca=ca, host=args.host, port=args.port)
+    server.start()
+    print(f"A2A task server listening on https://{args.host}:{args.port}")
+    try:
+        while True:
+            time.sleep(1)
+    except KeyboardInterrupt:
+        server.stop()
+
+
+def cmd_a2a(args) -> None:
+    command = getattr(args, "a2a_command", None) or "send"
+    if command == "send":
+        cmd_send(args)
+        return
+    if command == "status":
+        cmd_status(args)
+        return
+    if command == "serve":
+        cmd_serve(args)
+        return
+    raise SystemExit(f"Unknown a2a command: {command}")

hermes_cli/main.py

@@ -173,6 +173,13 @@ from hermes_constants import OPENROUTER_BASE_URL
 logger = logging.getLogger(__name__)
 
 
+def cmd_a2a(args):
+    """Dispatch A2A CLI subcommands lazily to avoid heavy imports at startup."""
+    from hermes_cli.a2a_cmd import cmd_a2a as _cmd_a2a
+
+    return _cmd_a2a(args)
+
+
 def _relative_time(ts) -> str:
     """Format a timestamp as relative time (e.g., '2h ago', 'yesterday')."""
     if not ts:
@@ -4781,6 +4788,45 @@ For more help on a command:
 
     gateway_parser.set_defaults(func=cmd_gateway)
     
+    # =========================================================================
+    # a2a command
+    # =========================================================================
+    a2a_parser = subparsers.add_parser(
+        "a2a",
+        help="A2A task delegation over mutual TLS",
+        description="Send, inspect, and serve structured A2A tasks between Hermes agents",
+    )
+    a2a_subparsers = a2a_parser.add_subparsers(dest="a2a_command")
+
+    a2a_send = a2a_subparsers.add_parser("send", help="Send an A2A task to another agent")
+    a2a_send.add_argument("--agent", required=True, help="Agent alias or URL (for example: allegro)")
+    a2a_send.add_argument("--task", required=True, help="Task text to delegate")
+    a2a_send.add_argument("--url", help="Explicit base URL for the remote agent")
+    a2a_send.add_argument("--requester", default=None, help="Requester label included in task metadata")
+    a2a_send.add_argument("--wait", action="store_true", help="Poll until the task reaches a terminal state")
+    a2a_send.add_argument("--timeout", type=float, default=30.0, help="Wait timeout in seconds (default: 30)")
+    a2a_send.add_argument("--poll-interval", type=float, default=0.5, help="Polling interval in seconds while waiting (default: 0.5)")
+    a2a_send.add_argument("--cert", default=None, help="Client certificate path (defaults from HERMES_A2A_CERT)")
+    a2a_send.add_argument("--key", default=None, help="Client private key path (defaults from HERMES_A2A_KEY)")
+    a2a_send.add_argument("--ca", default=None, help="Fleet CA certificate path (defaults from HERMES_A2A_CA)")
+
+    a2a_status = a2a_subparsers.add_parser("status", help="Fetch the current status of an A2A task")
+    a2a_status.add_argument("--agent", required=True, help="Agent alias or URL (for example: allegro)")
+    a2a_status.add_argument("--task-id", required=True, help="Task identifier returned by a2a send")
+    a2a_status.add_argument("--url", help="Explicit base URL for the remote agent")
+    a2a_status.add_argument("--cert", default=None, help="Client certificate path (defaults from HERMES_A2A_CERT)")
+    a2a_status.add_argument("--key", default=None, help="Client private key path (defaults from HERMES_A2A_KEY)")
+    a2a_status.add_argument("--ca", default=None, help="Fleet CA certificate path (defaults from HERMES_A2A_CA)")
+
+    a2a_serve = a2a_subparsers.add_parser("serve", help="Run the local A2A task server")
+    a2a_serve.add_argument("--host", default=os.environ.get("HERMES_A2A_HOST", "127.0.0.1"), help="Bind host (default: HERMES_A2A_HOST or 127.0.0.1)")
+    a2a_serve.add_argument("--port", type=int, default=int(os.environ.get("HERMES_A2A_PORT", "9443")), help="Bind port (default: HERMES_A2A_PORT or 9443)")
+    a2a_serve.add_argument("--cert", default=None, help="Server certificate path (defaults from HERMES_A2A_CERT)")
+    a2a_serve.add_argument("--key", default=None, help="Server private key path (defaults from HERMES_A2A_KEY)")
+    a2a_serve.add_argument("--ca", default=None, help="Fleet CA certificate path (defaults from HERMES_A2A_CA)")
+
+    a2a_parser.set_defaults(func=cmd_a2a)
+    
     # =========================================================================
     # setup command
     # =========================================================================

run_agent.py

@@ -20,7 +20,6 @@ Usage:
     response = agent.run_conversation("Tell me about the latest Python updates")
 """
 
-import ast
 import asyncio
 import base64
 import concurrent.futures
@@ -3329,119 +3328,6 @@ class AIAgent:
 
     _VALID_API_ROLES = frozenset({"system", "user", "assistant", "tool", "function", "developer"})
 
-    @staticmethod
-    def _normalize_tool_call_arguments(arguments: Any) -> tuple[str, bool]:
-        """Return ``(normalized_text, is_complete)`` for tool-call arguments.
-
-        Conservative by design: repairs harmless formatting quirks common in
-        Gemma 4 / Ollama output (whitespace, trailing commas, Python-style
-        single-quoted dicts, bare key/value pairs) but does NOT auto-close
-        truncated JSON objects. Truly incomplete fragments must remain marked
-        incomplete so the agent can retry instead of silently dropping fields.
-        """
-        if isinstance(arguments, (dict, list)):
-            return json.dumps(arguments, ensure_ascii=False, separators=(",", ":")), True
-        if arguments is None:
-            return "{}", True
-        if not isinstance(arguments, str):
-            arguments = str(arguments)
-
-        text = arguments.strip()
-        if not text:
-            return "{}", True
-
-        def _parse_candidate(candidate: str):
-            try:
-                return json.loads(candidate)
-            except (json.JSONDecodeError, TypeError, ValueError):
-                pass
-            try:
-                return ast.literal_eval(candidate)
-            except (SyntaxError, ValueError):
-                return None
-
-        candidates: list[str] = [text]
-
-        trimmed_trailing_commas = re.sub(r",\s*([}\]])", r"\1", text)
-        if trimmed_trailing_commas != text:
-            candidates.append(trimmed_trailing_commas)
-
-        if ":" in text and not text.startswith(("{", "[")):
-            wrapped = "{" + text + "}"
-            candidates.append(wrapped)
-            quoted_keys = re.sub(
-                r'([\{,]\s*)([A-Za-z_][A-Za-z0-9_\-]*)(\s*:)',
-                r'\1"\2"\3',
-                wrapped,
-            )
-            if quoted_keys != wrapped:
-                candidates.append(quoted_keys)
-                trimmed_quoted_keys = re.sub(r",\s*([}\]])", r"\1", quoted_keys)
-                if trimmed_quoted_keys != quoted_keys:
-                    candidates.append(trimmed_quoted_keys)
-
-        seen: set[str] = set()
-        for candidate in candidates:
-            if candidate in seen:
-                continue
-            seen.add(candidate)
-            parsed = _parse_candidate(candidate)
-            if isinstance(parsed, (dict, list)):
-                return json.dumps(parsed, ensure_ascii=False, separators=(",", ":")), True
-
-        return text, False
-
-    @staticmethod
-    def _merge_consecutive_assistant_tool_call_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-        """Merge adjacent assistant messages that each carry tool_calls.
-
-        Some providers emit parallel tool calls as multiple consecutive assistant
-        messages instead of a single assistant message with multiple tool calls.
-        Merge only adjacent assistant/tool-call messages; any non-assistant
-        boundary flushes the current batch.
-        """
-        merged: List[Dict[str, Any]] = []
-        pending: Optional[Dict[str, Any]] = None
-
-        def _flush_pending() -> None:
-            nonlocal pending
-            if pending is not None:
-                merged.append(pending)
-                pending = None
-
-        for msg in messages:
-            if not isinstance(msg, dict):
-                _flush_pending()
-                merged.append(msg)
-                continue
-
-            role = msg.get("role")
-            tool_calls = msg.get("tool_calls")
-            if role == "assistant" and isinstance(tool_calls, list) and tool_calls:
-                if pending is None:
-                    pending = copy.deepcopy(msg)
-                    continue
-
-                pending_tool_calls = pending.get("tool_calls")
-                if not isinstance(pending_tool_calls, list):
-                    pending_tool_calls = []
-                    pending["tool_calls"] = pending_tool_calls
-                pending_tool_calls.extend(copy.deepcopy(tool_calls))
-
-                pending_content = pending.get("content") or ""
-                current_content = msg.get("content") or ""
-                if pending_content and current_content:
-                    pending["content"] = pending_content + "\n" + current_content
-                elif current_content:
-                    pending["content"] = current_content
-                continue
-
-            _flush_pending()
-            merged.append(msg)
-
-        _flush_pending()
-        return merged
-
     @staticmethod
     def _sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """Fix orphaned tool_call / tool_result pairs before every LLM call.
@@ -3461,7 +3347,7 @@ class AIAgent:
                 )
                 continue
             filtered.append(msg)
-        messages = AIAgent._merge_consecutive_assistant_tool_call_messages(filtered)
+        messages = filtered
 
         surviving_call_ids: set = set()
         for msg in messages:
@@ -5368,9 +5254,12 @@ class AIAgent:
                 mock_tool_calls = []
                 for idx in sorted(tool_calls_acc):
                     tc = tool_calls_acc[idx]
-                    arguments, is_complete = self._normalize_tool_call_arguments(tc["function"]["arguments"])
-                    if not is_complete:
-                        has_truncated_tool_args = True
+                    arguments = tc["function"]["arguments"]
+                    if arguments and arguments.strip():
+                        try:
+                            json.loads(arguments)
+                        except json.JSONDecodeError:
+                            has_truncated_tool_args = True
                     mock_tool_calls.append(SimpleNamespace(
                         id=tc["id"],
                         type=tc["type"],
@@ -6674,7 +6563,6 @@ class AIAgent:
                     response_item_id if isinstance(response_item_id, str) else None,
                 )
 
-                normalized_args, _ = self._normalize_tool_call_arguments(tool_call.function.arguments)
                 tc_dict = {
                     "id": call_id,
                     "call_id": call_id,
@@ -6682,7 +6570,7 @@ class AIAgent:
                     "type": tool_call.type,
                     "function": {
                         "name": tool_call.function.name,
-                        "arguments": normalized_args,
+                        "arguments": tool_call.function.arguments
                     },
                 }
                 # Preserve extra_content (e.g. Gemini thought_signature) so it
@@ -10143,15 +10031,21 @@ class AIAgent:
                     # Handle empty strings as empty objects (common model quirk)
                     invalid_json_args = []
                     for tc in assistant_message.tool_calls:
-                        normalized_args, is_complete = self._normalize_tool_call_arguments(tc.function.arguments)
-                        tc.function.arguments = normalized_args
-                        if not is_complete:
-                            try:
-                                json.loads(normalized_args)
-                            except json.JSONDecodeError as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
-                            except Exception as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
+                        args = tc.function.arguments
+                        if isinstance(args, (dict, list)):
+                            tc.function.arguments = json.dumps(args)
+                            continue
+                        if args is not None and not isinstance(args, str):
+                            tc.function.arguments = str(args)
+                            args = tc.function.arguments
+                        # Treat empty/whitespace strings as empty object
+                        if not args or not args.strip():
+                            tc.function.arguments = "{}"
+                            continue
+                        try:
+                            json.loads(args)
+                        except json.JSONDecodeError as e:
+                            invalid_json_args.append((tc.function.name, str(e)))
                     
                     if invalid_json_args:
                         # Check if the invalid JSON is due to truncation rather

tests/agent/test_a2a_mtls.py

@@ -572,3 +572,94 @@ class TestA2AMTLSServerAndClient:
 
         assert not errors, f"Concurrent connection errors: {errors}"
         assert len(results) == 3
+
+
+@_requires_crypto
+class TestA2ATaskServerAndClient:
+    """Structured A2A task send/get flow over mTLS."""
+
+    @pytest.fixture(autouse=True)
+    def _pki(self, tmp_path):
+        ca_dir = tmp_path / "ca"
+        ca_dir.mkdir()
+        self.ca_crt, self.ca_key = _make_ca_keypair(ca_dir)
+        agent_dir = tmp_path / "agents"
+        agent_dir.mkdir()
+        self.srv_crt, self.srv_key = _make_agent_keypair(
+            agent_dir, "timmy", self.ca_crt, self.ca_key
+        )
+        self.cli_crt, self.cli_key = _make_agent_keypair(
+            agent_dir, "allegro", self.ca_crt, self.ca_key
+        )
+
+    @pytest.fixture()
+    def task_server(self):
+        from agent.a2a_mtls import A2ATaskServer
+
+        gate = threading.Event()
+
+        def analyze_executor(task: dict[str, object]) -> dict[str, object]:
+            gate.wait(timeout=2)
+            text = str(task.get("task", ""))
+            return {
+                "text": f"analysis:{text}",
+                "metadata": {"tool": "local-hermes-stub"},
+            }
+
+        port = _find_free_port()
+        server = A2ATaskServer(
+            cert=self.srv_crt,
+            key=self.srv_key,
+            ca=self.ca_crt,
+            host="127.0.0.1",
+            port=port,
+            executor=analyze_executor,
+        )
+        with server:
+            time.sleep(0.1)
+            yield server, port, gate
+
+    def test_task_send_get_and_completion_flow(self, task_server):
+        from agent.a2a_mtls import A2ATaskClient
+
+        server, port, gate = task_server
+        client = A2ATaskClient(cert=self.cli_crt, key=self.cli_key, ca=self.ca_crt)
+        base_url = f"https://127.0.0.1:{port}"
+
+        card = client.discover_card(base_url)
+        assert card["name"]
+
+        submitted = client.send_task(base_url, task="Analyze README.md", requester="timmy")
+        assert submitted["status"]["state"] in {"submitted", "working"}
+
+        in_flight = client.get_task(base_url, submitted["taskId"])
+        assert in_flight["status"]["state"] in {"submitted", "working"}
+
+        gate.set()
+        completed = client.wait_for_task(base_url, submitted["taskId"], timeout=5.0, poll_interval=0.05)
+        assert completed["status"]["state"] == "completed"
+        assert completed["artifacts"][0]["text"] == "analysis:Analyze README.md"
+
+    def test_failed_executor_marks_task_failed(self):
+        from agent.a2a_mtls import A2ATaskClient, A2ATaskServer
+
+        def failing_executor(task: dict[str, object]) -> dict[str, object]:
+            raise RuntimeError("boom")
+
+        port = _find_free_port()
+        server = A2ATaskServer(
+            cert=self.srv_crt,
+            key=self.srv_key,
+            ca=self.ca_crt,
+            host="127.0.0.1",
+            port=port,
+            executor=failing_executor,
+        )
+        with server:
+            time.sleep(0.1)
+            client = A2ATaskClient(cert=self.cli_crt, key=self.cli_key, ca=self.ca_crt)
+            base_url = f"https://127.0.0.1:{port}"
+            submitted = client.send_task(base_url, task="explode", requester="timmy")
+            failed = client.wait_for_task(base_url, submitted["taskId"], timeout=5.0, poll_interval=0.05)
+            assert failed["status"]["state"] == "failed"
+            assert "boom" in failed["status"]["message"]

tests/hermes_cli/test_a2a_cmd.py

@@ -0,0 +1,95 @@
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+
+def test_cmd_send_uses_registry_and_waits_for_terminal_task(tmp_path, monkeypatch, capsys):
+    hermes_home = tmp_path / ".hermes"
+    hermes_home.mkdir()
+    (hermes_home / "a2a_agents.json").write_text(
+        json.dumps({"allegro": {"url": "https://127.0.0.1:9443"}}),
+        encoding="utf-8",
+    )
+    monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+    from hermes_cli.a2a_cmd import cmd_a2a
+
+    class FakeClient:
+        def __init__(self, **kwargs):
+            self.kwargs = kwargs
+
+        def discover_card(self, base_url: str):
+            assert base_url == "https://127.0.0.1:9443"
+            return {"name": "allegro", "url": base_url}
+
+        def send_task(self, base_url: str, *, task: str, requester: str | None = None, metadata=None):
+            assert task == "analyze README"
+            return {"taskId": "task-123", "status": {"state": "submitted"}}
+
+        def wait_for_task(self, base_url: str, task_id: str, *, timeout: float, poll_interval: float):
+            assert task_id == "task-123"
+            return {
+                "taskId": task_id,
+                "status": {"state": "completed"},
+                "artifacts": [{"text": "README looks healthy"}],
+            }
+
+    args = argparse.Namespace(
+        a2a_command="send",
+        agent="allegro",
+        task="analyze README",
+        url=None,
+        wait=True,
+        timeout=5.0,
+        poll_interval=0.01,
+        requester="timmy",
+        cert="cert.pem",
+        key="key.pem",
+        ca="ca.pem",
+    )
+
+    with patch("hermes_cli.a2a_cmd.A2ATaskClient", FakeClient):
+        cmd_a2a(args)
+
+    result = json.loads(capsys.readouterr().out)
+    assert result["agent"] == "allegro"
+    assert result["card"]["name"] == "allegro"
+    assert result["task"]["status"]["state"] == "completed"
+    assert result["task"]["artifacts"][0]["text"] == "README looks healthy"
+
+
+def test_resolve_agent_url_supports_env_override(monkeypatch):
+    monkeypatch.setenv("HERMES_A2A_ALLEGRO_URL", "https://fleet-allegro:9443")
+    from hermes_cli.a2a_cmd import resolve_agent_url
+
+    assert resolve_agent_url("allegro") == "https://fleet-allegro:9443"
+
+
+def test_cmd_send_requires_known_agent(tmp_path, monkeypatch):
+    hermes_home = tmp_path / ".hermes"
+    hermes_home.mkdir()
+    monkeypatch.setenv("HERMES_HOME", str(hermes_home))
+
+    from hermes_cli.a2a_cmd import cmd_a2a
+
+    args = argparse.Namespace(
+        a2a_command="send",
+        agent="unknown",
+        task="do work",
+        url=None,
+        wait=False,
+        timeout=5.0,
+        poll_interval=0.05,
+        requester=None,
+        cert="cert.pem",
+        key="key.pem",
+        ca="ca.pem",
+    )
+
+    with pytest.raises(SystemExit):
+        cmd_a2a(args)

tests/run_agent/test_run_agent.py

@@ -1037,138 +1037,6 @@ class TestBuildAssistantMessage:
         result = agent._build_assistant_message(msg, "tool_calls")
         assert "extra_content" not in result["tool_calls"][0]
 
-    def test_tool_call_arguments_normalized_from_gemma4_whitespace(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments='  \n  {"path": "README.md"}  \n  ',
-            call_id="c4",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-    def test_tool_call_arguments_normalized_from_single_quotes_and_trailing_comma(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments="{'path': 'README.md',}",
-            call_id="c5",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-
-class TestNormalizeToolCallArguments:
-    @pytest.mark.parametrize(
-        ("raw_args", "expected"),
-        [
-            ('{"q":"test"}', '{"q":"test"}'),
-            ('  \n  {"q": "test"}  \n  ', '{"q":"test"}'),
-            ('{"q": "test",}', '{"q":"test"}'),
-            ("{'q': 'test'}", '{"q":"test"}'),
-            ("{'path': 'README.md', 'mode': 'read'}", '{"path":"README.md","mode":"read"}'),
-            ('"path": "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md", mode: "read"', '{"path":"README.md","mode":"read"}'),
-            ({"path": "README.md"}, '{"path":"README.md"}'),
-            (["README.md", "docs.md"], '["README.md","docs.md"]'),
-            ('\t\n  ', '{}'),
-            ('{"nested": {"path": "README.md"}}', '{"nested":{"path":"README.md"}}'),
-        ],
-    )
-    def test_complete_args_are_normalized(self, raw_args, expected):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is True
-        assert normalized == expected
-
-    @pytest.mark.parametrize(
-        "raw_args",
-        [
-            '{"path": "README.md"',
-            '{"a": 1, "b"',
-            '{"path": [1, 2}',
-            "{'path': 'README.md'",
-            'path: "README.md", mode:',
-            '{"command": "echo hello",',
-        ],
-    )
-    def test_incomplete_args_are_not_marked_complete(self, raw_args):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is False
-        assert isinstance(normalized, str)
-        assert normalized == raw_args.strip()
-
-
-class TestSanitizeApiMessages:
-    def test_merges_consecutive_assistant_tool_call_messages(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "first",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "second",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "search_files", "arguments": '{"pattern":"TODO"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {"role": "tool", "tool_call_id": "c2", "content": "matches"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert len(sanitized) == 3
-        assert sanitized[0]["role"] == "assistant"
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2"]
-        assert sanitized[0]["content"] == "first\nsecond"
-
-    def test_does_not_merge_assistant_tool_call_messages_across_non_assistant_boundary(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c2", "content": "b.py"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assistant_msgs = [m for m in sanitized if m.get("role") == "assistant"]
-        assert len(assistant_msgs) == 2
-        assert assistant_msgs[0]["tool_calls"][0]["id"] == "c1"
-        assert assistant_msgs[1]["tool_calls"][0]["id"] == "c2"
-
-    def test_merge_preserves_tool_call_order(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c3", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"c.py"}'}}],
-            },
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2", "c3"]
-
 
 class TestFormatToolsForSystemMessage:
     def test_no_tools_returns_empty_array(self, agent):
@@ -3599,59 +3467,6 @@ class TestStreamingApiCall:
         assert tc[0].function.arguments == '{"path":"x.txt","content":"hel'
         assert resp.choices[0].finish_reason == "length"
 
-    @pytest.mark.parametrize(
-        ("raw_arguments", "expected"),
-        [
-            ('  \n  {"path": "x.txt"}  \n  ', '{"path":"x.txt"}'),
-            ("{'path': 'x.txt',}", '{"path":"x.txt"}'),
-            ('path: "x.txt", mode: "read"', '{"path":"x.txt","mode":"read"}'),
-        ],
-    )
-    def test_repairable_tool_call_args_do_not_upgrade_finish_reason_to_length(self, agent, raw_arguments, expected):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", raw_arguments)]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.name == "read_file"
-        assert tc[0].function.arguments == expected
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_tool_call_args_single_quotes_across_chunks_normalized(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", "{'path':")]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, " 'x.txt',}")]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_split_json_chunks_still_reassemble(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", '{"path":')]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, ' "x.txt"}')]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
     def test_ollama_reused_index_separate_tool_calls(self, agent):
         """Ollama sends every tool call at index 0 with different ids.