Switch Agent0 from MCP to REST API (/api/api_message)

MCP endpoint returns 403 (wrong token) and the mcp library's TaskGroup
error is opaque. The REST API (X-API-KEY header, /api/api_message) is
already validated in connectivity tests and returns proper responses.

mcp_key field is reused as the X-API-KEY value.
context_id replaces chat_id for conversation continuity.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

server.py

@@ -268,60 +268,50 @@ async def handle_agent0_mcp(
     valid_poses: list[str],
 ) -> ChatResponse:
     """
-    Send a message to a live Agent Zero instance via its MCP streamable-http
-    server and return the response as a ChatResponse.
+    Send a message to a live Agent Zero instance via its REST API
+    (POST /api/api_message, X-API-KEY header).
 
-    MCP URL format: {endpoint}/mcp/t-{mcp_key}/http
-    Tool called: send_message (built into every Agent Zero instance)
-
-    Conversation continuity is maintained via Agent Zero's chat_id, which maps
-    to a gnommoweb conversation_id in _a0_sessions (in-memory).
+    Conversation continuity is maintained via Agent Zero's context_id, which
+    maps to a gnommoweb conversation_id in _a0_sessions (in-memory).
     """
-    from mcp.client.streamable_http import streamablehttp_client
-    from mcp import ClientSession
+    endpoint = (model.endpoint or "").strip().rstrip("/")
+    api_key  = model.mcp_key or ""   # stored in mcp_key field; used as X-API-KEY
+    api_url  = f"{endpoint}/api/api_message"
 
-    endpoint = (model.endpoint or "").rstrip("/")
-    mcp_key  = model.mcp_key or ""
-    mcp_url  = f"{endpoint}/mcp/t-{mcp_key}/http"
-
-    # Retrieve existing Agent0 chat_id for this conversation (if any)
-    a0_chat_id = _a0_sessions.get(req.conversation_id) if req.conversation_id else None
+    # Retrieve existing Agent0 context_id for this conversation (if any)
+    a0_context_id = _a0_sessions.get(req.conversation_id) if req.conversation_id else None
 
     log.info(
-        f"[{agent.name}] → Agent0 MCP  url={mcp_url}  "
-        f"conv={req.conversation_id}  a0_chat={a0_chat_id or 'new'}  "
+        f"[{agent.name}] → Agent0 REST  url={api_url}  "
+        f"conv={req.conversation_id}  a0_ctx={a0_context_id or 'new'}  "
         f"msg='{req.message[:60]}'"
     )
 
     try:
-        async with streamablehttp_client(mcp_url) as (read, write, _):
-            async with ClientSession(read, write) as session:
-                await session.initialize()
+        payload: dict = {
+            "message":       req.message,
+            "lifetime_hours": 24,
+        }
+        if a0_context_id:
+            payload["context_id"] = a0_context_id
 
-                tool_args: dict = {
-                    "message":        req.message,
-                    "persistent_chat": True,
-                }
-                if a0_chat_id:
-                    tool_args["chat_id"] = a0_chat_id
+        async with httpx.AsyncClient(timeout=120.0) as client:
+            r = await client.post(
+                api_url,
+                headers={"X-API-KEY": api_key, "Content-Type": "application/json"},
+                json=payload,
+            )
+            r.raise_for_status()
 
-                result = await session.call_tool("send_message", tool_args)
+        data = r.json()
+        log.info(f"[{agent.name}] ← Agent0 REST: {str(data)[:300]}")
 
-        # The tool returns a JSON-serialised ToolResponse / ToolError
-        raw = result.content[0].text if result.content else "{}"
-        log.info(f"[{agent.name}] ← Agent0 MCP raw: {raw[:300]}")
+        response_text  = data.get("response", "")
+        new_context_id = data.get("context_id", "")
 
-        parsed = json.loads(raw)
-
-        if parsed.get("status") == "error":
-            raise RuntimeError(parsed.get("error", "Unknown Agent0 error"))
-
-        response_text = parsed.get("response", "")
-        new_chat_id   = parsed.get("chat_id", "")
-
-        # Persist the Agent0 chat_id so the next turn continues the same context
-        if new_chat_id and req.conversation_id is not None:
-            _a0_sessions[req.conversation_id] = new_chat_id
+        # Persist context_id for conversation continuity
+        if new_context_id and req.conversation_id is not None:
+            _a0_sessions[req.conversation_id] = new_context_id
 
         pose = pick_pose(response_text, valid_poses)
 
@@ -334,7 +324,7 @@ async def handle_agent0_mcp(
         )
 
     except Exception as e:
-        log.error(f"[{agent.name}] Agent0 MCP error: {e}")
+        log.error(f"[{agent.name}] Agent0 REST error: {e}")
         fallback_pose = next(
             (p for p in ("sorry", "annoyed", "neutral") if p in valid_poses),
             valid_poses[0],
@@ -362,8 +352,10 @@ async def _fetch_hermes_token(endpoint: str) -> str:
     if cached:
         return cached
 
+    url = f"{endpoint}/"
+    log.info(f"[hermes] fetching token — GET {url!r} (len={len(url)}, bytes={url.encode()!r})")
     async with httpx.AsyncClient(timeout=10.0) as client:
-        resp = await client.get(f"{endpoint}/")
+        resp = await client.get(url)
         resp.raise_for_status()
         html = resp.text
 
@@ -405,7 +397,7 @@ async def handle_hermes(
     except ImportError:
         from websockets.client import connect as ws_connect  # type: ignore
 
-    endpoint  = (model.endpoint or "").rstrip("/")
+    endpoint  = (model.endpoint or "").strip().rstrip("/")
     hermes_session_id = _hermes_sessions.get(req.conversation_id) if req.conversation_id else None
 
     token = await _fetch_hermes_token(endpoint)
@@ -509,6 +501,9 @@ async def handle_hermes(
         if "401" in err_str or "403" in err_str or "Unauthorized" in err_str.lower():
             log.warning(f"[{agent.name}] Hermes token rejected, re-fetching…")
             _hermes_token_cache.pop(endpoint, None)
+            if req.conversation_id is not None:
+                _hermes_sessions.pop(req.conversation_id, None)
+            hermes_session_id = None  # force session.create on retry
             try:
                 token = await _fetch_hermes_token(endpoint)
                 return await _do_chat(token)
@@ -558,6 +553,12 @@ async def agent_chat(req: ChatRequest, authorization: str = Header(default="")):
     valid_poses = agent.poses if agent.poses else ["neutral"]
     model       = select_model(req.models)
 
+    log.info(
+        f"[route] agent={agent.name!r} agent_type={agent.agent_type!r} "
+        f"models_payload={req.models} "
+        f"selected_model={model}"
+    )
+
     # ── Route: Hermes Agent ───────────────────────────────────────────────────
     if agent.agent_type == "hermes":
         if not model or not model.endpoint:
@@ -569,7 +570,12 @@ async def agent_chat(req: ChatRequest, authorization: str = Header(default="")):
                 pose=valid_poses[0],
                 conversation_id=req.conversation_id,
             )
-        return await handle_hermes(req, agent, model, valid_poses)
+        try:
+            return await handle_hermes(req, agent, model, valid_poses)
+        except Exception as e:
+            import traceback
+            log.error(f"[hermes] unhandled exception:\n{traceback.format_exc()}")
+            raise
 
     # ── Route: Agent0 MCP ─────────────────────────────────────────────────────
     if agent.agent_type == "agent0" or (model and model.type == "agent0"):

test_connectivity.py

@@ -21,8 +21,8 @@ Configuration via environment variables (or a local .env file):
   LM_STUDIO_URL       — LM Studio base URL (default: http://localhost:1234)
   LM_STUDIO_MODEL     — model to use (default: first available)
   HERMES_URL          — Hermes dashboard URL (default: http://localhost:50007)
-  AGENT0_URL          — Agent Zero base URL
-  AGENT0_MCP_KEY      — Agent Zero MCP token
+  AGENT0_URL          — Agent Zero base URL (e.g. http://localhost:50003)
+  AGENT0_API_KEY      — Agent Zero X-API-KEY from the agent's dashboard
 
 Usage:
   python test_connectivity.py
@@ -287,27 +287,47 @@ async def test_hermes():
 # ── Agent Zero ────────────────────────────────────────────────────────────────
 
 async def test_agent0():
-    section("Agent Zero (MCP)")
+    section("Agent Zero (REST)")
     base_url = os.environ.get("AGENT0_URL", "")
-    mcp_key  = os.environ.get("AGENT0_MCP_KEY", "")
+    api_key  = os.environ.get("AGENT0_API_KEY", "")
 
-    if not base_url or not mcp_key:
-        skip("MCP endpoint", "AGENT0_URL and AGENT0_MCP_KEY not set")
+    if not base_url or not api_key:
+        skip("api_message endpoint", "AGENT0_URL and AGENT0_API_KEY not set")
         return
 
-    mcp_url = f"{base_url.rstrip('/')}/mcp/t-{mcp_key}/http"
+    url = f"{base_url.rstrip('/')}/api/api_message"
+
+    health_url = f"{base_url.rstrip('/')}/api/health"
 
     import httpx
     try:
+        # Step 1: health check (no auth needed, fast)
         async with httpx.AsyncClient(timeout=10.0) as client:
-            r = await client.get(mcp_url, headers={"Accept": "application/json"})
+            rh = await client.get(health_url)
 
-        if r.status_code in (200, 405):
-            ok("MCP endpoint reachable", f"HTTP {r.status_code} at {mcp_url}")
-        elif r.status_code == 401:
-            fail("Bad MCP key", f"HTTP {r.status_code}")
+        if rh.status_code != 200:
+            fail("Health endpoint", f"HTTP {rh.status_code}")
+            return
+
+        info = rh.json()
+        version = info.get("gitinfo", {}).get("version", "unknown")
+        ok(f"Reachable ({version})", health_url)
+
+        # Step 2: verify API key is accepted via api_reset_chat (fast, no LLM call)
+        reset_url = f"{base_url.rstrip('/')}/api/api_reset_chat"
+        async with httpx.AsyncClient(timeout=10.0) as client:
+            rk = await client.post(
+                reset_url,
+                headers={"X-API-KEY": api_key, "Content-Type": "application/json"},
+                json={},
+            )
+
+        if rk.status_code == 200:
+            ok("API key accepted", f"HTTP {rk.status_code}")
+        elif rk.status_code in (401, 403):
+            fail("Bad API key", f"HTTP {rk.status_code}")
         else:
-            ok("MCP endpoint reachable", f"HTTP {r.status_code}")
+            ok("API key accepted", f"HTTP {rk.status_code}")
 
     except httpx.ConnectError as e:
         fail("Not reachable", str(e)[:120])