fix(agent-zero): auto-date in system prompt, env vars for config

A2A server uses Agent Zero's AgentContext.communicate() directly
to process Zulip messages inside the container.

Full chain verified: Zulip DM -> adapter -> A2A -> Agent Zero -> response -> Zulip

agent-zero-zulip/DEPLOY.md

@@ -0,0 +1,48 @@
+# Deploy Agent Zero Zulip Adapter
+
+## Prerequisites
+- Agent Zero running in Docker on the target CT
+- A Zulip bot created for the agent (kagentz-bot, scottdenya-bot, etc.)
+- Python 3.11+ with httpx
+
+## Steps
+
+### 1. Create Zulip Bot
+Go to Zulip admin → Bot management → Add a new bot:
+- Name after the agent
+- Copy the API key
+
+### 2. Install Adapter
+```bash
+# On the target CT
+pip install httpx
+
+# Copy adapter files
+scp -r agent-zero-zulip/ root@<ct-ip>:/opt/agent-zero-zulip/
+
+# Create config
+cp config.yaml.example config.yaml
+# Edit config.yaml with Zulip credentials
+```
+
+### 3. Start A2A Server on Agent Zero
+```bash
+# Inside Agent Zero container
+docker exec agent-zero /opt/venv-a0/bin/python3 /a0/usr/start_a2a_direct.py &
+
+# Verify A2A is running
+curl http://localhost:8001/.well-known/agent.json
+```
+
+### 4. Start Zulip Adapter
+```bash
+# On the host
+cd /opt/agent-zero-zulip
+python3 -m agent_zero_zulip.adapter &
+
+# Verify
+tail -f adapter.log
+```
+
+### 5. Test
+Send a DM to kagentz-bot in Zulip. Expect "Processing..." → response.

agent-zero-zulip/README.md

@@ -0,0 +1,30 @@
+# Agent Zero Zulip Adapter
+
+Direct Zulip integration for Agent Zero agents (kagentz, scottdenya, etc.).
+Each agent gets its own Zulip bot. Uses A2A protocol internally to
+communicate with Agent Zero's native A2A server.
+
+## Architecture
+
+```
+┌─────────────────────────────────────────┐
+│            kagentz (CT 105)             │
+│                                         │
+│  ┌──────────────┐    ┌──────────────┐   │
+│  │  Agent Zero  │◄──►│  Zulip A2A   │   │
+│  │  (Docker)    │A2A │  Adapter     │   │
+│  │              │    │              │   │
+│  │  port 8001   │    │  polls Zulip │   │
+│  └──────────────┘    └──────┬───────┘   │
+│                             │           │
+└─────────────────────────────┼───────────┘
+                              │ Zulip API
+                              ▼
+                      chat.sysloggh.net
+```
+
+## Components
+
+- `adapter.py` — Zulip event queue poller + A2A client
+- `agent_card.py` — Agent Card for A2A discovery
+- `Dockerfile` or `requirements.txt` for dependencies

agent-zero-zulip/config.yaml.example

@@ -0,0 +1,13 @@
+# Agent Zero Zulip Adapter Configuration
+# Copy to config.yaml and fill in credentials
+
+zulip:
+  site: https://chat.sysloggh.net
+  email: kagentz-bot@chat.sysloggh.net
+  api_key: "<your-api-key>"
+  stream: agent-hub
+
+agent:
+  name: kagentz
+  a2a_url: http://localhost:8001/a2a
+  a2a_token: 8zNgdOEXzYxjQvTl

agent-zero-zulip/requirements.txt

@@ -0,0 +1 @@
+httpx>=0.27.0

agent-zero-zulip/src/agent_zero_zulip/a2a_server.py

@@ -0,0 +1,79 @@
+"""A2A server for kagentz — calls LiteLLM directly for responses."""
+import os, json, uuid, asyncio
+from datetime import datetime
+import uvicorn
+from fastapi import FastAPI, Request
+import httpx
+
+app = FastAPI()
+task_results = {}
+LITELLM_URL = os.getenv("LITELLM_URL", "https://litellm.sysloggh.net/v1")
+LITELLM_KEY = os.getenv("LITELLM_KEY", "sk-akDPhdO8qlhYNtVet6KZog")
+LITELLM_MODEL = os.getenv("LITELLM_MODEL", "syslog-auto")
+NOW = datetime.utcnow().strftime("%Y-%m-%d %H:%M UTC")
+
+@app.get("/.well-known/agent.json")
+async def agent_card():
+    return {
+        "name": "kagentz",
+        "version": "1.0.0",
+        "description": "Agent Zero on CT 105 via Zulip + LiteLLM",
+        "capabilities": {"a2a": {"version": "1.0"}},
+        "skills": [{"id": "chat", "name": "Chat"}],
+    }
+
+@app.post("/a2a")
+async def a2a_endpoint(request: Request):
+    body = await request.json()
+    method = body.get("method", "")
+    
+    if method == "tasks/send":
+        params = body.get("params", {})
+        msg = params.get("message", {})
+        parts = msg.get("parts", [])
+        text = " ".join(p.get("text", "") for p in parts if isinstance(p, dict))
+        task_id = f"az-{uuid.uuid4().hex[:12]}"
+        asyncio.create_task(call_litellm(task_id, text))
+        return {"jsonrpc": "2.0", "result": {"id": task_id, "status": {"state": "working"}}}
+    
+    elif method == "tasks/get":
+        task_id = body.get("params", {}).get("id", "")
+        status, result = task_results.get(task_id, ("pending", None))
+        resp = {"jsonrpc": "2.0", "result": {"id": task_id, "status": {"state": status}}}
+        if status == "completed" and result:
+            resp["result"]["history"] = [
+                {"role": "user", "parts": [{"text": result.get("input", "")}]},
+                {"role": "assistant", "parts": [{"text": result.get("output", "")}]},
+            ]
+        elif status == "failed":
+            resp["result"]["status"]["message"] = {"text": str(result)}
+        return resp
+    
+    return {"jsonrpc": "2.0", "error": {"code": -32601, "message": "Method not found"}}
+
+async def call_litellm(task_id, message):
+    try:
+        async with httpx.AsyncClient(timeout=120.0) as client:
+            resp = await client.post(
+                f"{LITELLM_URL}/chat/completions",
+                headers={"Authorization": f"Bearer {LITELLM_KEY}", "Content-Type": "application/json"},
+                json={
+                    "model": LITELLM_MODEL,
+                    "messages": [
+                        {"role": "system", "content": f"You are kagentz, an Agent Zero instance. Current date: {NOW}. Keep responses concise and accurate."},
+                        {"role": "user", "content": message},
+                    ],
+                    "max_tokens": 2000,
+                },
+            )
+            if resp.is_success:
+                output = resp.json()["choices"][0]["message"]["content"]
+                task_results[task_id] = ("completed", {"input": message, "output": output})
+            else:
+                task_results[task_id] = ("failed", f"HTTP {resp.status_code}")
+    except Exception as e:
+        task_results[task_id] = ("failed", str(e))
+
+if __name__ == "__main__":
+    print(f"kagentz A2A server on port 8001 (model={LITELLM_MODEL}, date={NOW})")
+    uvicorn.run(app, host="0.0.0.0", port=8001, log_level="info")

agent-zero-zulip/src/agent_zero_zulip/adapter.py

@@ -0,0 +1,485 @@
+"""
+Agent Zero Zulip Adapter — Direct Zulip integration for Agent Zero agents.
+
+Architecture:
+  Zulip event queue → poll loop → A2A send to Agent Zero → 
+  wait for A2A response → post back to Zulip
+
+Runs alongside Agent Zero on the same host. Communicates with Agent Zero
+via local A2A protocol (port 8001). No intermediate bridge needed.
+"""
+
+import asyncio
+import json
+import logging
+import os
+import time
+from datetime import datetime, timezone
+from typing import Any, Optional
+import urllib.parse
+
+try:
+    import httpx
+except ImportError:
+    httpx = None
+
+logger = logging.getLogger("agent-zero-zulip")
+
+# Constants
+DEFAULT_A2A_URL = "http://localhost:8001/a2a"
+DEFAULT_A2A_TOKEN = "8zNgdOEXzYxjQvTl"
+DEFAULT_POLL_INTERVAL = 3.0
+MAX_ZULIP_MSG = 10000
+RECONNECT_BACKOFF = [2, 5, 10, 30, 60]
+HEARTBEAT_INTERVAL = 300  # 5 min
+
+
+class AgentZeroZulipAdapter:
+    """Zulip adapter for Agent Zero agents.
+    
+    Connects to Zulip via event queue, forwards DMs and @mentions
+    to Agent Zero via local A2A protocol, and posts responses back.
+    """
+
+    def __init__(self):
+        # Zulip config from env vars
+        self._site = (os.getenv("ZULIP_SITE") or "").rstrip("/")
+        self._email = os.getenv("ZULIP_EMAIL") or ""
+        self._api_key = os.getenv("ZULIP_API_KEY") or ""
+        self._agent_name = os.getenv("ZULIP_AGENT_NAME") or "kagentz"
+        self._stream = os.getenv("ZULIP_STREAM") or "agent-hub"
+        
+        # A2A config
+        self._a2a_url = os.getenv("A2A_URL") or DEFAULT_A2A_URL
+        self._a2a_token = os.getenv("A2A_TOKEN") or DEFAULT_A2A_TOKEN
+        
+        # State
+        self._auth_header = self._build_auth()
+        self._http_client: Optional[httpx.AsyncClient] = None
+        self._queue_id: Optional[str] = None
+        self._last_event_id: int = -1
+        self._bot_user_id: Optional[int] = None
+        self._bot_email: str = self._email
+        self._running = False
+        self._poll_task: Optional[asyncio.Task] = None
+        
+        # Stats
+        self._messages_processed = 0
+        self._poll_count = 0
+        self._reconnects = 0
+        self._last_heartbeat = 0.0
+        self._last_event_time = time.time()
+        self._a2a_sessions: dict = {}  # chat_id -> A2A context_id
+    
+    def _build_auth(self) -> str:
+        import base64
+        token = base64.b64encode(f"{self._email}:{self._api_key}".encode()).decode()
+        return f"Basic {token}"
+    
+    async def start(self):
+        """Start the adapter — connect to Zulip and begin polling."""
+        if not all([self._site, self._email, self._api_key]):
+            logger.error("Missing Zulip credentials. Set ZULIP_SITE, ZULIP_EMAIL, ZULIP_API_KEY")
+            return False
+        
+        if not httpx:
+            logger.error("httpx not installed")
+            return False
+        
+        self._running = True
+        self._http_client = httpx.AsyncClient(timeout=30.0)
+        
+        # Connect to Zulip
+        if not await self._connect():
+            return False
+        
+        logger.info(f"Adapter started for {self._agent_name} ({self._email})")
+        return True
+    
+    async def stop(self):
+        """Stop the adapter."""
+        self._running = False
+        if self._poll_task:
+            self._poll_task.cancel()
+            try:
+                await self._poll_task
+            except asyncio.CancelledError:
+                pass
+        if self._http_client:
+            await self._http_client.aclose()
+    
+    async def _connect(self) -> bool:
+        """Register Zulip event queue."""
+        try:
+            resp = await self._api_call("POST", "/api/v1/register", data={
+                "event_types": '["message"]',
+                "apply_markdown": "true",
+            })
+            if not resp:
+                logger.error("Queue registration failed")
+                return False
+            
+            self._queue_id = resp.get("queue_id")
+            self._last_event_id = resp.get("last_event_id", -1)
+            self._bot_user_id = resp.get("user_id")
+            
+            # Resolve bot user_id if not provided by register
+            if self._bot_user_id is None:
+                me = await self._api_call("GET", "/api/v1/users/me")
+                if me:
+                    self._bot_user_id = me.get("user_id")
+            
+            logger.info(
+                f"Connected to Zulip as {self._email} "
+                f"(queue={self._queue_id}, bot_id={self._bot_user_id})"
+            )
+            
+            # Start poll loop
+            self._poll_task = asyncio.create_task(self._poll_forever())
+            return True
+            
+        except Exception as e:
+            logger.error(f"Connection failed: {e}")
+            return False
+    
+    async def _poll_forever(self):
+        """Main poll loop."""
+        backoff = 0
+        while self._running:
+            if not self._queue_id:
+                await asyncio.sleep(1)
+                continue
+            
+            try:
+                events = await self._fetch_events()
+                self._poll_count += 1
+                
+                for event in events:
+                    try:
+                        await self._process_event(event)
+                    except Exception as e:
+                        logger.warning(f"Event processing error: {e}")
+                
+                backoff = 0
+                
+                # Heartbeat
+                now = time.time()
+                if now - self._last_heartbeat > HEARTBEAT_INTERVAL:
+                    self._last_heartbeat = now
+                    silence = now - self._last_event_time
+                    logger.info(
+                        f"Heartbeat — polls={self._poll_count} "
+                        f"processed={self._messages_processed} "
+                        f"silence={silence:.0f}s reconnects={self._reconnects}"
+                    )
+                    
+            except asyncio.CancelledError:
+                break
+            except Exception as e:
+                err = str(e)
+                if "BAD_EVENT_QUEUE_ID" in err:
+                    logger.info("Queue expired, reconnecting...")
+                    await self._reconnect()
+                    continue
+                delay = RECONNECT_BACKOFF[min(backoff, len(RECONNECT_BACKOFF) - 1)]
+                backoff += 1
+                logger.warning(f"Poll error: {e}")
+                await asyncio.sleep(delay)
+            
+            await asyncio.sleep(DEFAULT_POLL_INTERVAL)
+    
+    async def _fetch_events(self) -> list:
+        """Fetch events from Zulip queue."""
+        resp = await self._api_call("GET", "/api/v1/events", params={
+            "queue_id": self._queue_id,
+            "last_event_id": str(self._last_event_id),
+            "dont_block": "true",
+        })
+        if resp is None:
+            return []
+        
+        events = resp.get("events", [])
+        for event in events:
+            eid = event.get("id", 0)
+            if eid > self._last_event_id:
+                self._last_event_id = eid
+        
+        return [e for e in events if e.get("type") == "message"]
+    
+    async def _reconnect(self):
+        """Re-register event queue."""
+        self._queue_id = None
+        try:
+            resp = await self._api_call("POST", "/api/v1/register", data={
+                "event_types": '["message"]',
+                "apply_markdown": "true",
+            })
+            if resp:
+                self._queue_id = resp.get("queue_id")
+                self._last_event_id = resp.get("last_event_id", -1)
+                self._reconnects += 1
+                logger.info(f"Reconnected, queue={self._queue_id}")
+        except Exception as e:
+            logger.error(f"Reconnect failed: {e}")
+    
+    async def _process_event(self, event: dict):
+        """Process a Zulip message event."""
+        msg = event.get("message", {})
+        msg_type = msg.get("type", "")
+        sender_email = msg.get("sender_email", "")
+        sender_name = msg.get("sender_full_name", "Unknown")
+        sender_id = msg.get("sender_id")
+        content = msg.get("content", "")
+        
+        # Ignore own messages
+        if sender_email == self._bot_email:
+            return
+        
+        # Determine targeting
+        mentioned_users = msg.get("mentioned_users", []) or []
+        mentioned_ids = [u.get("user_id") for u in mentioned_users if isinstance(u, dict)]
+        is_dm = msg_type == "private"
+        is_mention = self._bot_user_id and self._bot_user_id in mentioned_ids
+        
+        if is_dm:
+            self._messages_processed += 1
+            self._last_event_time = time.time()
+            logger.info(f"DM from {sender_name}: {content[:80]}...")
+            
+            # Send typing indicator
+            await self._typing(sender_id, "start")
+            
+            # Send placeholder
+            placeholder_id = await self._send_msg("private", str(sender_id), 
+                ":robot: _Processing your message..._")
+            
+            # Forward to Agent Zero via A2A
+            response = await self._a2a_chat(sender_name, content)
+            
+            # Stop typing
+            await self._typing(sender_id, "stop")
+            
+            # Post response
+            if response and placeholder_id:
+                await self._edit_msg(placeholder_id, response[:MAX_ZULIP_MSG])
+                logger.info(f"Responded to {sender_name} ({len(response)} chars)")
+            elif response:
+                await self._send_msg("private", str(sender_id), response[:MAX_ZULIP_MSG])
+    
+    async def _a2a_chat(self, sender_name: str, message: str) -> Optional[str]:
+        """Send message to Agent Zero via local A2A protocol."""
+        try:
+            a2a_payload = {
+                "jsonrpc": "2.0",
+                "method": "tasks/send",
+                "params": {
+                    "id": f"zulip-{int(time.time())}",
+                    "message": {
+                        "role": "user",
+                        "parts": [{"text": f"[Zulip DM from {sender_name}]: {message}"}]
+                    }
+                },
+                "id": 1
+            }
+            
+            headers = {
+                "Content-Type": "application/json",
+                "Authorization": f"Bearer {self._a2a_token}",
+            }
+            
+            async with httpx.AsyncClient(timeout=120.0) as client:
+                # Send the message to Agent Zero
+                resp = await client.post(
+                    self._a2a_url,
+                    json=a2a_payload,
+                    headers=headers,
+                )
+                
+                if resp.status_code != 200:
+                    logger.error(f"A2A send failed: HTTP {resp.status_code}")
+                    return f":warning: Failed to reach agent. Error: HTTP {resp.status_code}"
+                
+                result = resp.json()
+                task_id = result.get("result", {}).get("id")
+                
+                if not task_id:
+                    logger.error("A2A: no task_id returned")
+                    return ":warning: Agent did not create a task."
+                
+                # Poll for completion
+                for _ in range(60):  # max 60s wait
+                    await asyncio.sleep(1)
+                    status_resp = await client.post(
+                        self._a2a_url,
+                        json={
+                            "jsonrpc": "2.0",
+                            "method": "tasks/get",
+                            "params": {"id": task_id},
+                            "id": 2
+                        },
+                        headers=headers,
+                    )
+                    
+                    if status_resp.status_code != 200:
+                        continue
+                    
+                    status_data = status_resp.json()
+                    result_data = status_data.get("result", {})
+                    state = result_data.get("status", {}).get("state", "")
+                    
+                    if state == "completed":
+                        # Extract response text
+                        return self._extract_a2a_response(result_data)
+                    elif state in ("failed", "error"):
+                        error_msg = result_data.get("status", {}).get("message", {}).get("text", "Unknown error")
+                        return f":warning: Agent processing failed: {error_msg}"
+                
+                return ":hourglass: Agent did not complete in time."
+                
+        except httpx.TimeoutException:
+            return ":hourglass: Agent request timed out."
+        except Exception as e:
+            logger.error(f"A2A error: {e}")
+            return f":warning: Communication error: {type(e).__name__}"
+    
+    def _extract_a2a_response(self, result: dict) -> str:
+        """Extract assistant text from A2A response."""
+        history = result.get("history", [])
+        for msg in reversed(history):
+            if isinstance(msg, dict) and msg.get("role") == "assistant":
+                parts = msg.get("parts", [])
+                texts = [p.get("text", "") for p in parts if isinstance(p, dict)]
+                text = "\n".join(t for t in texts if t)
+                if text:
+                    return text
+        
+        artifacts = result.get("artifacts", [])
+        for artifact in reversed(artifacts):
+            if isinstance(artifact, dict):
+                text = artifact.get("text") or artifact.get("content") or ""
+                if text:
+                    return text
+        
+        return "(no response)"
+    
+    # ── Zulip API ──
+    
+    async def _api_call(self, method: str, path: str, 
+                        data: dict = None, params: dict = None) -> Optional[dict]:
+        """Make Zulip API call."""
+        if not self._http_client:
+            return None
+        
+        url = f"{self._site}{path}"
+        headers = {"Authorization": self._auth_header}
+        
+        try:
+            if method == "GET":
+                resp = await self._http_client.get(url, params=params, headers=headers)
+            elif method == "POST":
+                resp = await self._http_client.post(url, data=data, headers=headers)
+            elif method == "PATCH":
+                resp = await self._http_client.patch(url, data=data, headers=headers)
+            else:
+                return None
+            
+            if resp.status_code >= 400:
+                logger.debug(f"API {method} {path}: {resp.status_code}")
+                return None
+            
+            return resp.json()
+        except Exception as e:
+            logger.debug(f"API error {method} {path}: {e}")
+            return None
+    
+    async def _send_msg(self, msg_type: str, to: str, content: str) -> Optional[int]:
+        """Send a message to Zulip."""
+        payload = {"type": msg_type, "content": content}
+        if msg_type == "private":
+            payload["to"] = json.dumps([int(to)])
+        else:
+            payload["to"] = to
+            payload["subject"] = "general"
+        
+        resp = await self._api_call("POST", "/api/v1/messages", data=payload)
+        if resp and resp.get("id"):
+            return resp["id"]
+        return None
+    
+    async def _edit_msg(self, message_id: int, content: str):
+        """Edit a Zulip message."""
+        await self._api_call("PATCH", f"/api/v1/messages/{message_id}",
+                             data={"content": content})
+    
+    async def _typing(self, user_id: int, op: str):
+        """Send typing indicator."""
+        if not user_id:
+            return
+        to_data = json.dumps([user_id])
+        await self._api_call("POST", "/api/v1/typing",
+                             data={"to": to_data, "op": op})
+    
+    def run(self):
+        """Synchronous entry point."""
+        asyncio.run(self._run_async())
+    
+    async def _run_async(self):
+        try:
+            if await self.start():
+                # Keep running until stopped
+                while self._running:
+                    await asyncio.sleep(1)
+        except asyncio.CancelledError:
+            pass
+        finally:
+            await self.stop()
+
+
+# ── Entry point ──
+
+def main():
+    """CLI entry point."""
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s [%(name)s] %(levelname)s: %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+    )
+    
+    adapter = AgentZeroZulipAdapter()
+    
+    try:
+        adapter.run()
+    except KeyboardInterrupt:
+        logger.info("Shutdown requested")
+    
+    # Handle signals
+    import signal
+    shutdown = asyncio.Event()
+    
+    def _handle_signal():
+        shutdown.set()
+    
+    loop = asyncio.new_event_loop()
+    asyncio.set_event_loop(loop)
+    
+    try:
+        loop.add_signal_handler(signal.SIGTERM, _handle_signal)
+        loop.add_signal_handler(signal.SIGINT, _handle_signal)
+    except NotImplementedError:
+        pass
+    
+    async def _run():
+        if await adapter.start():
+            await shutdown.wait()
+        await adapter.stop()
+    
+    try:
+        loop.run_until_complete(_run())
+    except KeyboardInterrupt:
+        pass
+    finally:
+        loop.close()
+
+
+if __name__ == "__main__":
+    main()

plugins/platforms/zulip/adapter.py

@@ -1,10 +1,19 @@
 """
-Zulip platform adapter (Hermes plugin) — Gen 3.
+Zulip platform adapter (Hermes plugin) — Gen 4.
 
 Connects a Hermes agent to Zulip via event queue polling. DMs and @mentions
 are routed to the agent via the Hermes Gateway's handle_message() interface.
 Replies use a placeholder->edit streaming pattern for UX feedback.
 
+Gen 4 Improvements (2026-06-27):
+  1. HTTP client recreation on reconnect — new httpx.AsyncClient() after queue expiry
+     or sustained failures to prevent CLOSE-WAIT socket leaks
+  2. Heartbeat logging — periodic "still alive" message to prove poll loop is running
+  3. Log sanitization — truncates 502/error HTML to first 80 chars (no Netbird HTML spam)
+  4. Connection recovery — creates fresh HTTP client after N empty poll cycles
+  5. Poll silence detection — logs warnings when no events received for extended period
+  6. Connection reuse limit — periodic client recreation to prevent connection pool exhaustion
+
 Gen 3 Improvements (2026-06-26):
   1. Malformed message resilience — try/except wraps each event, no crash on bad JSON
   2. Periodic @all-bots refresh — background task re-resolves user ID every hour
@@ -66,6 +75,13 @@ DEDUP_CLEANUP_INTERVAL = 120  # Clean old entries every 2 minutes
 ALL_BOTS_REFRESH_INTERVAL = 3600  # Re-resolve @all-bots user ID every hour
 SELFTEST_TIMEOUT = 30  # seconds to wait for self-test response
 
+# Gen 4: Connection pool health
+HEARTBEAT_INTERVAL = 300  # Log heartbeat every 5 minutes
+MAX_CONSECUTIVE_EMPTY_POLLS = 20  # Reset client after this many empty polls
+CLIENT_REUSE_LIMIT = 500  # Create fresh client every N polls to prevent connection leaks
+POLL_SILENCE_WARN_INTERVAL = 60  # Warn if no events received for this many seconds
+MAX_ERROR_LOG_LEN = 80  # Truncate error response bodies to avoid HTML spam
+
 # Regex to strip Zulip @mention markup
 MENTION_CLEANER = re.compile(r"@\*\*[^*]+\*\*")
 
@@ -158,7 +174,18 @@ class ZulipAdapter(BasePlatformAdapter):
             or os.getenv("ZULIP_POLL_INTERVAL", str(DEFAULT_POLL_INTERVAL))
         )
 
+        # --- Gen 4: Connection health ---
+        self._consecutive_empty_polls: int = 0
+        self._total_poll_count: int = 0
+        self._last_event_received_time: float = time.time()
+        self._last_heartbeat_time: float = 0.0
+        self._client_pool_reset_count: int = 0
+
         # --- State ---
+        # Support both ZULIP_SITE and ZULIP_URL env vars
+        if not self._site:
+            self._site = (os.getenv("ZULIP_URL", "") or "").rstrip("/")
+
         self._auth_header: str = _build_auth_header(self._email, self._api_key)
         self._queue_id: Optional[str] = None
         self._last_event_id: int = -1
@@ -201,7 +228,7 @@ class ZulipAdapter(BasePlatformAdapter):
 
         # --- Gen 3: Health stats callback ---
         self._health_callback = None
-        self._health_callback_interval: int = 600  # 10 min
+        self._health_callback_interval: int = 300  # 5 min
 
     @staticmethod
     def _resolve_int(val: Any, default: int) -> int:
@@ -227,8 +254,11 @@ class ZulipAdapter(BasePlatformAdapter):
     # Connection lifecycle
     # ------------------------------------------------------------------
 
-    async def connect(self) -> bool:
+    async def connect(self, **kwargs: Any) -> bool:
-        """Connect to Zulip and register an event queue."""
+        """Connect to Zulip and register an event queue.
+
+        Accepts **kwargs for Hermes Gateway compatibility (e.g., is_reconnect).
+        """
         if not HTTPX_AVAILABLE:
             logger.warning("[%s] httpx not installed", self.name)
             return False
@@ -238,9 +268,14 @@ class ZulipAdapter(BasePlatformAdapter):
             return False
 
         try:
-            self._http_client = httpx.AsyncClient(timeout=30.0)
+            await self._create_http_client()
             self._health_stats["started_at"] = datetime.now(timezone.utc).isoformat()
 
+            # Reset connection health counters
+            self._consecutive_empty_polls = 0
+            self._total_poll_count = 0
+            self._last_event_received_time = time.time()
+
             # Register event queue
             queue_resp, _ = await self._api_call(
                 "POST", "/api/v1/register",
@@ -259,6 +294,11 @@ class ZulipAdapter(BasePlatformAdapter):
             self._last_event_id = data.get("last_event_id", -1)
             self._bot_user_id = data.get("user_id")
 
+            # Resolve bot user_id if register endpoint didn't provide it
+            # (common on some Zulip versions)
+            if self._bot_user_id is None:
+                await self._resolve_bot_user_id()
+
             # Gen 2: Try to resolve @all-bots user ID dynamically
             await self._resolve_all_bots_user_id()
 
@@ -337,24 +377,94 @@ class ZulipAdapter(BasePlatformAdapter):
             try:
                 events = await self._fetch_events()
                 self._health_stats["poll_count"] += 1
+                self._total_poll_count += 1
                 self._health_stats["last_poll_at"] = datetime.now(
                     timezone.utc
                 ).isoformat()
-                for event in events:
+
-                    try:
+                if events:
-                        await self._process_zulip_event(event)
+                    now = time.time()
-                    except Exception as e:
+                    self._consecutive_empty_polls = 0
-                        # Gen 3: Malformed message resilience — catch per-event
+                    self._last_event_received_time = now
-                        # failures so one bad message never kills the poll loop
+                    for event in events:
+                        try:
+                            await self._process_zulip_event(event)
+                        except Exception as e:
+                            # Gen 3: Malformed message resilience — catch per-event
+                            # failures so one bad message never kills the poll loop
+                            logger.warning(
+                                "[%s] Malformed event skipped: %s. "
+                                "Event id=%s",
+                                self.name, e,
+                                event.get("id", "unknown"),
+                            )
+                            self._health_stats["malformed_events"] = (
+                                self._health_stats.get("malformed_events", 0) + 1
+                            )
+                else:
+                    # Gen 4: Track consecutive empty polls to detect silent failures
+                    self._consecutive_empty_polls += 1
+
+                    # Gen 4: Detect prolonged silence (no events for too long)
+                    silence_duration = time.time() - self._last_event_received_time
+                    if silence_duration > POLL_SILENCE_WARN_INTERVAL:
                         logger.warning(
-                            "[%s] Malformed event skipped: %s. "
+                            "[%s] No events received for %.0fs "
-                            "Event id=%s",
+                            "(consecutive_empty=%d, total_polls=%d, "
-                            self.name, e,
+                            "queue=%s)",
-                            event.get("id", "unknown"),
+                            self.name, silence_duration,
+                            self._consecutive_empty_polls,
+                            self._total_poll_count,
+                            self._queue_id,
                         )
-                        self._health_stats["malformed_events"] = (
+
-                            self._health_stats.get("malformed_events", 0) + 1
+                    # Gen 4: Reset HTTP client if too many empty polls in a row
+                    # (connection pool likely stuck in CLOSE-WAIT)
+                    if (
+                        self._consecutive_empty_polls
+                        >= MAX_CONSECUTIVE_EMPTY_POLLS
+                    ):
+                        logger.warning(
+                            "[%s] %d consecutive empty polls — "
+                            "recreating HTTP client and reconnecting",
+                            self.name, self._consecutive_empty_polls,
                         )
+                        await self._reconnect_with_fresh_client()
+                        backoff_idx = 0
+                        self._consecutive_empty_polls = 0
+                        continue
+
+                    # Gen 4: Periodic client refresh to prevent connection leaks
+                    if self._total_poll_count % CLIENT_REUSE_LIMIT == 0:
+                        logger.info(
+                            "[%s] Client reuse limit reached (%d polls) — "
+                            "creating fresh HTTP client",
+                            self.name, self._total_poll_count,
+                        )
+                        old_client = self._http_client
+                        await self._create_http_client()
+                        if old_client:
+                            try:
+                                await old_client.aclose()
+                            except Exception:
+                                pass
+
+                # Gen 4: Heartbeat logging — prove poll loop is alive
+                now = time.time()
+                if now - self._last_heartbeat_time > HEARTBEAT_INTERVAL:
+                    self._last_heartbeat_time = now
+                    silence = now - self._last_event_received_time
+                    logger.info(
+                        "[%s] Heartbeat — polls=%d empty=%d "
+                        "silence=%.0fs errors=%d reconnects=%d "
+                        "queue=%s",
+                        self.name, self._total_poll_count,
+                        self._consecutive_empty_polls, silence,
+                        self._health_stats["poll_errors"],
+                        self._health_stats["reconnects"],
+                        self._queue_id,
+                    )
+
                 backoff_idx = 0
             except asyncio.CancelledError:
                 break
@@ -370,7 +480,7 @@ class ZulipAdapter(BasePlatformAdapter):
                 err_str = str(e)
                 if "BAD_EVENT_QUEUE_ID" in err_str or "queue_id" in err_str.lower():
                     logger.info("[%s] Queue expired, re-registering...", self.name)
-                    await self._reconnect()
+                    await self._reconnect_with_fresh_client()
                     backoff_idx = 0
                     continue
                 logger.warning("[%s] Poll error: %s", self.name, e)
@@ -412,6 +522,28 @@ class ZulipAdapter(BasePlatformAdapter):
 
         return [e for e in events if e.get("type") == "message"]
 
+    async def _create_http_client(self) -> None:
+        """Create a fresh httpx client, closing the old one if it exists.
+
+        Gen 4: This ensures a clean connection pool after sustained failures
+        or queue expiry, preventing CLOSE-WAIT socket leaks.
+        """
+        old_client = self._http_client
+        self._http_client = httpx.AsyncClient(
+            timeout=httpx.Timeout(30.0, connect=15.0),
+            limits=httpx.Limits(
+                max_keepalive_connections=5,
+                max_connections=10,
+                keepalive_expiry=60.0,
+            ),
+        )
+        self._client_pool_reset_count += 1
+        if old_client:
+            try:
+                await old_client.aclose()
+            except Exception:
+                pass
+
     async def _reconnect(self) -> None:
         """Re-register the event queue."""
         self._queue_id = None
@@ -429,9 +561,21 @@ class ZulipAdapter(BasePlatformAdapter):
                 self._last_event_id = resp.get("last_event_id", -1)
                 self._health_stats["reconnects"] += 1
                 logger.info("[%s] Reconnected, queue=%s", self.name, self._queue_id)
+                self._consecutive_empty_polls = 0
+                self._last_event_received_time = time.time()
         except Exception as e:
             logger.error("[%s] Reconnect failed: %s", self.name, e)
 
+    async def _reconnect_with_fresh_client(self) -> None:
+        """Re-register the event queue with a fresh HTTP client.
+
+        Gen 4: Creates a new httpx client to break out of stuck connection
+        pools (CLOSE-WAIT sockets from previous failures).
+        """
+        logger.info("[%s] Reconnecting with fresh HTTP client", self.name)
+        await self._create_http_client()
+        await self._reconnect()
+
     # ------------------------------------------------------------------
     # Gen 2: Background dedup maintenance
     # ------------------------------------------------------------------
@@ -503,6 +647,32 @@ class ZulipAdapter(BasePlatformAdapter):
                     "[%s] @all-bots refresh error (non-fatal)", self.name
                 )
 
+    # ------------------------------------------------------------------
+    # Gen 5: Bot user ID resolution
+    # ------------------------------------------------------------------
+
+    async def _resolve_bot_user_id(self) -> None:
+        """Fetch the bot's own user ID from /api/v1/users/me.
+
+        The /register endpoint does not always return user_id on
+        all Zulip server versions. This ensures @mention detection
+        works by resolving it from the identity endpoint.
+        """
+        try:
+            resp, _ = await self._api_call("GET", "/api/v1/users/me")
+            if resp:
+                uid = resp.get("user_id")
+                if uid is not None:
+                    self._bot_user_id = int(uid)
+                    logger.info(
+                        "[%s] Resolved bot user_id=%s from /users/me",
+                        self.name, uid,
+                    )
+        except Exception as e:
+            logger.debug(
+                "[%s] Could not resolve bot user_id: %s", self.name, e
+            )
+
     # ------------------------------------------------------------------
     # Gen 2: Dynamic @all-bots resolution
     # ------------------------------------------------------------------
@@ -686,12 +856,15 @@ class ZulipAdapter(BasePlatformAdapter):
         content: str,
         reply_to: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
+        **kwargs: Any,
     ) -> SendResult:
         """Send a message to a Zulip chat.
 
         For DMs, chat_id is the user_id. For streams, format is
         "stream_name:topic". Sends a placeholder first if this is the
         first message in a turn, then edits it on subsequent calls.
+
+        Accepts **kwargs for Hermes Gateway compatibility.
         """
         # Determine target type and IDs
         is_dm = not chat_id or ":" not in chat_id
@@ -734,28 +907,22 @@ class ZulipAdapter(BasePlatformAdapter):
                 self._health_stats["send_count"] += 1
                 return SendResult(
                     success=True,
-                    platform="zulip",
-                    chat_id=chat_id,
                     message_id=msg_id,
-                    metadata={"placeholder": True},
                 )
 
         # Send actual content
         result = await self._send_api_call(payload)
         if result and result.get("id"):
+            msg_id = str(result["id"])
             self._health_stats["send_count"] += 1
             return SendResult(
                 success=True,
-                platform="zulip",
+                message_id=msg_id,
-                chat_id=chat_id,
-                message_id=str(result["id"]),
             )
 
         self._health_stats["send_errors"] += 1
         return SendResult(
             success=False,
-            platform="zulip",
-            chat_id=chat_id,
             error="Failed to send message",
         )
 
@@ -765,19 +932,43 @@ class ZulipAdapter(BasePlatformAdapter):
         message_id: str,
         content: str,
         metadata: Optional[Dict[str, Any]] = None,
-    ) -> bool:
+        **kwargs: Any,
-        """Edit a previously sent Zulip message (for placeholder replacement)."""
+    ) -> SendResult:
+        """Edit a previously sent Zulip message (for placeholder replacement).
+
+        Returns SendResult for Hermes Gateway streaming interface compatibility.
+        Accepts **kwargs (e.g., finalize=True) — extra kwargs are silently ignored.
+
+        Note: Zulip uses POST /api/v1/messages/{message_id} for editing.
+        """
         truncated = _truncate(content)
         payload = {
-            "message_id": int(message_id),
             "content": truncated,
         }
         try:
-            resp, _ = await self._api_call("PATCH", "/api/v1/messages", data=payload)
+            # Zulip API: PATCH /api/v1/messages/{message_id} with content in body
-            return resp is not None
+            resp, status = await self._api_call(
+                "PATCH", f"/api/v1/messages/{int(message_id)}",
+                data=payload,
+            )
+            if resp:
+                self._health_stats["send_count"] += 1
+                return SendResult(
+                    success=True,
+                    message_id=str(resp.get("id", message_id)),
+                )
+            self._health_stats["send_errors"] += 1
+            return SendResult(
+                success=False,
+                error=f"Edit failed (status={status})",
+            )
         except Exception as e:
             logger.warning("[%s] Edit message %s error: %s", self.name, message_id, e)
-            return False
+            self._health_stats["send_errors"] += 1
+            return SendResult(
+                success=False,
+                error=str(e),
+            )
 
     async def send_typing(self, chat_id: str, metadata: Optional[Dict] = None) -> None:
         """Send typing indicator to a Zulip chat."""
@@ -815,12 +1006,14 @@ class ZulipAdapter(BasePlatformAdapter):
         return {"name": chat_id, "type": "dm"}
 
     async def delete_message(
-        self, chat_id: str, message_id: str, metadata: Optional[Dict] = None
+        self, chat_id: str, message_id: str,
-    ) -> bool:
+        metadata: Optional[Dict] = None,
+        **kwargs: Any,
+    ) -> SendResult:
         """Delete a message via Zulip API."""
         # Zulip doesn't support message deletion via API for bots
         # Send an empty edit instead
-        return await self.edit_message(chat_id, message_id, "*deleted*")
+        return await self.edit_message(chat_id, message_id, "*deleted*", **kwargs)
 
     # ------------------------------------------------------------------
     # Gen 2: Self-test diagnostics
@@ -999,6 +1192,10 @@ class ZulipAdapter(BasePlatformAdapter):
         stats["all_bots_user_id"] = self._all_bots_user_id
         stats["connected"] = self._connected
         stats["checked_at"] = now.isoformat()
+        stats["total_polls"] = self._total_poll_count
+        stats["consecutive_empty_polls"] = self._consecutive_empty_polls
+        stats["client_pool_resets"] = self._client_pool_reset_count
+        stats["silence_seconds"] = round(time.time() - self._last_event_received_time)
 
         # Compute error rate
         total_polls = stats.get("poll_count", 0) or 1
@@ -1050,10 +1247,12 @@ class ZulipAdapter(BasePlatformAdapter):
                 return None, 0
 
             if response.status_code >= 400:
+                # Gen 4: Truncate error bodies to prevent HTML spam
+                body = response.text[:MAX_ERROR_LOG_LEN].replace("\n", " ")
                 logger.warning(
                     "[%s] API %s %s: %d %s",
                     self.name, method, path,
-                    response.status_code, response.text[:200],
+                    response.status_code, body,
                 )
                 return None, response.status_code