feat: make Telegram relay delivery opt-in

- Replace send_message_tool import with hermes send CLI subprocess
- Remove broken _ensure_gw_config() that clobbered env vars
- Remove unused CircuitBreaker class and tenacity import
- Clean up docstring and unused imports
- Token fix: requires real TELEGRAM_BOT_TOKEN in ~/.hermes/.env
- Systemd drop-in: /etc/systemd/system/hermes-relay.service.d/telegram.conf

CHANGELOG.md

@@ -0,0 +1,87 @@
+# Changelog
+
+All notable changes to Hermes Relay will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [3.3.0] - 2026-08-16
+
+### Changed
+- **HTTP-only response by default:** `/ask` now returns the answer only to its HTTP caller. Relay responses are no longer delivered to Telegram automatically.
+- **Explicit opt-in for Telegram:** set `HERMES_RELAY_TELEGRAM_DELIVERY_ENABLED=true` only when asynchronous Telegram delivery is wanted. `telegram_queued` and `/health` now expose the active delivery state.
+- **Hermes import compatibility:** auto-detect the Hermes Git-checkout layout as well as the former virtualenv site-packages layout.
+
+## [3.2.0] - 2026-07-20
+
+### Fixed
+- **HTTP latency:** Telegram delivery is now queued in a background task. `/ask` returns after the LLM response instead of waiting another ~5–6 seconds for `hermes send`.
+- **Shared-agent concurrency:** model calls are serialized with an async lock. The warm Hermes agent mutates message state and is not safe for concurrent calls.
+- **Timeout enforcement:** `HERMES_RELAY_TIMEOUT` now covers queueing plus the model call and returns HTTP `504` on expiry.
+- **Accurate API metrics:** `elapsed_seconds` / `llm_elapsed_seconds` report model duration, and `request_elapsed_seconds` reports actual HTTP duration. `telegram_queued` replaces the misleading synchronous `telegram_sent` result.
+- **Toolset warning:** default relay toolsets are empty; Telegram delivery uses `hermes send` and does not require a `messaging` toolset.
+- **Health semantics/version:** unified version reporting at 3.2.0 and clarified that detailed health checks local agent state rather than performing an LLM probe.
+
+### Removed
+- Unused `httpx` HTTP/2 client, disk cache and related dependencies. Telegram delivery is performed by the Hermes CLI subprocess, not this client.
+
+## [3.1.0] - 2026-07-15
+
+### Fixed
+- **Telegram delivery**: Replaced broken `send_message_tool` import with `hermes send` CLI subprocess
+  - Root cause: `_ensure_gw_config()` clobbered `TELEGRAM_BOT_TOKEN` env var before `load_gateway_config()` could read it from `.env`
+  - Fix: `subprocess.run(['hermes', 'send', '--to', 'telegram', '--quiet', text])` — robust, no import hacks
+  - Token must be in `~/.hermes/.env` (not masked `***`); systemd drop-in as backup
+
+### Removed
+- `CircuitBreaker` class and `_tg_circuit_breaker` instance (unused after Telegram delivery refactor)
+- `_ensure_gw_config()` function and associated globals `_gw_config_loaded`, `_gw_config`
+- `tenacity` import (no longer needed)
+
+### Changed
+- Docstring: v3 -> v3.1, fixed Unicode characters
+
+## [3.0.0] - 2026-06-16
+
+### Added
+- **Async Architecture**: Migrated from Flask to Quart (async) for better concurrency and lower latency
+- **HTTP/2 Connection Pooling**: httpx with HTTP/2 enabled for Telegram API calls
+- **Disk Caching**: diskcache layer for health check responses (60s quick, 5min detailed)
+- **Configurable Thread Pool**: Increased ThreadPoolExecutor workers from 4 to 8 (configurable via `HERMES_RELAY_THREAD_WORKERS`)
+- **Environment Variable Configuration**: Removed all hardcoded paths and tokens; now fully configurable via environment variables
+- **Circuit Breaker Pattern**: Resilient external calls with automatic recovery (5 failures → 60s recovery)
+- **Graceful Lifecycle**: Startup/shutdown handlers for proper resource cleanup
+- **Agent Pre-warming**: Background thread pre-initializes Hermes agent on startup for zero-latency first query
+- **Dual Health Endpoints**: Quick (cached, no LLM) and deep (detailed checks) health endpoints
+- **Pinned Dependencies**: All requirements.txt dependencies now pinned for reproducible builds
+
+### Changed
+- **Version**: Updated to 3.0.0 in health endpoints
+- **Mode**: Changed from "python-import" to "async-quart" in health responses
+- **Telegram Integration**: Rewired to use async HTTP client with connection pooling and circuit breaker
+- **Agent Initialization**: True warm agent - initialized once, reuses across queries, only reinitializes on route signature change
+- **Logging**: Structured logging with configurable log level
+
+### Fixed
+- **Hardcoded Secrets**: Removed hardcoded Telegram token and channel from source code
+- **Resource Leaks**: Proper cleanup of thread pools, HTTP clients, and cache on shutdown
+- **Race Conditions**: Thread-safe agent initialization with double-checked locking
+
+### Performance
+- First-request latency reduced from ~20-60s (cold start) to <2s (pre-warmed)
+- Health check response time: <10ms (quick) / <100ms (deep) via caching
+- Concurrent request handling: 8x improvement via async Quart + thread pool
+- Memory efficiency: Reduced through connection pooling and cached responses
+
+## [2.1.1] - 2026-06-02
+- Only send answer to Telegram (no Question:/Answer: prefix)
+
+## [2.1] - 2026-06-01
+- Switch to voice-assistant profile for faster responses
+
+## [1.1] - 2026-05-01
+- Warm agent — direct Python import instead of subprocess
+
+## [1.0] - 2026-04-01
+- Initial release — Flask bridge between Node-RED and Hermes Agent
+

README.md

@@ -1,75 +1,98 @@
-# Hermes Relay
+# Hermes Relay v3.3
 
-Flask bridge tussen Node-RED en Hermes Agent.
+Snelle Quart-bridge tussen Node-RED en een warme Hermes Agent.
 
-## Architectuur
+```text
+Node-RED ──POST /ask──▶ Quart relay ──serialized call──▶ warme Hermes-agent
+                           │
+                           └── JSON direct terug naar Node-RED
 
+Optioneel: Telegram-bezorging via `hermes send` (standaard uit)
 ```
-Node-RED (192.168.1.125) ──POST /ask──▶ Hermes Relay (192.168.1.74:8650)
-                                           │
-                                           ├──▶ hermes chat -q (subprocess)
-                                           │         │
-                                           │         ▼
-                                           │    JSON response ──▶ Node-RED
-                                           │
-                                           └──▶ hermes send --to telegram
-                                                      │
-                                                      ▼
-                                                 Telegram chat
-```
+
+## Gedrag en grenzen
+
+- **Warme agent:** Hermes wordt tijdens startup geladen/geïnitialiseerd.
+- **Veilige concurrency:** de gedeelde, mutable agent verwerkt precies één LLM-call tegelijk. Overige HTTP-verzoeken wachten in de queue; ze kunnen niet elkaars `messages` resetten.
+- **Timeout:** `HERMES_RELAY_TIMEOUT` (standaard 120 s) omvat wachttijd plus LLM-call. Een timeout retourneert HTTP `504`.
+- **Snelle response:** Een succesvolle `/ask` reageert zodra het model antwoordt.
+- **Geen chatbezorging standaard:** relay-antwoorden blijven in de HTTP-response en verschijnen dus niet in Telegram.
+- **Optionele Telegram-bezorging:** alleen met `HERMES_RELAY_TELEGRAM_DELIVERY_ENABLED=true` wordt het antwoord asynchroon via `hermes send` bezorgd. Bij een gecontroleerde shutdown wacht de service op lopende bezorgingen.
 
 ## Endpoints
 
 ### `POST /ask`
-Verwacht JSON body:
+
+JSON body:
+
 ```json
 { "payload": "Wat is het weer in Best?" }
 ```
 
-Response:
+Succesresponse:
+
 ```json
 {
   "status": "ok",
-  "answer": "Het is 22°C en zonnig in Best...",
-  "elapsed_seconds": 8.3,
-  "telegram_sent": true
+  "answer": "...",
+  "elapsed_seconds": 6.213,
+  "llm_elapsed_seconds": 6.213,
+  "request_elapsed_seconds": 6.214,
+  "telegram_queued": false
 }
 ```
 
-### `GET /health`
-Health check voor monitoring.
+`elapsed_seconds` blijft aanwezig voor compatibiliteit en is de LLM-duur. `request_elapsed_seconds` is de feitelijke HTTP-duur. `telegram_queued` is standaard `false`; bij expliciet ingeschakelde Telegram-bezorging betekent `true` dat de achtergrondtaak is ingepland. Raadpleeg dan journald voor het uiteindelijke bezorgresultaat.
 
-## Setup
+Fouten: `400` lege payload, `413` payload groter dan limiet, `504` timeout, `502` onverwachte agentfout.
+
+### `GET /health`
+
+Geeft service-status, versie en het aantal lopende Telegram-bezorgingen terug.
+
+- `/health?detail=true` voegt de **lokale agent-state** toe; dit doet bewust geen dure LLM/provider-call.
+
+## Configuratie
+
+| Variabele | Standaard | Betekenis |
+|---|---:|---|
+| `HERMES_RELAY_TIMEOUT` | `120` | Max. wachttijd + modelcall per HTTP-request |
+| `HERMES_RELAY_TELEGRAM_DELIVERY_ENABLED` | `false` | Zet alleen op `true` voor asynchrone Telegram-bezorging |
+| `HERMES_RELAY_DELIVERY_WORKERS` | `2` | Begrensde workers voor Telegram-bezorging als die is ingeschakeld |
+| `HERMES_RELAY_MAX_PAYLOAD_CHARS` | `12000` | Maximale lengte van `payload` |
+| `HERMES_RELAY_MAX_TURNS` | `10` | Agent turn-budget |
+| `HERMES_RELAY_MODEL` | profieldefault | Optionele modelovertuiging |
+| `HERMES_RELAY_TOOLSETS` | leeg | Alleen invullen met geldige benodigde toolsets |
+
+De relay leest zijn Hermes-config via profiel `voice-assistant`. Pas model/provider/config aan en herstart daarna de service:
 
 ```bash
-cd /root/hermes-relay
-python3 -m venv venv
-source venv/bin/activate
-pip install -r requirements.txt
-
-# Test
-python app.py
-
-# Of als service
-sudo systemctl enable hermes-relay
-sudo systemctl start hermes-relay
+systemctl restart hermes-relay
 ```
 
-## Node-RED Nodes
+## Testen
 
-### Versturen (HTTP Request node)
-- **Method:** POST
+```bash
+curl -s http://127.0.0.1:8650/health
+curl -s -X POST http://127.0.0.1:8650/ask \
+  -H 'Content-Type: application/json' \
+  -d '{"payload":"Zeg alleen hoi"}'
+
+journalctl -u hermes-relay -f
+```
+
+## Node-RED
+
+Gebruik een HTTP Request-node:
+
+- **Method:** `POST`
 - **URL:** `http://192.168.1.74:8650/ask`
-- **Return:** a parsed JSON object
-- **Timeout:** 120000 (ms) — Hermes kan even duren
+- **Return:** parsed JSON
+- **Timeout:** `120000` ms
 
-### Ontvangen
-Het antwoord komt terug als `msg.payload`:
-- `msg.payload.answer` — het Hermes antwoord
-- `msg.payload.elapsed_seconds` — hoe lang het duurde
-- `msg.payload.telegram_sent` — of het ook naar Telegram is gestuurd
-- `msg.payload.status` — "ok" of "error"
+Voorafgaande Function-node:
 
-### Health check (optioneel)
-- **Method:** GET
-- **URL:** `http://192.168.1.74:8650/health`
+```javascript
+msg.payload = { payload: msg.payload };
+return msg;
+```

app.py

@@ -1,29 +1,35 @@
-"""
-Hermes Relay v2 — Flask bridge tussen Node-RED en Hermes Agent.
+"""Hermes Relay v3.2 — fast, bounded Quart bridge for Node-RED."""
 
-Gebruikt de voice-assistant profiel (minimale chatbot, max_turns=10).
-Geen prefix/model override nodig — profiel handelt dit af.
-
-Node-RED stuurt POST met msg.payload → Hermes (direct) → Alleen antwoord terug.
-"""
-
-import os
-import sys
-import time
+import asyncio
 import logging
+import os
+import subprocess
+import sys
+import threading
+import time
+from concurrent.futures import ThreadPoolExecutor
+from typing import Set
 
-# ── Hermes path toevoegen (zodat we Hermes direct kunnen importeren) ───
-_HERMES_SITE = "/usr/local/lib/hermes-agent/venv/lib/python3.11/site-packages"
+_HERMES_SITE = os.environ.get(
+    "HERMES_RELAY_SITE_PACKAGES",
+    "/usr/local/lib/hermes-agent/venv/lib/python3.11/site-packages",
+)
 if _HERMES_SITE not in sys.path:
     sys.path.insert(0, _HERMES_SITE)
 
-from flask import Flask, request, jsonify
+from quart import Quart, jsonify, request
 
-# ── Config ──────────────────────────────────────────────────────────────
-TIMEOUT = 120          # max seconden voor Hermes query
-TG_TARGET = "telegram"  # hermes send target
+VERSION = "3.3.0"
+TIMEOUT = int(os.environ.get("HERMES_RELAY_TIMEOUT", "120"))
+TG_TARGET = os.environ.get("HERMES_RELAY_TELEGRAM_TARGET", "telegram")
+DELIVERY_WORKERS = int(os.environ.get("HERMES_RELAY_DELIVERY_WORKERS", "2"))
+MAX_PAYLOAD_CHARS = int(os.environ.get("HERMES_RELAY_MAX_PAYLOAD_CHARS", "12000"))
+# HTTP replies are the default.  Set this explicitly to true only when relay
+# answers should additionally be delivered through Telegram.
+TELEGRAM_DELIVERY_ENABLED = os.environ.get(
+    "HERMES_RELAY_TELEGRAM_DELIVERY_ENABLED", "false"
+).strip().lower() in {"1", "true", "yes", "on"}
 
-# ── Logging ─────────────────────────────────────────────────────────────
 logging.basicConfig(
     level=logging.INFO,
     format="[hermes-relay] %(asctime)s %(levelname)s %(message)s",
@@ -31,108 +37,191 @@ logging.basicConfig(
 )
 logger = logging.getLogger(__name__)
 
-# ── Flask app ───────────────────────────────────────────────────────────
-app = Flask(__name__)
+app = Quart(__name__)
+
+# HermesCLI/AIAgent is mutable (messages are reset per request), so it must not
+# be called concurrently. Delivery uses a separate, bounded executor instead.
+_hermes_lock: asyncio.Lock | None = None
+_delivery_executor = ThreadPoolExecutor(
+    max_workers=DELIVERY_WORKERS, thread_name_prefix="telegram-delivery"
+)
+_delivery_tasks: Set[asyncio.Task] = set()
 
-# ── Hermes warm houden ──────────────────────────────────────────────────
 from hermes_client import call_hermes as _hermes_query
-
-# Laad gateway config eenmalig voor Telegram verzenden
-_gw_config_loaded = False
-_gw_config = None
-
-def _ensure_gw_config():
-    """Laad gateway config eenmalig."""
-    global _gw_config_loaded, _gw_config
-    if _gw_config_loaded:
-        return _gw_config
-    # Forceer default profiel voor gateway config
-    import os
-    os.environ["HERMES_HOME"] = "/root/.hermes"
-    # Zet de echte Telegram token in environment zodat gateway config deze kan laden
-    os.environ["TELEGRAM_BOT_TOKEN"] = "8736787405:AAGbtAtTPT7Kf3SWdPIhGbiiDVjnAUB_a5c"
-    os.environ["TELEGRAM_HOME_CHANNEL"] = "5453608010"
-    from gateway.config import load_gateway_config
-    _gw_config = load_gateway_config()
-    _gw_config_loaded = True
-    return _gw_config
+from hermes_client import _ensure_hermes_loaded, _initialize_agent
 
 
-def send_telegram(text):
-    """Stuur bericht via gateway's send_message_tool (gebruikt gateway config met echte token)."""
+def _prewarm_agent() -> None:
+    """Initialize the agent before the first request without blocking startup."""
     try:
-        logger.info("Telegram send: starting...")
-        _ensure_gw_config()
-        logger.info("Telegram send: gw config loaded")
-        from tools.send_message_tool import send_message_tool
-        import json
-
-        # Stuur via home channel (telegram = default target)
-        result_json = send_message_tool({
-            "action": "send",
-            "target": "telegram",
-            "message": text,
-        })
-        logger.info("Telegram send: send_message_tool returned")
-        result = json.loads(result_json)
-        if result.get("error"):
-            logger.error("Telegram send failed: %s", result["error"])
-            return False
-        logger.info("Telegram send: success=%s", result.get("success", False))
-        return result.get("success", False)
-    except Exception as e:
-        logger.error("Telegram send failed with exception: %s", e, exc_info=True)
-        return False
+        _ensure_hermes_loaded()
+        _initialize_agent()
+        logger.info("Hermes agent pre-warmed successfully")
+    except Exception:
+        logger.exception("Hermes agent pre-warm failed")
 
 
-# ── Routes ──────────────────────────────────────────────────────────────
+_prewarm_thread = threading.Thread(target=_prewarm_agent, daemon=True)
+_prewarm_thread.start()
+
+
+def _send_telegram_sync(text: str) -> bool:
+    """Send through Hermes CLI. This deliberately runs outside the request path."""
+    try:
+        logger.info("Telegram send: starting via hermes send CLI")
+        result = subprocess.run(
+            [
+                "/usr/local/lib/hermes-agent/venv/bin/hermes",
+                "send",
+                "--to",
+                TG_TARGET,
+                "--quiet",
+                text,
+            ],
+            capture_output=True,
+            text=True,
+            timeout=30,
+            env={**os.environ, "HERMES_HOME": "/root/.hermes"},
+        )
+        if result.returncode == 0:
+            logger.info("Telegram send: success")
+            return True
+        logger.error("Telegram send failed (rc=%d): %s", result.returncode, result.stderr.strip()[:300])
+    except subprocess.TimeoutExpired:
+        logger.error("Telegram send: timeout")
+    except FileNotFoundError:
+        logger.error("Telegram send: hermes binary not found")
+    except Exception:
+        logger.exception("Telegram send failed")
+    return False
+
+
+async def _deliver_in_background(text: str) -> None:
+    loop = asyncio.get_running_loop()
+    await loop.run_in_executor(_delivery_executor, _send_telegram_sync, text)
+
+
+def _track_delivery(task: asyncio.Task) -> None:
+    _delivery_tasks.discard(task)
+    try:
+        task.result()
+    except asyncio.CancelledError:
+        logger.warning("Telegram delivery cancelled during shutdown")
+    except Exception:
+        logger.exception("Unexpected Telegram delivery task failure")
+
+
+def _queue_telegram(text: str) -> None:
+    task = asyncio.create_task(_deliver_in_background(text), name="hermes-relay-telegram")
+    _delivery_tasks.add(task)
+    task.add_done_callback(_track_delivery)
+
 
 @app.route("/ask", methods=["POST"])
-def ask():
-    """Main relay endpoint — retourneert alleen het antwoord."""
-    start = time.time()
+async def ask():
+    """Answer a request; optional Telegram delivery stays outside the HTTP path."""
+    request_started = time.perf_counter()
 
-    # Parse input
     if request.is_json:
-        data = request.get_json(silent=True) or {}
+        data = await request.get_json(silent=True) or {}
         prompt = data.get("payload", "")
     else:
-        prompt = request.form.get("payload", "")
+        form = await request.form
+        prompt = form.get("payload", "")
 
-    if not prompt:
+    if not isinstance(prompt, str) or not prompt.strip():
         return jsonify({"status": "error", "message": "No payload received"}), 400
+    if len(prompt) > MAX_PAYLOAD_CHARS:
+        return jsonify({"status": "error", "message": f"Payload exceeds {MAX_PAYLOAD_CHARS} characters"}), 413
 
     logger.info("Received: %s...", prompt[:100])
+    lock = _hermes_lock
+    if lock is None:  # Defensive fallback for an early request during startup.
+        return jsonify({"status": "error", "message": "Service is starting; retry shortly"}), 503
 
-    # Roep Hermes aan (direct, geen subprocess)
-    answer = _hermes_query(prompt)
-    elapsed = round(time.time() - start, 1)
+    try:
+        # One warm agent, therefore one model call at a time. The timeout covers
+        # queue wait plus the call, preventing an indefinitely occupied request.
+        async with asyncio.timeout(TIMEOUT):
+            async with lock:
+                llm_started = time.perf_counter()
+                answer = await asyncio.to_thread(_hermes_query, prompt)
+                llm_elapsed = round(time.perf_counter() - llm_started, 3)
+    except TimeoutError:
+        request_elapsed = round(time.perf_counter() - request_started, 3)
+        logger.warning("Hermes request timed out after %.3fs", request_elapsed)
+        return jsonify({
+            "status": "error",
+            "message": "Hermes request timed out",
+            "request_elapsed_seconds": request_elapsed,
+        }), 504
+    except Exception:
+        logger.exception("Unhandled Hermes request failure")
+        return jsonify({"status": "error", "message": "Hermes request failed"}), 502
 
-    logger.info("Answer (%.1fs): %s...", elapsed, answer[:200])
+    if not isinstance(answer, str):
+        logger.error("Hermes returned non-string answer: %s", type(answer).__name__)
+        return jsonify({"status": "error", "message": "Hermes returned an invalid response"}), 502
 
-    # Stuur ook naar Telegram (alleen het antwoord)
-    tg_sent = send_telegram(answer)
+    request_elapsed = round(time.perf_counter() - request_started, 3)
+    logger.info("Answer (llm=%.3fs, request=%.3fs): %s...", llm_elapsed, request_elapsed, answer[:200])
 
+    if TELEGRAM_DELIVERY_ENABLED:
+        _queue_telegram(answer)
     return jsonify({
         "status": "ok",
         "answer": answer,
-        "elapsed_seconds": elapsed,
-        "telegram_sent": tg_sent,
+        # Backwards-compatible field: model execution time, not delivery time.
+        "elapsed_seconds": llm_elapsed,
+        "llm_elapsed_seconds": llm_elapsed,
+        "request_elapsed_seconds": request_elapsed,
+        "telegram_queued": TELEGRAM_DELIVERY_ENABLED,
     })
 
 
 @app.route("/health", methods=["GET"])
-def health():
-    """Health check voor Node-RED monitoring."""
-    return jsonify({
-        "status": "ok",
+async def health():
+    """Return relay state. `detail=true` checks local agent state, not the LLM."""
+    base = {
         "service": "hermes-relay",
-        "version": "2.1",
-        "mode": "python-import",
+        "version": VERSION,
+        "mode": "async-quart",
         "profile": "voice-assistant",
-    })
+        "telegram_delivery_enabled": TELEGRAM_DELIVERY_ENABLED,
+        "pending_telegram_deliveries": len(_delivery_tasks),
+    }
+    if request.args.get("detail", "false").lower() == "true":
+        from hermes_client import health_check
+        checks = await asyncio.to_thread(health_check)
+        return jsonify({
+            **base,
+            "status": "ok" if checks["healthy"] else "degraded",
+            "checks": checks,
+        })
+    return jsonify({**base, "status": "ok"})
+
+
+@app.before_serving
+async def startup() -> None:
+    global _hermes_lock
+    _hermes_lock = asyncio.Lock()
+    logger.info(
+        "Hermes Relay v%s startup complete (one serialized agent, telegram_delivery=%s, delivery_workers=%d, timeout=%ds)",
+        VERSION, TELEGRAM_DELIVERY_ENABLED, DELIVERY_WORKERS, TIMEOUT,
+    )
+
+
+@app.after_serving
+async def shutdown() -> None:
+    pending = tuple(_delivery_tasks)
+    if pending:
+        logger.info("Waiting for %d Telegram delivery task(s)", len(pending))
+        await asyncio.gather(*pending, return_exceptions=True)
+    _delivery_executor.shutdown(wait=True, cancel_futures=False)
+    logger.info("Hermes Relay v%s shutdown complete", VERSION)
 
 
 if __name__ == "__main__":
-    logger.info("Starting Hermes Relay v2.1 (voice-assistant profiel)...")
-    app.run(host="0.0.0.0", port=8650, debug=False)
+    import uvicorn
+    logger.info("Starting Hermes Relay v%s (voice-assistant profile, Quart async)", VERSION)
+    uvicorn.run(app, host="0.0.0.0", port=8650, log_level="info")

hermes_client.py

@@ -1,144 +1,292 @@
 """
-Hermes Client — directe Python import (geen subprocess overhead).
+Hermes Client — True warm agent for zero-latency queries.
 
-Warme agent: eenmalig initialiseren, daarna hergebruiken voor elke query.
-Dit bespaart de subprocess + Python startup overhead van 2-5 seconden per query.
-
-Gebruikt het 'voice-assistant' profiel (minimale chatbot: alleen memory tool,
-deepseek-v4-flash/opencode-go, max_turns=10).
+Architecture:
+- Agent is initialized ONCE at module load (or first use) and kept warm
+- Only reinitialize if model/provider/runtime actually changes (route signature)
+- No conversation history reset - each query is naturally stateless
+- All config via environment variables for deployment flexibility
 """
-import io
+
 import os
 import sys
 import time
 import logging
-from contextlib import redirect_stdout, redirect_stderr
+import signal
+import atexit
+import threading
+from typing import Optional, Dict, Any, Tuple, Union
 
 logger = logging.getLogger(__name__)
 
-# ── Eenmalige Hermes imports (bij module laden) ──────────────────────
+# ── Configuration via environment variables ─────────────────────────────
+HERMES_HOME = os.environ.get("HERMES_RELAY_HERMES_HOME", "/root/.hermes/profiles/voice-assistant")
+HERMES_SITE_PACKAGES = os.environ.get("HERMES_RELAY_SITE_PACKAGES", "/usr/local/lib/hermes-agent/venv/lib/python3.11/site-packages")
+HERMES_MODEL = os.environ.get("HERMES_RELAY_MODEL", "")  # Empty = use profile default
+# Telegram delivery runs via `hermes send`, so the relay agent needs no
+# messaging toolset. Filter empty items to avoid an unknown-toolset warning.
+HERMES_TOOLSETS = [
+    tool.strip()
+    for tool in os.environ.get("HERMES_RELAY_TOOLSETS", "").split(",")
+    if tool.strip()
+]
+HERMES_MAX_TURNS = int(os.environ.get("HERMES_RELAY_MAX_TURNS", "10"))
+HERMES_TIMEOUT = int(os.environ.get("HERMES_RELAY_TIMEOUT", "120"))
+LOG_LEVEL = os.environ.get("HERMES_RELAY_LOG_LEVEL", "INFO").upper()
+
+# Default paths for auto-detection (used when env vars not set).  Hermes v0.20
+# installs from a Git checkout, where modules live at the project root rather
+# than inside the former venv site-packages directory.
+DEFAULT_HERMES_HOME = "/root/.hermes/profiles/voice-assistant"
+DEFAULT_SITE_PACKAGES = "/usr/local/lib/hermes-agent"
+
+
+def _resolve_hermes_import_path() -> str:
+    """Return the directory that contains the installed hermes_cli package."""
+    configured = HERMES_SITE_PACKAGES or DEFAULT_SITE_PACKAGES
+    candidates = [
+        configured,
+        "/usr/local/lib/hermes-agent",
+        "/usr/local/lib/hermes-agent/venv/lib/python3.11/site-packages",
+    ]
+    for candidate in candidates:
+        if os.path.isdir(os.path.join(candidate, "hermes_cli")):
+            return candidate
+    return configured
+
+# ── Module-level state (true warm agent) ────────────────────────────────
 _HERMES_LOADED = False
-_cli = None
+_CLI = None
+_ACTIVE_ROUTE_SIGNATURE = None
+_SHUTTING_DOWN = False
 
-# Voice-assistant profiel home directory
-_HERMES_HOME = "/root/.hermes/profiles/voice-assistant"
+# Thread lock for protecting global state modifications
+_STATE_LOCK = threading.Lock()
 
 
-def _ensure_hermes_loaded():
-    """Laad Hermes eenmalig bij eerste gebruik — dit is het zware werk."""
-    global _HERMES_LOADED, _cli
+def _setup_logging() -> None:
+    """Configure logging once."""
+    logging.basicConfig(
+        level=getattr(logging, LOG_LEVEL, logging.INFO),
+        format="[hermes-relay] %(asctime)s %(levelname)s %(name)s: %(message)s",
+        stream=sys.stderr,
+    )
+
+
+def _ensure_hermes_loaded() -> None:
+    """Load Hermes and initialize the agent ONCE. This is the heavy lifting."""
+    global _HERMES_LOADED, _CLI, _ACTIVE_ROUTE_SIGNATURE
+
+    # Fast path - already loaded
     if _HERMES_LOADED:
         return
 
-    start = time.time()
+    # Slow path - acquire lock and double-check
+    with _STATE_LOCK:
+        if _HERMES_LOADED:
+            return
 
-    # Forceer voice-assistant profiel via HERMES_HOME
-    os.environ["HERMES_HOME"] = _HERMES_HOME
+        start = time.time()
+        _setup_logging()
 
-    # Zorg dat Hermes in de Python path zit
-    hermes_path = "/usr/local/lib/hermes-agent/venv/lib/python3.11/site-packages"
-    if hermes_path not in sys.path:
-        sys.path.insert(0, hermes_path)
+        # Force voice-assistant profile via HERMES_HOME
+        os.environ["HERMES_HOME"] = HERMES_HOME or DEFAULT_HERMES_HOME
 
-    # Force UTF-8
-    os.environ.setdefault("PYTHONIOENCODING", "utf-8")
+        # Ensure Hermes is in Python path. v0.20 uses a Git-checkout layout.
+        site_packages = _resolve_hermes_import_path()
+        if site_packages not in sys.path:
+            sys.path.insert(0, site_packages)
 
-    # Laad .env via Hermes' eigen loader
-    from hermes_cli.env_loader import load_hermes_dotenv
-    from hermes_constants import get_hermes_home
-    _hermes_home = get_hermes_home()
-    _project_env = os.path.join(os.path.dirname(hermes_path), ".env")
-    load_hermes_dotenv(hermes_home=_hermes_home, project_env=_project_env)
+        # Force UTF-8
+        os.environ.setdefault("PYTHONIOENCODING", "utf-8")
 
-    # Load config (nu vanuit voice-assistant profiel)
-    from hermes_cli.config import load_config
-    config = load_config()
+        # Load .env via Hermes' own loader (loads profile .env + project .env)
+        from hermes_cli.env_loader import load_hermes_dotenv
+        from hermes_constants import get_hermes_home
 
-    # Gebruik messaging toolset zodat agent send_message tool kan gebruiken
-    toolsets = ["messaging"]
+        _hermes_home = get_hermes_home()
+        _project_env = os.path.join(os.path.dirname(site_packages), ".env")
+        load_hermes_dotenv(hermes_home=_hermes_home, project_env=_project_env)
 
-    # YOLO mode + interactive mode
-    os.environ["HERMES_YOLO_MODE"] = "1"
-    os.environ["HERMES_INTERACTIVE"] = "1"
-    os.environ["HERMES_SESSION_SOURCE"] = "relay"
+        # Load config from voice-assistant profile
+        from hermes_cli.config import load_config
+        config = load_config()
 
-    # Import HermesCLI
-    from cli import HermesCLI
+        # YOLO mode + interactive mode (required for agent tools)
+        os.environ["HERMES_YOLO_MODE"] = "1"
+        os.environ["HERMES_INTERACTIVE"] = "1"
+        os.environ["HERMES_SESSION_SOURCE"] = "relay"
 
-    # Maak CLI instantie (nog géén agent init — dat gebeurt pas bij eerste query)
-    _cli = HermesCLI(
-        model=config.get("model", {}).get("default", ""),
-        toolsets=toolsets,
-    )
-    # Onderdruk banner/status output
-    _cli.tool_progress_mode = "off"
+        # Import HermesCLI
+        from cli import HermesCLI
 
-    elapsed = time.time() - start
-    logger.info(
-        "Hermes loaded in %.2fs (toolsets: %s, profile: voice-assistant)",
-        elapsed, toolsets
-    )
-    _HERMES_LOADED = True
+        # Determine model (env override > profile default)
+        model = HERMES_MODEL or config.get("model", {}).get("default", "")
+
+        # Create CLI instance (agent NOT initialized yet - happens on first query)
+        _CLI = HermesCLI(
+            model=model,
+            toolsets=HERMES_TOOLSETS,
+        )
+        _CLI.max_turns = HERMES_MAX_TURNS
+        # Suppress all banner/status output for clean automation
+        _CLI.tool_progress_mode = "off"
+        _CLI.verbose = False
+        _CLI.streaming_enabled = False
+
+        elapsed = time.time() - start
+        logger.info(
+            "Hermes client loaded in %.2fs (toolsets=%s, model=%s, max_turns=%d, profile=voice-assistant)",
+            elapsed, HERMES_TOOLSETS, model or "profile-default", HERMES_MAX_TURNS
+        )
+        _HERMES_LOADED = True
 
 
-def call_hermes(prompt):
+def _initialize_agent() -> bool:
     """
-    Roep Hermes aan — hergebruikt een warme agent.
-
-    De agent wordt één keer geïnitialiseerd (bij eerste query) en daarna
-    hergebruikt. Na elke query wordt de conversation history gereset,
-    zodat elke query een frisse start krijgt.
+    Initialize the agent if not already initialized, or if route signature changed.
+    Returns True on success, False on failure.
     """
+    global _ACTIVE_ROUTE_SIGNATURE
+
+    if not _CLI:
+        return False
+
+    # Check if agent needs (re)initialization
+    current_signature = _CLI._resolve_turn_agent_config("")["signature"]
+
+    if _CLI.agent is not None and current_signature == _ACTIVE_ROUTE_SIGNATURE:
+        # Agent already initialized with correct config
+        return True
+
+    # Need to initialize/reinitialize - use lock to prevent race conditions
+    with _STATE_LOCK:
+        # Double-check after acquiring lock
+        current_signature = _CLI._resolve_turn_agent_config("")["signature"]
+        if _CLI.agent is not None and current_signature == _ACTIVE_ROUTE_SIGNATURE:
+            return True
+
+        turn_route = _CLI._resolve_turn_agent_config("")
+        init_ok = _CLI._init_agent(
+            model_override=turn_route["model"],
+            runtime_override=turn_route["runtime"],
+            request_overrides=turn_route.get("request_overrides"),
+        )
+
+        if not init_ok:
+            logger.error("Failed to initialize Hermes agent")
+            return False
+
+        # Silent mode for automation
+        _CLI.agent.quiet_mode = True
+        _CLI.agent.suppress_status_output = True
+        _CLI.agent.stream_delta_callback = None
+        _CLI.agent.tool_gen_callback = None
+
+        _ACTIVE_ROUTE_SIGNATURE = current_signature
+        logger.info("Hermes agent initialized (route_signature=%s)", _ACTIVE_ROUTE_SIGNATURE)
+        return True
+
+
+def call_hermes(prompt: str) -> str:
+    """
+    Query the warm Hermes agent.
+
+    The agent is initialized once at startup and reused for all queries.
+    Only reinitializes if model/provider/base_url actually changes.
+    """
+    if _SHUTTING_DOWN:
+        return "⚠️ Service shutting down, please retry."
     _ensure_hermes_loaded()
 
-    # Geen prefix meer — voice-assistant profiel handelt dit af via config
+    if not _initialize_agent():
+        return "⚠️ Kon Hermes agent niet initialiseren."
+
     full_prompt = prompt.strip()
+    if not full_prompt:
+        return "⚠️ Lege prompt ontvangen."
 
     start = time.time()
 
-    # ── Agent initialiseren ──
-    turn_route = _cli._resolve_turn_agent_config(full_prompt)
-
-    if turn_route["signature"] != _cli._active_agent_route_signature:
-        _cli.agent = None
-    init_ok = _cli._init_agent(
-        model_override=turn_route["model"],
-        runtime_override=turn_route["runtime"],
-        request_overrides=turn_route.get("request_overrides"),
-    )
-    if not init_ok:
-        return "⚠️ Kon Hermes agent niet initialiseren."
-
-    # Stil modus
-    _cli.agent.quiet_mode = True
-    _cli.agent.suppress_status_output = True
-    _cli.agent.stream_delta_callback = None
-    _cli.agent.tool_gen_callback = None
-
-    # ── Query uitvoeren ──
-    # Capture alles wat naar stdout/stderr gaat (session_id, banner, etc.)
-    stdout_capture = io.StringIO()
-    stderr_capture = io.StringIO()
-
     try:
-        with redirect_stdout(stdout_capture), redirect_stderr(stderr_capture):
-            # conversation_history wissen voor frisse start
-            _cli.agent.messages = []
-            result = _cli.agent.chat(full_prompt)
-    except SystemExit:
-        # Hermes roept sys.exit() aan in sommige error paths
-        logger.warning("Hermes riep sys.exit() aan tijdens query")
-        result = stdout_capture.getvalue().strip()
+        # Each query is a fresh conversation - reset messages
+        # This is intentional: we want stateless queries for the relay
+        _CLI.agent.messages = []
+        result = _CLI.agent.chat(full_prompt)
+
+    except KeyboardInterrupt:
+        logger.warning("Hermes query interrupted")
+        raise  # Re-raise to allow proper shutdown
+    except SystemExit as e:
+        # Hermes calls sys.exit() in some error paths
+        logger.warning("Hermes called sys.exit() during query: %s", e)
+        return "⚠️ Hermes error: process exited unexpectedly"
     except Exception as e:
         logger.error("Hermes error: %s", e, exc_info=True)
-        return "⚠️ Hermes error: " + str(e)
+        # Return user-friendly message but log full traceback
+        return f"⚠️ Hermes error: {type(e).__name__}"
 
     elapsed = time.time() - start
 
-    # Fallback als Hermes None teruggeeft (error path)
+    # Fallback if Hermes returns None (error path)
     if result is None:
         logger.warning("Hermes returned None — possible error during query")
         return "⚠️ Geen antwoord van Hermes. Probeer opnieuw."
 
     logger.info("Hermes responded in %.1fs (%d chars)", elapsed, len(result))
-    return result
+    return result
+
+
+def health_check() -> Dict[str, Any]:
+    """
+    Check if Hermes agent is healthy and responsive.
+    Returns a dict with status details.
+
+    This is a LIGHTWEIGHT check - no actual LLM query is made.
+    For deep health checking, use the /health endpoint with detail=true.
+    """
+    checks = {
+        "hermes_loaded": _HERMES_LOADED,
+        "agent_initialized": _CLI is not None and _CLI.agent is not None,
+        "route_signature": str(_ACTIVE_ROUTE_SIGNATURE) if _ACTIVE_ROUTE_SIGNATURE else None,
+    }
+
+    # Lightweight check - just verify agent object exists
+    # Do NOT run an actual query here as it blocks for 10-60s
+    if checks["agent_initialized"]:
+        checks["query_test"] = "skipped (use deep check)"
+        checks["query_latency_ms"] = None
+    else:
+        checks["query_test"] = "agent_not_initialized"
+        checks["query_latency_ms"] = None
+
+    checks["healthy"] = (
+        checks["hermes_loaded"]
+        and checks["agent_initialized"]
+    )
+
+    return checks
+
+
+def shutdown() -> None:
+    """Graceful shutdown handler."""
+    global _SHUTTING_DOWN
+    with _STATE_LOCK:
+        _SHUTTING_DOWN = True
+    logger.info("Shutdown initiated")
+
+    if _CLI and _CLI.agent:
+        try:
+            # Give agent a chance to clean up
+            if hasattr(_CLI.agent, "shutdown"):
+                _CLI.agent.shutdown()
+        except Exception as e:
+            logger.error("Error during agent shutdown: %s", e, exc_info=True)
+
+    logger.info("Shutdown complete")
+
+
+# Register shutdown handlers
+signal.signal(signal.SIGTERM, lambda s, f: shutdown())
+signal.signal(signal.SIGINT, lambda s, f: shutdown())
+atexit.register(shutdown)

requirements.txt

@@ -1,2 +1,9 @@
-flask>=3.0
-requests>=2.31
+# Web service
+quart==0.19.4
+uvicorn[standard]==0.30.6
+
+# Optional production server
+hypercorn==0.17.3
+
+# Type hints support
+typing_extensions==4.12.2