Merge SyslogSolution/syslog-harness: accept current state (Phase 0 + Redis lazy reconnect + dashboard fix)

.env.example

@@ -0,0 +1,8 @@
+# Syslog Harness Environment
+REDIS_HOST=192.168.68.8
+REDIS_PORT=6379
+AMDPVE_ENDPOINT=http://192.168.68.15:8080
+LLMGPU_ENDPOINT=http://192.168.68.8:8080
+OCU_LLM_ENDPOINT=http://192.168.68.110:8080
+CIRCUIT_BREAKER_THRESHOLD=5
+CIRCUIT_BREAKER_TIMEOUT=30

MIGRATION_PLAN.md

@@ -0,0 +1,71 @@
+# Syslog Harness — Production Migration Plan
+
+## Current State (Development)
+- **Host:** CT 114 (192.168.68.123)
+- **Docker containers:** `syslog-queue` (:8091), `syslog-dashboard` (:3001)
+- **Nginx:** Local on CT 114, routing to GPUs + Docker services
+- **Status:** All components verified and operational
+
+## Target State (Production)
+- **Host:** New CT (e.g., `docker-vm` on 192.168.68.x)
+- **Docker containers:** Same queue + dashboard services
+- **Nginx:** Containerized on production CT
+- **GPU backends:** Same (192.168.68.15, .8, .110)
+
+## Migration Steps
+
+### 1. Prepare Production CT
+```bash
+# Create new CT on Proxmox
+# Install Docker
+apt update && apt install -y docker.io docker-compose-plugin
+
+# Pull/cloned harness repo
+git clone <repo-url> /root/syslog-harness
+cd /root/syslog-harness
+```
+
+### 2. Update docker-compose.yml for Production
+- Change `REDIS_HOST` to production Redis IP
+- Update GPU endpoint env vars if IPs change
+- Add volume mounts for persistence
+
+### 3. Build & Deploy
+```bash
+# Build images
+docker compose build
+
+# Start services
+docker compose up -d
+
+# Verify health
+curl http://localhost:8091/health
+curl http://localhost:3001/api/status
+```
+
+### 4. Configure Nginx
+- Copy `/etc/nginx/conf.d/gpu-router.conf` to production CT
+- Update upstream IPs if needed
+- Test and reload
+
+### 5. DNS / Routing Update
+- Point agent traffic to new CT IP
+- Update Hermes config `inference_api_url`
+- Test agent routing
+
+### 6. Verification Checklist
+- [ ] Queue service health check passes
+- [ ] Dashboard API returns GPU health
+- [ ] Nginx routes to correct GPU based on header
+- [ ] Circuit breaker triggers on excess load
+- [ ] Queue fallback works when GPUs down
+- [ ] Agent requests reach correct model
+
+## Rollback Plan
+- Keep CT 114 running as backup
+- Revert DNS/routing to .123 if issues
+- Docker containers can be stopped/started instantly
+
+---
+*Created: May 15, 2026*
+*Status: Development verified, ready for production migration*

README.md

@@ -0,0 +1,75 @@
+# syslog-harness — Inference API Harness
+
+CT 116 Docker stack for routing local GPU models through a unified OpenAI-compatible API.
+
+## Architecture
+
+```
+nginx :80 → router :9000 → GPU backends
+                ├─ qwen3.6-35B-A3B (MoE) @ 192.168.68.15:8080  [2 slots, 262K ctx]
+                ├─ qwen3.6-27B-code (Dense) @ 192.168.68.8:8080  [2 slots, 262K ctx]
+                └─ gemma-4-12b (VLM) @ 192.168.68.110:8080    [2 slots, 262K ctx]
+                                     Total: 6 concurrent slots
+
+LiteLLM :8081 (fallback) | Dashboard :3000 | Redis :6379 (local)
+```
+
+## Deploy
+
+```bash
+cd /opt/inference-harness
+docker compose up -d
+```
+
+## Endpoints
+
+| URL | Purpose |
+|-----|---------|
+| `/v1/chat/completions` | Inference API (OpenAI-compatible) — **API key required** |
+| `/v1/models` | Available models |
+| `/` | Dashboard (GPU health, routing, agents, timeseries) |
+
+## Authentication
+
+**All `/v1/chat/completions` requests require a valid API key** via `Authorization: Bearer <key>`. Missing or invalid keys return **401 Unauthorized**.
+
+## Agent API Keys
+
+| Agent | Key |
+|-------|-----|
+| Abiba | `sk-syslog-abiba` |
+| Mumuni | `sk-syslog-mumuni` |
+| Tanko | `sk-syslog-tanko` |
+| Koby | `sk-syslog-koby` |
+| Kagenz0 | `sk-syslog-kagenz0` |
+| Koonimo | `sk-syslog-koonimo` |
+
+## Routing Tiers
+
+| Tier | Trigger | Priority |
+|------|---------|----------|
+| Lightweight | No system prompt, ≤1 turn, ≤100 words | VLM → MoE → Dense |
+| Simple Conv | ≤1000 tokens, ≤4 turns | VLM → MoE → Dense |
+| Heavy | >4000 tokens OR >8 turns | Dense → MoE → VLM |
+| Default | Everything else | MoE → VLM → Dense |
+
+## Queue
+
+When all GPUs are saturated, requests enter a polling queue (500ms intervals) instead of returning 503 immediately. Timeout: 30s (configurable via `QUEUE_TIMEOUT` env or `X-Queue-Timeout` header).
+
+## Models
+
+| GPU | Model | VRAM | Slots | Context | Best For |
+|-----|-------|------|-------|
+| Strix Halo | qwen3.6-35B-A3B (MoE) | 65GB | 2 | 262K | General quality |
+| RTX 3090 | qwen3.6-27B-code (Dense) | 24GB | 2 | 262K | Code, reasoning |
+| RTX 5070 | gemma-4-12b (VLM) | 12GB | 2 | 262K | Speed, vision |
+
+## Maintenance
+
+Automated cron job runs daily at 3:00 AM UTC (`/opt/inference-harness/maintenance.sh`):
+- Cleans Redis timeseries keys >60 days
+- Prunes Docker build cache >7 days
+- Logs container health and Redis memory
+
+Logs: `/var/log/harness-maintenance.log`

gpu-router-docker.conf

@@ -0,0 +1,106 @@
+## Syslog GPU Router — Nginx Configuration (Docker-internal)
+## Routes incoming agent requests to the appropriate GPU backend
+## based on the X-Syslog-Model header.
+
+upstream amdpve_pool {
+    ## Strix Halo 395 — qwen3.6-35B-A3B (MoE) — Default workhorse
+    server 192.168.68.15:8080;
+}
+
+upstream llmgpu_pool {
+    ## RTX 3090 — qwen3.5-27B (Dense) — Heavy reasoning
+    server 192.168.68.8:8080;
+}
+
+upstream ocu_llm_pool {
+    ## New Backend — gemma-4-12b (VLM) — Vision + light tasks
+    server 192.168.68.110:8080;
+}
+
+upstream queue_service {
+    ## Agent queue with circuit breaker (Docker container)
+    server queue-service:8091;
+}
+
+upstream dashboard_service {
+    ## Harness dashboard (Docker container)
+    server dashboard:3001;
+}
+
+## ------------------------------------------------------------------
+## Mapping: X-Syslog-Model header → upstream backend
+## ------------------------------------------------------------------
+map $http_x_syslog_model $gpu_upstream {
+    default          amdpve_pool;
+    "standard"       amdpve_pool;
+    "heavy"          llmgpu_pool;
+    "qwen3.5-27B"    llmgpu_pool;
+    "light"          ocu_llm_pool;
+    "gemma-4-12b"        ocu_llm_pool;
+}
+
+## Rate limit zone — 10 req/s per IP, burst of 20
+limit_req_zone $binary_remote_addr zone=perip:10m rate=10r/s;
+
+server {
+    listen 80;
+    server_name _;
+
+    ## ------------------------------------------------------------------
+    ## Dashboard — observability UI (MUST be before / catch-all)
+    ## ------------------------------------------------------------------
+    location /dashboard {
+        proxy_pass http://dashboard_service/;
+        proxy_set_header Host              $host;
+        proxy_set_header X-Real-IP         $remote_addr;
+        proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
+    }
+
+    ## ------------------------------------------------------------------
+    ## Main location — proxy to selected upstream
+    ## ------------------------------------------------------------------
+    location / {
+        limit_req zone=perip burst=20 nodelay;
+        limit_req_status 503;
+        proxy_pass http://$gpu_upstream;
+
+        ## Preserve original host and headers
+        proxy_set_header Host              $host;
+        proxy_set_header X-Real-IP         $remote_addr;
+        proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+
+        ## Pass through the model header so backends can log it
+        proxy_pass_header X-Syslog-Model;
+
+        ## Streaming support (SSE for LLM responses)
+        proxy_buffering off;
+        proxy_cache     off;
+        proxy_read_timeout  300s;
+        proxy_send_timeout  300s;
+
+        ## Basic failover — retry on error or timeout
+        proxy_next_upstream error timeout http_502 http_503;
+        proxy_next_upstream_tries 2;
+
+        ## Add a response header for observability
+        add_header X-Routed-To $gpu_upstream always;
+
+        ## Fallback to queue when all GPU upstreams are down
+        error_page 502 503 504 = @queue_fallback;
+    }
+
+    ## ------------------------------------------------------------------
+    ## Queue fallback — enqueue when GPUs are unavailable
+    ## ------------------------------------------------------------------
+    location @queue_fallback {
+        rewrite ^ /enqueue break;
+        proxy_pass http://queue_service;
+        proxy_set_header Host              $host;
+        proxy_set_header X-Real-IP         $remote_addr;
+        proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_set_header Content-Type      $content_type;
+        proxy_pass_request_body            on;
+    }
+}

gpu-router.conf

@@ -0,0 +1,106 @@
+## Syslog GPU Router — Nginx Configuration
+## Routes incoming agent requests to the appropriate GPU backend
+## based on the X-Syslog-Model header.
+
+upstream amdpve_pool {
+    ## Strix Halo 395 — qwen3.6-35B-A3B (MoE) — Default workhorse
+    server 192.168.68.15:8080;
+}
+
+upstream llmgpu_pool {
+    ## RTX 3090 — qwen3.5-27B (Dense) — Heavy reasoning
+    server 192.168.68.8:8080;
+}
+
+upstream ocu_llm_pool {
+    ## New Backend — gemma-4-12b (VLM) — Vision + light tasks
+    server 192.168.68.110:8080;
+}
+
+upstream queue_service {
+    ## Agent queue with circuit breaker (Docker container)
+    server 127.0.0.1:8091;
+}
+
+upstream dashboard_service {
+    ## Harness dashboard (Docker container)
+    server 127.0.0.1:3001;
+}
+
+## ------------------------------------------------------------------
+## Mapping: X-Syslog-Model header → upstream backend
+## ------------------------------------------------------------------
+map $http_x_syslog_model $gpu_upstream {
+    default          amdpve_pool;   # missing header → default workhorse
+    "standard"       amdpve_pool;
+    "heavy"          llmgpu_pool;
+    "qwen3.5-27B"    llmgpu_pool;
+    "light"          ocu_llm_pool;
+    "gemma-4-12b"        ocu_llm_pool;
+}
+
+server {
+    listen 8080;
+    server_name _;
+
+    # Rate limit zone — 10 req/s per IP, burst of 20
+    limit_req_zone $binary_remote_addr zone=perip:10m rate=10r/s;
+
+    ## ------------------------------------------------------------------
+    ## Dashboard — observability UI (MUST be before / catch-all)
+    ## ------------------------------------------------------------------
+    location /dashboard {
+        proxy_pass http://dashboard_service/;
+        proxy_set_header Host              $host;
+        proxy_set_header X-Real-IP         $remote_addr;
+        proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
+    }
+
+    ## ------------------------------------------------------------------
+    ## Main location — proxy to selected upstream
+    ## ------------------------------------------------------------------
+    location / {
+        limit_req zone=perip burst=20 nodelay;
+        limit_req_status 503;
+        proxy_pass http://$gpu_upstream;
+
+        ## Preserve original host and headers
+        proxy_set_header Host              $host;
+        proxy_set_header X-Real-IP         $remote_addr;
+        proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+
+        ## Pass through the model header so backends can log it
+        proxy_pass_header X-Syslog-Model;
+
+        ## Streaming support (SSE for LLM responses)
+        proxy_buffering off;
+        proxy_cache     off;
+        proxy_read_timeout  300s;
+        proxy_send_timeout  300s;
+
+        ## Basic failover — retry on error or timeout
+        proxy_next_upstream error timeout http_502 http_503;
+        proxy_next_upstream_tries 2;
+
+        ## Add a response header for observability
+        add_header X-Routed-To $gpu_upstream always;
+
+        ## Fallback to queue when all GPU upstreams are down
+        error_page 502 503 504 = @queue_fallback;
+    }
+
+    ## ------------------------------------------------------------------
+    ## Queue fallback — enqueue when GPUs are unavailable
+    ## ------------------------------------------------------------------
+    location @queue_fallback {
+        rewrite ^ /enqueue break;
+        proxy_pass http://queue_service;
+        proxy_set_header Host              $host;
+        proxy_set_header X-Real-IP         $remote_addr;
+        proxy_set_header X-Forwarded-For   $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_set_header Content-Type      $content_type;
+        proxy_pass_request_body            on;
+    }
+}

phase0-dual-keys.json

@@ -0,0 +1,20 @@
+{
+  "sk-syslog-local-master-key": {"tier": "enterprise", "agent": "admin", "deprecated": true},
+  "sk-33a0d6e6-a6da12fb483770d6f63f543b7e16a742": {"tier": "enterprise", "agent": "admin"},
+  "sk-syslog-abiba": {"tier": "enterprise", "agent": "Abiba", "deprecated": true},
+  "sk-6a6e49d0-e35e27c524ce0ba2de8f862a0e966b0c": {"tier": "enterprise", "agent": "Abiba"},
+  "sk-syslog-mumuni": {"tier": "enterprise", "agent": "Mumuni", "deprecated": true},
+  "sk-ba249c99-29e3b163a8d8ab7abb8663cae0dceb37": {"tier": "enterprise", "agent": "Mumuni"},
+  "sk-syslog-tanko": {"tier": "enterprise", "agent": "Tanko", "deprecated": true},
+  "sk-4a67112a-d9b62caf5a5881a15837df506f782708": {"tier": "enterprise", "agent": "Tanko"},
+  "sk-syslog-koby": {"tier": "enterprise", "agent": "Koby", "deprecated": true},
+  "sk-9dd66bba-316f616ec40cb51f3b489a0146bb986f": {"tier": "enterprise", "agent": "Koby"},
+  "sk-syslog-kagenz0": {"tier": "enterprise", "agent": "Kagenz0", "deprecated": true},
+  "sk-5a05020f-a984a8a20f731c5a65dfc07d51495bb3": {"tier": "enterprise", "agent": "Kagenz0"},
+  "sk-syslog-koonimo": {"tier": "enterprise", "agent": "Koonimo", "deprecated": true},
+  "sk-b2f0e1b2-70b7cf1c5a5e402a49b5469e88d9e335": {"tier": "enterprise", "agent": "Koonimo"},
+  "sk-starter-abc123": {"tier": "starter", "agent": "test-starter", "deprecated": true},
+  "sk-a8307a00-cd76daa07dd61c41aeb4545656f8b8c4": {"tier": "starter", "agent": "test-starter"},
+  "sk-professional-xyz789": {"tier": "professional", "agent": "test-pro", "deprecated": true},
+  "sk-b519a91e-4d2774c6c7da3fc31326e69ccc426781": {"tier": "professional", "agent": "test-pro"}
+}

queue-service/queue-service.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""Syslog Inference Queue Service — Circuit breaker + request queuing.
+
+Ports: 8091
+Endpoints:
+  /health          — liveness probe (Nginx upstream check)
+  /enqueue         — POST inference request into queue (fallback from Nginx)
+  /status          — GET queue depth + circuit breaker state
+"""
+
+import json
+import os
+import sys
+import time
+import urllib.request
+from flask import Flask, request, jsonify
+
+app = Flask(__name__)
+
+# Configuration
+REDIS_HOST = os.getenv("REDIS_HOST", "192.168.68.7")
+REDIS_PORT = int(os.getenv("REDIS_PORT", "6379"))
+QUEUE_KEY = "inference:requests"
+CIRCUIT_OPEN_THRESHOLD = 50
+CIRCUIT_WARN_THRESHOLD = 30
+
+# GPU endpoints for draining
+GPUS = {
+    "amdpve": "192.168.68.15:8080",
+    "llmgpu": "192.168.68.8:8080",
+    "ocu_llm": "192.168.68.110:8080",
+}
+
+
+def get_redis():
+    try:
+        import redis
+        return redis.Redis(host=REDIS_HOST, port=REDIS_PORT, decode_responses=True)
+    except Exception:
+        return None
+
+
+def get_queue_depth(r):
+    try:
+        return r.llen(QUEUE_KEY)
+    except Exception:
+        return 0
+
+
+def check_gpu_health(endpoint):
+    try:
+        req = urllib.request.Request(f"http://{endpoint}/v1/models")
+        req.add_header("User-Agent", "queue-service/1.0")
+        resp = urllib.request.urlopen(req, timeout=3)
+        return resp.status == 200
+    except Exception:
+        return False
+
+
+@app.route("/health")
+def health():
+    """Nginx upstream health probe. Returns 200 if service is alive."""
+    return jsonify({"status": "ok", "service": "queue-service"}), 200
+
+
+@app.route("/enqueue", methods=["POST"])
+def enqueue():
+    """Fallback endpoint — Nginx calls this when all GPU upstreams are down."""
+    r = get_redis()
+    if not r:
+        return jsonify({"error": "Redis unavailable"}), 503
+
+    depth = get_queue_depth(r)
+    if depth >= CIRCUIT_OPEN_THRESHOLD:
+        return jsonify({
+            "error": "Circuit breaker OPEN",
+            "queue_depth": depth,
+            "threshold": CIRCUIT_OPEN_THRESHOLD
+        }), 503
+
+    # Store the request in queue
+    payload = request.get_data(as_text=True)
+    headers = {k: v for k, v in request.headers if k.startswith("X-")}
+    r.rpush(QUEUE_KEY, json.dumps({
+        "payload": payload,
+        "headers": headers,
+        "queued_at": time.time()
+    }))
+
+    new_depth = get_queue_depth(r)
+    return jsonify({
+        "status": "queued",
+        "position": new_depth,
+        "circuit": "warn" if new_depth >= CIRCUIT_WARN_THRESHOLD else "closed"
+    }), 202
+
+
+@app.route("/status")
+def status():
+    """GET queue depth + circuit breaker state + GPU health."""
+    r = get_redis()
+    depth = get_queue_depth(r) if r else -1
+    circuit = "open" if depth >= CIRCUIT_OPEN_THRESHOLD else ("warn" if depth >= CIRCUIT_WARN_THRESHOLD else "closed")
+
+    gpu_health = {}
+    for name, endpoint in GPUS.items():
+        gpu_health[name] = "up" if check_gpu_health(endpoint) else "down"
+
+    return jsonify({
+        "queue_depth": depth,
+        "circuit_breaker": circuit,
+        "gpu_health": gpu_health,
+        "thresholds": {
+            "warn": CIRCUIT_WARN_THRESHOLD,
+            "open": CIRCUIT_OPEN_THRESHOLD
+        }
+    })
+
+
+if __name__ == "__main__":
+    app.run(host="0.0.0.0", port=8091)

router/route_v2.py

@@ -0,0 +1,77 @@
+def route(rd, tier, agent=""):
+    msgs = rd.get("messages",[]); t = estimate_tokens(msgs)
+    sys = any(m.get("role")=="system" for m in msgs)
+    turns = len([m for m in msgs if m.get("role") in ("user","assistant")])
+    hints = rd.get("routing_hints",{})
+    allowed = TIER_MODELS.get(tier, ["gemma-4-12b"])
+    avail = [m for m in available_models() if m in allowed]
+    if not avail: return {"model": allowed[0], "reason": "all_saturated", "saturated": True}
+    if all(is_gpu_busy(m) for m in avail):
+        return {"model": avail[0], "reason": "all_saturated", "saturated": True}
+
+    # GUARD: multimodal -> VLM only (sole vision model)
+    has_image = any(
+        isinstance(m.get("content"), list) and
+        any(p.get("type") == "image_url" for p in m["content"] if isinstance(p, dict))
+        for m in msgs
+    )
+    if has_image:
+        if "gemma-4-12b" in avail and not is_gpu_busy("gemma-4-12b"):
+            return {"model": "gemma-4-12b", "reason": "vision"}
+        elif "gemma-4-12b" in avail:
+            return {"model": "gemma-4-12b", "reason": "vision_saturated", "saturated": True}
+        else:
+            return {"model": allowed[0], "reason": "vision_unavailable"}
+
+    req = rd.get("model","auto")
+    if req != "auto":
+        target = req if req in avail else avail[0]
+        if is_gpu_busy(target) and req in allowed:
+            alts = [m for m in avail if m != target and m in allowed]
+            if alts:
+                alt = select_best_gpu(alts, "explicit", agent)
+                if alt: return alt
+        return {"model": target, "reason": "explicit"}
+
+    if hints:
+        if hints.get("priority")=="speed" and "gemma-4-12b" in avail:
+            return select_best_gpu(["gemma-4-12b"], "hint_speed", agent) or {"model":"gemma-4-12b","reason":"hint_speed"}
+        if hints.get("priority")=="quality" and "qwen3.6-35B-A3B" in avail:
+            return select_best_gpu(["qwen3.6-35B-A3B"], "hint_quality", agent) or {"model":"qwen3.6-35B-A3B","reason":"hint_quality"}
+        if hints.get("priority")=="code" and "qwen3.6-27B-code" in avail:
+            return select_best_gpu(["qwen3.6-27B-code"], "hint_code", agent) or {"model":"qwen3.6-27B-code","reason":"hint_code"}
+
+    first_msg = msgs[0].get("content","") if msgs else ""
+    words = len(first_msg.split()) if isinstance(first_msg, str) else 99
+
+    # TIER 1: Tiny - single-turn micro queries -> VLM (fastest)
+    if not sys and turns <= 1 and t <= 300 and words <= 100 and "gemma-4-12b" in avail:
+        if not is_gpu_busy("gemma-4-12b"):
+            return {"model":"gemma-4-12b","reason":"tiny"}
+        fallback = [m for m in ["qwen3.6-27B-code","qwen3.6-35B-A3B"] if m in avail]
+        result = select_best_gpu(fallback, "tiny_fallback", agent)
+        if result: return result
+
+    # TIER 2: Light - moderate chat -> Dense primary, saves VLM for vision/speed
+    if t <= 5000 and turns <= 4:
+        candidates = [m for m in ["qwen3.6-27B-code","gemma-4-12b","qwen3.6-35B-A3B"] if m in avail]
+        result = select_best_gpu(candidates, "light", agent)
+        if result: return result
+
+    # TIER 3: Medium - quality matters -> MoE primary, Dense fallback
+    if t <= 30000:
+        candidates = [m for m in ["qwen3.6-35B-A3B","qwen3.6-27B-code","gemma-4-12b"] if m in avail]
+        result = select_best_gpu(candidates, "medium", agent)
+        if result: return result
+
+    # TIER 4: Heavy - big context needs big model -> MoE->Dense->VLM
+    if t > 30000:
+        candidates = [m for m in ["qwen3.6-35B-A3B","qwen3.6-27B-code","gemma-4-12b"] if m in avail]
+        result = select_best_gpu(candidates, "heavy", agent)
+        if result: return result
+
+    # TIER 5: Default - best quality first
+    candidates = [m for m in ["qwen3.6-35B-A3B","qwen3.6-27B-code","gemma-4-12b"] if m in avail]
+    result = select_best_gpu(candidates, "default", agent)
+    if result: return result
+    return {"model":avail[0],"reason":"last_resort"}

router/route_v3.py

@@ -0,0 +1,92 @@
+def route(rd, tier, agent=""):
+    msgs = rd.get("messages",[]); t = estimate_tokens(msgs)
+    sys = any(m.get("role")=="system" for m in msgs)
+    turns = len([m for m in msgs if m.get("role") in ("user","assistant")])
+    hints = rd.get("routing_hints",{})
+    allowed = TIER_MODELS.get(tier, ["gemma-4-12b"])
+    avail = [m for m in available_models() if m in allowed]
+    if not avail: return {"model": allowed[0], "reason": "all_saturated", "saturated": True}
+    if all(is_gpu_busy(m) for m in avail):
+        return {"model": avail[0], "reason": "all_saturated", "saturated": True}
+
+    # GUARD: multimodal -> VLM only (sole vision model)
+    has_image = any(
+        isinstance(m.get("content"), list) and
+        any(p.get("type") == "image_url" for p in m["content"] if isinstance(p, dict))
+        for m in msgs
+    )
+    if has_image:
+        if "gemma-4-12b" in avail and not is_gpu_busy("gemma-4-12b"):
+            return {"model": "gemma-4-12b", "reason": "vision"}
+        elif "gemma-4-12b" in avail:
+            return {"model": "gemma-4-12b", "reason": "vision_saturated", "saturated": True}
+        else:
+            return {"model": allowed[0], "reason": "vision_unavailable"}
+
+    req = rd.get("model","auto")
+    if req != "auto":
+        target = req if req in avail else avail[0]
+        if is_gpu_busy(target) and req in allowed:
+            alts = [m for m in avail if m != target and m in allowed]
+            if alts:
+                alt = select_best_gpu(alts, "explicit", agent)
+                if alt: return alt
+        return {"model": target, "reason": "explicit"}
+
+    if hints:
+        if hints.get("priority")=="speed" and "gemma-4-12b" in avail:
+            return select_best_gpu(["gemma-4-12b"], "hint_speed", agent) or {"model":"gemma-4-12b","reason":"hint_speed"}
+        if hints.get("priority")=="quality" and "qwen3.6-35B-A3B" in avail:
+            return select_best_gpu(["qwen3.6-35B-A3B"], "hint_quality", agent) or {"model":"qwen3.6-35B-A3B","reason":"hint_quality"}
+        if hints.get("priority")=="code" and "qwen3.6-27B-code" in avail:
+            return select_best_gpu(["qwen3.6-27B-code"], "hint_code", agent) or {"model":"qwen3.6-27B-code","reason":"hint_code"}
+
+    first_msg = msgs[0].get("content","") if msgs else ""
+    words = len(first_msg.split()) if isinstance(first_msg, str) else 99
+
+    # TIER 1: Tiny - single-turn micro queries -> VLM (fastest)
+    if not sys and turns <= 1 and t <= 300 and words <= 100 and "gemma-4-12b" in avail:
+        if not is_gpu_busy("gemma-4-12b"):
+            return {"model":"gemma-4-12b","reason":"tiny"}
+        fallback = [m for m in ["qwen3.6-27B-code","qwen3.6-35B-A3B"] if m in avail]
+        result = select_best_gpu(fallback, "tiny_fallback", agent)
+        if result: return result
+
+    # TIER 2: Light - moderate chat -> Dense primary, saves VLM for vision/speed
+    if t <= 5000 and turns <= 4:
+        candidates = [m for m in ["qwen3.6-27B-code","gemma-4-12b","qwen3.6-35B-A3B"] if m in avail]
+        result = select_best_gpu(candidates, "light", agent)
+        if result: return result
+
+    # TIER 3: Medium - quality matters -> MoE primary (60%), Dense spillover (40%)
+    if t <= 30000:
+        candidates = moe_spillover(avail, ["qwen3.6-35B-A3B","qwen3.6-27B-code","gemma-4-12b"])
+        result = select_best_gpu(candidates, "medium", agent)
+        if result: return result
+
+    # TIER 4: Heavy - big context -> MoE primary (60%), Dense spillover (40%)
+    if t > 30000:
+        candidates = moe_spillover(avail, ["qwen3.6-35B-A3B","qwen3.6-27B-code","gemma-4-12b"])
+        result = select_best_gpu(candidates, "heavy", agent)
+        if result: return result
+
+    # TIER 5: Default - MoE primary (60%), Dense spillover (40%)
+    candidates = moe_spillover(avail, ["qwen3.6-35B-A3B","qwen3.6-27B-code","gemma-4-12b"])
+    result = select_best_gpu(candidates, "default", agent)
+    if result: return result
+    return {"model":avail[0],"reason":"last_resort"}
+
+
+def moe_spillover(avail, default_order):
+    """Spill 40% of MoE-first traffic to Dense to prevent Strix Halo overheating.
+    Only applies when MoE is first candidate, available, and not busy."""
+    import random
+    if (default_order[0] == "qwen3.6-35B-A3B" 
+        and "qwen3.6-35B-A3B" in avail 
+        and not is_gpu_busy("qwen3.6-35B-A3B")
+        and "qwen3.6-27B-code" in avail
+        and not is_gpu_busy("qwen3.6-27B-code")
+        and random.random() < 0.4):
+        # Swap: Dense first, MoE second
+        return ["qwen3.6-27B-code","qwen3.6-35B-A3B"] + [m for m in default_order[2:] if m in avail and m not in ("qwen3.6-27B-code","qwen3.6-35B-A3B")]
+    return [m for m in default_order if m in avail]