From 4e18f3cff21a3fba3640c74d16090539adfd056e Mon Sep 17 00:00:00 2001
From: "tropibot[bot]" <285933081+tropibot[bot]@users.noreply.github.com>
Date: Sat, 24 Oct 2026 19:11:36 +0000
Subject: [PATCH 1/2] chore: bump ollama to v0.40.11

---
 dappnode_package.json                      |  2 +-
 docker-compose.yml                         |  2 +-
 package_variants/cpu/dappnode_package.json | 11 +++++++----
 3 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/dappnode_package.json b/dappnode_package.json
index 657dd58..df0bf49 100644
--- a/dappnode_package.json
+++ b/dappnode_package.json
@@ -19,7 +19,7 @@
     {
       "arg": "OLLAMA_VERSION",
       "repo": "ollama/ollama",
-      "version": "v0.32.5"
+      "version": "v0.40.11"
     }
   ],
   "version": "0.1.0"
diff --git a/docker-compose.yml b/docker-compose.yml
index 1e328ab..82ee11c 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -4,7 +4,7 @@ services:
     build:
       context: ollama
       args:
-        OLLAMA_VERSION: v0.32.5
+        OLLAMA_VERSION: v0.40.11
     container_name: ollama.ollama.dnp.dappnode.eth
     volumes:
       - ollama:/root/.ollama
diff --git a/package_variants/cpu/dappnode_package.json b/package_variants/cpu/dappnode_package.json
index 7a5d9a1..400d234 100644
--- a/package_variants/cpu/dappnode_package.json
+++ b/package_variants/cpu/dappnode_package.json
@@ -1,10 +1,13 @@
 {
-  "name": "ollama-cpu.dnp.dappnode.eth",
-  "version": "0.1.1",
+  "architectures": [
+    "linux/amd64",
+    "linux/arm64"
+  ],
+  "description": "Run large language models locally on your DAppNode. Ollama inference engine with CPU-only support.\n\n**Features:**\n- CPU acceleration for inference\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n",
   "links": {
     "Model Manager": "http://ui.ollama-cpu.dappnode",
     "api": "http://ollama-cpu.dappnode:11434"
   },
-  "architectures": ["linux/amd64", "linux/arm64"],
-  "description": "Run large language models locally on your DAppNode. Ollama inference engine with CPU-only support.\n\n**Features:**\n- CPU acceleration for inference\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n"
+  "name": "ollama-cpu.dnp.dappnode.eth",
+  "version": "0.1.2"
 }

From 5bd187802abf29695400430b2bddfc9b40e3fb65 Mon Sep 17 00:00:00 2001
From: hcastc00 <hcastc00@estudiantes.unileon.es>
Date: Wed, 28 Oct 2026 17:16:18 +0100
Subject: [PATCH 2/2] feat: support decision models and recover UI after
 restarts

---
 dappnode_package.json                         |   2 +-
 getting-started.md                            |  11 ++
 package_variants/amd/dappnode_package.json    |   2 +-
 package_variants/cpu/dappnode_package.json    |   2 +-
 package_variants/nvidia/dappnode_package.json |   2 +-
 ui/index.html                                 | 128 +++++++++++++-----
 ui/nginx.conf                                 |  19 ++-
 ui/tests/check-restarts.py                    |  96 +++++++++++++
 8 files changed, 226 insertions(+), 36 deletions(-)
 create mode 100644 ui/tests/check-restarts.py

diff --git a/dappnode_package.json b/dappnode_package.json
index df0bf49..b722be9 100644
--- a/dappnode_package.json
+++ b/dappnode_package.json
@@ -6,7 +6,7 @@
   "categories": [
     "AI"
   ],
-  "description": "Run large language models locally on your DAppNode with GPU acceleration. Ollama provides a fast and efficient LLM inference engine with AMD ROCm support.\n\n**Features:**\n- AMD GPU acceleration via ROCm\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- AMD GPU with ROCm support\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)",
+  "description": "Run large language models locally on your DAppNode with GPU acceleration. Ollama provides a fast and efficient LLM inference engine with AMD ROCm support.\n\n**Features:**\n- AMD GPU acceleration via ROCm\n- Typed decisions with probabilities via the Jev-compatible System One API\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- AMD GPU with ROCm support\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)",
   "license": "GPL-3.0",
   "links": {
     "Models library": "https://ollama.com/library"
diff --git a/getting-started.md b/getting-started.md
index 398eada..b7859e9 100644
--- a/getting-started.md
+++ b/getting-started.md
@@ -5,6 +5,7 @@ Now you can:
 - Open the **Model Manager** from your package links to browse, pull, and manage your models.
 - Browse the full [Ollama model library](https://ollama.com/library) to find models.
 - Use the Ollama API to integrate with your apps.
+- Run Jev-style decisions locally: choose a label, estimate a yes/no probability, or score text against a rubric.
 
 ### Quick start
 
@@ -32,6 +33,16 @@ curl http://<ollama-api-url>:11434/api/tags
 
 Point any Ollama-compatible app (Open WebUI, OpenClaw, etc.) to the API URL shown in your package links.
 
+### Decision models (System One)
+
+Use **Decision models only** in Model Manager to filter the catalog. Decision models show a **decision** badge; installed models are identified through their reported capabilities.
+
+Pull a compatible model such as `tev1:0.8b`, `tev1`, `nimble`, `clef`, or `clef-flash`, then use **`POST /v1/systemone`** from your application for choices, yes/no probabilities, and scoring. This is the Jev-compatible API for supported Ollama models; it does not install TypeSafe's hosted Jev model.
+
+The endpoint is available directly at `http://ollama-<variant>.dappnode:11434/v1/systemone` and through Model Manager at `http://ui.ollama-<variant>.dappnode/v1/systemone`. Replace `<variant>` with `cpu`, `amd`, or `nvidia`. Local requests do not require an API key.
+
+See the [decision guide](https://docs.ollama.com/capabilities/decision) and [System One API reference](https://docs.ollama.com/api/systemone) for request formats and limits.
+
 ### Resources
 
 - [Ollama documentation](https://github.com/ollama/ollama/blob/main/README.md)
diff --git a/package_variants/amd/dappnode_package.json b/package_variants/amd/dappnode_package.json
index 9b63971..769b547 100644
--- a/package_variants/amd/dappnode_package.json
+++ b/package_variants/amd/dappnode_package.json
@@ -5,5 +5,5 @@
     "Model Manager": "http://ui.ollama-amd.dappnode",
     "api": "http://ollama-amd.dappnode:11434"
   },
-  "description": "Run large language models locally on your DAppNode with GPU acceleration. Ollama with AMD ROCm support for GPU inference.\n\n**Features:**\n- AMD GPU acceleration via ROCm\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- AMD GPU with ROCm support\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n"
+  "description": "Run large language models locally on your DAppNode with GPU acceleration. Ollama with AMD ROCm support for GPU inference.\n\n**Features:**\n- AMD GPU acceleration via ROCm\n- Typed decisions with probabilities via the Jev-compatible System One API\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- AMD GPU with ROCm support\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n"
 }
diff --git a/package_variants/cpu/dappnode_package.json b/package_variants/cpu/dappnode_package.json
index 400d234..8509d06 100644
--- a/package_variants/cpu/dappnode_package.json
+++ b/package_variants/cpu/dappnode_package.json
@@ -3,7 +3,7 @@
     "linux/amd64",
     "linux/arm64"
   ],
-  "description": "Run large language models locally on your DAppNode. Ollama inference engine with CPU-only support.\n\n**Features:**\n- CPU acceleration for inference\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n",
+  "description": "Run large language models locally on your DAppNode. Ollama inference engine with CPU-only support.\n\n**Features:**\n- CPU acceleration for inference\n- Typed decisions with probabilities via the Jev-compatible System One API\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n",
   "links": {
     "Model Manager": "http://ui.ollama-cpu.dappnode",
     "api": "http://ollama-cpu.dappnode:11434"
diff --git a/package_variants/nvidia/dappnode_package.json b/package_variants/nvidia/dappnode_package.json
index ba38726..c733703 100644
--- a/package_variants/nvidia/dappnode_package.json
+++ b/package_variants/nvidia/dappnode_package.json
@@ -6,5 +6,5 @@
     "api": "http://ollama-nvidia.dappnode:11434"
   },
   "architectures": ["linux/amd64", "linux/arm64"],
-  "description": "Run large language models locally on your DAppNode with GPU acceleration. Ollama with NVIDIA GPU support for GPU inference.\n\n**Features:**\n- NVIDIA GPU acceleration\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- NVIDIA GPU with CUDA support\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n"
+  "description": "Run large language models locally on your DAppNode with GPU acceleration. Ollama with NVIDIA GPU support for GPU inference.\n\n**Features:**\n- NVIDIA GPU acceleration\n- Typed decisions with probabilities via the Jev-compatible System One API\n- Complete privacy - all processing stays local\n- Support for multiple LLM models (Llama, Mistral, CodeLlama, etc.)\n\n**Requirements:**\n- NVIDIA GPU with CUDA support\n- At least 8GB RAM (16GB+ recommended)\n- Sufficient storage for models (10GB+ recommended)\n"
 }
diff --git a/ui/index.html b/ui/index.html
index 4f43bde..c578566 100644
--- a/ui/index.html
+++ b/ui/index.html
@@ -642,6 +642,24 @@
       opacity: 0.3;
       cursor: not-allowed;
     }
+    .decision-badge {
+      display: inline-block;
+      font-size: 11px;
+      font-weight: 500;
+      color: var(--accent-hover);
+      background: var(--surface2);
+      padding: 2px 7px;
+      border-radius: 5px;
+    }
+
+    .decision-filter {
+      display: flex;
+      align-items: center;
+      gap: 8px;
+      font-size: 13px;
+      color: var(--text2);
+      margin: 0 0 14px;
+    }
   </style>
 </head>
 
@@ -691,6 +709,7 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
       <div class="catalog-search">
         <input type="text" id="catalogSearch" placeholder="Search models..." />
       </div>
+      <label class="decision-filter"><input type="checkbox" id="decisionCatalogOnly"> Decision models only</label>
       <div id="catalogList"></div>
       <div class="catalog-pagination" id="catalogPagination"></div>
     </div>
@@ -703,6 +722,7 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
     const API = '/api';
     let runningModels = new Set();
     let installedModelNames = new Set();
+    const decisionCapabilities = new Map();
     let catalogCache = [];
     let catalogQuery = '';
     let catalogPage = 1;
@@ -710,10 +730,14 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
     const tagsCacheMap = {}; // modelName -> [{tag, size}]
 
     // --- Init ---
-    async function init() {
-      await checkStatus();
-      await loadModels();
+    function init() {
+      checkStatus();
+      loadModels();
       loadCatalog();
+      document.getElementById('decisionCatalogOnly').addEventListener('change', () => {
+        catalogPage = 1;
+        loadCatalog();
+      });
       setInterval(() => { checkStatus(); loadModels(); }, 10000);
 
       let searchTimeout;
@@ -735,10 +759,11 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
       pagination.innerHTML = '';
 
       try {
-        const url = catalogQuery
-          ? `/ollama-search?q=${encodeURIComponent(catalogQuery)}`
-          : '/ollama-search';
-        const res = await fetch(url);
+        const params = new URLSearchParams();
+        if (catalogQuery) params.set('q', catalogQuery);
+        if (document.getElementById('decisionCatalogOnly').checked) params.set('c', 'decision');
+        const url = '/ollama-search?' + params.toString();
+        const res = await fetch(url, { signal: AbortSignal.timeout(10000) });
         if (!res.ok) throw new Error(`HTTP ${res.status}`);
         const html = await res.text();
         const parsed = parseCatalogHTML(html);
@@ -756,6 +781,10 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
 
     // Fallback curated list when ollama.com is unreachable
     const FALLBACK_MODELS = [
+      { name: 'tev1', desc: 'Together AI decision models for classification, yes/no probabilities, and scoring', pulls: '', decision: true },
+      { name: 'nimble', desc: 'Bespoke Labs 9B model for typed decisions', pulls: '', decision: true },
+      { name: 'clef', desc: 'Cloudflare decision model with image input support', pulls: '', decision: true },
+      { name: 'clef-flash', desc: 'Compact decision model with image input support', pulls: '', decision: true },
       { name: 'llama3.2', desc: 'Meta Llama 3.2 - lightweight model for diverse tasks', pulls: '' },
       { name: 'llama3.3', desc: 'Meta Llama 3.3 - 70B instruction-tuned model', pulls: '' },
       { name: 'gemma3', desc: 'Google Gemma 3 - lightweight open model', pulls: '' },
@@ -779,9 +808,11 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
     ];
 
     function filterFallback(query) {
-      if (!query) return FALLBACK_MODELS;
+      const models = document.getElementById('decisionCatalogOnly').checked
+        ? FALLBACK_MODELS.filter(m => m.decision) : FALLBACK_MODELS;
+      if (!query) return models;
       const q = query.toLowerCase();
-      return FALLBACK_MODELS.filter(m =>
+      return models.filter(m =>
         m.name.toLowerCase().includes(q) || m.desc.toLowerCase().includes(q)
       );
     }
@@ -812,7 +843,7 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
         const pullMatch = allText.match(/([\d.]+[KMB]?\s*Pulls?)/i);
         if (pullMatch) pulls = pullMatch[1];
 
-        models.push({ name, desc, pulls });
+        models.push({ name, desc, pulls, decision: /\bdecision\b/i.test(allText) });
       });
 
       // Strategy 2: search result IDs
@@ -861,6 +892,7 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
               <div class="catalog-card-info">
                 <div class="catalog-card-name">
                   ${esc(m.name)}
+                  ${m.decision ? '<span class="decision-badge">decision</span>' : ''}
                   ${isInstalled ? '<span class="installed-check">&#10003;</span>' : ''}
                 </div>
                 ${m.desc ? `<div class="catalog-card-desc">${esc(m.desc)}</div>` : ''}
@@ -912,7 +944,9 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
 
       // Fetch tags from ollama.com
       try {
-        const res = await fetch(`/ollama-library/${encodeURIComponent(modelName)}/tags`);
+        const res = await fetch(`/ollama-library/${encodeURIComponent(modelName)}/tags`, {
+          signal: AbortSignal.timeout(10000)
+        });
         if (!res.ok) throw new Error(`HTTP ${res.status}`);
         const html = await res.text();
         const tags = parseTagsHTML(html, modelName);
@@ -1005,7 +1039,7 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
       const dot = document.getElementById('statusDot');
       const text = document.getElementById('statusText');
       try {
-        const res = await fetch(API + '/tags');
+        const res = await fetch(API + '/tags', { signal: AbortSignal.timeout(5000) });
         if (res.ok) {
           dot.className = 'dot';
           text.textContent = 'Connected';
@@ -1016,18 +1050,46 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
       }
       // Check running models
       try {
-        const res = await fetch(API + '/ps');
+        const res = await fetch(API + '/ps', { signal: AbortSignal.timeout(5000) });
         const data = await res.json();
         runningModels = new Set((data.models || []).map(m => m.name));
       } catch { runningModels = new Set(); }
     }
 
+    // Check capabilities once per model digest; retry failed checks on refresh.
+    async function markDecisionModels(models) {
+      const cards = document.querySelectorAll('#modelsList .model-card');
+      await Promise.all(models.map(async (model, index) => {
+        try {
+          const key = model.digest || model.name;
+          if (!decisionCapabilities.has(key)) {
+            const res = await fetch(API + '/show', {
+              method: 'POST',
+              headers: { 'Content-Type': 'application/json' },
+              body: JSON.stringify({ model: model.name }),
+              signal: AbortSignal.timeout(15000)
+            });
+            if (!res.ok) return;
+            const data = await res.json();
+            decisionCapabilities.set(key, data.capabilities?.includes('decision') === true);
+          }
+          if (decisionCapabilities.get(key)) {
+            const badge = document.createElement('span');
+            badge.className = 'decision-badge';
+            badge.textContent = 'decision';
+            cards[index].querySelector('.model-meta').append(badge);
+          }
+        } catch { /* Leave the model usable if its capabilities are unavailable. */ }
+      }));
+    }
+
     // --- Load models ---
     async function loadModels() {
       const list = document.getElementById('modelsList');
       const countEl = document.getElementById('modelCount');
       try {
-        const res = await fetch(API + '/tags');
+        const res = await fetch(API + '/tags', { signal: AbortSignal.timeout(5000) });
+        if (!res.ok) throw new Error(`HTTP ${res.status}`);
         const data = await res.json();
         const models = data.models || [];
         installedModelNames = new Set(models.map(m => m.name));
@@ -1070,6 +1132,7 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
               </div>
             </div>`;
         }).join('');
+        markDecisionModels(models);
       } catch {
         list.innerHTML = `
           <div class="empty-state">
@@ -1098,10 +1161,27 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify({ name, stream: true })
         });
+        if (!res.ok) {
+          const data = await res.json().catch(() => ({}));
+          throw new Error(data.error || `HTTP ${res.status}`);
+        }
 
         const reader = res.body.getReader();
         const decoder = new TextDecoder();
         let buffer = '';
+        let succeeded = false;
+        function processLine(line) {
+          if (!line.trim()) return;
+          const json = JSON.parse(line);
+          if (json.error) throw new Error(json.error);
+          if (json.status === 'success') succeeded = true;
+          if (json.total && json.completed) {
+            const pct = Math.round((json.completed / json.total) * 100);
+            updateProgress(json.status || 'Downloading...', pct);
+          } else {
+            updateProgress(json.status || 'Processing...', -1);
+          }
+        }
 
         while (true) {
           const { done, value } = await reader.read();
@@ -1111,24 +1191,10 @@ <h2>Available models <a href="https://ollama.com/search" target="_blank" style="
           const lines = buffer.split('\n');
           buffer = lines.pop();
 
-          for (const line of lines) {
-            if (!line.trim()) continue;
-            try {
-              const json = JSON.parse(line);
-              if (json.error) {
-                throw new Error(json.error);
-              }
-              if (json.total && json.completed) {
-                const pct = Math.round((json.completed / json.total) * 100);
-                updateProgress(json.status || 'Downloading...', pct);
-              } else {
-                updateProgress(json.status || 'Processing...', -1);
-              }
-            } catch (e) {
-              if (e.message && !e.message.includes('JSON')) throw e;
-            }
-          }
+          for (const line of lines) processLine(line);
         }
+        processLine(buffer + decoder.decode());
+        if (!succeeded) throw new Error('Download ended before Ollama confirmed success. Please retry.');
 
         toast(`Model "${name}" pulled successfully`, 'success');
         input.value = '';
diff --git a/ui/nginx.conf b/ui/nginx.conf
index e3882a1..b0d890c 100644
--- a/ui/nginx.conf
+++ b/ui/nginx.conf
@@ -5,6 +5,11 @@ server {
     root /usr/share/nginx/html;
     index index.html;
 
+    # Resolve at request time so UI startup and backend recreation are independent.
+    resolver 127.0.0.11 valid=5s ipv6=off;
+    resolver_timeout 3s;
+    set $ollama_backend ollama:11434;
+
     # Serve the UI
     location / {
         try_files $uri $uri/ /index.html;
@@ -31,8 +36,20 @@ server {
     }
 
     # Proxy Ollama API requests
+    # System One and other OpenAI-compatible endpoints use /v1/, not /api/.
+    location /v1/ {
+        proxy_pass http://$ollama_backend$request_uri;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_buffering off;
+        proxy_cache off;
+        proxy_read_timeout 600s;
+        proxy_connect_timeout 60s;
+        client_max_body_size 32m;
+    }
+
     location /api/ {
-        proxy_pass http://ollama:11434/api/;
+        proxy_pass http://$ollama_backend$request_uri;
         proxy_set_header Host $host;
         proxy_set_header X-Real-IP $remote_addr;
         proxy_buffering off;
diff --git a/ui/tests/check-restarts.py b/ui/tests/check-restarts.py
new file mode 100644
index 0000000..56a6ed7
--- /dev/null
+++ b/ui/tests/check-restarts.py
@@ -0,0 +1,96 @@
+"""Run with python3 ui/tests/check-restarts.py; requires Docker. No model inference."""
+import json
+import http.client
+from pathlib import Path
+import subprocess
+import tempfile
+import time
+import urllib.error
+import urllib.request
+import uuid
+
+
+def docker(*args):
+    return subprocess.check_output(['docker', *args], text=True).strip()
+
+
+name = 'ollama-ui-check-' + uuid.uuid4().hex[:8]
+containers = []
+ui_dir = Path(__file__).resolve().parents[1]
+image = name + ':test'
+network_created = False
+try:
+    docker('build', '-q', '-t', image, str(ui_dir))
+    docker('network', 'create', name)
+    network_created = True
+    # Make Docker's allocated subnet explicit so the old IP can be held during the test.
+    subnet = json.loads(docker('network', 'inspect', name))[0]['IPAM']['Config'][0]['Subnet']
+    docker('network', 'rm', name)
+    docker('network', 'create', '--subnet', subnet, name)
+    ui = name + '-ui'
+    containers.append(ui)
+    docker('run', '-d', '--name', ui, '--network', name, '-p', '127.0.0.1::80', image)
+    port = docker('port', ui, '80/tcp').rsplit(':', 1)[1]
+    origin = 'http://127.0.0.1:' + port
+
+    def wait_for(path, expected=None):
+        deadline = time.monotonic() + 20
+        while time.monotonic() < deadline:
+            try:
+                with urllib.request.urlopen(origin + path, timeout=2) as response:
+                    body = response.read().decode()
+                    if expected is None or expected in body:
+                        return body
+            except (urllib.error.URLError, http.client.HTTPException, OSError):
+                pass
+            time.sleep(0.25)
+        raise AssertionError('UI did not recover at ' + path)
+
+    assert 'Model Manager' in wait_for('/')  # Starts before backend DNS exists.
+    try:
+        urllib.request.urlopen(origin + '/api/tags', timeout=5)
+        raise AssertionError('Missing backend unexpectedly responded')
+    except urllib.error.HTTPError as error:
+        assert error.code == 502
+
+    with tempfile.TemporaryDirectory(prefix='ollama-ui-check-') as directory:
+        config = Path(directory) / 'default.conf'
+        config.write_text('''server {
+          listen 11434;
+          default_type application/json;
+          location /api/tags { return 200 '{"models":[],"uri":"$request_uri"}'; }
+          location /v1/models { return 200 '{"object":"list","data":[]}'; }
+        }''')
+        backend = name + '-backend'
+        containers.append(backend)
+
+        def start_backend():
+            docker('run', '-d', '--name', backend, '--network', name,
+                   '--network-alias', 'ollama', '-v',
+                   str(config) + ':/etc/nginx/conf.d/default.conf:ro', 'nginx:1.27-alpine')
+            return next(iter(json.loads(docker('inspect', backend))[0]
+                             ['NetworkSettings']['Networks'].values()))['IPAddress']
+
+        old_ip = start_backend()
+        assert json.loads(wait_for('/api/tags?check=1', 'models'))['uri'] == '/api/tags?check=1'
+        assert json.loads(wait_for('/v1/models', 'list'))['object'] == 'list'
+        started_at = json.loads(docker('inspect', ui))[0]['State']['StartedAt']
+        docker('rm', '-f', backend)
+        blocker = name + '-old-ip'
+        containers.append(blocker)
+        docker('run', '-d', '--name', blocker, '--network', name, '--ip', old_ip,
+               'nginx:1.27-alpine')
+        new_ip = start_backend()
+        assert new_ip != old_ip
+        assert 'models' in wait_for('/api/tags', 'models')
+        assert 'list' in wait_for('/v1/models', 'list')
+        assert json.loads(docker('inspect', ui))[0]['State']['StartedAt'] == started_at
+    print('PASS: UI starts without backend and recovers after backend IP changes; both API prefixes work.')
+finally:
+    for container in reversed(containers):
+        subprocess.run(['docker', 'rm', '-fv', container], stdout=subprocess.DEVNULL,
+                       stderr=subprocess.DEVNULL)
+    if network_created:
+        subprocess.run(['docker', 'network', 'rm', name], stdout=subprocess.DEVNULL)
+    subprocess.run(['docker', 'image', 'rm', image], stdout=subprocess.DEVNULL,
+                   stderr=subprocess.DEVNULL)