From 0d0c4e5231419fe6867a221cece8cc0d5d7e8ebf Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Mon, 16 Feb 2026 13:25:43 +0100
Subject: [PATCH 01/13] Make cellpose UI template-safe for static rendering

---
 bioengine_apps/cellpose_finetuning/index.html | 110 +++++++++---------
 1 file changed, 56 insertions(+), 54 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 40dd8202..648cb345 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -53,7 +53,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                     <div class="text-xs text-slate-500 font-medium flex items-center gap-1">
                         <span v-if="isConnected" class="text-green-600 flex items-center gap-1"><i class="fas fa-circle text-[8px]"></i> Connected</span>
                         <span v-else class="text-red-500 flex items-center gap-1"><i class="fas fa-circle text-[8px]"></i> Disconnected</span>
-                        <span v-if="isConnected" class="opacity-75">• {{ workspace }}</span>
+                        <span v-if="isConnected" class="opacity-75">• [[ workspace ]]</span>
                     </div>
                 </div>
             </div>
@@ -86,7 +86,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                         <i class="fas fa-server text-slate-400 text-xs"></i>
                         <select v-model="selectedServiceId" class="bg-transparent border-none text-sm focus:ring-0 w-48 font-medium">
                             <option value="" disabled>Select Service</option>
-                            <option v-for="id in availableServices" :key="id" :value="id">{{ id }}</option>
+                            <option v-for="id in availableServices" :key="id" :value="id">[[ id ]]</option>
                         </select>
                         <button @click="listServices" class="text-slate-400 hover:text-blue-600 transition" title="Refresh Services">
                             <i class="fas fa-sync-alt" :class="{'fa-spin': isScanning}"></i>
@@ -94,7 +94,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                     </div>
                     
                     <div v-else class="flex items-center gap-2 bg-green-50 px-3 py-1.5 rounded-lg border border-green-200 text-green-800 text-sm font-medium">
-                        <i class="fas fa-check-circle"></i> {{ selectedServiceId }}
+                        <i class="fas fa-check-circle"></i> [[ selectedServiceId ]]
                     </div>
 
                     <button @click="disconnect" class="text-slate-500 hover:text-red-600 text-sm font-medium p-2 rounded-lg hover:bg-red-50 transition" title="Disconnect">
@@ -111,7 +111,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                     <i class="fas fa-wifi"></i>
                     <div class="text-xs">
                         <div class="font-bold">Connection Issue</div>
-                        <div>{{ syncError }}</div>
+                        <div>[[ syncError ]]</div>
                     </div>
                     <button @click="syncError = null" class="ml-2 text-white/50 hover:text-white"><i class="fas fa-times"></i></button>
                 </div>
@@ -128,7 +128,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                                     @click="currentTab = tab.id"
                                     :class="currentTab === tab.id ? 'bg-blue-600 text-white shadow-lg shadow-blue-900/50' : 'hover:bg-slate-800 hover:text-white'"
                                     class="w-full text-left px-4 py-3 rounded-lg flex items-center gap-3 transition-all duration-200 font-medium">
-                                    <i :class="tab.icon" class="w-5 text-center"></i> {{ tab.name }}
+                                    <i :class="tab.icon" class="w-5 text-center"></i> [[ tab.name ]]
                                 </button>
                             </nav>
                         </div>
@@ -141,7 +141,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                         </button>
                         <div class="flex items-center justify-between mb-3">
                             <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
-                                {{ monitoredSession.status_type === 'running' ? 'Active Training' : 'Training Status' }}
+                                [[ monitoredSession.status_type === 'running' ? 'Active Training' : 'Training Status' ]]
                             </h3>
                             <span class="w-2 h-2 rounded-full transition-colors duration-300"
                                   :class="{
@@ -154,7 +154,7 @@ <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
                         <div class="bg-slate-800 rounded-lg p-3 border border-slate-700 transition-colors duration-300" 
                              :class="{'border-red-500/50': monitoredSession.status_type === 'failed'}">
                             <div class="flex justify-between items-center mb-1">
-                                <span class="text-white text-sm font-mono font-bold">{{ monitoringSessionId.slice(0, 8) }}...</span>
+                                <span class="text-white text-sm font-mono font-bold">[[ monitoringSessionId.slice(0, 8) ]]...</span>
                                 <span class="text-xs px-1.5 py-0.5 rounded capitalize" 
                                       :class="{
                                           'bg-blue-900 text-blue-300': monitoredSession.status_type === 'running',
@@ -162,7 +162,7 @@ <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
                                           'bg-red-900 text-red-300': monitoredSession.status_type === 'failed',
                                           'bg-gray-700 text-gray-300': monitoredSession.status_type === 'stopped' || monitoredSession.status_type === 'waiting'
                                       }">
-                                    {{ monitoredSession.status_type || 'Run' }}
+                                    [[ monitoredSession.status_type || 'Run' ]]
                                 </span>
                             </div>
                             
@@ -171,12 +171,12 @@ <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
                                      :style="{ width: ((monitoredSession.current_epoch || 0) / (monitoredSession.total_epochs || 1) * 100) + '%' }"></div>
                             </div>
                             <div v-if="monitoredSession.status_type !== 'failed'" class="flex justify-between text-[10px] text-slate-400">
-                                <span>Epoch {{ monitoredSession.current_epoch || 0 }}</span>
-                                <span>{{ monitoredSession.total_epochs || '?' }} Total</span>
+                                <span>Epoch [[ monitoredSession.current_epoch || 0 ]]</span>
+                                <span>[[ monitoredSession.total_epochs || '?' ]] Total</span>
                             </div>
 
                             <div v-if="monitoredSession.status_type === 'failed'" class="mt-2 text-[10px] text-red-300 leading-tight">
-                                {{ monitoredSession.message }}
+                                [[ monitoredSession.message ]]
                             </div>
                         </div>
                     </div>
@@ -195,12 +195,12 @@ <h2 class="text-2xl font-bold text-gray-800 mb-2">Welcome to Cellpose Fine-Tunin
                         <p class="text-gray-500 mb-8" v-else>Select a Cellpose Fine-Tuning Service to continue.</p>
                         
                         <div v-if="isConnected && availableServices.length === 0 && !isScanning" class="bg-orange-50 text-orange-800 p-4 rounded-lg text-sm mb-4">
-                            <i class="fas fa-exclamation-triangle mr-2"></i> No services found in workspace "{{ workspace }}".
+                            <i class="fas fa-exclamation-triangle mr-2"></i> No services found in workspace "[[ workspace ]]".
                             <br>Make sure the Worker is running and you are logged into the correct workspace.
                         </div>
 
                         <div v-if="!isConnected && autoConnectError" class="bg-red-50 text-red-800 p-4 rounded-lg text-sm mb-4">
-                             {{ autoConnectError }}
+                             [[ autoConnectError ]]
                         </div>
 
                          <button v-if="isConnected && !service" @click="listServices" class="text-blue-600 hover:text-blue-800 font-medium">
@@ -214,7 +214,7 @@ <h2 class="text-2xl font-bold text-gray-800 mb-2">Welcome to Cellpose Fine-Tunin
                      <div class="flex justify-between items-end">
                         <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                         <div v-if="syncError" class="text-red-500 text-sm px-4 py-1 bg-red-50 rounded-full border border-red-200 flex items-center" title="Connection to worker is unstable">
-                            <i class="fas fa-exclamation-triangle mr-2"></i> {{ syncError }}
+                            <i class="fas fa-exclamation-triangle mr-2"></i> [[ syncError ]]
                         </div>
                         <button @click="refreshSessions(false)" class="text-gray-500 hover:text-blue-600 transition text-sm font-medium bg-white px-3 py-1.5 rounded-lg border border-gray-200 shadow-sm">
                             <i class="fas fa-sync-alt mr-1" :class="{'fa-spin': refreshLoading}"></i> Refresh
@@ -237,7 +237,7 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                                 <tr v-for="(session, id) in sessions" :key="id" class="hover:bg-slate-50 transition-colors cursor-pointer group" @click="viewSession(id)">
                                     <td class="px-6 py-4 whitespace-nowrap">
                                         <div class="flex items-center">
-                                            <span class="font-mono text-sm text-blue-600 font-medium group-hover:underline">{{ id.slice(0, 8) }}</span>
+                                            <span class="font-mono text-sm text-blue-600 font-medium group-hover:underline">[[ id.slice(0, 8) ]]</span>
                                         </div>
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap">
@@ -248,12 +248,12 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                                                   'bg-red-100 text-red-800': session.status_type === 'failed',
                                                   'bg-gray-100 text-gray-800': session.status_type === 'waiting' || session.status_type === 'stopped'
                                               }">
-                                            {{ session.status_type }}
+                                            [[ session.status_type ]]
                                         </span>
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-sm text-gray-500">
                                          <div v-if="session.total_epochs">
-                                            <div class="flex justify-between mb-1 text-xs"><span>{{ session.current_epoch || 0 }}</span><span>{{ session.total_epochs }}</span></div>
+                                            <div class="flex justify-between mb-1 text-xs"><span>[[ session.current_epoch || 0 ]]</span><span>[[ session.total_epochs ]]</span></div>
                                             <div class="w-24 bg-gray-200 rounded-full h-1.5">
                                                 <div class="bg-blue-500 h-1.5 rounded-full" :style="{ width: ((session.current_epoch || 0) / session.total_epochs * 100) + '%' }"></div>
                                             </div>
@@ -261,10 +261,10 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                                          <span v-else>-</span>
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-sm text-gray-500">
-                                        {{ getLatestLoss(session) }}
+                                        [[ getLatestLoss(session) ]]
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-sm text-gray-500">
-                                        {{ formatDate(session.start_time) }}
+                                        [[ formatDate(session.start_time) ]]
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-right text-sm font-medium" @click.stop>
                                          <button @click="viewSession(id)" class="text-blue-600 hover:text-blue-900 mr-3">Details</button>
@@ -375,13 +375,13 @@ <h3 class="text-lg font-semibold text-gray-800 mb-4 flex items-center gap-2"><i
                              <button type="button" @click="currentTab = 'dashboard'" class="px-6 py-2.5 text-gray-600 font-medium hover:bg-gray-100 rounded-lg transition">Cancel</button>
                              <button type="submit" :disabled="isStartingTraining" class="bg-blue-600 text-white px-8 py-2.5 rounded-lg shadow-sm hover:bg-blue-700 hover:shadow-md transition-all font-medium disabled:opacity-50 flex items-center gap-2">
                                 <i v-if="isStartingTraining" class="fas fa-circle-notch fa-spin"></i>
-                                {{ isStartingTraining ? 'Starting...' : 'Start Training' }}
+                                [[ isStartingTraining ? 'Starting...' : 'Start Training' ]]
                             </button>
                          </div>
 
                         <div v-if="trainingError" class="bg-red-50 text-red-800 p-4 rounded-lg text-sm border border-red-200 flex items-start gap-2">
                             <i class="fas fa-exclamation-circle mt-0.5"></i>
-                            <span class="font-mono whitespace-pre-wrap">{{ trainingError }}</span>
+                            <span class="font-mono whitespace-pre-wrap">[[ trainingError ]]</span>
                         </div>
 
                      </form>
@@ -396,7 +396,7 @@ <h3 class="text-lg font-semibold text-gray-800 mb-4 flex items-center gap-2"><i
                         <div class="flex-1">
                              <h2 class="text-2xl font-bold text-gray-900 flex items-center gap-3">
                                  Session Analysis
-                                 <span class="text-sm font-mono font-normal text-gray-500 bg-gray-100 px-2 py-0.5 rounded">{{ selectedSessionId }}</span>
+                                 <span class="text-sm font-mono font-normal text-gray-500 bg-gray-100 px-2 py-0.5 rounded">[[ selectedSessionId ]]</span>
                              </h2>
                         </div>
                         <div class="flex items-center gap-2">
@@ -420,32 +420,32 @@ <h2 class="text-2xl font-bold text-gray-900 flex items-center gap-3">
                     <div class="grid grid-cols-1 md:grid-cols-4 gap-6 mb-8">
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                             <div class="text-sm text-gray-500 mb-1">Status</div>
-                            <div class="text-lg font-bold capitalize" :class="statusTextColor(currentSession?.status_type)">{{ currentSession?.status_type }}</div>
+                            <div class="text-lg font-bold capitalize" :class="statusTextColor(currentSession?.status_type)">[[ currentSession?.status_type ]]</div>
                         </div>
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                             <div class="text-sm text-gray-500 mb-1">Progress</div>
-                            <div class="text-lg font-bold text-gray-800">{{ currentSession?.current_epoch || 0 }} <span class="text-sm text-gray-400 font-normal">/ {{ currentSession?.total_epochs }} Epochs</span></div>
+                            <div class="text-lg font-bold text-gray-800">[[ currentSession?.current_epoch || 0 ]] <span class="text-sm text-gray-400 font-normal">/ [[ currentSession?.total_epochs ]] Epochs</span></div>
                         </div>
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                             <div class="text-sm text-gray-500 mb-1">Samples</div>
-                            <div class="text-lg font-bold text-gray-800">{{ currentSession?.n_train || '-' }} <span class="text-sm text-gray-400 font-normal">Train</span></div>
+                            <div class="text-lg font-bold text-gray-800">[[ currentSession?.n_train || '-' ]] <span class="text-sm text-gray-400 font-normal">Train</span></div>
                         </div>
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                              <div class="text-sm text-gray-500 mb-1">Duration</div>
-                            <div class="text-lg font-bold text-gray-800">{{ formatDuration(currentSession?.elapsed_seconds) }}</div>
+                            <div class="text-lg font-bold text-gray-800">[[ formatDuration(currentSession?.elapsed_seconds) ]]</div>
                         </div>
                     </div>
 
                     <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200 mb-8">
                         <h3 class="font-bold text-gray-700 mb-3 px-2">Training Hyperparameters</h3>
                         <div class="grid grid-cols-1 md:grid-cols-2 lg:grid-cols-4 gap-3 text-sm">
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Model</span><div class="font-mono text-gray-800">{{ currentSession?.model || '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Epochs</span><div class="font-mono text-gray-800">{{ currentSession?.n_epochs ?? currentSession?.total_epochs ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Samples</span><div class="font-mono text-gray-800">{{ currentSession?.n_samples ?? 'all' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Learning Rate</span><div class="font-mono text-gray-800">{{ currentSession?.learning_rate ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Weight Decay</span><div class="font-mono text-gray-800">{{ currentSession?.weight_decay ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Min Train Masks</span><div class="font-mono text-gray-800">{{ currentSession?.min_train_masks ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Validation Interval</span><div class="font-mono text-gray-800">{{ currentSession?.validation_interval ?? 'default (10)' }}</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Model</span><div class="font-mono text-gray-800">[[ currentSession?.model || '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Epochs</span><div class="font-mono text-gray-800">[[ currentSession?.n_epochs ?? currentSession?.total_epochs ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Samples</span><div class="font-mono text-gray-800">[[ currentSession?.n_samples ?? 'all' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Learning Rate</span><div class="font-mono text-gray-800">[[ currentSession?.learning_rate ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Weight Decay</span><div class="font-mono text-gray-800">[[ currentSession?.weight_decay ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Min Train Masks</span><div class="font-mono text-gray-800">[[ currentSession?.min_train_masks ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Validation Interval</span><div class="font-mono text-gray-800">[[ currentSession?.validation_interval ?? 'default (10)' ]]</div></div>
                         </div>
                     </div>
                 
@@ -470,11 +470,11 @@ <h3 class="font-bold text-gray-700 mb-4 px-2">Validation Metrics</h3>
                         <div class="bg-slate-900 rounded-xl overflow-hidden text-slate-300 font-mono text-xs flex flex-col h-64 shadow-sm border border-slate-800">
                             <div class="bg-slate-950 px-4 py-2 border-b border-slate-800 font-bold text-slate-500 flex justify-between">
                                 <span>SYSTEM LOG</span>
-                                <span class="text-[10px]">{{ currentSession?.status_type === 'running' ? 'LIVE' : 'OFFLINE' }}</span>
+                                <span class="text-[10px]">[[ currentSession?.status_type === 'running' ? 'LIVE' : 'OFFLINE' ]]</span>
                             </div>
                             <div class="p-4 overflow-y-auto log-area flex-grow">
-                                <p class="mb-2"><span class="text-slate-500">Session ID:</span> {{ selectedSessionId }}</p>
-                                <p class="mb-2 text-yellow-500">> {{ currentSession?.message }}</p>
+                                <p class="mb-2"><span class="text-slate-500">Session ID:</span> [[ selectedSessionId ]]</p>
+                                <p class="mb-2 text-yellow-500">> [[ currentSession?.message ]]</p>
                                 <p v-if="currentSession?.status_type === 'failed'" class="text-red-400 mt-2">Error: Process terminated unexpectedly.</p>
                             </div>
                         </div>
@@ -490,13 +490,13 @@ <h3 class="font-bold text-green-900 mb-2">Ready for Export</h3>
                                  </div>
                                  <button @click="exportModel" :disabled="isExporting" class="w-full bg-green-600 text-white px-4 py-3 rounded-lg hover:bg-green-700 disabled:opacity-50 font-medium shadow-sm transition">
                                      <i v-if="isExporting" class="fas fa-spinner fa-spin mr-1"></i>
-                                     {{ isExporting ? 'Exporting Package...' : 'Export to Artifact Manager' }}
+                                     [[ isExporting ? 'Exporting Package...' : 'Export to Artifact Manager' ]]
                                  </button>
                              </div>
                              
                              <div v-if="exportResult" class="mt-4 p-3 bg-white rounded border border-green-200 text-sm">
                                  <div class="flex items-center text-green-600 font-bold mb-1"><i class="fas fa-check-circle mr-2"></i> Export Success</div>
-                                 <a :href="exportResult.artifact_url" target="_blank" class="text-blue-600 hover:underline break-all block">{{ exportResult.model_name }}</a>
+                                 <a :href="exportResult.artifact_url" target="_blank" class="text-blue-600 hover:underline break-all block">[[ exportResult.model_name ]]</a>
                              </div>
                         </div>
                     </div>
@@ -518,11 +518,11 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Model Source</label>
                                     <select v-model="inferenceParams.model" class="w-full bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm">
                                          <optgroup label="Pretrained Models">
-                                             <option v-for="m in pretrainedModels" :key="m" :value="m">{{ m }}</option>
+                                             <option v-for="m in pretrainedModels" :key="m" :value="m">[[ m ]]</option>
                                          </optgroup>
                                          <optgroup label="Your Sessions">
                                              <option v-for="(s, id) in sessions" :key="id" :value="id" v-show="s.status_type==='completed' || s.status_type==='running'">
-                                                 {{ s.status_type==='running' ? '🟢' : '⚫' }} {{ id.slice(0,8) }}
+                                                 [[ s.status_type==='running' ? '🟢' : '⚫' ]] [[ id.slice(0,8) ]]
                                              </option>
                                          </optgroup>
                                     </select>
@@ -538,7 +538,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
 
                                 <div v-if="runningTrainingSessionsCount > 0" class="bg-amber-50 border border-amber-200 text-amber-800 rounded-lg p-3 text-xs leading-relaxed">
                                     <i class="fas fa-triangle-exclamation mr-1"></i>
-                                    {{ runningTrainingSessionsCount }} training session(s) are running. Inference may fail with GPU memory errors while training is active.
+                                    [[ runningTrainingSessionsCount ]] training session(s) are running. Inference may fail with GPU memory errors while training is active.
                                 </div>
 
                                 <div>
@@ -550,7 +550,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                                             <p class="text-xs text-gray-500">Click or Drag Image Here</p>
                                         </div>
                                         <div v-else class="text-sm font-medium text-blue-600 truncate px-2">
-                                            {{ inferenceFile.name }}
+                                            [[ inferenceFile.name ]]
                                         </div>
                                     </div>
                                 </div>
@@ -559,7 +559,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                             <div class="p-4 border-t border-gray-100 bg-gray-50/50 flex-none">
                                 <button @click="runInference" :disabled="!inferenceFile || isInferring" class="w-full bg-purple-600 text-white px-4 py-3 rounded-lg hover:bg-purple-700 transition disabled:opacity-50 font-medium shadow-sm flex justify-center items-center gap-2">
                                     <i v-if="isInferring" class="fas fa-spinner fa-spin"></i>
-                                    {{ isInferring ? 'Processing...' : 'Run Segmentation' }}
+                                    [[ isInferring ? 'Processing...' : 'Run Segmentation' ]]
                                 </button>
                             </div>
                          </div>
@@ -579,7 +579,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                              <!-- Legend -->
                              <transition name="fade">
                                  <div v-if="inferenceResult" class="absolute bottom-4 right-4 bg-black/80 backdrop-blur text-white text-xs px-3 py-1.5 rounded-full border border-white/10 shadow-lg">
-                                     <i class="fas fa-cubes text-purple-400 mr-1"></i> Found <b>{{ inferenceObjectCount }}</b> objects
+                                     <i class="fas fa-cubes text-purple-400 mr-1"></i> Found <b>[[ inferenceObjectCount ]]</b> objects
                                  </div>
                              </transition>
                          </div>
@@ -596,7 +596,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                 <div class="flex justify-between items-center mb-4 border-b border-gray-100 pb-4 flex-none">
                     <div>
                         <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
-                        <p class="text-xs text-gray-500">{{ trainingParams.artifact || 'No artifact selected' }}</p>
+                        <p class="text-xs text-gray-500">[[ trainingParams.artifact || 'No artifact selected' ]]</p>
                     </div>
                     <button @click="showFileBrowser = false" class="text-gray-400 hover:text-gray-600"><i class="fas fa-times"></i></button>
                 </div>
@@ -604,7 +604,7 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
                 <div class="flex-none bg-gray-50 p-2 rounded-lg mb-4 flex items-center gap-2 text-sm text-gray-600 border border-gray-200">
                     <button @click="browserNavigateUp" class="hover:bg-gray-200 p-1 rounded transition" :disabled="!browserPath"><i class="fas fa-arrow-up"></i></button>
                     <span class="font-mono text-gray-400">/</span>
-                    <span class="font-mono">{{ browserPath }}</span>
+                    <span class="font-mono">[[ browserPath ]]</span>
                 </div>
 
                 <div class="flex-grow overflow-y-auto min-h-0 bg-white border border-gray-100 rounded-lg">
@@ -612,7 +612,7 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
                         <i class="fas fa-circle-notch fa-spin text-2xl mb-2"></i><br>Loading files...
                     </div>
                     <div v-else-if="browserError" class="p-8 text-center text-red-400">
-                        <i class="fas fa-exclamation-circle text-2xl mb-2"></i><br>{{ browserError }}
+                        <i class="fas fa-exclamation-circle text-2xl mb-2"></i><br>[[ browserError ]]
                     </div>
                     <ul v-else class="divide-y divide-gray-50">
                         <li v-if="browserPath" @click="browserNavigateUp" class="px-4 py-2 hover:bg-gray-50 cursor-pointer flex items-center gap-3 text-gray-500">
@@ -624,7 +624,7 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
                             
                             <div class="flex items-center gap-3 overflow-hidden">
                                 <i :class="item.type === 'folder' ? 'fas fa-folder text-yellow-400' : 'fas fa-file text-gray-400'"></i>
-                                <span class="text-sm truncate" :class="item.type === 'folder' ? 'font-medium text-gray-700' : 'text-gray-500'">{{ item.name }}</span>
+                                <span class="text-sm truncate" :class="item.type === 'folder' ? 'font-medium text-gray-700' : 'text-gray-500'">[[ item.name ]]</span>
                             </div>
 
                             <div v-if="item.type === 'folder'" class="flex gap-2 opacity-0 group-hover:opacity-100 transition-opacity">
@@ -640,10 +640,10 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
 
                 <div class="mt-4 pt-4 border-t border-gray-100 flex justify-between items-center text-xs text-gray-400">
                     <div class="flex items-center gap-3">
-                        <span>Current: /{{ browserPath }}</span>
+                        <span>Current: /[[ browserPath ]]</span>
                         <transition name="fade">
                             <span v-if="browserFeedback" class="text-green-600 font-bold bg-green-50 px-2 py-0.5 rounded border border-green-200">
-                                <i class="fas fa-check"></i> {{ browserFeedback }}
+                                <i class="fas fa-check"></i> [[ browserFeedback ]]
                             </span>
                         </transition>
                     </div>
@@ -675,12 +675,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     <div v-if="isUploading" class="space-y-2 pt-2">
                          <div class="flex justify-between text-xs font-medium text-gray-600">
                             <span>Uploading...</span>
-                            <span>{{ uploadProgress }}%</span>
+                            <span>[[ uploadProgress ]]%</span>
                         </div>
                         <div class="w-full bg-gray-100 rounded-full h-2 overflow-hidden">
                             <div class="bg-green-500 h-2 rounded-full transition-all duration-300" :style="{ width: uploadProgress + '%' }"></div>
                         </div>
-                        <div class="text-xs text-gray-400 truncate">{{ uploadStatus }}</div>
+                        <div class="text-xs text-gray-400 truncate">[[ uploadStatus ]]</div>
                     </div>
 
                     <div class="flex justify-end gap-3 pt-4 border-t border-gray-100">
@@ -696,7 +696,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
     <script>
         const { createApp, ref, reactive, computed, onMounted, watch } = Vue;
 
-        createApp({
+        const app = createApp({
             setup() {
                 // Connection State
                 const serverUrl = ref('https://hypha.aicell.io');
@@ -1954,7 +1954,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     statusTextColor, formatDuration, handleFolderSelect, getLatestLoss, formatDate,
                 };
             }
-        }).mount('#app');
+        });
+        app.config.compilerOptions.delimiters = ['[[', ']]'];
+        app.mount('#app');
     </script>
 </body>
 </html>
\ No newline at end of file

From 7d9ae4d5ab681ad90d6f43bb53b216a28f751b40 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Mon, 16 Feb 2026 13:25:43 +0100
Subject: [PATCH 02/13] Make cellpose UI template-safe for static rendering

---
 bioengine_apps/cellpose_finetuning/index.html | 110 +++++++++---------
 1 file changed, 56 insertions(+), 54 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 40dd8202..648cb345 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -53,7 +53,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                     <div class="text-xs text-slate-500 font-medium flex items-center gap-1">
                         <span v-if="isConnected" class="text-green-600 flex items-center gap-1"><i class="fas fa-circle text-[8px]"></i> Connected</span>
                         <span v-else class="text-red-500 flex items-center gap-1"><i class="fas fa-circle text-[8px]"></i> Disconnected</span>
-                        <span v-if="isConnected" class="opacity-75">• {{ workspace }}</span>
+                        <span v-if="isConnected" class="opacity-75">• [[ workspace ]]</span>
                     </div>
                 </div>
             </div>
@@ -86,7 +86,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                         <i class="fas fa-server text-slate-400 text-xs"></i>
                         <select v-model="selectedServiceId" class="bg-transparent border-none text-sm focus:ring-0 w-48 font-medium">
                             <option value="" disabled>Select Service</option>
-                            <option v-for="id in availableServices" :key="id" :value="id">{{ id }}</option>
+                            <option v-for="id in availableServices" :key="id" :value="id">[[ id ]]</option>
                         </select>
                         <button @click="listServices" class="text-slate-400 hover:text-blue-600 transition" title="Refresh Services">
                             <i class="fas fa-sync-alt" :class="{'fa-spin': isScanning}"></i>
@@ -94,7 +94,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                     </div>
                     
                     <div v-else class="flex items-center gap-2 bg-green-50 px-3 py-1.5 rounded-lg border border-green-200 text-green-800 text-sm font-medium">
-                        <i class="fas fa-check-circle"></i> {{ selectedServiceId }}
+                        <i class="fas fa-check-circle"></i> [[ selectedServiceId ]]
                     </div>
 
                     <button @click="disconnect" class="text-slate-500 hover:text-red-600 text-sm font-medium p-2 rounded-lg hover:bg-red-50 transition" title="Disconnect">
@@ -111,7 +111,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                     <i class="fas fa-wifi"></i>
                     <div class="text-xs">
                         <div class="font-bold">Connection Issue</div>
-                        <div>{{ syncError }}</div>
+                        <div>[[ syncError ]]</div>
                     </div>
                     <button @click="syncError = null" class="ml-2 text-white/50 hover:text-white"><i class="fas fa-times"></i></button>
                 </div>
@@ -128,7 +128,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                                     @click="currentTab = tab.id"
                                     :class="currentTab === tab.id ? 'bg-blue-600 text-white shadow-lg shadow-blue-900/50' : 'hover:bg-slate-800 hover:text-white'"
                                     class="w-full text-left px-4 py-3 rounded-lg flex items-center gap-3 transition-all duration-200 font-medium">
-                                    <i :class="tab.icon" class="w-5 text-center"></i> {{ tab.name }}
+                                    <i :class="tab.icon" class="w-5 text-center"></i> [[ tab.name ]]
                                 </button>
                             </nav>
                         </div>
@@ -141,7 +141,7 @@ <h1 class="font-bold text-lg text-slate-800 leading-tight">Cellpose Fine-Tuning<
                         </button>
                         <div class="flex items-center justify-between mb-3">
                             <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
-                                {{ monitoredSession.status_type === 'running' ? 'Active Training' : 'Training Status' }}
+                                [[ monitoredSession.status_type === 'running' ? 'Active Training' : 'Training Status' ]]
                             </h3>
                             <span class="w-2 h-2 rounded-full transition-colors duration-300"
                                   :class="{
@@ -154,7 +154,7 @@ <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
                         <div class="bg-slate-800 rounded-lg p-3 border border-slate-700 transition-colors duration-300" 
                              :class="{'border-red-500/50': monitoredSession.status_type === 'failed'}">
                             <div class="flex justify-between items-center mb-1">
-                                <span class="text-white text-sm font-mono font-bold">{{ monitoringSessionId.slice(0, 8) }}...</span>
+                                <span class="text-white text-sm font-mono font-bold">[[ monitoringSessionId.slice(0, 8) ]]...</span>
                                 <span class="text-xs px-1.5 py-0.5 rounded capitalize" 
                                       :class="{
                                           'bg-blue-900 text-blue-300': monitoredSession.status_type === 'running',
@@ -162,7 +162,7 @@ <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
                                           'bg-red-900 text-red-300': monitoredSession.status_type === 'failed',
                                           'bg-gray-700 text-gray-300': monitoredSession.status_type === 'stopped' || monitoredSession.status_type === 'waiting'
                                       }">
-                                    {{ monitoredSession.status_type || 'Run' }}
+                                    [[ monitoredSession.status_type || 'Run' ]]
                                 </span>
                             </div>
                             
@@ -171,12 +171,12 @@ <h3 class="text-xs font-bold text-slate-400 uppercase tracking-wider">
                                      :style="{ width: ((monitoredSession.current_epoch || 0) / (monitoredSession.total_epochs || 1) * 100) + '%' }"></div>
                             </div>
                             <div v-if="monitoredSession.status_type !== 'failed'" class="flex justify-between text-[10px] text-slate-400">
-                                <span>Epoch {{ monitoredSession.current_epoch || 0 }}</span>
-                                <span>{{ monitoredSession.total_epochs || '?' }} Total</span>
+                                <span>Epoch [[ monitoredSession.current_epoch || 0 ]]</span>
+                                <span>[[ monitoredSession.total_epochs || '?' ]] Total</span>
                             </div>
 
                             <div v-if="monitoredSession.status_type === 'failed'" class="mt-2 text-[10px] text-red-300 leading-tight">
-                                {{ monitoredSession.message }}
+                                [[ monitoredSession.message ]]
                             </div>
                         </div>
                     </div>
@@ -195,12 +195,12 @@ <h2 class="text-2xl font-bold text-gray-800 mb-2">Welcome to Cellpose Fine-Tunin
                         <p class="text-gray-500 mb-8" v-else>Select a Cellpose Fine-Tuning Service to continue.</p>
                         
                         <div v-if="isConnected && availableServices.length === 0 && !isScanning" class="bg-orange-50 text-orange-800 p-4 rounded-lg text-sm mb-4">
-                            <i class="fas fa-exclamation-triangle mr-2"></i> No services found in workspace "{{ workspace }}".
+                            <i class="fas fa-exclamation-triangle mr-2"></i> No services found in workspace "[[ workspace ]]".
                             <br>Make sure the Worker is running and you are logged into the correct workspace.
                         </div>
 
                         <div v-if="!isConnected && autoConnectError" class="bg-red-50 text-red-800 p-4 rounded-lg text-sm mb-4">
-                             {{ autoConnectError }}
+                             [[ autoConnectError ]]
                         </div>
 
                          <button v-if="isConnected && !service" @click="listServices" class="text-blue-600 hover:text-blue-800 font-medium">
@@ -214,7 +214,7 @@ <h2 class="text-2xl font-bold text-gray-800 mb-2">Welcome to Cellpose Fine-Tunin
                      <div class="flex justify-between items-end">
                         <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                         <div v-if="syncError" class="text-red-500 text-sm px-4 py-1 bg-red-50 rounded-full border border-red-200 flex items-center" title="Connection to worker is unstable">
-                            <i class="fas fa-exclamation-triangle mr-2"></i> {{ syncError }}
+                            <i class="fas fa-exclamation-triangle mr-2"></i> [[ syncError ]]
                         </div>
                         <button @click="refreshSessions(false)" class="text-gray-500 hover:text-blue-600 transition text-sm font-medium bg-white px-3 py-1.5 rounded-lg border border-gray-200 shadow-sm">
                             <i class="fas fa-sync-alt mr-1" :class="{'fa-spin': refreshLoading}"></i> Refresh
@@ -237,7 +237,7 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                                 <tr v-for="(session, id) in sessions" :key="id" class="hover:bg-slate-50 transition-colors cursor-pointer group" @click="viewSession(id)">
                                     <td class="px-6 py-4 whitespace-nowrap">
                                         <div class="flex items-center">
-                                            <span class="font-mono text-sm text-blue-600 font-medium group-hover:underline">{{ id.slice(0, 8) }}</span>
+                                            <span class="font-mono text-sm text-blue-600 font-medium group-hover:underline">[[ id.slice(0, 8) ]]</span>
                                         </div>
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap">
@@ -248,12 +248,12 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                                                   'bg-red-100 text-red-800': session.status_type === 'failed',
                                                   'bg-gray-100 text-gray-800': session.status_type === 'waiting' || session.status_type === 'stopped'
                                               }">
-                                            {{ session.status_type }}
+                                            [[ session.status_type ]]
                                         </span>
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-sm text-gray-500">
                                          <div v-if="session.total_epochs">
-                                            <div class="flex justify-between mb-1 text-xs"><span>{{ session.current_epoch || 0 }}</span><span>{{ session.total_epochs }}</span></div>
+                                            <div class="flex justify-between mb-1 text-xs"><span>[[ session.current_epoch || 0 ]]</span><span>[[ session.total_epochs ]]</span></div>
                                             <div class="w-24 bg-gray-200 rounded-full h-1.5">
                                                 <div class="bg-blue-500 h-1.5 rounded-full" :style="{ width: ((session.current_epoch || 0) / session.total_epochs * 100) + '%' }"></div>
                                             </div>
@@ -261,10 +261,10 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                                          <span v-else>-</span>
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-sm text-gray-500">
-                                        {{ getLatestLoss(session) }}
+                                        [[ getLatestLoss(session) ]]
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-sm text-gray-500">
-                                        {{ formatDate(session.start_time) }}
+                                        [[ formatDate(session.start_time) ]]
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-right text-sm font-medium" @click.stop>
                                          <button @click="viewSession(id)" class="text-blue-600 hover:text-blue-900 mr-3">Details</button>
@@ -375,13 +375,13 @@ <h3 class="text-lg font-semibold text-gray-800 mb-4 flex items-center gap-2"><i
                              <button type="button" @click="currentTab = 'dashboard'" class="px-6 py-2.5 text-gray-600 font-medium hover:bg-gray-100 rounded-lg transition">Cancel</button>
                              <button type="submit" :disabled="isStartingTraining" class="bg-blue-600 text-white px-8 py-2.5 rounded-lg shadow-sm hover:bg-blue-700 hover:shadow-md transition-all font-medium disabled:opacity-50 flex items-center gap-2">
                                 <i v-if="isStartingTraining" class="fas fa-circle-notch fa-spin"></i>
-                                {{ isStartingTraining ? 'Starting...' : 'Start Training' }}
+                                [[ isStartingTraining ? 'Starting...' : 'Start Training' ]]
                             </button>
                          </div>
 
                         <div v-if="trainingError" class="bg-red-50 text-red-800 p-4 rounded-lg text-sm border border-red-200 flex items-start gap-2">
                             <i class="fas fa-exclamation-circle mt-0.5"></i>
-                            <span class="font-mono whitespace-pre-wrap">{{ trainingError }}</span>
+                            <span class="font-mono whitespace-pre-wrap">[[ trainingError ]]</span>
                         </div>
 
                      </form>
@@ -396,7 +396,7 @@ <h3 class="text-lg font-semibold text-gray-800 mb-4 flex items-center gap-2"><i
                         <div class="flex-1">
                              <h2 class="text-2xl font-bold text-gray-900 flex items-center gap-3">
                                  Session Analysis
-                                 <span class="text-sm font-mono font-normal text-gray-500 bg-gray-100 px-2 py-0.5 rounded">{{ selectedSessionId }}</span>
+                                 <span class="text-sm font-mono font-normal text-gray-500 bg-gray-100 px-2 py-0.5 rounded">[[ selectedSessionId ]]</span>
                              </h2>
                         </div>
                         <div class="flex items-center gap-2">
@@ -420,32 +420,32 @@ <h2 class="text-2xl font-bold text-gray-900 flex items-center gap-3">
                     <div class="grid grid-cols-1 md:grid-cols-4 gap-6 mb-8">
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                             <div class="text-sm text-gray-500 mb-1">Status</div>
-                            <div class="text-lg font-bold capitalize" :class="statusTextColor(currentSession?.status_type)">{{ currentSession?.status_type }}</div>
+                            <div class="text-lg font-bold capitalize" :class="statusTextColor(currentSession?.status_type)">[[ currentSession?.status_type ]]</div>
                         </div>
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                             <div class="text-sm text-gray-500 mb-1">Progress</div>
-                            <div class="text-lg font-bold text-gray-800">{{ currentSession?.current_epoch || 0 }} <span class="text-sm text-gray-400 font-normal">/ {{ currentSession?.total_epochs }} Epochs</span></div>
+                            <div class="text-lg font-bold text-gray-800">[[ currentSession?.current_epoch || 0 ]] <span class="text-sm text-gray-400 font-normal">/ [[ currentSession?.total_epochs ]] Epochs</span></div>
                         </div>
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                             <div class="text-sm text-gray-500 mb-1">Samples</div>
-                            <div class="text-lg font-bold text-gray-800">{{ currentSession?.n_train || '-' }} <span class="text-sm text-gray-400 font-normal">Train</span></div>
+                            <div class="text-lg font-bold text-gray-800">[[ currentSession?.n_train || '-' ]] <span class="text-sm text-gray-400 font-normal">Train</span></div>
                         </div>
                         <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200">
                              <div class="text-sm text-gray-500 mb-1">Duration</div>
-                            <div class="text-lg font-bold text-gray-800">{{ formatDuration(currentSession?.elapsed_seconds) }}</div>
+                            <div class="text-lg font-bold text-gray-800">[[ formatDuration(currentSession?.elapsed_seconds) ]]</div>
                         </div>
                     </div>
 
                     <div class="bg-white p-4 rounded-xl shadow-sm border border-gray-200 mb-8">
                         <h3 class="font-bold text-gray-700 mb-3 px-2">Training Hyperparameters</h3>
                         <div class="grid grid-cols-1 md:grid-cols-2 lg:grid-cols-4 gap-3 text-sm">
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Model</span><div class="font-mono text-gray-800">{{ currentSession?.model || '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Epochs</span><div class="font-mono text-gray-800">{{ currentSession?.n_epochs ?? currentSession?.total_epochs ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Samples</span><div class="font-mono text-gray-800">{{ currentSession?.n_samples ?? 'all' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Learning Rate</span><div class="font-mono text-gray-800">{{ currentSession?.learning_rate ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Weight Decay</span><div class="font-mono text-gray-800">{{ currentSession?.weight_decay ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Min Train Masks</span><div class="font-mono text-gray-800">{{ currentSession?.min_train_masks ?? '-' }}</div></div>
-                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Validation Interval</span><div class="font-mono text-gray-800">{{ currentSession?.validation_interval ?? 'default (10)' }}</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Model</span><div class="font-mono text-gray-800">[[ currentSession?.model || '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Epochs</span><div class="font-mono text-gray-800">[[ currentSession?.n_epochs ?? currentSession?.total_epochs ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Samples</span><div class="font-mono text-gray-800">[[ currentSession?.n_samples ?? 'all' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Learning Rate</span><div class="font-mono text-gray-800">[[ currentSession?.learning_rate ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Weight Decay</span><div class="font-mono text-gray-800">[[ currentSession?.weight_decay ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Min Train Masks</span><div class="font-mono text-gray-800">[[ currentSession?.min_train_masks ?? '-' ]]</div></div>
+                            <div class="bg-gray-50 rounded p-2"><span class="text-gray-500">Validation Interval</span><div class="font-mono text-gray-800">[[ currentSession?.validation_interval ?? 'default (10)' ]]</div></div>
                         </div>
                     </div>
                 
@@ -470,11 +470,11 @@ <h3 class="font-bold text-gray-700 mb-4 px-2">Validation Metrics</h3>
                         <div class="bg-slate-900 rounded-xl overflow-hidden text-slate-300 font-mono text-xs flex flex-col h-64 shadow-sm border border-slate-800">
                             <div class="bg-slate-950 px-4 py-2 border-b border-slate-800 font-bold text-slate-500 flex justify-between">
                                 <span>SYSTEM LOG</span>
-                                <span class="text-[10px]">{{ currentSession?.status_type === 'running' ? 'LIVE' : 'OFFLINE' }}</span>
+                                <span class="text-[10px]">[[ currentSession?.status_type === 'running' ? 'LIVE' : 'OFFLINE' ]]</span>
                             </div>
                             <div class="p-4 overflow-y-auto log-area flex-grow">
-                                <p class="mb-2"><span class="text-slate-500">Session ID:</span> {{ selectedSessionId }}</p>
-                                <p class="mb-2 text-yellow-500">> {{ currentSession?.message }}</p>
+                                <p class="mb-2"><span class="text-slate-500">Session ID:</span> [[ selectedSessionId ]]</p>
+                                <p class="mb-2 text-yellow-500">> [[ currentSession?.message ]]</p>
                                 <p v-if="currentSession?.status_type === 'failed'" class="text-red-400 mt-2">Error: Process terminated unexpectedly.</p>
                             </div>
                         </div>
@@ -490,13 +490,13 @@ <h3 class="font-bold text-green-900 mb-2">Ready for Export</h3>
                                  </div>
                                  <button @click="exportModel" :disabled="isExporting" class="w-full bg-green-600 text-white px-4 py-3 rounded-lg hover:bg-green-700 disabled:opacity-50 font-medium shadow-sm transition">
                                      <i v-if="isExporting" class="fas fa-spinner fa-spin mr-1"></i>
-                                     {{ isExporting ? 'Exporting Package...' : 'Export to Artifact Manager' }}
+                                     [[ isExporting ? 'Exporting Package...' : 'Export to Artifact Manager' ]]
                                  </button>
                              </div>
                              
                              <div v-if="exportResult" class="mt-4 p-3 bg-white rounded border border-green-200 text-sm">
                                  <div class="flex items-center text-green-600 font-bold mb-1"><i class="fas fa-check-circle mr-2"></i> Export Success</div>
-                                 <a :href="exportResult.artifact_url" target="_blank" class="text-blue-600 hover:underline break-all block">{{ exportResult.model_name }}</a>
+                                 <a :href="exportResult.artifact_url" target="_blank" class="text-blue-600 hover:underline break-all block">[[ exportResult.model_name ]]</a>
                              </div>
                         </div>
                     </div>
@@ -518,11 +518,11 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Model Source</label>
                                     <select v-model="inferenceParams.model" class="w-full bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm">
                                          <optgroup label="Pretrained Models">
-                                             <option v-for="m in pretrainedModels" :key="m" :value="m">{{ m }}</option>
+                                             <option v-for="m in pretrainedModels" :key="m" :value="m">[[ m ]]</option>
                                          </optgroup>
                                          <optgroup label="Your Sessions">
                                              <option v-for="(s, id) in sessions" :key="id" :value="id" v-show="s.status_type==='completed' || s.status_type==='running'">
-                                                 {{ s.status_type==='running' ? '🟢' : '⚫' }} {{ id.slice(0,8) }}
+                                                 [[ s.status_type==='running' ? '🟢' : '⚫' ]] [[ id.slice(0,8) ]]
                                              </option>
                                          </optgroup>
                                     </select>
@@ -538,7 +538,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
 
                                 <div v-if="runningTrainingSessionsCount > 0" class="bg-amber-50 border border-amber-200 text-amber-800 rounded-lg p-3 text-xs leading-relaxed">
                                     <i class="fas fa-triangle-exclamation mr-1"></i>
-                                    {{ runningTrainingSessionsCount }} training session(s) are running. Inference may fail with GPU memory errors while training is active.
+                                    [[ runningTrainingSessionsCount ]] training session(s) are running. Inference may fail with GPU memory errors while training is active.
                                 </div>
 
                                 <div>
@@ -550,7 +550,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                                             <p class="text-xs text-gray-500">Click or Drag Image Here</p>
                                         </div>
                                         <div v-else class="text-sm font-medium text-blue-600 truncate px-2">
-                                            {{ inferenceFile.name }}
+                                            [[ inferenceFile.name ]]
                                         </div>
                                     </div>
                                 </div>
@@ -559,7 +559,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                             <div class="p-4 border-t border-gray-100 bg-gray-50/50 flex-none">
                                 <button @click="runInference" :disabled="!inferenceFile || isInferring" class="w-full bg-purple-600 text-white px-4 py-3 rounded-lg hover:bg-purple-700 transition disabled:opacity-50 font-medium shadow-sm flex justify-center items-center gap-2">
                                     <i v-if="isInferring" class="fas fa-spinner fa-spin"></i>
-                                    {{ isInferring ? 'Processing...' : 'Run Segmentation' }}
+                                    [[ isInferring ? 'Processing...' : 'Run Segmentation' ]]
                                 </button>
                             </div>
                          </div>
@@ -579,7 +579,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                              <!-- Legend -->
                              <transition name="fade">
                                  <div v-if="inferenceResult" class="absolute bottom-4 right-4 bg-black/80 backdrop-blur text-white text-xs px-3 py-1.5 rounded-full border border-white/10 shadow-lg">
-                                     <i class="fas fa-cubes text-purple-400 mr-1"></i> Found <b>{{ inferenceObjectCount }}</b> objects
+                                     <i class="fas fa-cubes text-purple-400 mr-1"></i> Found <b>[[ inferenceObjectCount ]]</b> objects
                                  </div>
                              </transition>
                          </div>
@@ -596,7 +596,7 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                 <div class="flex justify-between items-center mb-4 border-b border-gray-100 pb-4 flex-none">
                     <div>
                         <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
-                        <p class="text-xs text-gray-500">{{ trainingParams.artifact || 'No artifact selected' }}</p>
+                        <p class="text-xs text-gray-500">[[ trainingParams.artifact || 'No artifact selected' ]]</p>
                     </div>
                     <button @click="showFileBrowser = false" class="text-gray-400 hover:text-gray-600"><i class="fas fa-times"></i></button>
                 </div>
@@ -604,7 +604,7 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
                 <div class="flex-none bg-gray-50 p-2 rounded-lg mb-4 flex items-center gap-2 text-sm text-gray-600 border border-gray-200">
                     <button @click="browserNavigateUp" class="hover:bg-gray-200 p-1 rounded transition" :disabled="!browserPath"><i class="fas fa-arrow-up"></i></button>
                     <span class="font-mono text-gray-400">/</span>
-                    <span class="font-mono">{{ browserPath }}</span>
+                    <span class="font-mono">[[ browserPath ]]</span>
                 </div>
 
                 <div class="flex-grow overflow-y-auto min-h-0 bg-white border border-gray-100 rounded-lg">
@@ -612,7 +612,7 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
                         <i class="fas fa-circle-notch fa-spin text-2xl mb-2"></i><br>Loading files...
                     </div>
                     <div v-else-if="browserError" class="p-8 text-center text-red-400">
-                        <i class="fas fa-exclamation-circle text-2xl mb-2"></i><br>{{ browserError }}
+                        <i class="fas fa-exclamation-circle text-2xl mb-2"></i><br>[[ browserError ]]
                     </div>
                     <ul v-else class="divide-y divide-gray-50">
                         <li v-if="browserPath" @click="browserNavigateUp" class="px-4 py-2 hover:bg-gray-50 cursor-pointer flex items-center gap-3 text-gray-500">
@@ -624,7 +624,7 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
                             
                             <div class="flex items-center gap-3 overflow-hidden">
                                 <i :class="item.type === 'folder' ? 'fas fa-folder text-yellow-400' : 'fas fa-file text-gray-400'"></i>
-                                <span class="text-sm truncate" :class="item.type === 'folder' ? 'font-medium text-gray-700' : 'text-gray-500'">{{ item.name }}</span>
+                                <span class="text-sm truncate" :class="item.type === 'folder' ? 'font-medium text-gray-700' : 'text-gray-500'">[[ item.name ]]</span>
                             </div>
 
                             <div v-if="item.type === 'folder'" class="flex gap-2 opacity-0 group-hover:opacity-100 transition-opacity">
@@ -640,10 +640,10 @@ <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
 
                 <div class="mt-4 pt-4 border-t border-gray-100 flex justify-between items-center text-xs text-gray-400">
                     <div class="flex items-center gap-3">
-                        <span>Current: /{{ browserPath }}</span>
+                        <span>Current: /[[ browserPath ]]</span>
                         <transition name="fade">
                             <span v-if="browserFeedback" class="text-green-600 font-bold bg-green-50 px-2 py-0.5 rounded border border-green-200">
-                                <i class="fas fa-check"></i> {{ browserFeedback }}
+                                <i class="fas fa-check"></i> [[ browserFeedback ]]
                             </span>
                         </transition>
                     </div>
@@ -675,12 +675,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     <div v-if="isUploading" class="space-y-2 pt-2">
                          <div class="flex justify-between text-xs font-medium text-gray-600">
                             <span>Uploading...</span>
-                            <span>{{ uploadProgress }}%</span>
+                            <span>[[ uploadProgress ]]%</span>
                         </div>
                         <div class="w-full bg-gray-100 rounded-full h-2 overflow-hidden">
                             <div class="bg-green-500 h-2 rounded-full transition-all duration-300" :style="{ width: uploadProgress + '%' }"></div>
                         </div>
-                        <div class="text-xs text-gray-400 truncate">{{ uploadStatus }}</div>
+                        <div class="text-xs text-gray-400 truncate">[[ uploadStatus ]]</div>
                     </div>
 
                     <div class="flex justify-end gap-3 pt-4 border-t border-gray-100">
@@ -696,7 +696,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
     <script>
         const { createApp, ref, reactive, computed, onMounted, watch } = Vue;
 
-        createApp({
+        const app = createApp({
             setup() {
                 // Connection State
                 const serverUrl = ref('https://hypha.aicell.io');
@@ -1954,7 +1954,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     statusTextColor, formatDuration, handleFolderSelect, getLatestLoss, formatDate,
                 };
             }
-        }).mount('#app');
+        });
+        app.config.compilerOptions.delimiters = ['[[', ']]'];
+        app.mount('#app');
     </script>
 </body>
 </html>
\ No newline at end of file

From 016a1ac25dd92606f1fb1b88dc5f5a93c92485a4 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Thu, 19 Feb 2026 22:56:15 +0100
Subject: [PATCH 03/13] Fix inference render and add workspace-aware artifact
 autocomplete error state

---
 bioengine_apps/cellpose_finetuning/index.html | 621 ++++++++++++++++--
 tests/cellpose/test_ui_e2e.py                 | 255 ++++++-
 2 files changed, 801 insertions(+), 75 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 648cb345..48c58658 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -295,21 +295,84 @@ <h2 class="text-2xl font-bold text-gray-900">New Training Session</h2>
                          <div class="bg-white p-6 rounded-xl shadow-sm border border-gray-200">
                             <div class="flex justify-between items-center mb-4">
                                 <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class="fas fa-database text-gray-400"></i> Dataset & Model</h3>
-                                <button type="button" @click="showUploadModal = true" class="text-sm bg-green-50 text-green-700 px-3 py-1.5 rounded-lg border border-green-200 hover:bg-green-100 transition flex items-center gap-1 font-medium">
-                                    <i class="fas fa-cloud-upload-alt"></i> Upload New
-                                </button>
                             </div>
                             
                             <div class="grid grid-cols-1 md:grid-cols-2 gap-6">
-                                <div>
+                                <div class="md:col-span-2">
+                                    <label class="block text-sm font-medium text-gray-700 mb-1">Dataset Source</label>
+                                    <div class="relative">
+                                        <i class="fas fa-database absolute left-3 top-1/2 -translate-y-1/2 text-gray-400 text-xs pointer-events-none"></i>
+                                        <select v-model="trainingParams.dataset_source" class="w-full appearance-none bg-white border border-gray-300 rounded-lg pl-9 pr-9 py-2.5 text-sm font-medium text-gray-700 shadow-sm hover:border-gray-400 focus:ring-2 focus:ring-blue-500 focus:border-blue-500 outline-none transition">
+                                            <option value="Artifact">Artifact</option>
+                                            <option value="Upload">Upload</option>
+                                            <option value="BioImage Archive">BioImage Archive</option>
+                                        </select>
+                                        <i class="fas fa-chevron-down absolute right-3 top-1/2 -translate-y-1/2 text-gray-400 text-xs pointer-events-none"></i>
+                                    </div>
+                                </div>
+                                
+                                <div v-if="trainingParams.dataset_source === 'Artifact'">
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Artifact ID <span class="text-red-500">*</span></label>
+                                    <div class="space-y-1">
+                                        <div class="flex gap-2 relative">
+                                            <input 
+                                                v-model="trainingParams.artifact" 
+                                                @input="handleArtifactInput"
+                                                @blur="closeSuggestions"
+                                                @keydown.tab.prevent="acceptSuggestion"
+                                                required 
+                                                :class="[
+                                                    'flex-grow rounded-lg px-3 py-2 text-sm outline-none transition border focus:ring-2',
+                                                    artifactInputError
+                                                        ? 'bg-red-50 border-red-500 text-red-700 placeholder-red-300 focus:border-red-500 focus:ring-red-300'
+                                                        : 'bg-gray-50 border-gray-300 focus:border-blue-500 focus:ring-blue-500'
+                                                ]"
+                                                placeholder="workspace/dataset-alias"
+                                            >
+                                            <button type="button" @click="openFileBrowser" class="px-3 py-2 bg-slate-100 border border-slate-300 rounded-lg hover:bg-slate-200 text-slate-600 transition" title="Browse Files">
+                                                <i class="fas fa-folder-open"></i>
+                                            </button>
+                                            
+                                            <!-- Autocomplete Suggestions -->
+                                            <div v-if="showSuggestions && suggestions.length > 0" class="absolute top-full left-0 w-full bg-white border border-gray-200 rounded-lg shadow-lg z-50 max-h-48 overflow-y-auto mt-1">
+                                                <div 
+                                                    v-for="(suggestion, index) in suggestions" 
+                                                    :key="index"
+                                                    @mousedown.prevent
+                                                    @click="selectSuggestion(suggestion)"
+                                                    class="px-3 py-2 text-sm hover:bg-blue-50 cursor-pointer flex justify-between items-center"
+                                                    :class="{'bg-blue-50': index === 0}"
+                                                >
+                                                    <span>[[ suggestion ]]</span>
+                                                    <span v-if="index === 0" class="text-xs text-gray-400">Tab</span>
+                                                </div>
+                                            </div>
+                                        </div>
+                                        <p v-if="artifactInputError" class="text-xs text-red-600">[[ artifactInputError ]]</p>
+                                    </div>
+                                </div>
+
+                                <div v-if="trainingParams.dataset_source === 'Upload'">
+                                    <label class="block text-sm font-medium text-gray-700 mb-1">Upload Data</label>
                                     <div class="flex gap-2">
-                                        <input v-model="trainingParams.artifact" required class="flex-grow bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm focus:ring-2 focus:ring-blue-500 focus:border-blue-500 outline-none transition" placeholder="workspace/dataset-alias">
-                                        <button type="button" @click="openFileBrowser" class="px-3 py-2 bg-slate-100 border border-slate-300 rounded-lg hover:bg-slate-200 text-slate-600 transition" title="Browse Files">
-                                            <i class="fas fa-folder-open"></i>
+                                        <div v-if="trainingParams.artifact" class="flex-grow bg-green-50 border border-green-200 rounded-lg px-3 py-2 text-sm text-green-800 flex items-center">
+                                            <i class="fas fa-check-circle mr-2"></i> [[ trainingParams.artifact ]]
+                                        </div>
+                                        <div v-else class="flex-grow bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm text-gray-400 italic">
+                                            No dataset uploaded yet
+                                        </div>
+                                        <button type="button" @click="showUploadModal = true" class="px-4 py-2 bg-green-600 text-white rounded-lg hover:bg-green-700 transition shadow-sm font-medium whitespace-nowrap">
+                                            <i class="fas fa-cloud-upload-alt mr-1"></i> Upload Folder
                                         </button>
                                     </div>
                                 </div>
+
+                                <div v-if="trainingParams.dataset_source === 'BioImage Archive'">
+                                    <label class="block text-sm font-medium text-gray-700 mb-1">Archive URL <span class="text-red-500">*</span></label>
+                                    <input v-model="trainingParams.bioimage_archive_url" required class="w-full bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm focus:ring-2 focus:ring-blue-500 focus:border-blue-500 outline-none transition" placeholder="https://www.ebi.ac.uk/biostudies/bioimages/studies/...">
+                                </div>
+                                
+                                <div class="md:col-span-2 space-y-4">
                                 <div>
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Base Model <span class="text-red-500">*</span></label>
                                     <div class="relative">
@@ -351,6 +414,7 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                 </div>
                             </div>
                          </div>
+                          </div>
 
                          <!-- Card: Hyperparameters -->
                          <div class="bg-white p-6 rounded-xl shadow-sm border border-gray-200">
@@ -502,11 +566,11 @@ <h3 class="font-bold text-green-900 mb-2">Ready for Export</h3>
                     </div>
                 </div>
 
-                <!-- Inference Tab -->
-                <div v-if="service && currentTab === 'inference'" class="max-w-7xl mx-auto lg:h-[calc(100vh-8rem)] h-auto flex flex-col">
+                 <!-- Inference Tab -->
+                 <div v-if="currentTab === 'inference'" class="max-w-7xl mx-auto lg:h-[calc(100vh-8rem)] h-auto flex flex-col">
                      <h2 class="text-2xl font-bold text-gray-900 mb-6 flex-none">Live Inference</h2>
-                     
-                     <div class="flex-grow grid grid-cols-1 lg:grid-cols-12 gap-6 lg:min-h-0 h-auto">
+
+                     <div v-if="service" class="flex-grow grid grid-cols-1 lg:grid-cols-12 gap-6 lg:min-h-0 h-auto">
                          <!-- Left: Controls -->
                          <div class="lg:col-span-4 bg-white rounded-xl shadow-sm border border-gray-200 flex flex-col overflow-hidden">
                             <div class="p-4 border-b border-gray-100 flex-none bg-gray-50/50">
@@ -584,6 +648,19 @@ <h3 class="font-bold text-gray-700">Configuration</h3>
                              </transition>
                          </div>
                      </div>
+                     <div v-else class="bg-white rounded-xl border border-gray-200 shadow-sm p-8 text-center">
+                        <div class="inline-flex items-center justify-center w-14 h-14 rounded-full bg-purple-50 text-purple-600 mb-4">
+                            <i class="fas fa-circle-notch fa-spin text-xl"></i>
+                        </div>
+                        <h3 class="text-lg font-semibold text-gray-900 mb-2">Initializing inference service</h3>
+                        <p class="text-sm text-gray-600 mb-5">
+                            The worker connection is active, but the inference backend is not ready yet.
+                        </p>
+                        <button @click="connect" class="inline-flex items-center gap-2 px-4 py-2 rounded-lg bg-purple-600 text-white text-sm font-medium hover:bg-purple-700 transition">
+                            <i class="fas fa-rotate-right"></i>
+                            Retry Connection
+                        </button>
+                    </div>
                 </div>
 
             </div>
@@ -694,7 +771,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
     </div>
 
     <script>
-        const { createApp, ref, reactive, computed, onMounted, watch } = Vue;
+        const { createApp, ref, reactive, computed, onMounted, watch, nextTick } = Vue;
 
         const app = createApp({
             setup() {
@@ -730,7 +807,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 
                 // Form Data
                 const trainingParams = reactive({
+                    dataset_source: 'Artifact',
                     artifact: '',
+                    bioimage_archive_url: '',
                     model: 'cpsam',
                     train_images: 'images/',
                     train_annotations: 'annotations/',
@@ -764,6 +843,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const canvasStyle = reactive({ top: '0', left: '0', width: '100%', height: '100%' });
                 const inferenceObjectCount = ref(0);
 
+                // Export
+                const exportModelName = ref('');
+                const isExporting = ref(false);
+                const exportResult = ref(null);
+
                 // File Browser
                 const showFileBrowser = ref(false);
                 const browserFiles = ref([]);
@@ -771,10 +855,334 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const browserLoading = ref(false);
                 const browserError = ref('');
 
-                // Export Data
-                const exportModelName = ref('');
-                const isExporting = ref(false);
-                const exportResult = ref(null);
+
+                // Autocomplete Logic
+                const suggestions = ref([]);
+                const showSuggestions = ref(false);
+                const artifactInputError = ref('');
+
+                const workspaceSuggestionCache = ref([]);
+                const artifactInputVersion = ref(0);
+                const workspaceClients = reactive({});
+                const artifactDebug = reactive({
+                    cacheSize: 0,
+                    lastShape: '',
+                    lastWorkspace: '',
+                    lastQuery: '',
+                    globalMatches: 0,
+                    scopedMatches: 0,
+                    sessionMatches: 0,
+                    suggestions: 0,
+                    lastError: '',
+                });
+
+                const normalizeWorkspaceId = (entry) => {
+                    if (!entry) return null;
+                    if (typeof entry === 'string') return entry;
+                    return entry.id || entry.workspace || entry.name || null;
+                };
+
+                const isWorkspacePermissionError = (error) => {
+                    const message = String(error?.message || error || '').toLowerCase();
+                    return (
+                        message.includes('permission denied') ||
+                        message.includes('access denied') ||
+                        message.includes('not authorized') ||
+                        message.includes('unauthorized')
+                    );
+                };
+
+                const fetchWorkspaces = async () => {
+                    const workspaces = new Set();
+                    if (workspace.value) workspaces.add(workspace.value);
+                    const userWs = getUserWorkspace();
+                    if (userWs) workspaces.add(userWs);
+                    workspaces.add('bioimage-io');
+
+                    if (!window.hyphaClient) return Array.from(workspaces);
+
+                    if (workspaceSuggestionCache.value.length > 0) {
+                        workspaceSuggestionCache.value.forEach((ws) => workspaces.add(ws));
+                        return Array.from(workspaces);
+                    }
+
+                    try {
+                        const listed = await window.hyphaClient.listWorkspaces();
+                        const normalized = (Array.isArray(listed) ? listed : [])
+                            .map(normalizeWorkspaceId)
+                            .filter(Boolean);
+                        workspaceSuggestionCache.value = normalized;
+                        normalized.forEach((ws) => workspaces.add(ws));
+                    } catch (e) {
+                        console.warn('Failed to fetch workspaces via server.listWorkspaces():', e);
+                    }
+
+                    return Array.from(workspaces);
+                };
+
+                const getWorkspaceClient = async (targetWorkspace) => {
+                    if (!targetWorkspace || targetWorkspace === workspace.value) return window.hyphaClient;
+                    if (workspaceClients[targetWorkspace]) return workspaceClients[targetWorkspace];
+
+                    const config = {
+                        server_url: serverUrl.value,
+                        workspace: targetWorkspace,
+                        method_timeout: 120000,
+                    };
+                    if (authToken.value) config.token = authToken.value;
+
+                    const client = await hyphaWebsocketClient.connectToServer(config);
+                    workspaceClients[targetWorkspace] = client;
+                    return client;
+                };
+
+                const normalizeArtifactListResponse = (listed) => {
+                    if (Array.isArray(listed)) {
+                        artifactDebug.lastShape = 'array';
+                        return listed;
+                    }
+                    if (listed?.items && Array.isArray(listed.items)) {
+                        artifactDebug.lastShape = 'items-array';
+                        return listed.items;
+                    }
+                    if (listed?.artifacts && Array.isArray(listed.artifacts)) {
+                        artifactDebug.lastShape = 'artifacts-array';
+                        return listed.artifacts;
+                    }
+                    if (listed && typeof listed === 'object') {
+                        artifactDebug.lastShape = 'object-map';
+                        if (listed.items && typeof listed.items === 'object') return Object.values(listed.items);
+                        if (listed.artifacts && typeof listed.artifacts === 'object') return Object.values(listed.artifacts);
+                        return Object.values(listed);
+                    }
+                    artifactDebug.lastShape = typeof listed;
+                    return [];
+                };
+
+                const queryWorkspaceArtifacts = async (workspaceId, queryText) => {
+                    const query = String(queryText || '').trim();
+                    const keywords = query ? [query] : [];
+                    let am = null;
+                    try {
+                        const client = await getWorkspaceClient(workspaceId);
+                        am = await client.getService('public/artifact-manager');
+                    } catch (error) {
+                        artifactDebug.lastError = String(error?.message || error);
+                        if (isWorkspacePermissionError(error)) {
+                            artifactInputError.value = `Permission denied to workspace '${workspaceId}'.`;
+                        }
+                        return [];
+                    }
+
+                    const attempts = [
+                        () => am.list(workspaceId, { keywords, limit: 30 }),
+                        () => am.list(workspaceId, { keywords }),
+                        () => am.list({ workspace: workspaceId, keywords, limit: 30 }),
+                        () => am.list(workspaceId),
+                    ];
+
+                    let lastAttemptError = null;
+
+                    for (const attempt of attempts) {
+                        try {
+                            const listed = await attempt();
+                            const rawItems = normalizeArtifactListResponse(listed);
+                            const normalized = rawItems
+                                .map((item) => {
+                                    const entry = normalizeArtifactEntry(item);
+                                    if (!entry) return null;
+                                    if (entry.workspace) return entry;
+                                    return {
+                                        workspace: workspaceId,
+                                        alias: entry.alias,
+                                        fullId: `${workspaceId}/${entry.alias}`,
+                                    };
+                                })
+                                .filter(Boolean);
+
+                            artifactDebug.cacheSize = normalized.length;
+                            artifactDebug.lastError = '';
+                            return normalized;
+                        } catch (error) {
+                            lastAttemptError = error;
+                            artifactDebug.lastError = String(error?.message || error);
+                            if (isWorkspacePermissionError(error)) {
+                                artifactInputError.value = `Permission denied to workspace '${workspaceId}'.`;
+                            }
+                        }
+                    }
+
+                    if (lastAttemptError && !artifactInputError.value) {
+                        const message = String(lastAttemptError?.message || lastAttemptError || '').trim();
+                        artifactInputError.value = `Unable to access workspace '${workspaceId}'. ${message}`.trim();
+                    }
+
+                    return [];
+                };
+
+                const normalizeArtifactEntry = (item) => {
+                    if (!item) return null;
+                    if (typeof item === 'string') {
+                        if (!item.includes('/')) {
+                            return {
+                                workspace: null,
+                                alias: item,
+                                fullId: item,
+                            };
+                        }
+                        const firstSlash = item.indexOf('/');
+                        return {
+                            workspace: item.slice(0, firstSlash),
+                            alias: item.slice(firstSlash + 1),
+                            fullId: item,
+                        };
+                    }
+
+                    const workspaceId = item.workspace || item?.manifest?.workspace || item.parent_id || null;
+                    const alias = item.alias || item.name || item?.manifest?.alias || null;
+                    const itemId = item.id || item.path || item.key || null;
+
+                    if (workspaceId && alias) {
+                        const cleanWorkspace = String(workspaceId).split('/')[0];
+                        return {
+                            workspace: cleanWorkspace,
+                            alias,
+                            fullId: `${cleanWorkspace}/${alias}`,
+                        };
+                    }
+
+                    if (typeof itemId === 'string' && itemId.includes('/')) {
+                        const firstSlash = itemId.indexOf('/');
+                        return {
+                            workspace: itemId.slice(0, firstSlash),
+                            alias: itemId.slice(firstSlash + 1),
+                            fullId: itemId,
+                        };
+                    }
+
+                    if (alias) {
+                        return {
+                            workspace: null,
+                            alias,
+                            fullId: alias,
+                        };
+                    }
+
+                    return null;
+                };
+
+                const collectSessionArtifactIds = () => {
+                    const artifactIds = new Set();
+                    for (const session of Object.values(sessions.value || {})) {
+                        if (!session || typeof session !== 'object') continue;
+
+                        const candidates = [
+                            session.artifact,
+                            session.dataset,
+                            session.dataset_id,
+                            session?.params?.artifact,
+                            session?.training_params?.artifact,
+                            session?.config?.artifact,
+                        ];
+
+                        for (const candidate of candidates) {
+                            if (typeof candidate === 'string' && candidate.includes('/')) {
+                                artifactIds.add(candidate);
+                            }
+                        }
+                    }
+                    return Array.from(artifactIds);
+                };
+
+                const handleArtifactInput = async () => {
+                     const requestVersion = ++artifactInputVersion.value;
+                     const val = String(trainingParams.artifact || '').trim();
+                     artifactDebug.lastQuery = val;
+                     artifactInputError.value = '';
+
+                     if (!val) {
+                         suggestions.value = [];
+                         showSuggestions.value = false;
+                         artifactDebug.suggestions = 0;
+                         return;
+                     }
+
+                     if (!val.includes('/')) {
+                         const workspaces = await fetchWorkspaces();
+                         if (requestVersion !== artifactInputVersion.value) return;
+                         suggestions.value = workspaces
+                            .filter((w) => w.toLowerCase().startsWith(val.toLowerCase()))
+                            .slice(0, 20);
+                         showSuggestions.value = suggestions.value.length > 0;
+                                 artifactDebug.suggestions = suggestions.value.length;
+                                 console.info('[ArtifactAutocomplete] workspace suggestions', {
+                                     query: val,
+                                     count: suggestions.value.length,
+                                 });
+                         return;
+                     }
+
+                     const parts = val.split('/');
+                     if (parts.length === 2) {
+                         const ws = parts[0];
+                         const query = parts[1].toLowerCase();
+                         artifactDebug.lastWorkspace = ws;
+                         const remoteMatches = await queryWorkspaceArtifacts(ws, query);
+                         if (requestVersion !== artifactInputVersion.value) return;
+
+                         const scopedMatches = remoteMatches
+                            .filter((entry) => entry.alias.toLowerCase().startsWith(query))
+                            .map((entry) => entry.fullId);
+
+                        const sessionMatches = collectSessionArtifactIds()
+                            .filter((artifactId) => artifactId.startsWith(`${ws}/`))
+                            .filter((artifactId) => artifactId.slice(ws.length + 1).toLowerCase().startsWith(query));
+
+                        artifactDebug.globalMatches = 0;
+                        artifactDebug.scopedMatches = scopedMatches.length;
+                        artifactDebug.sessionMatches = sessionMatches.length;
+                        suggestions.value = Array.from(new Set([...scopedMatches, ...sessionMatches])).slice(0, 30);
+                        showSuggestions.value = suggestions.value.length > 0;
+                        artifactDebug.suggestions = suggestions.value.length;
+                        console.info('[ArtifactAutocomplete] artifact suggestions', {
+                            workspace: ws,
+                            query,
+                            globalMatches: 0,
+                            scopedMatches: scopedMatches.length,
+                            sessionMatches: sessionMatches.length,
+                            finalSuggestions: suggestions.value.length,
+                            sample: suggestions.value.slice(0, 5),
+                        });
+                     } else {
+                        suggestions.value = [];
+                        showSuggestions.value = false;
+                        artifactDebug.suggestions = 0;
+                     }
+                };
+
+                const selectSuggestion = (s) => {
+                     // If it's a workspace (no slash), add slash
+                     if (!s.includes('/')) {
+                         trainingParams.artifact = s + '/';
+                         // Refocus input to keep typing?
+                         // Ensure handleArtifactInput runs for the new value
+                         handleArtifactInput();
+                     } else {
+                         trainingParams.artifact = s;
+                         showSuggestions.value = false;
+                     }
+                };
+                
+                const acceptSuggestion = () => {
+                     if (showSuggestions.value && suggestions.value.length > 0) {
+                         selectSuggestion(suggestions.value[0]);
+                     }
+                };
+                
+                const closeSuggestions = () => {
+                    // Timeout to allow click event to propagate
+                     setTimeout(() => showSuggestions.value = false, 200);
+                };
 
                 // Computed Properties
                 const currentSession = computed(() => selectedSessionId.value ? sessions.value[selectedSessionId.value] : null);
@@ -926,7 +1334,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     isConnecting.value = true;
                     autoConnectError.value = '';
                     try {
-                        const config = { "server_url": serverUrl.value, method_timeout: 20000 };
+                        const config = { "server_url": serverUrl.value, method_timeout: 120000 };
                         if (authToken.value) config.token = authToken.value;
                         
                         const client = await hyphaWebsocketClient.connectToServer(config);
@@ -982,6 +1390,16 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
 
                 const disconnect = () => {
                     if (window.hyphaClient) window.hyphaClient.disconnect();
+                    for (const [workspaceId, client] of Object.entries(workspaceClients)) {
+                        if (client && client !== window.hyphaClient) {
+                            try {
+                                client.disconnect();
+                            } catch (e) {
+                                console.warn(`Failed to disconnect workspace client ${workspaceId}:`, e);
+                            }
+                        }
+                        delete workspaceClients[workspaceId];
+                    }
                     isConnected.value = false;
                     service.value = null;
                     sessions.value = {};
@@ -992,9 +1410,73 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     isScanning.value = true;
                     try {
                         const client = window.hyphaClient;
-                        let services = [];
+                        const attemptList = [
+                            () => client.listServices({ workspace: workspace.value || client?.config?.workspace }),
+                            () => client.listServices(),
+                            () => client.list_services({ workspace: workspace.value || client?.config?.workspace }),
+                            () => client.list_services(),
+                        ];
+
+                        let listed = [];
+                        for (const attempt of attemptList) {
+                            try {
+                                const result = await attempt();
+                                if (Array.isArray(result)) {
+                                    listed = result;
+                                } else if (result && typeof result === 'object') {
+                                    listed = Object.values(result);
+                                }
+                                if (listed.length > 0) break;
+                            } catch (err) {
+                                // Try next strategy
+                            }
+                        }
+
+                        const serviceIds = listed
+                            .map((entry) => {
+                                if (!entry) return null;
+                                if (typeof entry === 'string') return entry;
+                                return entry.id || entry.service_id || null;
+                            })
+                            .filter(Boolean);
+
+                        const cellposeIds = serviceIds.filter((id) => id.includes('cellpose-finetuning'));
+                        availableServices.value = cellposeIds;
+
+                        const preferredIds = [
+                            `${workspace.value}/cellpose-finetuning`,
+                            `${workspace.value}/cellpose-finetuning-test`,
+                            `${workspace.value}/cellpose-finetuning-live`,
+                            'ri-scale/cellpose-finetuning',
+                            'ri-scale/cellpose-finetuning-test',
+                            'bioimage-io/cellpose-finetuning',
+                        ].filter(Boolean);
+
+                        if (!availableServices.value.includes(selectedServiceId.value)) {
+                            selectedServiceId.value = '';
+                        }
+
+                        if (!selectedServiceId.value && availableServices.value.length > 0) {
+                            const preferred = preferredIds.find((id) => availableServices.value.includes(id));
+                            selectedServiceId.value = preferred || availableServices.value[0];
+                        }
 
-                        selectedServiceId.value = "bioimage-io/cellpose-finetuning";
+                        if (availableServices.value.length === 0) {
+                            for (const fallbackId of preferredIds) {
+                                try {
+                                    await client.getService(fallbackId);
+                                    availableServices.value = [fallbackId];
+                                    selectedServiceId.value = fallbackId;
+                                    break;
+                                } catch (_) {
+                                    // try next fallback id
+                                }
+                            }
+                        }
+
+                        if (availableServices.value.length === 0) {
+                            service.value = null;
+                        }
                     } catch (e) {
                         console.warn("Service scan failed", e);
                     } finally {
@@ -1025,38 +1507,40 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                   // --- Dashboard Methods ---
                 const refreshSessions = async (background = false) => {
                     if (!service.value) return;
+                    if (refreshSessions._inFlight) {
+                        if (background) return;
+                        await refreshSessions._inFlight;
+                        return;
+                    }
                     if (!background) refreshLoading.value = true;
                     
-                    try {
-                        // Use a timeout for the call to prevent hanging
-                        let result = null;
+                    refreshSessions._inFlight = (async () => {
                         try {
-                             result = await Promise.race([
-                                service.value.list_training_sessions(),
-                                new Promise((_, reject) => setTimeout(() => reject(new Error("Timeout")), 20000))
-                            ]);
-                        } catch (err) {
-                            console.warn("Training sessions list timed out or failed:", err);
-                            throw err;
-                        }
-                        
-                        if (result) {
-                            sessions.value = result;
-                            syncError.value = null;
-
-                            // Find active session
-                            const running = Object.entries(sessions.value).find(([_, s]) => s && s.status_type === 'running');
-                            if (running) {
-                                monitoringSessionId.value = running[0];
-                            } else {
-                                monitoringSessionId.value = null;
+                            const result = await service.value.list_training_sessions();
+                            
+                            if (result) {
+                                sessions.value = result;
+                                syncError.value = null;
+
+                                // Find active session
+                                const running = Object.entries(sessions.value).find(([_, s]) => s && s.status_type === 'running');
+                                if (running) {
+                                    monitoringSessionId.value = running[0];
+                                } else {
+                                    monitoringSessionId.value = null;
+                                }
                             }
+                        } catch (e) {
+                            console.error(e);
+                            // Show sync error but preserve old sessions so user can see last known state
+                            syncError.value = "Sync failed: " + (e.message || "Timeout") + ". Displayed data may be outdated.";
                         }
-                    } catch (e) {
-                        console.error(e);
-                        // Show sync error but preserve old sessions so user can see last known state
-                        syncError.value = "Sync failed: " + (e.message || "Timeout") + ". Displayed data may be outdated.";
+                    })();
+
+                    try {
+                        await refreshSessions._inFlight;
                     } finally {
+                        refreshSessions._inFlight = null;
                         if (!background) refreshLoading.value = false;
                     }
                 };
@@ -1078,6 +1562,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                                 }
                             } catch (e) {
                                 console.warn("Auto-update loop error:", e);
+                                // If token expired, try to nudge user?
+                                if (e.message && e.message.includes('expired')) {
+                                    syncError.value = "Session expired. Please refresh page.";
+                                    isConnected.value = false;
+                                }
                             }
                         }
                         updateTimer = setTimeout(runLoop, 4000);
@@ -1085,6 +1574,23 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     runLoop();
                 };
 
+                watch([currentTab, service, isConnected], async () => {
+                    if (currentTab.value !== 'inference') return;
+                    await nextTick();
+                    const heading = document.querySelector('h2');
+                    const headingText = heading ? heading.textContent : null;
+                    const hasConfigCard = !!document.querySelector('h3');
+                    const hasFileInput = !!document.querySelector('input[type="file"]');
+                    console.info('[InferenceTab] render diagnostics', {
+                        isConnected: isConnected.value,
+                        hasService: !!service.value,
+                        currentTab: currentTab.value,
+                        headingText,
+                        hasConfigCard,
+                        hasFileInput,
+                    });
+                });
+
                 const viewSession = (id) => {
                     if (!sessions.value[id]) return;
                     selectedSessionId.value = id;
@@ -1207,11 +1713,27 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         const hasTrainAnnotations = !!(trainingParams.train_annotations && trainingParams.train_annotations.trim());
 
                         if (!hasMetadataDir && (!hasTrainImages || !hasTrainAnnotations)) {
-                            throw new Error("Provide either metadata folder, or both train images and train annotations.");
+                           // If using BioImage Archive, we might not need explicit images/annotations if the archive has a standard structure OR we assume default?
+                           // But usually we still need them.
+                           // Actually, for BioImage Archive, the prompt says "allow the user to post a URL to bioimage archive".
+                           // Does the backend handle downloading?
+                           // If so, we pass the URL as 'artifact'? Or a new param?
+                           // The prompt implies "Dataset source" choice.
+                        }
+                        
+                        // Determine the artifact value based on source
+                        let artifactValue = trainingParams.artifact;
+                        if (trainingParams.dataset_source === 'BioImage Archive') {
+                            artifactValue = trainingParams.bioimage_archive_url;
+                            if (!artifactValue) throw new Error("Please enter a BioImage Archive URL.");
+                        } else if (trainingParams.dataset_source === 'Upload') {
+                             if (!artifactValue) throw new Error("Please upload a dataset first.");
+                        } else {
+                             if (!artifactValue) throw new Error("Please specify an Artifact ID.");
                         }
 
                         console.log("Invoking service.start_training with:", {
-                            artifact: trainingParams.artifact,
+                            artifact: artifactValue,
                             train_images: trainingParams.train_images,
                             train_annotations: trainingParams.train_annotations,
                             metadata_dir: trainingParams.metadata_dir || null,
@@ -1227,7 +1749,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         
                         const result = await service.value.start_training(
                             {
-                                artifact: trainingParams.artifact,
+                                artifact: artifactValue,
                                 train_images: trainingParams.train_images || null,
                                 train_annotations: trainingParams.train_annotations || null,
                                 metadata_dir: trainingParams.metadata_dir || null,
@@ -1950,6 +2472,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     showFileBrowser, openFileBrowser, browserFiles, browserPath, browserLoading, browserError,
                     browserVisibleItems, browserNavigate, browserNavigateUp, selectFolderFor, browserFeedback,
 
+                    // Autocomplete
+                    suggestions, showSuggestions, artifactInputError, handleArtifactInput, selectSuggestion, acceptSuggestion, closeSuggestions,
+
                     // Helpers
                     statusTextColor, formatDuration, handleFolderSelect, getLatestLoss, formatDate,
                 };
diff --git a/tests/cellpose/test_ui_e2e.py b/tests/cellpose/test_ui_e2e.py
index ef1a576a..52bdcc86 100644
--- a/tests/cellpose/test_ui_e2e.py
+++ b/tests/cellpose/test_ui_e2e.py
@@ -8,18 +8,22 @@
 
 SERVER_URL = "https://hypha.aicell.io"
 APP_WORKSPACE = os.environ.get("HYPHA_TEST_WORKSPACE", "ri-scale")
-APP_URL = f"{SERVER_URL}/{APP_WORKSPACE}/view/cellpose-finetuning"
+APP_ID = os.environ.get("HYPHA_TEST_APP_ID", "cellpose-finetuning")
+APP_URL = f"{SERVER_URL}/{APP_WORKSPACE}/view/{APP_ID}"
+APP_FALLBACK_URL = f"{SERVER_URL}/{APP_WORKSPACE}/view/cellpose-finetuning-test"
 RUN_LIVE_REGRESSION_TESTS = os.environ.get("RUN_LIVE_REGRESSION_TESTS") == "1"
 
 
 def open_and_connect(page: Page) -> None:
     token = os.environ.get("HYPHA_TOKEN")
     if token:
-        page.add_init_script(
-            f"window.localStorage.setItem('hypha_token', {token!r});"
-        )
+        page.add_init_script(f"window.localStorage.setItem('hypha_token', {token!r});")
     response = page.goto(APP_URL, timeout=60000)
-    assert response and response.ok, "Failed to load app"
+    if not (response and response.ok) and APP_URL != APP_FALLBACK_URL:
+        response = page.goto(APP_FALLBACK_URL, timeout=60000)
+    assert (
+        response and response.ok
+    ), f"Failed to load app at {APP_URL} (fallback attempted: {APP_FALLBACK_URL})"
     expect(page).to_have_title(re.compile(r"Cellpose Fine-Tuning"), timeout=15000)
     expect(page.locator("text=Connected").first).to_be_visible(timeout=30000)
 
@@ -34,57 +38,82 @@ def _handle(dialog):
     page.on("dialog", _handle)
     return captured
 
+
+def capture_console_logs(page: Page, marker: str):
+    captured = []
+
+    def _on_console(msg):
+        try:
+            text = msg.text
+        except Exception:
+            text = str(msg)
+        if marker in text:
+            captured.append(text)
+
+    page.on("console", _on_console)
+    return captured
+
+
 def test_app_loads(page: Page):
     """App loads and connects to Hypha."""
     open_and_connect(page)
 
+
 def test_navigation(page: Page):
     """Basic tab navigation works."""
     open_and_connect(page)
-    
+
     # Default is Dashboard
     expect(page.locator("h2:has-text('Dashboard')")).to_be_visible()
-    
+
     # Click New Training
     page.click("text=New Training")
     expect(page.locator("h2:has-text('New Training Session')")).to_be_visible()
-    
+
     # Click Inference
     page.click("text=Inference")
     expect(page.locator("h2:has-text('Live Inference')")).to_be_visible()
 
+
 def test_file_browser(page: Page):
     """File browser opens and lists artifact content."""
     open_and_connect(page)
-    
+
     # Go to Training
     page.click("text=New Training")
-    
+
     artifact_id = "ri-scale/cellpose-finetuning"
     page.fill("input[placeholder='workspace/dataset-alias']", artifact_id)
-    
+
     # Open Browser
     page.click("button[title='Browse Files']")
-    
+
     # Wait for modal
     expect(page.locator("text=Artifact Explorer")).to_be_visible()
-    
+
     # Wait for loading to finish
     expect(page.locator("text=Loading files...")).not_to_be_visible()
-    
+
     # Should see "index.html"
     expect(page.locator("text=index.html")).to_be_visible()
-    
+
     page.locator("div:has(h3:has-text('Artifact Explorer')) button").last.click()
-    
+
+
 def test_start_training_button_and_result(page: Page):
     """Clicks Start Training and validates the returned UI result."""
     open_and_connect(page)
 
     page.click("text=New Training")
     page.fill("input[placeholder='workspace/dataset-alias']", "ri-scale/zarr-demo")
-    page.fill("input[placeholder='e.g. images/*/*.tif']", "images/108bb69d-2e52-4382-8100-e96173db24ee/*.ome.tif")
-    page.fill("input[placeholder='e.g. annotations/*/*_mask.ome.tif']", "annotations/108bb69d-2e52-4382-8100-e96173db24ee/*_mask.ome.tif")
+    page.fill(
+        "input[placeholder='e.g. images/*/*.tif']",
+        "images/108bb69d-2e52-4382-8100-e96173db24ee/*.ome.tif",
+    )
+    page.fill(
+        "input[placeholder='e.g. annotations/*/*_mask.ome.tif']",
+        "annotations/108bb69d-2e52-4382-8100-e96173db24ee/*_mask.ome.tif",
+    )
     page.fill("input[placeholder='Optional']", "")
     page.locator("input[placeholder='Optional']").nth(1).fill("")
     page.fill("input[placeholder='Pretrained ID or Session ID']", "cpsam")
@@ -107,13 +136,17 @@ def test_infer_button_and_result(page: Page):
     open_and_connect(page)
     page.click("text=Inference")
 
-    image_path = Path("/Users/hugokallander/github-repos/bioengine-worker/tests/cellpose_legacy_scripts/t0000.ome.tif")
+    image_path = Path(
+        "/Users/hugokallander/github-repos/bioengine-worker/tests/cellpose_legacy_scripts/t0000.ome.tif"
+    )
     assert image_path.exists(), f"Inference image not found: {image_path}"
 
     infer_input = page.locator("input[type='file']").first
     infer_input.set_input_files(str(image_path))
     expect(page.locator("text=t0000.ome.tif")).to_be_visible(timeout=10000)
-    expect(page.locator("img[src^='blob:'], img[src^='data:image/png;base64']").first).to_be_visible(timeout=10000)
+    expect(
+        page.locator("img[src^='blob:'], img[src^='data:image/png;base64']").first
+    ).to_be_visible(timeout=10000)
 
     dialogs = capture_dialog(page)
     page.click("button:has-text('Run Segmentation')")
@@ -129,7 +162,9 @@ def test_infer_button_and_result(page: Page):
 def test_no_raw_template_markers(page: Page):
     """The rendered UI should not leak template raw markers."""
     if not RUN_LIVE_REGRESSION_TESTS:
-        pytest.skip("Set RUN_LIVE_REGRESSION_TESTS=1 to run live template-marker regression")
+        pytest.skip(
+            "Set RUN_LIVE_REGRESSION_TESTS=1 to run live template-marker regression"
+        )
 
     open_and_connect(page)
     html = page.content()
@@ -140,14 +175,22 @@ def test_no_raw_template_markers(page: Page):
 def test_stop_training_persists_after_reload(page: Page):
     """Start then stop a session, and ensure status does not revert to running after refresh/reload."""
     if not RUN_LIVE_REGRESSION_TESTS:
-        pytest.skip("Set RUN_LIVE_REGRESSION_TESTS=1 to run live stop-persistence regression")
+        pytest.skip(
+            "Set RUN_LIVE_REGRESSION_TESTS=1 to run live stop-persistence regression"
+        )
 
     open_and_connect(page)
 
     page.click("text=New Training")
     page.fill("input[placeholder='workspace/dataset-alias']", "ri-scale/zarr-demo")
-    page.fill("input[placeholder='e.g. images/*/*.tif']", "images/108bb69d-2e52-4382-8100-e96173db24ee/*.ome.tif")
-    page.fill("input[placeholder='e.g. annotations/*/*_mask.ome.tif']", "annotations/108bb69d-2e52-4382-8100-e96173db24ee/*_mask.ome.tif")
+    page.fill(
+        "input[placeholder='e.g. images/*/*.tif']",
+        "images/108bb69d-2e52-4382-8100-e96173db24ee/*.ome.tif",
+    )
+    page.fill(
+        "input[placeholder='e.g. annotations/*/*_mask.ome.tif']",
+        "annotations/108bb69d-2e52-4382-8100-e96173db24ee/*_mask.ome.tif",
+    )
     page.fill("input[placeholder='Optional']", "")
     page.locator("input[placeholder='Optional']").nth(1).fill("")
     page.fill("input[placeholder='Pretrained ID or Session ID']", "cpsam")
@@ -161,7 +204,9 @@ def test_stop_training_persists_after_reload(page: Page):
     assert start_dialogs["message"] is not None, "No Start Training dialog received"
 
     match = re.search(r"Session ID:\s*([a-zA-Z0-9-]+)", start_dialogs["message"] or "")
-    assert match is not None, f"Could not parse session id from: {start_dialogs['message']}"
+    assert (
+        match is not None
+    ), f"Could not parse session id from: {start_dialogs['message']}"
     session_id = match.group(1)
 
     page.click("text=Dashboard")
@@ -190,6 +235,162 @@ def test_stop_training_persists_after_reload(page: Page):
     page.click("text=Dashboard")
     row_after_reload = page.locator("tr", has=page.locator(f"text={session_id[:8]}"))
     expect(row_after_reload.first).to_be_visible(timeout=20000)
-    status_after_reload = row_after_reload.first.locator("td").nth(1).inner_text().strip().lower()
-    assert "running" not in status_after_reload, f"Status reverted to running: {status_after_reload}"
+    status_after_reload = (
+        row_after_reload.first.locator("td").nth(1).inner_text().strip().lower()
+    )
+    assert (
+        "running" not in status_after_reload
+    ), f"Status reverted to running: {status_after_reload}"
+
+
+def test_dataset_source_options(page: Page):
+    """Test switching between Artifact, Upload, and BioImage Archive options."""
+    open_and_connect(page)
+    page.click("text=New Training")
+
+    # Locate the dataset source dropdown
+    # Try finding by label text and getting the select in the parent container
+    # Or simply finding the select that contains 'BioImage Archive' option
+    dropdown = page.locator("select:has(option:has-text('BioImage Archive'))")
+    expect(dropdown).to_be_visible()
+
+    # 1. Default should be Artifact
+    expect(dropdown).to_have_value("Artifact")
+    # Verify Artifact ID input is visible
+    expect(page.locator("input[placeholder='workspace/dataset-alias']")).to_be_visible()
+
+    # 2. Switch to BioImage Archive
+    dropdown.select_option("BioImage Archive")
+    # Verify Archive URL input is visible and Artifact ID input is hidden
+    archive_input = page.locator("input[placeholder*='ebi.ac.uk']")
+    expect(archive_input).to_be_visible()
+    expect(
+        page.locator("input[placeholder='workspace/dataset-alias']")
+    ).not_to_be_visible()
+
+    # 3. Switch to Upload
+    dropdown.select_option("Upload")
+    # Verify Upload button is visible and Archive URL input is hidden
+    upload_btn = page.locator("button:has-text('Upload Folder')")
+    expect(upload_btn).to_be_visible()
+    expect(archive_input).not_to_be_visible()
+
+    # Check Upload Modal opens
+    upload_btn.click()
+    expect(page.locator("h3:has-text('Upload Dataset')")).to_be_visible()
+    page.click("button:has-text('Cancel')")
+    expect(page.locator("h3:has-text('Upload Dataset')")).not_to_be_visible()
+
+
+def test_inference_panel_not_blank_live(page: Page):
+    """Inference tab should render configuration controls, not an empty panel."""
+    if not RUN_LIVE_REGRESSION_TESTS:
+        pytest.skip(
+            "Set RUN_LIVE_REGRESSION_TESTS=1 to run live inference-panel regression"
+        )
+
+    inference_logs = capture_console_logs(page, "[InferenceTab]")
+    open_and_connect(page)
+    page.click("text=Inference")
+
+    expect(page.locator("h2:has-text('Live Inference')")).to_be_visible(timeout=20000)
+
+    has_config = False
+    has_init_state = False
+    for _ in range(8):
+        page.wait_for_timeout(1500)
+        has_config = page.locator("h3:has-text('Configuration')").count() > 0
+        has_init_state = (
+            page.locator("h3:has-text('Initializing inference service')").count() > 0
+        )
+        if has_config or has_init_state:
+            break
+
+    if not (has_config or has_init_state):
+        body_preview = page.locator("body").inner_text()[:1200]
+        raise AssertionError(
+            "Inference tab rendered blank (neither configuration nor initializing state visible). "
+            f"Captured logs: {inference_logs[-12:]}. Body preview: {body_preview}"
+        )
+
+    if has_config:
+        expect(page.locator("button:has-text('Run Segmentation')")).to_be_visible(
+            timeout=15000
+        )
+    if has_init_state:
+        expect(page.locator("button:has-text('Retry Connection')")).to_be_visible(
+            timeout=15000
+        )
+
+    page.wait_for_timeout(1000)
+    assert (
+        inference_logs
+    ), "No [InferenceTab] diagnostic logs captured after opening Inference tab"
+
+
+def test_workspace_autocomplete_suggestions_live(page: Page):
+    """Workspace autocomplete should show options while typing workspace prefix."""
+    if not RUN_LIVE_REGRESSION_TESTS:
+        pytest.skip(
+            "Set RUN_LIVE_REGRESSION_TESTS=1 to run live workspace-autocomplete regression"
+        )
+
+    artifact_logs = capture_console_logs(page, "[ArtifactAutocomplete]")
+    open_and_connect(page)
+    page.click("text=New Training")
+
+    artifact_input = page.locator("input[placeholder='workspace/dataset-alias']")
+    expect(artifact_input).to_be_visible(timeout=15000)
+    artifact_input.fill(APP_WORKSPACE[:2])
+
+    suggestions = page.locator(
+        "input[placeholder='workspace/dataset-alias'] + button + div > div"
+    )
+    try:
+        expect(suggestions.first).to_be_visible(timeout=30000)
+    except AssertionError as exc:
+        raise AssertionError(
+            "Workspace suggestions were not visible. "
+            f"Captured diagnostic logs: {artifact_logs[-12:]}"
+        ) from exc
+
+    first_text = suggestions.first.inner_text().strip()
+    assert (
+        "/" not in first_text
+    ), f"Expected workspace-only suggestion, got: {first_text}"
+
+
+def test_artifact_autocomplete_suggestions_live(page: Page):
+    """Artifact autocomplete should show suggestions after entering workspace/ prefix."""
+    if not RUN_LIVE_REGRESSION_TESTS:
+        pytest.skip(
+            "Set RUN_LIVE_REGRESSION_TESTS=1 to run live artifact-autocomplete regression"
+        )
+
+    artifact_logs = capture_console_logs(page, "[ArtifactAutocomplete]")
+    open_and_connect(page)
+    page.click("text=New Training")
 
+    artifact_input = page.locator("input[placeholder='workspace/dataset-alias']")
+    expect(artifact_input).to_be_visible(timeout=15000)
+
+    artifact_input.fill(f"{APP_WORKSPACE}/c")
+
+    suggestions = page.locator(
+        "input[placeholder='workspace/dataset-alias'] + button + div > div"
+    )
+    try:
+        expect(suggestions.first).to_be_visible(timeout=30000)
+    except AssertionError as exc:
+        raise AssertionError(
+            "Autocomplete suggestions were not visible. "
+            f"Captured diagnostic logs: {artifact_logs[-12:]}"
+        ) from exc
+
+    first_text = suggestions.first.inner_text()
+    assert first_text.startswith(
+        f"{APP_WORKSPACE}/"
+    ), f"Unexpected suggestion: {first_text}"
+    assert (
+        artifact_logs
+    ), "No [ArtifactAutocomplete] diagnostic logs captured during autocomplete flow"

From ad23395da65b350b4824b928207d8c225bcd8163 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Thu, 19 Feb 2026 23:11:53 +0100
Subject: [PATCH 04/13] cleanup artifact autocomplete workspace client handling

---
 bioengine_apps/cellpose_finetuning/index.html | 69 +++++++++----------
 1 file changed, 32 insertions(+), 37 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 48c58658..c4281e8c 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -921,7 +921,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 };
 
                 const getWorkspaceClient = async (targetWorkspace) => {
-                    if (!targetWorkspace || targetWorkspace === workspace.value) return window.hyphaClient;
+                    if (!targetWorkspace) {
+                        throw new Error('Workspace is required for artifact lookup.');
+                    }
                     if (workspaceClients[targetWorkspace]) return workspaceClients[targetWorkspace];
 
                     const config = {
@@ -929,7 +931,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         workspace: targetWorkspace,
                         method_timeout: 120000,
                     };
-                    if (authToken.value) config.token = authToken.value;
+                    if (authToken.value) {
+                        config.token = authToken.value;
+                    } else {
+                        throw new Error('Missing auth token for cross-workspace artifact lookup.');
+                    }
 
                     const client = await hyphaWebsocketClient.connectToServer(config);
                     workspaceClients[targetWorkspace] = client;
@@ -973,42 +979,31 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         }
                         return [];
                     }
+                    
+                    try {
+                        const listed = await am.list({keywords, limit: 30, _rkwargs: true });
+                        const rawItems = normalizeArtifactListResponse(listed);
+                        const normalized = rawItems
+                            .map((item) => {
+                                const entry = normalizeArtifactEntry(item);
+                                if (!entry) return null;
+                                if (entry.workspace) return entry;
+                                return {
+                                    workspace: workspaceId,
+                                    alias: entry.alias,
+                                    fullId: `${workspaceId}/${entry.alias}`,
+                                };
+                            })
+                            .filter(Boolean);
 
-                    const attempts = [
-                        () => am.list(workspaceId, { keywords, limit: 30 }),
-                        () => am.list(workspaceId, { keywords }),
-                        () => am.list({ workspace: workspaceId, keywords, limit: 30 }),
-                        () => am.list(workspaceId),
-                    ];
-
-                    let lastAttemptError = null;
-
-                    for (const attempt of attempts) {
-                        try {
-                            const listed = await attempt();
-                            const rawItems = normalizeArtifactListResponse(listed);
-                            const normalized = rawItems
-                                .map((item) => {
-                                    const entry = normalizeArtifactEntry(item);
-                                    if (!entry) return null;
-                                    if (entry.workspace) return entry;
-                                    return {
-                                        workspace: workspaceId,
-                                        alias: entry.alias,
-                                        fullId: `${workspaceId}/${entry.alias}`,
-                                    };
-                                })
-                                .filter(Boolean);
-
-                            artifactDebug.cacheSize = normalized.length;
-                            artifactDebug.lastError = '';
-                            return normalized;
-                        } catch (error) {
-                            lastAttemptError = error;
-                            artifactDebug.lastError = String(error?.message || error);
-                            if (isWorkspacePermissionError(error)) {
-                                artifactInputError.value = `Permission denied to workspace '${workspaceId}'.`;
-                            }
+                        artifactDebug.cacheSize = normalized.length;
+                        artifactDebug.lastError = '';
+                        return normalized;
+                    } catch (error) {
+                        lastAttemptError = error;
+                        artifactDebug.lastError = String(error?.message || error);
+                        if (isWorkspacePermissionError(error)) {
+                            artifactInputError.value = `Permission denied to workspace '${workspaceId}'.`;
                         }
                     }
 

From accfd6815e5559adcfb93b686786472f2681e064 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Thu, 19 Feb 2026 23:54:01 +0100
Subject: [PATCH 05/13] fix artifact browser gating, nested pair matching, and
 add BIA URL dataset ingestion

---
 bioengine_apps/cellpose_finetuning/index.html | 193 ++++++++---
 bioengine_apps/cellpose_finetuning/main.py    | 301 +++++++++++++++++-
 .../tests/test_metadata_and_glob.py           |  78 +++++
 3 files changed, 523 insertions(+), 49 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index c4281e8c..6de4c7d6 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -316,8 +316,9 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                     <div class="space-y-1">
                                         <div class="flex gap-2 relative">
                                             <input 
-                                                v-model="trainingParams.artifact" 
+                                                v-model="trainingParams.artifact_input" 
                                                 @input="handleArtifactInput"
+                                                @focus="handleArtifactInput"
                                                 @blur="closeSuggestions"
                                                 @keydown.tab.prevent="acceptSuggestion"
                                                 required 
@@ -329,7 +330,12 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                                 ]"
                                                 placeholder="workspace/dataset-alias"
                                             >
-                                            <button type="button" @click="openFileBrowser" class="px-3 py-2 bg-slate-100 border border-slate-300 rounded-lg hover:bg-slate-200 text-slate-600 transition" title="Browse Files">
+                                            <button type="button" @click="openFileBrowser" :disabled="artifactLookupPending || !artifactIdExists" :class="[
+                                                'px-3 py-2 border rounded-lg transition',
+                                                (artifactLookupPending || !artifactIdExists)
+                                                    ? 'bg-slate-100 border-slate-200 text-slate-300 cursor-not-allowed'
+                                                    : 'bg-slate-100 border-slate-300 hover:bg-slate-200 text-slate-600'
+                                            ]" :title="artifactLookupPending ? 'Checking artifact...' : (artifactIdExists ? 'Browse Files' : 'Enter a valid existing artifact ID first')">
                                                 <i class="fas fa-folder-open"></i>
                                             </button>
                                             
@@ -338,8 +344,7 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                                 <div 
                                                     v-for="(suggestion, index) in suggestions" 
                                                     :key="index"
-                                                    @mousedown.prevent
-                                                    @click="selectSuggestion(suggestion)"
+                                                    @mousedown.prevent="selectSuggestion(suggestion)"
                                                     class="px-3 py-2 text-sm hover:bg-blue-50 cursor-pointer flex justify-between items-center"
                                                     :class="{'bg-blue-50': index === 0}"
                                                 >
@@ -355,8 +360,8 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                 <div v-if="trainingParams.dataset_source === 'Upload'">
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Upload Data</label>
                                     <div class="flex gap-2">
-                                        <div v-if="trainingParams.artifact" class="flex-grow bg-green-50 border border-green-200 rounded-lg px-3 py-2 text-sm text-green-800 flex items-center">
-                                            <i class="fas fa-check-circle mr-2"></i> [[ trainingParams.artifact ]]
+                                        <div v-if="trainingParams.uploaded_artifact" class="flex-grow bg-green-50 border border-green-200 rounded-lg px-3 py-2 text-sm text-green-800 flex items-center">
+                                            <i class="fas fa-check-circle mr-2"></i> [[ trainingParams.uploaded_artifact ]]
                                         </div>
                                         <div v-else class="flex-grow bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm text-gray-400 italic">
                                             No dataset uploaded yet
@@ -369,7 +374,12 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
 
                                 <div v-if="trainingParams.dataset_source === 'BioImage Archive'">
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Archive URL <span class="text-red-500">*</span></label>
-                                    <input v-model="trainingParams.bioimage_archive_url" required class="w-full bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm focus:ring-2 focus:ring-blue-500 focus:border-blue-500 outline-none transition" placeholder="https://www.ebi.ac.uk/biostudies/bioimages/studies/...">
+                                    <div class="flex gap-2">
+                                        <input v-model="trainingParams.bioimage_archive_url" required class="flex-grow bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm focus:ring-2 focus:ring-blue-500 focus:border-blue-500 outline-none transition" placeholder="https://www.ebi.ac.uk/biostudies/bioimages/studies/...">
+                                        <a href="https://www.ebi.ac.uk/biostudies/bioimages" target="_blank" rel="noopener noreferrer" class="px-4 py-2 bg-slate-100 border border-slate-300 rounded-lg hover:bg-slate-200 text-slate-700 transition text-sm whitespace-nowrap">
+                                            <i class="fas fa-external-link-alt mr-1"></i> Search EBI
+                                        </a>
+                                    </div>
                                 </div>
                                 
                                 <div class="md:col-span-2 space-y-4">
@@ -390,7 +400,7 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                         <input v-model="trainingParams.train_images" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. images/*/*.tif">
                                         <div class="mt-1 text-xs text-gray-500 flex flex-wrap items-center gap-2">
                                             <span>Examples:</span>
-                                            <button type="button" @click="applyTrainingGlobExample" class="text-blue-600 hover:underline">images/*/*.tif</button>
+                                            <button type="button" @click="applyTrainingGlobExample" class="text-blue-600 hover:underline">images/**/*.tif</button>
                                             <button type="button" @click="applyTrainingFolderExample" class="text-blue-600 hover:underline">images/ + annotations/</button>
                                         </div>
                                     </div>
@@ -673,7 +683,7 @@ <h3 class="text-lg font-semibold text-gray-900 mb-2">Initializing inference serv
                 <div class="flex justify-between items-center mb-4 border-b border-gray-100 pb-4 flex-none">
                     <div>
                         <h3 class="text-lg font-bold text-gray-900">Artifact Explorer</h3>
-                        <p class="text-xs text-gray-500">[[ trainingParams.artifact || 'No artifact selected' ]]</p>
+                        <p class="text-xs text-gray-500">[[ getSelectedArtifactId() || 'No artifact selected' ]]</p>
                     </div>
                     <button @click="showFileBrowser = false" class="text-gray-400 hover:text-gray-600"><i class="fas fa-times"></i></button>
                 </div>
@@ -808,7 +818,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 // Form Data
                 const trainingParams = reactive({
                     dataset_source: 'Artifact',
-                    artifact: '',
+                    artifact_input: '',
+                    uploaded_artifact: '',
                     bioimage_archive_url: '',
                     model: 'cpsam',
                     train_images: 'images/',
@@ -860,6 +871,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const suggestions = ref([]);
                 const showSuggestions = ref(false);
                 const artifactInputError = ref('');
+                const artifactIdExists = ref(false);
+                const artifactLookupPending = ref(false);
+                const artifactLookupVersion = ref(0);
 
                 const workspaceSuggestionCache = ref([]);
                 const artifactInputVersion = ref(0);
@@ -924,6 +938,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     if (!targetWorkspace) {
                         throw new Error('Workspace is required for artifact lookup.');
                     }
+
+                    const activeWorkspace = workspace.value || window.hyphaClient?.config?.workspace || '';
+                    if (activeWorkspace && targetWorkspace === activeWorkspace && window.hyphaClient) {
+                        return window.hyphaClient;
+                    }
+
                     if (workspaceClients[targetWorkspace]) return workspaceClients[targetWorkspace];
 
                     const config = {
@@ -931,8 +951,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         workspace: targetWorkspace,
                         method_timeout: 120000,
                     };
-                    if (authToken.value) {
-                        config.token = authToken.value;
+                    const tokenForLookup = authToken.value || window.hyphaClient?.config?.token || null;
+                    if (tokenForLookup) {
+                        config.token = tokenForLookup;
                     } else {
                         throw new Error('Missing auth token for cross-workspace artifact lookup.');
                     }
@@ -969,6 +990,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     const query = String(queryText || '').trim();
                     const keywords = query ? [query] : [];
                     let am = null;
+                    let lastAttemptError = null;
                     try {
                         const client = await getWorkspaceClient(workspaceId);
                         am = await client.getService('public/artifact-manager');
@@ -980,30 +1002,39 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         return [];
                     }
                     
-                    try {
-                        const listed = await am.list({keywords, limit: 30, _rkwargs: true });
-                        const rawItems = normalizeArtifactListResponse(listed);
-                        const normalized = rawItems
-                            .map((item) => {
-                                const entry = normalizeArtifactEntry(item);
-                                if (!entry) return null;
-                                if (entry.workspace) return entry;
-                                return {
-                                    workspace: workspaceId,
-                                    alias: entry.alias,
-                                    fullId: `${workspaceId}/${entry.alias}`,
-                                };
-                            })
-                            .filter(Boolean);
-
-                        artifactDebug.cacheSize = normalized.length;
-                        artifactDebug.lastError = '';
-                        return normalized;
-                    } catch (error) {
-                        lastAttemptError = error;
-                        artifactDebug.lastError = String(error?.message || error);
-                        if (isWorkspacePermissionError(error)) {
-                            artifactInputError.value = `Permission denied to workspace '${workspaceId}'.`;
+                    const attempts = [
+                        () => am.list({ keywords, limit: 30, _rkwargs: true }),
+                        () => am.list({ keywords, limit: 30 }),
+                        () => am.list(workspaceId, { keywords, limit: 30 }),
+                        () => am.list(workspaceId, { keywords }),
+                    ];
+
+                    for (const attempt of attempts) {
+                        try {
+                            const listed = await attempt();
+                            const rawItems = normalizeArtifactListResponse(listed);
+                            const normalized = rawItems
+                                .map((item) => {
+                                    const entry = normalizeArtifactEntry(item);
+                                    if (!entry) return null;
+                                    if (entry.workspace) return entry;
+                                    return {
+                                        workspace: workspaceId,
+                                        alias: entry.alias,
+                                        fullId: `${workspaceId}/${entry.alias}`,
+                                    };
+                                })
+                                .filter(Boolean);
+
+                            artifactDebug.cacheSize = normalized.length;
+                            artifactDebug.lastError = '';
+                            return normalized;
+                        } catch (error) {
+                            lastAttemptError = error;
+                            artifactDebug.lastError = String(error?.message || error);
+                            if (isWorkspacePermissionError(error)) {
+                                artifactInputError.value = `Permission denied to workspace '${workspaceId}'.`;
+                            }
                         }
                     }
 
@@ -1015,6 +1046,56 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     return [];
                 };
 
+                const checkArtifactExists = async (workspaceId, alias) => {
+                    const lookupVersion = ++artifactLookupVersion.value;
+                    const normalizedWorkspace = String(workspaceId || '').trim();
+                    const normalizedAlias = String(alias || '').trim();
+
+                    if (!normalizedWorkspace || !normalizedAlias) {
+                        artifactIdExists.value = false;
+                        artifactLookupPending.value = false;
+                        return false;
+                    }
+
+                    const artifactId = `${normalizedWorkspace}/${normalizedAlias}`;
+                    artifactLookupPending.value = true;
+
+                    try {
+                        const client = await getWorkspaceClient(normalizedWorkspace);
+                        const am = await client.getService('public/artifact-manager');
+                        const attempts = [
+                            () => am.get(artifactId, { _rkwargs: true }),
+                            () => am.get(artifactId),
+                            () => (typeof am.read === 'function' ? am.read(artifactId) : Promise.reject(new Error('read unavailable'))),
+                        ];
+
+                        for (const attempt of attempts) {
+                            try {
+                                const result = await attempt();
+                                if (lookupVersion !== artifactLookupVersion.value) return false;
+                                const exists = !!result;
+                                artifactIdExists.value = exists;
+                                return exists;
+                            } catch (error) {
+                                if (isWorkspacePermissionError(error)) {
+                                    artifactInputError.value = `Permission denied to workspace '${normalizedWorkspace}'.`;
+                                }
+                            }
+                        }
+                    } catch (_e) {
+                        // Keep exists=false; detailed errors are handled in autocomplete path.
+                    } finally {
+                        if (lookupVersion === artifactLookupVersion.value) {
+                            artifactLookupPending.value = false;
+                        }
+                    }
+
+                    if (lookupVersion === artifactLookupVersion.value) {
+                        artifactIdExists.value = false;
+                    }
+                    return false;
+                };
+
                 const normalizeArtifactEntry = (item) => {
                     if (!item) return null;
                     if (typeof item === 'string') {
@@ -1091,7 +1172,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
 
                 const handleArtifactInput = async () => {
                      const requestVersion = ++artifactInputVersion.value;
-                     const val = String(trainingParams.artifact || '').trim();
+                     const val = String(trainingParams.artifact_input || '').trim();
                      artifactDebug.lastQuery = val;
                      artifactInputError.value = '';
 
@@ -1099,10 +1180,14 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                          suggestions.value = [];
                          showSuggestions.value = false;
                          artifactDebug.suggestions = 0;
+                         artifactIdExists.value = false;
+                         artifactLookupPending.value = false;
                          return;
                      }
 
                      if (!val.includes('/')) {
+                         artifactIdExists.value = false;
+                         artifactLookupPending.value = false;
                          const workspaces = await fetchWorkspaces();
                          if (requestVersion !== artifactInputVersion.value) return;
                          suggestions.value = workspaces
@@ -1121,6 +1206,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                      if (parts.length === 2) {
                          const ws = parts[0];
                          const query = parts[1].toLowerCase();
+                         checkArtifactExists(ws, parts[1]);
                          artifactDebug.lastWorkspace = ws;
                          const remoteMatches = await queryWorkspaceArtifacts(ws, query);
                          if (requestVersion !== artifactInputVersion.value) return;
@@ -1149,6 +1235,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             sample: suggestions.value.slice(0, 5),
                         });
                      } else {
+                                artifactIdExists.value = false;
+                                artifactLookupPending.value = false;
                         suggestions.value = [];
                         showSuggestions.value = false;
                         artifactDebug.suggestions = 0;
@@ -1158,12 +1246,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const selectSuggestion = (s) => {
                      // If it's a workspace (no slash), add slash
                      if (!s.includes('/')) {
-                         trainingParams.artifact = s + '/';
+                         trainingParams.artifact_input = s + '/';
                          // Refocus input to keep typing?
                          // Ensure handleArtifactInput runs for the new value
                          handleArtifactInput();
                      } else {
-                         trainingParams.artifact = s;
+                         trainingParams.artifact_input = s;
                          showSuggestions.value = false;
                      }
                 };
@@ -1259,6 +1347,13 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     trainingParams.train_annotations = 'annotations/*/*_mask.ome.tif';
                 };
 
+                const getSelectedArtifactId = () => {
+                    if (trainingParams.dataset_source === 'Upload') {
+                        return trainingParams.uploaded_artifact;
+                    }
+                    return trainingParams.artifact_input;
+                };
+
                 const applyTrainingFolderExample = () => {
                     trainingParams.train_images = 'images/';
                     trainingParams.train_annotations = 'annotations/';
@@ -1717,11 +1812,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         }
                         
                         // Determine the artifact value based on source
-                        let artifactValue = trainingParams.artifact;
+                            let artifactValue = trainingParams.artifact_input;
                         if (trainingParams.dataset_source === 'BioImage Archive') {
                             artifactValue = trainingParams.bioimage_archive_url;
                             if (!artifactValue) throw new Error("Please enter a BioImage Archive URL.");
                         } else if (trainingParams.dataset_source === 'Upload') {
+                                artifactValue = trainingParams.uploaded_artifact;
                              if (!artifactValue) throw new Error("Please upload a dataset first.");
                         } else {
                              if (!artifactValue) throw new Error("Please specify an Artifact ID.");
@@ -1857,7 +1953,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         await am.commit(artifact.id);
                         
                         // Set fields in form
-                        trainingParams.artifact = artifact.id;
+                        trainingParams.uploaded_artifact = artifact.id;
                         showUploadModal.value = false;
                         alert(`Dataset '${artifact.id}' uploaded successfully!`);
                         
@@ -2161,7 +2257,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
 
                 // --- File Browser ---
                 const openFileBrowser = async () => {
-                   if (!trainingParams.artifact) {
+                   const selectedArtifactId = getSelectedArtifactId();
+                   if (!selectedArtifactId) {
                        alert("Please enter an Artifact ID first (e.g. from the dashboard or your workspace).");
                        return;
                    }
@@ -2197,8 +2294,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                            // But `browserFiles` is the state for the whole tree currently.
                            // Let's optimize: When opening browser, fetch EVERYTHING.
                            
-                           console.log("Fetching file list for artifact:", trainingParams.artifact);
-                           rawFiles = await am.list_files(trainingParams.artifact);
+                           console.log("Fetching file list for artifact:", selectedArtifactId);
+                           rawFiles = await am.list_files(selectedArtifactId);
                            
                            // If rawFiles is just ["images", "annotations"], that means the artifact
                            // literally only contains those entries (maybe as empty file placeholders for folders?)
@@ -2208,7 +2305,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         // Fallback: If list_files didn't work as expected or returned empty
                         if (!rawFiles || rawFiles.length === 0) {
                             console.log("list_files returned empty, trying to read manifest directly...");
-                            const art = await am.read(trainingParams.artifact);
+                            const art = await am.read(selectedArtifactId);
                             if (art.manifest && art.manifest.files) {
                                 rawFiles = art.manifest.files;
                             } else if (art.files) {
@@ -2276,7 +2373,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                              let newFiles = [];
                              try {
                                  // Try positional: list_files(id, path)
-                                 newFiles = await am.list_files(trainingParams.artifact, newPath);
+                                 newFiles = await am.list_files(getSelectedArtifactId(), newPath);
                              } catch (err) {
                                  console.warn("list_files(id, path) failed. Trying without path filtering or keyword args if object", err);
                              }
@@ -2468,10 +2565,10 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     browserVisibleItems, browserNavigate, browserNavigateUp, selectFolderFor, browserFeedback,
 
                     // Autocomplete
-                    suggestions, showSuggestions, artifactInputError, handleArtifactInput, selectSuggestion, acceptSuggestion, closeSuggestions,
+                    suggestions, showSuggestions, artifactInputError, artifactIdExists, artifactLookupPending, handleArtifactInput, selectSuggestion, acceptSuggestion, closeSuggestions,
 
                     // Helpers
-                    statusTextColor, formatDuration, handleFolderSelect, getLatestLoss, formatDate,
+                    statusTextColor, formatDuration, handleFolderSelect, getLatestLoss, formatDate, getSelectedArtifactId,
                 };
             }
         });
diff --git a/bioengine_apps/cellpose_finetuning/main.py b/bioengine_apps/cellpose_finetuning/main.py
index ba6a4442..dab7837d 100644
--- a/bioengine_apps/cellpose_finetuning/main.py
+++ b/bioengine_apps/cellpose_finetuning/main.py
@@ -9,6 +9,7 @@
 import asyncio
 import base64
 import fnmatch
+import hashlib
 import io
 import json
 import logging
@@ -47,6 +48,8 @@
 TRAINING_PARAMS_FILENAME = "training_params.json"
 STOP_REQUESTED_FILENAME = "stop.requested"
 STATUS_STALE_SECONDS = 300
+BIA_FTS_ENDPOINT = "https://beta.bioimagearchive.org/search/search/fts"
+BIA_IMAGE_ENDPOINT = "https://beta.bioimagearchive.org/search/search/fts/image"
 
 # Model template for BioImage.io export
 MODEL_TEMPLATE_PY = '''"""BioImage.io Model Wrapper for Cellpose 4.0.7 (Cellpose-SAM).
@@ -2003,6 +2006,226 @@ def _glob_base_folder(path_pattern: str) -> str:
     return normalized[: slash_before + 1]
 
 
+def _is_bioimage_archive_url(candidate: str) -> bool:
+    parsed = urlparse(str(candidate or ""))
+    return (
+        parsed.scheme in {"http", "https"}
+        and "bioimagearchive.org" in parsed.netloc.lower()
+    )
+
+
+def _extract_bia_accession(url: str) -> str | None:
+    match = re.search(r"(S-BIAD\d+)", str(url or ""), flags=re.IGNORECASE)
+    return match.group(1).upper() if match else None
+
+
+def _iter_string_values(payload: Any) -> list[str]:
+    out: list[str] = []
+    stack = [payload]
+    seen: set[int] = set()
+
+    while stack:
+        current = stack.pop()
+        current_id = id(current)
+        if current_id in seen:
+            continue
+        seen.add(current_id)
+
+        if isinstance(current, str):
+            out.append(current)
+            continue
+        if isinstance(current, dict):
+            stack.extend(current.values())
+            continue
+        if isinstance(current, list):
+            stack.extend(current)
+            continue
+
+    return out
+
+
+def _looks_like_image_url(url: str) -> bool:
+    path = urlparse(url).path.lower()
+    return path.endswith((".tif", ".tiff", ".ome.tif", ".ome.tiff", ".png", ".jpg", ".jpeg"))
+
+
+def _looks_like_mask_url(url: str) -> bool:
+    path = urlparse(url).path.lower()
+    return any(token in path for token in ("_mask", "-mask", "_label", "-label", "annotation", "segmentation"))
+
+
+def _pair_key_from_url(url: str, *, is_mask: bool) -> str:
+    path = urlparse(url).path.lower().strip("/")
+    for suffix in (
+        ".ome.tiff",
+        ".ome.tif",
+        ".tiff",
+        ".tif",
+        ".png",
+        ".jpg",
+        ".jpeg",
+    ):
+        if path.endswith(suffix):
+            path = path[: -len(suffix)]
+            break
+
+    if is_mask:
+        for marker in (
+            "_mask",
+            "-mask",
+            "_label",
+            "-label",
+            "_annotation",
+            "-annotation",
+        ):
+            if path.endswith(marker):
+                path = path[: -len(marker)]
+                break
+
+    for marker in ("/images/", "/annotations/", "/masks/", "/labels/"):
+        if marker in path:
+            path = path.split(marker, 1)[1]
+            break
+
+    return path
+
+
+async def _fetch_bia_payload(client: Any, endpoint: str, accession: str) -> Any:
+    attempts: list[tuple[str, dict[str, Any]]] = [
+        ("get", {"params": {"q": accession, "size": 1000}}),
+        ("get", {"params": {"query": accession, "size": 1000}}),
+        ("post", {"json": {"query": accession, "size": 1000}}),
+        ("post", {"json": {"q": accession, "size": 1000}}),
+    ]
+
+    for method, kwargs in attempts:
+        try:
+            if method == "get":
+                response = await client.get(endpoint, **kwargs)
+            else:
+                response = await client.post(endpoint, **kwargs)
+            if response.status_code // 100 == 2:
+                return response.json()
+        except Exception:
+            continue
+
+    return None
+
+
+def _local_path_for_remote_url(root: Path, category: str, remote_url: str) -> Path:
+    parsed = urlparse(remote_url)
+    basename = Path(parsed.path).name or "asset.bin"
+    stem = Path(basename).stem
+    suffix = "".join(Path(basename).suffixes) or ".bin"
+    digest = hashlib.md5(remote_url.encode("utf-8")).hexdigest()[:10]
+    filename = f"{stem}_{digest}{suffix}"
+    return root / category / filename
+
+
+def _is_test_url(url: str) -> bool:
+    path = urlparse(url).path.lower()
+    return "/test/" in path or "_test" in path or "-test" in path
+
+
+async def make_training_pairs_from_bioimage_archive_url(
+    config: TrainingParams,
+    save_path: Path,
+) -> tuple[list[TrainingPair], list[TrainingPair]]:
+    import httpx
+
+    archive_url = config["artifact_id"]
+    accession = _extract_bia_accession(archive_url)
+    if not accession:
+        raise ValueError(
+            "Could not parse BioImage Archive accession (e.g. S-BIAD1234) from URL."
+        )
+
+    async with httpx.AsyncClient(timeout=120) as client:
+        payloads = [
+            await _fetch_bia_payload(client, BIA_FTS_ENDPOINT, accession),
+            await _fetch_bia_payload(client, BIA_IMAGE_ENDPOINT, accession),
+        ]
+
+        candidates: set[str] = set()
+        for payload in payloads:
+            if payload is None:
+                continue
+            for value in _iter_string_values(payload):
+                if isinstance(value, str) and value.startswith(("http://", "https://")):
+                    if _looks_like_image_url(value):
+                        candidates.add(value)
+
+        if not candidates:
+            raise ValueError(
+                "No downloadable image assets were found from BioImage Archive APIs for this accession."
+            )
+
+        image_urls = sorted([url for url in candidates if not _looks_like_mask_url(url)])
+        mask_urls = sorted([url for url in candidates if _looks_like_mask_url(url)])
+
+        if not image_urls or not mask_urls:
+            raise ValueError(
+                "Found assets in BioImage Archive response, but could not identify both image and mask files."
+            )
+
+        mask_map: dict[str, str] = {}
+        for mask_url in mask_urls:
+            key = _pair_key_from_url(mask_url, is_mask=True)
+            if key and key not in mask_map:
+                mask_map[key] = mask_url
+
+        paired_urls: list[tuple[str, str]] = []
+        for image_url in image_urls:
+            key = _pair_key_from_url(image_url, is_mask=False)
+            mask_url = mask_map.get(key)
+            if mask_url:
+                paired_urls.append((image_url, mask_url))
+
+        if not paired_urls:
+            raise ValueError(
+                "No image/mask pairs could be inferred from BioImage Archive assets."
+            )
+
+        if config["n_samples"] is not None and config["n_samples"] < len(paired_urls):
+            subset_idx = np.random.default_rng().permutation(len(paired_urls))[ : config["n_samples"] ]
+            paired_urls = [paired_urls[i] for i in subset_idx]
+
+        bia_cache_root = save_path / "bia_download"
+        train_pairs: list[TrainingPair] = []
+        test_pairs: list[TrainingPair] = []
+
+        for image_url, mask_url in paired_urls:
+            local_image = _local_path_for_remote_url(bia_cache_root, "images", image_url)
+            local_mask = _local_path_for_remote_url(bia_cache_root, "annotations", mask_url)
+            local_image.parent.mkdir(parents=True, exist_ok=True)
+            local_mask.parent.mkdir(parents=True, exist_ok=True)
+
+            if not local_image.exists() or local_image.stat().st_size <= 0:
+                resp = await client.get(image_url)
+                resp.raise_for_status()
+                local_image.write_bytes(resp.content)
+
+            if not local_mask.exists() or local_mask.stat().st_size <= 0:
+                resp = await client.get(mask_url)
+                resp.raise_for_status()
+                local_mask.write_bytes(resp.content)
+
+            pair = TrainingPair(image=local_image, annotation=local_mask)
+            if _is_test_url(image_url) or _is_test_url(mask_url):
+                test_pairs.append(pair)
+            else:
+                train_pairs.append(pair)
+
+        if not train_pairs and test_pairs:
+            train_pairs = test_pairs
+            test_pairs = []
+
+        if not train_pairs:
+            raise ValueError("No training pairs found after downloading BioImage Archive assets")
+
+        return train_pairs, test_pairs
+
+
 async def list_artifact_files_recursive(
     artifact: AsyncHyphaArtifact,
     folder_path: str,
@@ -2042,6 +2265,19 @@ async def list_artifact_files_recursive(
                 "directory",
                 "dir",
             }
+            if not is_dir and not entry_type and normalized:
+                # Some artifact backends return directory entries without trailing '/'
+                # and without explicit type metadata. Probe such paths as directories.
+                probe_dir = normalized if normalized.endswith("/") else normalized + "/"
+                if probe_dir not in visited:
+                    try:
+                        probe_entries = await artifact.ls(probe_dir)
+                        if isinstance(probe_entries, list):
+                            queue.append(probe_dir)
+                            continue
+                    except Exception:
+                        pass
+
             if is_dir:
                 next_dir = normalized if normalized.endswith("/") else normalized + "/"
                 if next_dir not in visited:
@@ -2215,6 +2451,64 @@ def _annotation_key(path: str) -> str:
                 break
         return key
 
+    image_root = _normalize_artifact_relpath(_glob_base_folder(normalized_image_pattern))
+    annotation_root = _normalize_artifact_relpath(
+        _glob_base_folder(normalized_annotation_pattern)
+    )
+
+    def _relative_key(path: str, root: str, remove_label_suffix: bool) -> str:
+        normalized = _normalize_artifact_relpath(path).lower()
+        relative = normalized
+        if root and normalized.startswith(root):
+            relative = normalized[len(root) :]
+
+        for suffix in (
+            ".ome.tiff",
+            ".ome.tif",
+            ".tiff",
+            ".tif",
+            ".png",
+            ".jpg",
+            ".jpeg",
+        ):
+            if relative.endswith(suffix):
+                relative = relative[: -len(suffix)]
+                break
+
+        if remove_label_suffix:
+            for marker in (
+                "_mask",
+                "-mask",
+                "_label",
+                "-label",
+                "_annotation",
+                "-annotation",
+            ):
+                if relative.endswith(marker):
+                    relative = relative[: -len(marker)]
+                    break
+
+        return relative.strip("/")
+
+    relative_ann_map: dict[str, str] = {}
+    for annot_file in annotation_files:
+        key = _relative_key(annot_file, annotation_root, True)
+        if key and key not in relative_ann_map:
+            relative_ann_map[key] = annot_file
+
+    for image_file in image_files:
+        key = _relative_key(image_file, image_root, False)
+        annot_file = relative_ann_map.get(key)
+        if annot_file:
+            pairs.append((image_file, annot_file))
+
+    if pairs:
+        logger.info(
+            f"Matched {len(pairs)} pairs from {len(image_files)} images "
+            f"and {len(annotation_files)} annotations"
+        )
+        return pairs
+
     # Fallback matcher for mixed conventions like image '*.tif' vs annotation '*_mask.ome.tif'.
     fallback_ann_map: dict[str, str] = {}
     for annot_file in annotation_files:
@@ -2586,8 +2880,13 @@ async def make_training_pairs(
     Returns:
         Tuple of (train_pairs, test_pairs). test_pairs is empty if test folders not specified.
     """
+    artifact_id = config["artifact_id"]
+
+    if _is_bioimage_archive_url(artifact_id):
+        return await make_training_pairs_from_bioimage_archive_url(config, save_path)
+
     artifact = await make_artifact_client(
-        config["artifact_id"],
+        artifact_id,
         config["server_url"],
     )
 
diff --git a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py b/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
index 8e479f60..f10b5b86 100644
--- a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
+++ b/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
@@ -4,6 +4,8 @@
 from pathlib import Path
 
 from bioengine_apps.cellpose_finetuning.main import (
+    _extract_bia_accession,
+    _pair_key_from_url,
     list_matching_artifact_paths,
     make_training_pairs_from_metadata,
     match_image_annotation_pairs,
@@ -64,6 +66,35 @@ def test_match_image_annotation_pairs_with_mixed_ome_suffix_convention() -> None
     ]
 
 
+def test_match_image_annotation_pairs_with_many_nested_folders() -> None:
+    image_files = [
+        "images/folder1/sub1/img001.tif",
+        "images/folder2/sub2/img002.tif",
+    ]
+    annotation_files = [
+        "annotations/folder1/sub1/img001_mask.ome.tif",
+        "annotations/folder2/sub2/img002_mask.ome.tif",
+    ]
+
+    pairs = match_image_annotation_pairs(
+        image_files,
+        annotation_files,
+        "images/**/*.tif",
+        "annotations/**/*_mask.ome.tif",
+    )
+
+    assert pairs == [
+        (
+            "images/folder1/sub1/img001.tif",
+            "annotations/folder1/sub1/img001_mask.ome.tif",
+        ),
+        (
+            "images/folder2/sub2/img002.tif",
+            "annotations/folder2/sub2/img002_mask.ome.tif",
+        ),
+    ]
+
+
 class _FakeArtifact:
     async def ls(self, folder_path: str):
         await asyncio.sleep(0)
@@ -137,6 +168,33 @@ async def get(self, remote_paths, local_paths, on_error="ignore"):
                                 local_path.write_bytes(b"dummy")
 
 
+class _FakeArtifactAmbiguousDirs:
+    async def ls(self, folder_path: str):
+        await asyncio.sleep(0)
+        folder = folder_path.rstrip("/") + "/"
+        if folder == "images/":
+            return [{"path": "images/108bb69d-2e52-4382-8100-e96173db24ee"}]
+        if folder == "images/108bb69d-2e52-4382-8100-e96173db24ee/":
+            return [
+                {
+                    "path": "images/108bb69d-2e52-4382-8100-e96173db24ee/t0000.ome.tif",
+                    "type": "file",
+                },
+                {
+                    "path": "images/108bb69d-2e52-4382-8100-e96173db24ee/t0001.ome.tif",
+                    "type": "file",
+                },
+            ]
+        return []
+
+    async def get(self, remote_paths, local_paths, on_error="ignore"):
+        await asyncio.sleep(0)
+        for local in local_paths:
+            local_path = Path(local)
+            local_path.parent.mkdir(parents=True, exist_ok=True)
+            local_path.write_bytes(b"dummy")
+
+
 def test_make_training_pairs_from_metadata(tmp_path: Path) -> None:
     artifact = _FakeArtifact()
     train_pairs, test_pairs = asyncio.run(
@@ -165,6 +223,15 @@ def test_list_matching_artifact_paths_directory_without_trailing_slash() -> None
     ]
 
 
+def test_list_matching_artifact_paths_with_ambiguous_directory_entries() -> None:
+    artifact = _FakeArtifactAmbiguousDirs()
+    matches = asyncio.run(list_matching_artifact_paths(artifact, "images/*/*.tif"))
+    assert matches == [
+        "images/108bb69d-2e52-4382-8100-e96173db24ee/t0000.ome.tif",
+        "images/108bb69d-2e52-4382-8100-e96173db24ee/t0001.ome.tif",
+    ]
+
+
 def test_make_training_pairs_from_nested_metadata_formats(tmp_path: Path) -> None:
     artifact = _FakeArtifactNestedMetadata()
     train_pairs, test_pairs = asyncio.run(
@@ -182,3 +249,14 @@ def test_make_training_pairs_from_nested_metadata_formats(tmp_path: Path) -> Non
     assert train_pairs[0]["annotation"].as_posix().endswith("annotations/train/t0010_mask.ome.tif")
     assert test_pairs[0]["image"].as_posix().endswith("images/test/t0011.tif")
     assert test_pairs[0]["annotation"].as_posix().endswith("annotations/test/t0011_mask.ome.tif")
+
+
+def test_extract_bia_accession_from_gallery_url() -> None:
+    url = "https://beta.bioimagearchive.org/bioimage-archive/galleries/ai/ai-ready-study/S-BIAD1392"
+    assert _extract_bia_accession(url) == "S-BIAD1392"
+
+
+def test_pair_key_from_url_handles_mask_suffix() -> None:
+    img = "https://example.org/data/images/folder1/t0001.ome.tif"
+    ann = "https://example.org/data/annotations/folder1/t0001_mask.ome.tif"
+    assert _pair_key_from_url(img, is_mask=False) == _pair_key_from_url(ann, is_mask=True)

From 95e816320ab50c4f81da7ebc935f8cb376016ac8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Fri, 20 Feb 2026 00:13:53 +0100
Subject: [PATCH 06/13] Add auto split mode and robust folder detection in
 cellpose UI

---
 bioengine_apps/cellpose_finetuning/index.html | 92 ++++++++++++++++---
 bioengine_apps/cellpose_finetuning/main.py    | 84 ++++++++++++++++-
 2 files changed, 160 insertions(+), 16 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 6de4c7d6..b2aefcb2 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -395,6 +395,24 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                 </div>
 
                                 <div class="md:col-span-2 grid grid-cols-1 md:grid-cols-2 gap-4 pt-2">
+                                    <div class="md:col-span-2">
+                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-2">Train/Test Split</label>
+                                        <div class="flex items-center gap-5 text-sm text-gray-700">
+                                            <label class="inline-flex items-center gap-2 cursor-pointer">
+                                                <input type="radio" value="manual" v-model="trainingParams.split_mode" class="text-blue-600 focus:ring-blue-500">
+                                                <span>Manual test paths</span>
+                                            </label>
+                                            <label class="inline-flex items-center gap-2 cursor-pointer">
+                                                <input type="radio" value="auto" v-model="trainingParams.split_mode" class="text-blue-600 focus:ring-blue-500">
+                                                <span>Auto split</span>
+                                            </label>
+                                            <div v-if="trainingParams.split_mode === 'auto'" class="inline-flex items-center gap-2">
+                                                <span class="text-xs text-gray-500">Train ratio</span>
+                                                <input v-model.number="trainingParams.train_split_ratio" type="number" min="0.1" max="0.95" step="0.05" class="w-24 bg-white border border-gray-300 rounded-md px-2 py-1 text-sm">
+                                            </div>
+                                        </div>
+                                        <p class="mt-1 text-xs text-gray-500" v-if="trainingParams.split_mode === 'auto'">Uses train paths only and splits into train/test (default 0.8).</p>
+                                    </div>
                                      <div>
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Train Images</label>
                                         <input v-model="trainingParams.train_images" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. images/*/*.tif">
@@ -413,11 +431,11 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                         <input v-model="trainingParams.metadata_dir" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. metadata/">
                                         <p class="text-xs text-gray-500 mt-1">Use this to load image/annotation pairs from metadata JSON files; then explicit annotation masks are optional.</p>
                                     </div>
-                                    <div>
+                                    <div v-if="trainingParams.split_mode === 'manual'">
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Images (Opt)</label>
                                         <input v-model="trainingParams.test_images" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
                                     </div>
-                                    <div>
+                                    <div v-if="trainingParams.split_mode === 'manual'">
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Annotations (Opt)</label>
                                         <input v-model="trainingParams.test_annotations" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
                                     </div>
@@ -825,6 +843,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     train_images: 'images/',
                     train_annotations: 'annotations/',
                     metadata_dir: '',
+                    split_mode: 'manual',
+                    train_split_ratio: 0.8,
                     test_images: '',
                     test_annotations: '',
                     n_epochs: 50,
@@ -862,6 +882,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 // File Browser
                 const showFileBrowser = ref(false);
                 const browserFiles = ref([]);
+                const browserPathTypes = reactive({});
                 const browserPath = ref(''); 
                 const browserLoading = ref(false);
                 const browserError = ref('');
@@ -1828,8 +1849,10 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             train_images: trainingParams.train_images,
                             train_annotations: trainingParams.train_annotations,
                             metadata_dir: trainingParams.metadata_dir || null,
-                            test_images: trainingParams.test_images || null,
-                            test_annotations: trainingParams.test_annotations || null,
+                            split_mode: trainingParams.split_mode,
+                            train_split_ratio: trainingParams.train_split_ratio,
+                            test_images: trainingParams.split_mode === 'manual' ? (trainingParams.test_images || null) : null,
+                            test_annotations: trainingParams.split_mode === 'manual' ? (trainingParams.test_annotations || null) : null,
                             model: trainingParams.model,
                             n_epochs: trainingParams.n_epochs,
                             learning_rate: trainingParams.learning_rate,
@@ -1844,8 +1867,10 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                                 train_images: trainingParams.train_images || null,
                                 train_annotations: trainingParams.train_annotations || null,
                                 metadata_dir: trainingParams.metadata_dir || null,
-                                test_images: trainingParams.test_images || null,
-                                test_annotations: trainingParams.test_annotations || null,
+                                split_mode: trainingParams.split_mode || 'manual',
+                                train_split_ratio: Number(trainingParams.train_split_ratio || 0.8),
+                                test_images: trainingParams.split_mode === 'manual' ? (trainingParams.test_images || null) : null,
+                                test_annotations: trainingParams.split_mode === 'manual' ? (trainingParams.test_annotations || null) : null,
                                 model: trainingParams.model || 'cpsam',
                                 n_samples: null,
                                 n_epochs: trainingParams.n_epochs,
@@ -2266,6 +2291,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                    browserPath.value = '';
                    browserError.value = '';
                    browserFiles.value = [];
+                   Object.keys(browserPathTypes).forEach((k) => delete browserPathTypes[k]);
                    browserLoading.value = true;
                    
                    try {
@@ -2320,12 +2346,33 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             console.log("Service keys:", Object.keys(service.value));
                         }
                            
-                        browserFiles.value = rawFiles.map(f => {
+                           browserFiles.value = rawFiles.map(f => {
                             // Handle object format {name: "...", ...} or string
                             let p = typeof f === 'string' ? f : (f.path || f.name || f.key || "");
                             // Remove leading ./ or /
                             return p.replace(/^\.?\//, '');
                         }).filter(f => f && typeof f === 'string');
+
+                        const detectPathType = async (pathStr) => {
+                            const normalizedPath = String(pathStr || '').replace(/^\.?\//, '').replace(/\/$/, '');
+                            if (!normalizedPath) return;
+
+                            try {
+                                const listed = await am.list_files(selectedArtifactId, normalizedPath);
+                                if (Array.isArray(listed) && listed.length > 0) {
+                                    browserPathTypes[normalizedPath] = 'folder';
+                                    return;
+                                }
+                            } catch (_e) {
+                                // Ignore and try other heuristics
+                            }
+
+                            if (browserPathTypes[normalizedPath] !== 'folder') {
+                                browserPathTypes[normalizedPath] = 'file';
+                            }
+                        };
+
+                        await Promise.all(browserFiles.value.slice(0, 200).map((pathStr) => detectPathType(pathStr)));
                             
                         console.log("Loaded browser files:", browserFiles.value.length, browserFiles.value.slice(0, 10));
 
@@ -2401,6 +2448,22 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                                          browserFiles.value.push(pathStr);
                                          added++;
                                      }
+
+                                     const normalizedPath = String(pathStr || '').replace(/\/$/, '');
+                                     if (normalizedPath) {
+                                         try {
+                                             const listed = await am.list_files(getSelectedArtifactId(), normalizedPath);
+                                             if (Array.isArray(listed) && listed.length > 0) {
+                                                 browserPathTypes[normalizedPath] = 'folder';
+                                             } else if (browserPathTypes[normalizedPath] !== 'folder') {
+                                                 browserPathTypes[normalizedPath] = 'file';
+                                             }
+                                         } catch (_e) {
+                                             if (browserPathTypes[normalizedPath] !== 'folder') {
+                                                 browserPathTypes[normalizedPath] = 'file';
+                                             }
+                                         }
+                                     }
                                  }
                                  console.log(`Fetched ${added} new items for ${newPath}`);
                              }
@@ -2444,7 +2507,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                          if (parts.length > 1 && parts[1] !== "") {
                              // It is a directory containing stuff
                              if (!items.has(name)) {
-                                 items.set(name, { name, type: 'folder', fullPath: p + name });
+                                 const fullPath = (p + name).replace(/\/$/, '');
+                                 browserPathTypes[fullPath] = 'folder';
+                                 items.set(name, { name, type: 'folder', fullPath });
                              }
                          } else {
                              // It is a file or empty directory placeholder
@@ -2452,20 +2517,23 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                              // Check if it's a file
                              if (!items.has(name)) {
                                 const isExplicitFolder = f.endsWith('/');
+                                const fullPath = (p + name).replace(/\/$/, '');
+                                const knownType = browserPathTypes[fullPath];
                                 items.set(name, { 
                                     name, 
-                                    type: isExplicitFolder ? 'folder' : 'file', 
-                                    fullPath: p + name 
+                                    type: knownType || (isExplicitFolder ? 'folder' : 'file'), 
+                                    fullPath
                                 });
                              }
                          }
                     }
 
-                    // Heuristic: If we have files named "images", "masks" etc without extensions, treat as folders
+                    // Heuristic fallback only when type is still unknown
                     for (const [name, item] of items) {
-                        if (item.type === 'file' && !name.includes('.')) {
+                        if (!browserPathTypes[item.fullPath] && item.type === 'file' && !name.includes('.')) {
                              if (['images', 'annotations', 'masks', 'models', 'data', 'train', 'test', 'val'].includes(name.toLowerCase())) {
                                  item.type = 'folder';
+                                 browserPathTypes[item.fullPath] = 'folder';
                              }
                         }
                     }
diff --git a/bioengine_apps/cellpose_finetuning/main.py b/bioengine_apps/cellpose_finetuning/main.py
index dab7837d..c264d973 100644
--- a/bioengine_apps/cellpose_finetuning/main.py
+++ b/bioengine_apps/cellpose_finetuning/main.py
@@ -259,6 +259,8 @@ class TrainingParams(TypedDict):
     metadata_dir: str | None
     test_images: str | None
     test_annotations: str | None
+    split_mode: str
+    train_split_ratio: float
     model: str | Path
     n_epochs: int
     learning_rate: float
@@ -2871,6 +2873,36 @@ def get_training_subset(
     return image_paths, annotation_paths
 
 
+def split_training_pairs(
+    train_pairs: list[TrainingPair],
+    train_ratio: float,
+) -> tuple[list[TrainingPair], list[TrainingPair]]:
+    """Split training pairs into train/test subsets using the provided ratio.
+
+    Keeps at least one sample in train; for datasets with >=2 samples, keeps at
+    least one sample in test.
+    """
+    if not train_pairs:
+        return [], []
+
+    n_total = len(train_pairs)
+    if n_total == 1:
+        return train_pairs, []
+
+    ratio = float(train_ratio)
+    ratio = max(0.05, min(0.95, ratio))
+
+    n_train = int(round(n_total * ratio))
+    n_train = max(1, min(n_total - 1, n_train))
+
+    indices = np.random.default_rng().permutation(n_total)
+    train_idx = set(indices[:n_train])
+
+    split_train = [pair for i, pair in enumerate(train_pairs) if i in train_idx]
+    split_test = [pair for i, pair in enumerate(train_pairs) if i not in train_idx]
+    return split_train, split_test
+
+
 async def make_training_pairs(
     config: TrainingParams,
     save_path: Path,
@@ -2983,6 +3015,13 @@ async def make_training_pairs(
             test_annotation_paths,
         )
 
+    split_mode = str(config.get("split_mode", "manual") or "manual").lower()
+    if split_mode == "auto" and not test_pairs:
+        train_pairs, test_pairs = split_training_pairs(
+            train_pairs,
+            float(config.get("train_split_ratio", 0.8)),
+        )
+
     return train_pairs, test_pairs
 
 
@@ -3538,6 +3577,21 @@ async def start_training(
             ),
             examples=["annotations/test/", "annotations/test/*_mask.ome.tif"],
         ),
+        split_mode: str = Field(
+            "manual",
+            description=(
+                "Dataset split mode. 'manual' uses test_images/test_annotations; "
+                "'auto' creates test split from training pairs using train_split_ratio."
+            ),
+            examples=["manual", "auto"],
+        ),
+        train_split_ratio: float = Field(
+            0.8,
+            description=(
+                "Train ratio used only when split_mode='auto'. "
+                "Example: 0.8 means 80% train / 20% test."
+            ),
+        ),
         model: str = Field(
             PretrainedModel.CPSAM.value,
             description=(
@@ -3596,6 +3650,8 @@ async def start_training(
             metadata_dir = wrapped.get("metadata_dir", metadata_dir)
             test_images = wrapped.get("test_images", test_images)
             test_annotations = wrapped.get("test_annotations", test_annotations)
+            split_mode = wrapped.get("split_mode", split_mode)
+            train_split_ratio = wrapped.get("train_split_ratio", train_split_ratio)
             model = wrapped.get("model", model)
             n_samples = wrapped.get("n_samples", n_samples)
             n_epochs = wrapped.get("n_epochs", n_epochs)
@@ -3612,6 +3668,8 @@ async def start_training(
         metadata_dir = normalize_optional_param(metadata_dir)
         test_images = normalize_optional_param(test_images)
         test_annotations = normalize_optional_param(test_annotations)
+        split_mode = normalize_optional_param(split_mode)
+        train_split_ratio = normalize_optional_param(train_split_ratio)
         model = normalize_optional_param(model)
         n_samples = normalize_optional_param(n_samples)
         n_epochs = normalize_optional_param(n_epochs)
@@ -3633,10 +3691,24 @@ async def start_training(
             if not isinstance(metadata_dir, str) or not metadata_dir:
                 raise ValueError("metadata_dir must be a non-empty string")
 
-        if (test_images is None) ^ (test_annotations is None):
-            raise ValueError(
-                "test_images and test_annotations must be provided together"
-            )
+        if not isinstance(split_mode, str) or split_mode not in {"manual", "auto"}:
+            split_mode = "manual"
+
+        if train_split_ratio is None:
+            train_split_ratio = 0.8
+        train_split_ratio = float(train_split_ratio)
+        if not (0.0 < train_split_ratio < 1.0):
+            raise ValueError("train_split_ratio must be between 0 and 1")
+
+        if split_mode == "manual":
+            if (test_images is None) ^ (test_annotations is None):
+                raise ValueError(
+                    "test_images and test_annotations must be provided together in manual split mode"
+                )
+        else:
+            # In auto split mode, ignore explicitly provided test inputs.
+            test_images = None
+            test_annotations = None
 
         if not isinstance(model, str) or not model:
             model = PretrainedModel.CPSAM.value
@@ -3684,6 +3756,8 @@ async def start_training(
             metadata_dir=metadata_dir,
             test_images=test_images,
             test_annotations=test_annotations,
+            split_mode=split_mode,
+            train_split_ratio=train_split_ratio,
             model=model_id,
             n_epochs=n_epochs,
             learning_rate=learning_rate,
@@ -3917,6 +3991,8 @@ async def restart_training(
             metadata_dir=params.get("metadata_dir"),
             test_images=params.get("test_images"),
             test_annotations=params.get("test_annotations"),
+            split_mode=params.get("split_mode", "manual"),
+            train_split_ratio=params.get("train_split_ratio", 0.8),
             model=restart_model,
             n_samples=params.get("n_samples"),
             n_epochs=restart_epochs,

From 069c1197008f9a72a3a8f1bc5bd3019fe1eb815e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Fri, 20 Feb 2026 00:36:46 +0100
Subject: [PATCH 07/13] Simplify training form in BioImage Archive mode

---
 bioengine_apps/cellpose_finetuning/index.html | 45 +++++++++----------
 1 file changed, 22 insertions(+), 23 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index b2aefcb2..884fc6d2 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -394,7 +394,7 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                     </div>
                                 </div>
 
-                                <div class="md:col-span-2 grid grid-cols-1 md:grid-cols-2 gap-4 pt-2">
+                                <div v-if="trainingParams.dataset_source !== 'BioImage Archive'" class="md:col-span-2 grid grid-cols-1 md:grid-cols-2 gap-4 pt-2">
                                     <div class="md:col-span-2">
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-2">Train/Test Split</label>
                                         <div class="flex items-center gap-5 text-sm text-gray-700">
@@ -440,6 +440,9 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                         <input v-model="trainingParams.test_annotations" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
                                     </div>
                                 </div>
+                                <div v-else class="md:col-span-2 rounded-lg border border-blue-200 bg-blue-50 px-4 py-3 text-sm text-blue-900">
+                                    BioImage Archive mode discovers image/mask pairs from the archive URL automatically, so path and split fields are not needed.
+                                </div>
                             </div>
                          </div>
                           </div>
@@ -1819,22 +1822,18 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     isStartingTraining.value = true;
                     trainingError.value = '';
                     try {
+                        const isBioImageArchive = trainingParams.dataset_source === 'BioImage Archive';
                         const hasMetadataDir = !!(trainingParams.metadata_dir && trainingParams.metadata_dir.trim());
                         const hasTrainImages = !!(trainingParams.train_images && trainingParams.train_images.trim());
                         const hasTrainAnnotations = !!(trainingParams.train_annotations && trainingParams.train_annotations.trim());
 
-                        if (!hasMetadataDir && (!hasTrainImages || !hasTrainAnnotations)) {
-                           // If using BioImage Archive, we might not need explicit images/annotations if the archive has a standard structure OR we assume default?
-                           // But usually we still need them.
-                           // Actually, for BioImage Archive, the prompt says "allow the user to post a URL to bioimage archive".
-                           // Does the backend handle downloading?
-                           // If so, we pass the URL as 'artifact'? Or a new param?
-                           // The prompt implies "Dataset source" choice.
+                        if (!isBioImageArchive && !hasMetadataDir && (!hasTrainImages || !hasTrainAnnotations)) {
+                           throw new Error("Please provide either metadata folder or both train image and annotation paths.");
                         }
                         
                         // Determine the artifact value based on source
                             let artifactValue = trainingParams.artifact_input;
-                        if (trainingParams.dataset_source === 'BioImage Archive') {
+                        if (isBioImageArchive) {
                             artifactValue = trainingParams.bioimage_archive_url;
                             if (!artifactValue) throw new Error("Please enter a BioImage Archive URL.");
                         } else if (trainingParams.dataset_source === 'Upload') {
@@ -1846,13 +1845,13 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
 
                         console.log("Invoking service.start_training with:", {
                             artifact: artifactValue,
-                            train_images: trainingParams.train_images,
-                            train_annotations: trainingParams.train_annotations,
-                            metadata_dir: trainingParams.metadata_dir || null,
-                            split_mode: trainingParams.split_mode,
-                            train_split_ratio: trainingParams.train_split_ratio,
-                            test_images: trainingParams.split_mode === 'manual' ? (trainingParams.test_images || null) : null,
-                            test_annotations: trainingParams.split_mode === 'manual' ? (trainingParams.test_annotations || null) : null,
+                            train_images: isBioImageArchive ? null : (trainingParams.train_images || null),
+                            train_annotations: isBioImageArchive ? null : (trainingParams.train_annotations || null),
+                            metadata_dir: isBioImageArchive ? null : (trainingParams.metadata_dir || null),
+                            split_mode: isBioImageArchive ? 'manual' : trainingParams.split_mode,
+                            train_split_ratio: isBioImageArchive ? 0.8 : trainingParams.train_split_ratio,
+                            test_images: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_images || null) : null,
+                            test_annotations: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_annotations || null) : null,
                             model: trainingParams.model,
                             n_epochs: trainingParams.n_epochs,
                             learning_rate: trainingParams.learning_rate,
@@ -1864,13 +1863,13 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         const result = await service.value.start_training(
                             {
                                 artifact: artifactValue,
-                                train_images: trainingParams.train_images || null,
-                                train_annotations: trainingParams.train_annotations || null,
-                                metadata_dir: trainingParams.metadata_dir || null,
-                                split_mode: trainingParams.split_mode || 'manual',
-                                train_split_ratio: Number(trainingParams.train_split_ratio || 0.8),
-                                test_images: trainingParams.split_mode === 'manual' ? (trainingParams.test_images || null) : null,
-                                test_annotations: trainingParams.split_mode === 'manual' ? (trainingParams.test_annotations || null) : null,
+                                train_images: isBioImageArchive ? null : (trainingParams.train_images || null),
+                                train_annotations: isBioImageArchive ? null : (trainingParams.train_annotations || null),
+                                metadata_dir: isBioImageArchive ? null : (trainingParams.metadata_dir || null),
+                                split_mode: isBioImageArchive ? 'manual' : (trainingParams.split_mode || 'manual'),
+                                train_split_ratio: isBioImageArchive ? 0.8 : Number(trainingParams.train_split_ratio || 0.8),
+                                test_images: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_images || null) : null,
+                                test_annotations: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_annotations || null) : null,
                                 model: trainingParams.model || 'cpsam',
                                 n_samples: null,
                                 n_epochs: trainingParams.n_epochs,

From b96ad1f4423c67629dd0f8a71aefcec5e424ec12 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Fri, 20 Feb 2026 00:48:08 +0100
Subject: [PATCH 08/13] Fix training base-model selector to dropdown-only

---
 bioengine_apps/cellpose_finetuning/index.html | 22 +++++++++++++------
 1 file changed, 15 insertions(+), 7 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 884fc6d2..5895f907 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -385,13 +385,21 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                 <div class="md:col-span-2 space-y-4">
                                 <div>
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Base Model <span class="text-red-500">*</span></label>
-                                    <div class="relative">
-                                        <input v-model="trainingParams.model" required class="w-full bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm focus:ring-2 focus:ring-blue-500 focus:border-blue-500 outline-none transition" placeholder="Pretrained ID or Session ID">
-                                        <div class="absolute right-2 top-2 text-xs text-gray-400 space-x-1">
-                                            <span class="cursor-pointer hover:text-blue-600" @click="trainingParams.model = 'cpsam'">cpsam</span>
-                                            <span class="cursor-pointer hover:text-blue-600" @click="trainingParams.model = 'cyto3'">cyto3</span>
-                                        </div>
-                                    </div>
+                                    <select v-model="trainingParams.model" required class="w-full bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm">
+                                        <optgroup label="Pretrained Models">
+                                            <option v-for="m in pretrainedModels" :key="`train-${m}`" :value="m">[[ m ]]</option>
+                                        </optgroup>
+                                        <optgroup label="Your Completed Sessions" v-if="Object.values(sessions).some(s => s.status_type === 'completed')">
+                                            <option
+                                                v-for="(s, id) in sessions"
+                                                :key="id"
+                                                :value="id"
+                                                v-show="s.status_type === 'completed'"
+                                            >
+                                                ⚫ [[ id.slice(0,8) ]]
+                                            </option>
+                                        </optgroup>
+                                    </select>
                                 </div>
 
                                 <div v-if="trainingParams.dataset_source !== 'BioImage Archive'" class="md:col-span-2 grid grid-cols-1 md:grid-cols-2 gap-4 pt-2">

From cd47449f212ab6c2da6078c7815b926d66b6fe55 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Sat, 21 Feb 2026 18:44:25 +0100
Subject: [PATCH 09/13] Fix BioImage Archive dataset-source training flow

---
 bioengine_apps/cellpose_finetuning/index.html | 905 ++++++++++++++++--
 bioengine_apps/cellpose_finetuning/main.py    | 176 +++-
 .../tests/test_metadata_and_glob.py           |  60 +-
 3 files changed, 971 insertions(+), 170 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 5895f907..216856dd 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -318,7 +318,7 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                             <input 
                                                 v-model="trainingParams.artifact_input" 
                                                 @input="handleArtifactInput"
-                                                @focus="handleArtifactInput"
+                                                @focus="handleArtifactFocus"
                                                 @blur="closeSuggestions"
                                                 @keydown.tab.prevent="acceptSuggestion"
                                                 required 
@@ -326,7 +326,9 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                                     'flex-grow rounded-lg px-3 py-2 text-sm outline-none transition border focus:ring-2',
                                                     artifactInputError
                                                         ? 'bg-red-50 border-red-500 text-red-700 placeholder-red-300 focus:border-red-500 focus:ring-red-300'
-                                                        : 'bg-gray-50 border-gray-300 focus:border-blue-500 focus:ring-blue-500'
+                                                        : (artifactIdExists && !artifactLookupPending
+                                                            ? 'bg-green-50 border-green-400 text-green-900 focus:border-green-500 focus:ring-green-400'
+                                                            : 'bg-gray-50 border-gray-300 focus:border-blue-500 focus:ring-blue-500')
                                                 ]"
                                                 placeholder="workspace/dataset-alias"
                                             >
@@ -354,29 +356,47 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                             </div>
                                         </div>
                                         <p v-if="artifactInputError" class="text-xs text-red-600">[[ artifactInputError ]]</p>
+                                        <p v-else-if="artifactLookupPending" class="text-xs text-blue-600">Checking artifact…</p>
+                                        <p v-else-if="artifactIdExists" class="text-xs text-green-700 flex items-center gap-1"><i class="fas fa-check-circle"></i> Artifact found</p>
                                     </div>
                                 </div>
 
-                                <div v-if="trainingParams.dataset_source === 'Upload'">
+                                <div v-if="trainingParams.dataset_source === 'Upload'" class="md:col-span-2">
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Upload Data</label>
-                                    <div class="flex gap-2">
-                                        <div v-if="trainingParams.uploaded_artifact" class="flex-grow bg-green-50 border border-green-200 rounded-lg px-3 py-2 text-sm text-green-800 flex items-center">
-                                            <i class="fas fa-check-circle mr-2"></i> [[ trainingParams.uploaded_artifact ]]
+                                    <div class="flex items-center gap-2">
+                                        <div v-if="trainingParams.uploaded_artifact" class="flex-grow min-w-0 h-10 bg-green-50 border border-green-200 rounded-lg px-3 text-sm text-green-800 flex items-center">
+                                            <i class="fas fa-check-circle mr-2 shrink-0"></i>
+                                            <span class="truncate" :title="trainingParams.uploaded_artifact">[[ trainingParams.uploaded_artifact ]]</span>
                                         </div>
-                                        <div v-else class="flex-grow bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm text-gray-400 italic">
+                                        <div v-else class="flex-grow min-w-0 h-10 bg-gray-50 border border-gray-300 rounded-lg px-3 text-sm text-gray-400 italic flex items-center">
                                             No dataset uploaded yet
                                         </div>
-                                        <button type="button" @click="showUploadModal = true" class="px-4 py-2 bg-green-600 text-white rounded-lg hover:bg-green-700 transition shadow-sm font-medium whitespace-nowrap">
+                                        <button type="button" @click="showUploadModal = true" class="h-10 px-4 inline-flex items-center bg-green-600 text-white rounded-lg hover:bg-green-700 transition shadow-sm font-medium whitespace-nowrap">
                                             <i class="fas fa-cloud-upload-alt mr-1"></i> Upload Folder
                                         </button>
+                                        <button
+                                            type="button"
+                                            @click="openFileBrowser"
+                                            :disabled="!trainingParams.uploaded_artifact"
+                                            :class="[
+                                                'h-10 px-4 inline-flex items-center border rounded-lg transition shadow-sm font-medium whitespace-nowrap',
+                                                trainingParams.uploaded_artifact
+                                                    ? 'bg-slate-100 border-slate-300 hover:bg-slate-200 text-slate-700'
+                                                    : 'bg-slate-100 border-slate-200 text-slate-300 cursor-not-allowed'
+                                            ]"
+                                            :title="trainingParams.uploaded_artifact ? 'Browse uploaded dataset files' : 'Upload a dataset first'"
+                                        >
+                                            <i class="fas fa-folder-open mr-1"></i> Browse Files
+                                        </button>
                                     </div>
+                                    <p class="mt-1 text-xs text-gray-500">Use Browse Files to pick image/annotation folders, same as Artifact source.</p>
                                 </div>
 
                                 <div v-if="trainingParams.dataset_source === 'BioImage Archive'">
                                     <label class="block text-sm font-medium text-gray-700 mb-1">Archive URL <span class="text-red-500">*</span></label>
                                     <div class="flex gap-2">
                                         <input v-model="trainingParams.bioimage_archive_url" required class="flex-grow bg-gray-50 border border-gray-300 rounded-lg px-3 py-2 text-sm focus:ring-2 focus:ring-blue-500 focus:border-blue-500 outline-none transition" placeholder="https://www.ebi.ac.uk/biostudies/bioimages/studies/...">
-                                        <a href="https://www.ebi.ac.uk/biostudies/bioimages" target="_blank" rel="noopener noreferrer" class="px-4 py-2 bg-slate-100 border border-slate-300 rounded-lg hover:bg-slate-200 text-slate-700 transition text-sm whitespace-nowrap">
+                                        <a href="https://beta.bioimagearchive.org/bioimage-archive/galleries/ai/ai-ready-studies" target="_blank" rel="noopener noreferrer" class="px-4 py-2 bg-slate-100 border border-slate-300 rounded-lg hover:bg-slate-200 text-slate-700 transition text-sm whitespace-nowrap">
                                             <i class="fas fa-external-link-alt mr-1"></i> Search EBI
                                         </a>
                                     </div>
@@ -396,13 +416,53 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                                 :value="id"
                                                 v-show="s.status_type === 'completed'"
                                             >
-                                                ⚫ [[ id.slice(0,8) ]]
+                                                ✅ [[ id.slice(0,8) ]]
                                             </option>
                                         </optgroup>
                                     </select>
                                 </div>
 
                                 <div v-if="trainingParams.dataset_source !== 'BioImage Archive'" class="md:col-span-2 grid grid-cols-1 md:grid-cols-2 gap-4 pt-2">
+                                     <div>
+                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Train Images</label>
+                                        <input v-model="trainingParams.train_images" @input="markPathFieldEdited('train_images')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. images/*/*.tif">
+                                        <div class="mt-1 text-xs text-gray-500 flex flex-wrap items-center gap-2">
+                                            <span>Examples:</span>
+                                            <button type="button" @click="applyTrainingGlobExample" class="text-blue-600 hover:underline">images/*/*.tif</button>
+                                            <button type="button" @click="applyTrainingFolderExample" class="text-blue-600 hover:underline">images/ + annotations/</button>
+                                            <button type="button" @click="validateTrainingPaths(true)" :disabled="isPathValidating" class="text-blue-600 hover:underline disabled:opacity-50">
+                                                [[ isPathValidating ? 'Checking…' : 'Validate paths' ]]
+                                            </button>
+                                        </div>
+                                    </div>
+                                    <div>
+                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Train Annotations</label>
+                                        <input v-model="trainingParams.train_annotations" @input="markPathFieldEdited('train_annotations')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. annotations/*/*_mask.ome.tif">
+                                    </div>
+                                    <div v-if="showAutoDetectButton" class="md:col-span-2 rounded-md border border-blue-200 bg-blue-50 px-3 py-2 flex items-center justify-between gap-3">
+                                        <div class="text-xs text-blue-900">
+                                            <i class="fas fa-magic mr-1"></i>
+                                            Auto-detect paths by scanning this artifact
+                                        </div>
+                                        <button type="button" @click="runPathAutoDetection(true)" :disabled="isPathDetecting" class="px-3 py-1.5 rounded bg-blue-600 text-white text-xs font-medium hover:bg-blue-700 disabled:opacity-50">
+                                            [[ isPathDetecting ? 'Detecting…' : 'Auto-detect paths' ]]
+                                        </button>
+                                    </div>
+                                    <div v-if="pathAutoError" class="md:col-span-2 rounded-md border border-red-200 bg-red-50 px-3 py-2 text-xs text-red-800">
+                                        <i class="fas fa-exclamation-triangle mr-1"></i> [[ pathAutoError ]]
+                                    </div>
+                                    <div v-else-if="pathValidationMessage" :class="[
+                                        'md:col-span-2 rounded-md border px-3 py-2 text-xs',
+                                        pathValidationOk ? 'border-emerald-200 bg-emerald-50 text-emerald-900' : 'border-red-200 bg-red-50 text-red-800'
+                                    ]">
+                                        <i :class="pathValidationOk ? 'fas fa-check-circle mr-1' : 'fas fa-exclamation-triangle mr-1'"></i>
+                                        [[ pathValidationMessage ]]
+                                    </div>
+                                    <div class="md:col-span-2">
+                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Metadata Folder (Opt)</label>
+                                        <input v-model="trainingParams.metadata_dir" @input="markPathFieldEdited('metadata_dir')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. metadata/">
+                                        <p class="text-xs text-gray-500 mt-1">Use this to load image/annotation pairs from metadata JSON files; then explicit annotation masks are optional.</p>
+                                    </div>
                                     <div class="md:col-span-2">
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-2">Train/Test Split</label>
                                         <div class="flex items-center gap-5 text-sm text-gray-700">
@@ -420,32 +480,14 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                             </div>
                                         </div>
                                         <p class="mt-1 text-xs text-gray-500" v-if="trainingParams.split_mode === 'auto'">Uses train paths only and splits into train/test (default 0.8).</p>
-                                    </div>
-                                     <div>
-                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Train Images</label>
-                                        <input v-model="trainingParams.train_images" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. images/*/*.tif">
-                                        <div class="mt-1 text-xs text-gray-500 flex flex-wrap items-center gap-2">
-                                            <span>Examples:</span>
-                                            <button type="button" @click="applyTrainingGlobExample" class="text-blue-600 hover:underline">images/**/*.tif</button>
-                                            <button type="button" @click="applyTrainingFolderExample" class="text-blue-600 hover:underline">images/ + annotations/</button>
-                                        </div>
-                                    </div>
-                                    <div>
-                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Train Annotations</label>
-                                        <input v-model="trainingParams.train_annotations" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. annotations/*/*_mask.ome.tif">
-                                    </div>
-                                    <div class="md:col-span-2">
-                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Metadata Folder (Opt)</label>
-                                        <input v-model="trainingParams.metadata_dir" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. metadata/">
-                                        <p class="text-xs text-gray-500 mt-1">Use this to load image/annotation pairs from metadata JSON files; then explicit annotation masks are optional.</p>
                                     </div>
                                     <div v-if="trainingParams.split_mode === 'manual'">
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Images (Opt)</label>
-                                        <input v-model="trainingParams.test_images" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
+                                        <input v-model="trainingParams.test_images" @input="markPathFieldEdited('test_images')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
                                     </div>
                                     <div v-if="trainingParams.split_mode === 'manual'">
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Annotations (Opt)</label>
-                                        <input v-model="trainingParams.test_annotations" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
+                                        <input v-model="trainingParams.test_annotations" @input="markPathFieldEdited('test_annotations')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
                                     </div>
                                 </div>
                                 <div v-else class="md:col-span-2 rounded-lg border border-blue-200 bg-blue-50 px-4 py-3 text-sm text-blue-900">
@@ -476,7 +518,7 @@ <h3 class="text-lg font-semibold text-gray-800 mb-4 flex items-center gap-2"><i
 
                          <div class="flex justify-end gap-4">
                              <button type="button" @click="currentTab = 'dashboard'" class="px-6 py-2.5 text-gray-600 font-medium hover:bg-gray-100 rounded-lg transition">Cancel</button>
-                             <button type="submit" :disabled="isStartingTraining" class="bg-blue-600 text-white px-8 py-2.5 rounded-lg shadow-sm hover:bg-blue-700 hover:shadow-md transition-all font-medium disabled:opacity-50 flex items-center gap-2">
+                             <button type="submit" :disabled="isTrainingStartDisabled" class="bg-blue-600 text-white px-8 py-2.5 rounded-lg shadow-sm hover:bg-blue-700 hover:shadow-md transition-all font-medium disabled:opacity-50 flex items-center gap-2">
                                 <i v-if="isStartingTraining" class="fas fa-circle-notch fa-spin"></i>
                                 [[ isStartingTraining ? 'Starting...' : 'Start Training' ]]
                             </button>
@@ -784,7 +826,18 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     </div>
                     <div>
                         <label class="block text-sm font-medium text-gray-700 mb-1">Select Folder</label>
-                        <input type="file" ref="folderInput" @change="handleFolderSelect" webkitdirectory directory multiple required class="w-full text-sm border border-gray-300 rounded-lg p-2 bg-gray-50">
+                        <input id="upload-folder-input" type="file" ref="folderInput" @change="handleFolderSelect" webkitdirectory directory multiple required class="hidden">
+                        <label for="upload-folder-input" class="inline-flex items-center gap-2 px-4 py-2 border border-gray-300 bg-gray-50 hover:bg-gray-100 text-gray-700 rounded-lg text-sm font-medium cursor-pointer transition">
+                            <i class="fas fa-folder-open text-gray-500"></i>
+                            Choose Folder
+                        </label>
+                        <p class="text-xs text-gray-600 mt-2">
+                            <span v-if="selectedFileCount > 0">
+                                <i class="fas fa-check-circle text-green-600"></i>
+                                [[ selectedFolderName ]] ([[ selectedFileCount ]] files)
+                            </span>
+                            <span v-else class="text-gray-400">No folder selected</span>
+                        </p>
                         <p class="text-xs text-gray-500 mt-1.5"><i class="fas fa-info-circle"></i> Folder should contain images and mask files.</p>
                     </div>
                     
@@ -872,10 +925,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const uploadProgress = ref(0);
                 const uploadStatus = ref('');
                 const folderInput = ref(null);
+                const selectedFolderName = ref('');
+                const selectedFileCount = ref(0);
 
                 // Inference Data
                 const inferenceParams = reactive({ model: 'cpsam', diameter: null });
-                const pretrainedModels = ['cpsam', 'cyto3', 'nuclei', 'tissuenet'];
+                const pretrainedModels = ['cpsam'];
                 const inferenceFile = ref(null);
                 const inferencePreview = ref(null);
                 const isInferring = ref(false);
@@ -897,6 +952,23 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const browserPath = ref(''); 
                 const browserLoading = ref(false);
                 const browserError = ref('');
+                const isPathDetecting = ref(false);
+                const isPathValidating = ref(false);
+                const pathAutoHint = ref('');
+                const pathValidationMessage = ref('');
+                const pathValidationOk = ref(false);
+                const pathFieldsDirty = reactive({
+                    train_images: false,
+                    train_annotations: false,
+                    metadata_dir: false,
+                    test_images: false,
+                    test_annotations: false,
+                });
+                const lastAutoPathArtifact = ref('');
+                const autoDetectedPathsPristine = ref(false);
+                let continuousValidationTimer = null;
+
+                const SUPPORTED_IMAGE_EXT_RE = /\.(ome\.)?(tif|tiff|png|jpg|jpeg)$/i;
 
 
                 // Autocomplete Logic
@@ -1021,6 +1093,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const queryWorkspaceArtifacts = async (workspaceId, queryText) => {
                     const query = String(queryText || '').trim();
                     const keywords = query ? [query] : [];
+                    const pageLimit = 200;
+                    const maxArtifacts = 2000;
                     let am = null;
                     let lastAttemptError = null;
                     try {
@@ -1033,34 +1107,58 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         }
                         return [];
                     }
-                    
+
+                    const normalizeEntries = (rawItems) => rawItems
+                        .map((item) => {
+                            const entry = normalizeArtifactEntry(item);
+                            if (!entry) return null;
+                            if (entry.workspace) return entry;
+                            return {
+                                workspace: workspaceId,
+                                alias: entry.alias,
+                                fullId: `${workspaceId}/${entry.alias}`,
+                            };
+                        })
+                        .filter(Boolean);
+
+                    const fetchPaged = async (buildAttempt) => {
+                        const collected = [];
+                        const seen = new Set();
+
+                        for (let offset = 0; offset < maxArtifacts; offset += pageLimit) {
+                            const listed = await buildAttempt(offset);
+                            const rawItems = normalizeArtifactListResponse(listed);
+                            const normalized = normalizeEntries(rawItems);
+
+                            if (normalized.length === 0) break;
+
+                            for (const entry of normalized) {
+                                const id = entry.fullId;
+                                if (!id || seen.has(id)) continue;
+                                seen.add(id);
+                                collected.push(entry);
+                            }
+
+                            if (normalized.length < pageLimit) break;
+                            if (collected.length >= maxArtifacts) break;
+                        }
+
+                        return collected;
+                    };
+
                     const attempts = [
-                        () => am.list({ keywords, limit: 30, _rkwargs: true }),
-                        () => am.list({ keywords, limit: 30 }),
-                        () => am.list(workspaceId, { keywords, limit: 30 }),
-                        () => am.list(workspaceId, { keywords }),
+                        (offset) => am.list({ keywords, limit: pageLimit, offset, stage: 'all', _rkwargs: true }),
+                        (offset) => am.list({ keywords, limit: pageLimit, offset, stage: 'all' }),
+                        (offset) => am.list({ keywords, limit: pageLimit, offset }),
                     ];
 
+                    let normalizedResult = [];
                     for (const attempt of attempts) {
                         try {
-                            const listed = await attempt();
-                            const rawItems = normalizeArtifactListResponse(listed);
-                            const normalized = rawItems
-                                .map((item) => {
-                                    const entry = normalizeArtifactEntry(item);
-                                    if (!entry) return null;
-                                    if (entry.workspace) return entry;
-                                    return {
-                                        workspace: workspaceId,
-                                        alias: entry.alias,
-                                        fullId: `${workspaceId}/${entry.alias}`,
-                                    };
-                                })
-                                .filter(Boolean);
-
-                            artifactDebug.cacheSize = normalized.length;
+                            normalizedResult = await fetchPaged(attempt);
+                            artifactDebug.cacheSize = normalizedResult.length;
                             artifactDebug.lastError = '';
-                            return normalized;
+                            break;
                         } catch (error) {
                             lastAttemptError = error;
                             artifactDebug.lastError = String(error?.message || error);
@@ -1070,6 +1168,33 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         }
                     }
 
+                    if (query && normalizedResult.length === 0) {
+                        const broadAttempts = [
+                            (offset) => am.list({ limit: pageLimit, offset, stage: 'all', _rkwargs: true }),
+                            (offset) => am.list({ limit: pageLimit, offset, stage: 'all' }),
+                            (offset) => am.list({ limit: pageLimit, offset }),
+                        ];
+
+                        for (const attempt of broadAttempts) {
+                            try {
+                                normalizedResult = await fetchPaged(attempt);
+                                artifactDebug.cacheSize = normalizedResult.length;
+                                artifactDebug.lastError = '';
+                                break;
+                            } catch (error) {
+                                lastAttemptError = error;
+                                artifactDebug.lastError = String(error?.message || error);
+                                if (isWorkspacePermissionError(error)) {
+                                    artifactInputError.value = `Permission denied to workspace '${workspaceId}'.`;
+                                }
+                            }
+                        }
+                    }
+
+                    if (normalizedResult.length > 0) {
+                        return normalizedResult;
+                    }
+
                     if (lastAttemptError && !artifactInputError.value) {
                         const message = String(lastAttemptError?.message || lastAttemptError || '').trim();
                         artifactInputError.value = `Unable to access workspace '${workspaceId}'. ${message}`.trim();
@@ -1238,7 +1363,6 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                      if (parts.length === 2) {
                          const ws = parts[0];
                          const query = parts[1].toLowerCase();
-                         checkArtifactExists(ws, parts[1]);
                          artifactDebug.lastWorkspace = ws;
                          const remoteMatches = await queryWorkspaceArtifacts(ws, query);
                          if (requestVersion !== artifactInputVersion.value) return;
@@ -1275,6 +1399,30 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                      }
                 };
 
+                const handleArtifactFocus = async () => {
+                    const requestVersion = ++artifactInputVersion.value;
+                    const val = String(trainingParams.artifact_input || '').trim();
+
+                    if (val) {
+                        await handleArtifactInput();
+                        return;
+                    }
+
+                    artifactInputError.value = '';
+                    artifactIdExists.value = false;
+                    artifactLookupPending.value = false;
+
+                    const workspaces = await fetchWorkspaces();
+                    if (requestVersion !== artifactInputVersion.value) return;
+
+                    suggestions.value = workspaces.slice(0, 100);
+                    showSuggestions.value = suggestions.value.length > 0;
+                    artifactDebug.suggestions = suggestions.value.length;
+                    console.info('[ArtifactAutocomplete] workspace suggestions on focus', {
+                        count: suggestions.value.length,
+                    });
+                };
+
                 const selectSuggestion = (s) => {
                      // If it's a workspace (no slash), add slash
                      if (!s.includes('/')) {
@@ -1377,6 +1525,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const applyTrainingGlobExample = () => {
                     trainingParams.train_images = 'images/*/*.tif';
                     trainingParams.train_annotations = 'annotations/*/*_mask.ome.tif';
+                    markPathFieldEdited('train_images');
+                    markPathFieldEdited('train_annotations');
                 };
 
                 const getSelectedArtifactId = () => {
@@ -1386,9 +1536,482 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     return trainingParams.artifact_input;
                 };
 
+                const isCompleteArtifactId = (artifactId) => {
+                    const normalized = String(artifactId || '').trim();
+                    if (!normalized) return false;
+                    const parts = normalized.split('/');
+                    return parts.length === 2 && !!parts[0] && !!parts[1];
+                };
+
+                const isAutoDetectReady = () => {
+                    const artifactId = String(getSelectedArtifactId() || '').trim();
+                    if (!isCompleteArtifactId(artifactId)) return false;
+
+                    if (trainingParams.dataset_source === 'Artifact') {
+                        return artifactIdExists.value && !artifactLookupPending.value;
+                    }
+
+                    if (trainingParams.dataset_source === 'Upload') {
+                        return !!trainingParams.uploaded_artifact && isCompleteArtifactId(trainingParams.uploaded_artifact);
+                    }
+
+                    return false;
+                };
+
+                const hasFreshAutoDetectedPaths = computed(() => {
+                    const artifactId = String(getSelectedArtifactId() || '').trim();
+                    return !!artifactId && autoDetectedPathsPristine.value && artifactId === lastAutoPathArtifact.value;
+                });
+                const showAutoDetectButton = computed(() => isAutoDetectReady() && !hasFreshAutoDetectedPaths.value);
+                const pathAutoError = computed(() => {
+                    const msg = String(pathAutoHint.value || '').trim();
+                    if (!msg) return '';
+                    if (/auto-detect failed|no supported image|no supported mask|does not exist|permission/i.test(msg)) {
+                        return msg;
+                    }
+                    return '';
+                });
+                const isTrainingStartDisabled = computed(() => {
+                    if (isStartingTraining.value) return true;
+                    if (trainingParams.dataset_source === 'BioImage Archive') {
+                        return !String(trainingParams.bioimage_archive_url || '').trim();
+                    }
+                    return !pathValidationOk.value || isPathValidating.value;
+                });
+
                 const applyTrainingFolderExample = () => {
                     trainingParams.train_images = 'images/';
                     trainingParams.train_annotations = 'annotations/';
+                    markPathFieldEdited('train_images');
+                    markPathFieldEdited('train_annotations');
+                };
+
+                const markPathFieldEdited = (field) => {
+                    autoDetectedPathsPristine.value = false;
+                    if (pathFieldsDirty[field] === false) pathFieldsDirty[field] = true;
+                };
+
+                const resetPathDirtyFlags = () => {
+                    pathFieldsDirty.train_images = false;
+                    pathFieldsDirty.train_annotations = false;
+                    pathFieldsDirty.metadata_dir = false;
+                    pathFieldsDirty.test_images = false;
+                    pathFieldsDirty.test_annotations = false;
+                };
+
+                const hasNestedUnderRoot = (paths, root) => {
+                    const prefix = `${String(root || '').toLowerCase()}/`;
+                    return paths.some((pathStr) => {
+                        const normalized = String(pathStr || '').toLowerCase();
+                        if (!normalized.startsWith(prefix)) return false;
+                        const rest = normalized.slice(prefix.length);
+                        return rest.includes('/');
+                    });
+                };
+
+                const hasDirectFilesUnderRoot = (paths, root) => {
+                    const prefix = `${String(root || '').toLowerCase()}/`;
+                    return paths.some((pathStr) => {
+                        const normalized = String(pathStr || '').toLowerCase();
+                        if (!normalized.startsWith(prefix)) return false;
+                        const rest = normalized.slice(prefix.length);
+                        return rest.length > 0 && !rest.includes('/');
+                    });
+                };
+
+                const choosePreferredSpec = (paths, roots) => {
+                    const candidates = Array.isArray(roots) ? roots : [];
+                    for (const root of candidates) {
+                        if (hasNestedUnderRoot(paths, root)) return `${root}/*/`;
+                    }
+                    for (const root of candidates) {
+                        if (hasDirectFilesUnderRoot(paths, root)) return `${root}/*`;
+                    }
+                    return '';
+                };
+
+                const normalizeListedPath = (entry) => {
+                    const raw = typeof entry === 'string'
+                        ? entry
+                        : (entry?.name || entry?.path || entry?.key || '');
+                    return String(raw || '').replace(/^\.?\//, '').trim();
+                };
+
+                const hasFileExtension = (pathStr) => baseName(pathStr).includes('.');
+
+                const normalizeListEntry = (entry, parentDir = '') => {
+                    let normalized = normalizeListedPath(entry);
+                    if (!normalized) return null;
+
+                    const parent = String(parentDir || '').replace(/^\.?\//, '').replace(/\/+$/, '');
+                    if (parent && !normalized.startsWith(parent + '/')) {
+                        normalized = `${parent}/${normalized}`;
+                    }
+
+                    const explicitFolder =
+                        (typeof entry === 'object' && (entry?.type === 'folder' || entry?.type === 'directory')) ||
+                        normalized.endsWith('/');
+                    const inferredFolder = !hasFileExtension(normalized);
+
+                    return {
+                        path: normalized.replace(/\/+$/, ''),
+                        isFolder: explicitFolder || inferredFolder,
+                    };
+                };
+
+                const listFilesPage = async (am, artifactId, dirPath = '', limit = 1000, offset = 0) => {
+                    const argsDir = dirPath ? dirPath.replace(/\/+$/, '') : null;
+                    const attempts = [
+                        () => am.list_files({ artifact_id: artifactId, dir_path: argsDir, limit, offset, _rkwargs: true }),
+                        () => am.list_files(artifactId, argsDir, limit, offset),
+                        () => am.list_files(artifactId, argsDir),
+                        () => am.list_files(artifactId),
+                    ];
+
+                    for (const attempt of attempts) {
+                        try {
+                            const result = await attempt();
+                            if (Array.isArray(result)) return result;
+                        } catch (_error) {
+                            // try next attempt
+                        }
+                    }
+                    return [];
+                };
+
+                const listFilesPaged = async (am, artifactId, dirPath = '') => {
+                    const limit = 1000;
+                    let offset = 0;
+                    const all = [];
+
+                    while (offset <= 50000) {
+                        const page = await listFilesPage(am, artifactId, dirPath, limit, offset);
+                        if (!Array.isArray(page) || page.length === 0) break;
+                        all.push(...page);
+                        if (page.length < limit) break;
+                        offset += limit;
+                    }
+
+                    return all;
+                };
+
+                const loadArtifactFiles = async (artifactId) => {
+                    const am = await getArtifactManager();
+                    const fileSet = new Set();
+                    const folderQueue = [];
+                    const visitedFolders = new Set();
+
+                    const rootItems = await listFilesPaged(am, artifactId, '');
+                    const normalizedRoot = rootItems
+                        .map((item) => normalizeListEntry(item, ''))
+                        .filter(Boolean);
+
+                    for (const item of normalizedRoot) {
+                        if (item.isFolder) folderQueue.push(item.path);
+                        else fileSet.add(item.path);
+                    }
+
+                    while (folderQueue.length > 0 && visitedFolders.size < 500) {
+                        const folder = folderQueue.shift();
+                        if (!folder || visitedFolders.has(folder)) continue;
+                        visitedFolders.add(folder);
+
+                        const folderItems = await listFilesPaged(am, artifactId, folder);
+                        const normalizedFolder = folderItems
+                            .map((item) => normalizeListEntry(item, folder))
+                            .filter(Boolean);
+
+                        for (const item of normalizedFolder) {
+                            if (item.path === folder) continue;
+                            if (item.isFolder) {
+                                if (!visitedFolders.has(item.path)) folderQueue.push(item.path);
+                            } else {
+                                fileSet.add(item.path);
+                            }
+                        }
+                    }
+
+                    let paths = Array.from(fileSet).filter((p) => p && !p.endsWith('/'));
+
+                    if (paths.length === 0) {
+                        const artifact = await am.read(artifactId);
+                        const fallback = artifact?.manifest?.files || artifact?.files || [];
+                        paths = (Array.isArray(fallback) ? fallback : []).map(normalizeListedPath).filter(Boolean);
+                    }
+
+                    return Array.from(new Set(paths.filter((p) => !p.endsWith('/'))));
+                };
+
+                const baseName = (pathStr) => String(pathStr || '').split('/').pop() || '';
+                const parentDir = (pathStr) => {
+                    const normalized = String(pathStr || '').replace(/\/+$/, '');
+                    const idx = normalized.lastIndexOf('/');
+                    return idx >= 0 ? normalized.slice(0, idx) : '';
+                };
+
+                const isSupportedImagePath = (pathStr) => SUPPORTED_IMAGE_EXT_RE.test(String(pathStr || ''));
+
+                const isMaskPath = (pathStr) => {
+                    const value = String(pathStr || '').toLowerCase();
+                    return (
+                        value.includes('/mask') || value.includes('/masks') ||
+                        value.includes('/annotation') || value.includes('/annotations') ||
+                        value.includes('/label') || value.includes('/labels') ||
+                        /(^|[\/_-])(mask|masks|label|labels|annotation|annotations|seg|segmentation)([\/_-]|$)/.test(value)
+                    );
+                };
+
+                const scoreDir = (dir, count, keywords) => {
+                    const text = String(dir || '').toLowerCase();
+                    let score = count;
+                    if (text.includes('/train/') || text.startsWith('train/')) score += 2;
+                    for (const keyword of keywords) {
+                        if (text.includes(keyword)) score += 3;
+                    }
+                    return score;
+                };
+
+                const chooseBestDir = (files, keywords) => {
+                    if (!files || files.length === 0) return '';
+                    const counts = new Map();
+                    files.forEach((pathStr) => {
+                        const dir = parentDir(pathStr);
+                        counts.set(dir, (counts.get(dir) || 0) + 1);
+                    });
+
+                    let bestDir = '';
+                    let bestScore = -Infinity;
+                    for (const [dir, count] of counts.entries()) {
+                        const score = scoreDir(dir, count, keywords);
+                        if (score > bestScore) {
+                            bestScore = score;
+                            bestDir = dir;
+                        }
+                    }
+                    return bestDir;
+                };
+
+                const globToRegex = (pattern) => {
+                    const normalized = String(pattern || '').replace(/^\.?\//, '').trim();
+                    const escaped = normalized
+                        .replace(/[.+^${}()|[\]\\]/g, '\\$&')
+                        .replace(/\*\*/g, '::DOUBLE_STAR::')
+                        .replace(/\*/g, '[^/]*')
+                        .replace(/\?/g, '.')
+                        .replace(/::DOUBLE_STAR::/g, '.*');
+                    return new RegExp(`^${escaped}$`, 'i');
+                };
+
+                const fileMatcherFromSpec = (spec) => {
+                    const value = String(spec || '').trim().replace(/^\.?\//, '');
+                    if (!value) return () => false;
+                    if (value.includes('*') || value.includes('?')) {
+                        const wildcardPattern = value.endsWith('/') ? `${value}**` : value;
+                        const regex = globToRegex(wildcardPattern);
+                        return (pathStr) => regex.test(pathStr);
+                    }
+
+                    const normalized = value.replace(/\/+$/, '');
+                    const isLikelyDirectory = value.endsWith('/') || !baseName(normalized).includes('.');
+                    if (isLikelyDirectory) {
+                        const prefix = normalized ? `${normalized}/` : '';
+                        return (pathStr) => String(pathStr).startsWith(prefix);
+                    }
+
+                    return (pathStr) => String(pathStr) === normalized;
+                };
+
+                const normalizePairStem = (pathStr) => {
+                    const file = baseName(pathStr).toLowerCase();
+                    const withoutExt = file.replace(/\.(ome\.)?(tif|tiff|png|jpg|jpeg)$/i, '');
+                    return withoutExt
+                        .replace(/([._-])(mask|masks|label|labels|annotation|annotations|seg|segmentation)$/i, '')
+                        .replace(/\s+/g, '');
+                };
+
+                const validatePairSpecs = (allFiles, imageSpec, annotationSpec, phaseLabel) => {
+                    const imageMatcher = fileMatcherFromSpec(imageSpec);
+                    const annotationMatcher = fileMatcherFromSpec(annotationSpec);
+                    const imageFiles = allFiles.filter((pathStr) => isSupportedImagePath(pathStr) && imageMatcher(pathStr) && !isMaskPath(pathStr));
+                    const annotationFiles = allFiles.filter((pathStr) => isSupportedImagePath(pathStr) && annotationMatcher(pathStr));
+
+                    if (imageFiles.length === 0) {
+                        throw new Error(`${phaseLabel}: no supported image files found for '${imageSpec}'.`);
+                    }
+                    if (annotationFiles.length === 0) {
+                        throw new Error(`${phaseLabel}: no supported mask/annotation files found for '${annotationSpec}'.`);
+                    }
+
+                    const annotationStemSet = new Set(annotationFiles.map((pathStr) => normalizePairStem(pathStr)));
+                    const pairedCount = imageFiles.filter((pathStr) => annotationStemSet.has(normalizePairStem(pathStr))).length;
+                    if (pairedCount === 0) {
+                        throw new Error(`${phaseLabel}: found files, but no image/mask pairs matched by filename stem.`);
+                    }
+
+                    return {
+                        imageCount: imageFiles.length,
+                        annotationCount: annotationFiles.length,
+                        pairedCount,
+                    };
+                };
+
+                const runPathAutoDetection = async (manualTrigger = false) => {
+                    const artifactId = String(getSelectedArtifactId() || '').trim();
+                    if (!isCompleteArtifactId(artifactId)) {
+                        if (manualTrigger) {
+                            pathAutoHint.value = 'Finish entering a valid artifact ID (workspace/alias) before auto-detecting paths.';
+                        }
+                        return;
+                    }
+
+                    if (!isAutoDetectReady()) {
+                        if (manualTrigger) {
+                            pathAutoHint.value = 'Auto-detect is available after the artifact ID is resolved to an existing artifact.';
+                        }
+                        return;
+                    }
+
+                    const userEditedAnyPath = Object.values(pathFieldsDirty).some(Boolean);
+                    if (userEditedAnyPath && !manualTrigger) {
+                        return;
+                    }
+
+                    isPathDetecting.value = true;
+                    pathAutoHint.value = '';
+                    autoDetectedPathsPristine.value = false;
+
+                    try {
+                        const files = await loadArtifactFiles(artifactId);
+                        const supported = files.filter((pathStr) => isSupportedImagePath(pathStr));
+                        const masks = supported.filter((pathStr) => isMaskPath(pathStr));
+                        const images = supported.filter((pathStr) => !isMaskPath(pathStr));
+
+                        if (images.length === 0 || masks.length === 0) {
+                            pathAutoHint.value = 'Auto-detect tried, but no supported image/mask files were found (supported: tif/tiff/png/jpg/jpeg).';
+                            return;
+                        }
+
+                        const trainImageDir = chooseBestDir(images, ['images', '/img', '/raw', '/train']);
+                        const trainMaskDir = chooseBestDir(masks, ['annotations', '/masks', '/labels', '/train']);
+                        const preferredTrainImages = choosePreferredSpec(images, ['images']);
+                        const preferredTrainAnnotations = choosePreferredSpec(masks, ['annotations', 'masks', 'labels']);
+                        const testImageDir = chooseBestDir(images.filter((pathStr) => /(^|\/)test(\/|$)|(^|\/)val(\/|$)|(^|\/)validation(\/|$)/i.test(pathStr)), ['test', 'val', 'validation']);
+                        const testMaskDir = chooseBestDir(masks.filter((pathStr) => /(^|\/)test(\/|$)|(^|\/)val(\/|$)|(^|\/)validation(\/|$)/i.test(pathStr)), ['test', 'val', 'validation']);
+
+                        trainingParams.train_images = preferredTrainImages || (trainImageDir ? `${trainImageDir}/` : '');
+                        trainingParams.train_annotations = preferredTrainAnnotations || (trainMaskDir ? `${trainMaskDir}/` : '');
+
+                        if (trainingParams.split_mode === 'manual' && testImageDir && testMaskDir) {
+                            trainingParams.test_images = `${testImageDir}/`;
+                            trainingParams.test_annotations = `${testMaskDir}/`;
+                        }
+
+                        resetPathDirtyFlags();
+                        autoDetectedPathsPristine.value = true;
+                        pathAutoHint.value = '';
+                        lastAutoPathArtifact.value = artifactId;
+                        await validateTrainingPaths(true);
+                    } catch (error) {
+                        autoDetectedPathsPristine.value = false;
+                        pathAutoHint.value = `Auto-detect failed: ${String(error?.message || error)}`;
+                    } finally {
+                        isPathDetecting.value = false;
+                    }
+                };
+
+                const validateTrainingPaths = async (showSuccess = true) => {
+                    const isBioImageArchive = trainingParams.dataset_source === 'BioImage Archive';
+                    if (isBioImageArchive) {
+                        pathValidationOk.value = true;
+                        pathValidationMessage.value = '';
+                        return true;
+                    }
+
+                    const artifactId = String(getSelectedArtifactId() || '').trim();
+                    if (!artifactId) {
+                        pathValidationOk.value = false;
+                        pathValidationMessage.value = showSuccess ? 'Select an artifact first.' : '';
+                        return false;
+                    }
+
+                    if (!isCompleteArtifactId(artifactId) || !isAutoDetectReady()) {
+                        pathValidationOk.value = false;
+                        pathValidationMessage.value = showSuccess
+                            ? 'Finish entering a valid, existing artifact ID before validating paths.'
+                            : '';
+                        return false;
+                    }
+
+                    const hasMetadataDir = !!String(trainingParams.metadata_dir || '').trim();
+                    const hasTrainImages = !!String(trainingParams.train_images || '').trim();
+                    const hasTrainAnnotations = !!String(trainingParams.train_annotations || '').trim();
+                    const hasTestImages = !!String(trainingParams.test_images || '').trim();
+                    const hasTestAnnotations = !!String(trainingParams.test_annotations || '').trim();
+
+                    if (!hasMetadataDir && (!hasTrainImages || !hasTrainAnnotations)) {
+                        pathValidationOk.value = false;
+                        pathValidationMessage.value = showSuccess ? 'Provide either metadata folder, or both train image and train annotation paths.' : '';
+                        return false;
+                    }
+
+                    if (trainingParams.split_mode === 'manual' && (hasTestImages !== hasTestAnnotations)) {
+                        pathValidationOk.value = false;
+                        pathValidationMessage.value = showSuccess ? 'For manual split, provide both test image and test annotation paths (or leave both empty).' : '';
+                        return false;
+                    }
+
+                    isPathValidating.value = true;
+                    pathValidationMessage.value = '';
+
+                    try {
+                        const files = await loadArtifactFiles(artifactId);
+                        const supportedFileCount = files.filter((pathStr) => isSupportedImagePath(pathStr)).length;
+
+                        if (!hasMetadataDir) {
+                            const trainResult = validatePairSpecs(files, trainingParams.train_images, trainingParams.train_annotations, 'Train set');
+                            let message = `Train pairs found: ${trainResult.pairedCount} (images: ${trainResult.imageCount}, masks: ${trainResult.annotationCount}).`;
+
+                            const hasManualTestPaths = trainingParams.split_mode === 'manual' &&
+                                String(trainingParams.test_images || '').trim() &&
+                                String(trainingParams.test_annotations || '').trim();
+
+                            if (hasManualTestPaths) {
+                                const testResult = validatePairSpecs(files, trainingParams.test_images, trainingParams.test_annotations, 'Test set');
+                                message += ` Test pairs found: ${testResult.pairedCount}.`;
+                            }
+
+                            pathValidationOk.value = true;
+                            pathValidationMessage.value = `${message} Supported files in artifact: ${supportedFileCount}.`;
+                        } else {
+                            if (supportedFileCount === 0) {
+                                throw new Error('No supported image/mask files found in artifact (supported: tif/tiff/png/jpg/jpeg).');
+                            }
+                            pathValidationOk.value = true;
+                            pathValidationMessage.value = `Metadata mode enabled. Found ${supportedFileCount} supported files in artifact.`;
+                        }
+
+                        return true;
+                    } catch (error) {
+                        pathValidationOk.value = false;
+                        pathValidationMessage.value = String(error?.message || error);
+                        return false;
+                    } finally {
+                        isPathValidating.value = false;
+                    }
+                };
+
+                const scheduleContinuousPathValidation = () => {
+                    if (continuousValidationTimer) {
+                        clearTimeout(continuousValidationTimer);
+                    }
+                    continuousValidationTimer = setTimeout(async () => {
+                        if (trainingParams.dataset_source === 'BioImage Archive') {
+                            pathValidationOk.value = true;
+                            pathValidationMessage.value = '';
+                            return;
+                        }
+                        await validateTrainingPaths(true);
+                    }, 250);
                 };
 
                 const getUserWorkspace = () => {
@@ -1713,6 +2336,77 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     });
                 });
 
+                watch(
+                    () => trainingParams.split_mode,
+                    (mode) => {
+                        if (mode === 'auto') {
+                            trainingParams.test_images = '';
+                            trainingParams.test_annotations = '';
+                        }
+                        autoDetectedPathsPristine.value = false;
+                        scheduleContinuousPathValidation();
+                    }
+                );
+
+                watch(
+                    () => [trainingParams.dataset_source, trainingParams.artifact_input, trainingParams.uploaded_artifact, artifactIdExists.value, artifactLookupPending.value],
+                    async () => {
+                        const artifactId = String(getSelectedArtifactId() || '').trim();
+                        if (artifactId !== lastAutoPathArtifact.value) {
+                            autoDetectedPathsPristine.value = false;
+                        }
+                        if (trainingParams.dataset_source === 'BioImage Archive') return;
+                        if (!isCompleteArtifactId(artifactId)) return;
+                        if (!isAutoDetectReady()) return;
+                        if (artifactId === lastAutoPathArtifact.value) return;
+                        await runPathAutoDetection(false);
+                    }
+                );
+
+                watch(
+                    () => [trainingParams.dataset_source, trainingParams.artifact_input],
+                    async () => {
+                        if (trainingParams.dataset_source !== 'Artifact') return;
+
+                        const artifactId = String(trainingParams.artifact_input || '').trim();
+                        const parts = artifactId.split('/');
+                        if (parts.length !== 2) {
+                            artifactIdExists.value = false;
+                            artifactLookupPending.value = false;
+                            return;
+                        }
+
+                        const workspaceId = String(parts[0] || '').trim();
+                        const alias = String(parts[1] || '').trim();
+                        if (!workspaceId || !alias) {
+                            artifactIdExists.value = false;
+                            artifactLookupPending.value = false;
+                            return;
+                        }
+
+                        await checkArtifactExists(workspaceId, alias);
+                    }
+                );
+
+                watch(
+                    () => [
+                        trainingParams.dataset_source,
+                        trainingParams.artifact_input,
+                        trainingParams.uploaded_artifact,
+                        trainingParams.train_images,
+                        trainingParams.train_annotations,
+                        trainingParams.metadata_dir,
+                        trainingParams.test_images,
+                        trainingParams.test_annotations,
+                        trainingParams.split_mode,
+                        artifactIdExists.value,
+                        artifactLookupPending.value,
+                    ],
+                    () => {
+                        scheduleContinuousPathValidation();
+                    }
+                );
+
                 const viewSession = (id) => {
                     if (!sessions.value[id]) return;
                     selectedSessionId.value = id;
@@ -1838,46 +2532,29 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         if (!isBioImageArchive && !hasMetadataDir && (!hasTrainImages || !hasTrainAnnotations)) {
                            throw new Error("Please provide either metadata folder or both train image and annotation paths.");
                         }
+
+                        if (!isBioImageArchive) {
+                            const pathsOk = await validateTrainingPaths(true);
+                            if (!pathsOk) {
+                                throw new Error(pathValidationMessage.value || 'Training path validation failed.');
+                            }
+                        }
                         
                         // Determine the artifact value based on source
-                            let artifactValue = trainingParams.artifact_input;
+                        let artifactValue = trainingParams.artifact_input;
                         if (isBioImageArchive) {
-                            artifactValue = trainingParams.bioimage_archive_url;
+                            artifactValue = String(trainingParams.bioimage_archive_url || '').trim();
                             if (!artifactValue) throw new Error("Please enter a BioImage Archive URL.");
                         } else if (trainingParams.dataset_source === 'Upload') {
-                                artifactValue = trainingParams.uploaded_artifact;
-                             if (!artifactValue) throw new Error("Please upload a dataset first.");
+                            artifactValue = trainingParams.uploaded_artifact;
+                            if (!artifactValue) throw new Error("Please upload a dataset first.");
                         } else {
-                             if (!artifactValue) throw new Error("Please specify an Artifact ID.");
+                            if (!artifactValue) throw new Error("Please specify an Artifact ID.");
                         }
 
-                        console.log("Invoking service.start_training with:", {
-                            artifact: artifactValue,
-                            train_images: isBioImageArchive ? null : (trainingParams.train_images || null),
-                            train_annotations: isBioImageArchive ? null : (trainingParams.train_annotations || null),
-                            metadata_dir: isBioImageArchive ? null : (trainingParams.metadata_dir || null),
-                            split_mode: isBioImageArchive ? 'manual' : trainingParams.split_mode,
-                            train_split_ratio: isBioImageArchive ? 0.8 : trainingParams.train_split_ratio,
-                            test_images: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_images || null) : null,
-                            test_annotations: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_annotations || null) : null,
-                            model: trainingParams.model,
-                            n_epochs: trainingParams.n_epochs,
-                            learning_rate: trainingParams.learning_rate,
-                            min_train_masks: trainingParams.min_train_masks
-                        });
-                        
-                        console.log("Calling start_training with keyword args via _rkwargs.");
-                        
-                        const result = await service.value.start_training(
-                            {
+                        const payload = isBioImageArchive
+                            ? {
                                 artifact: artifactValue,
-                                train_images: isBioImageArchive ? null : (trainingParams.train_images || null),
-                                train_annotations: isBioImageArchive ? null : (trainingParams.train_annotations || null),
-                                metadata_dir: isBioImageArchive ? null : (trainingParams.metadata_dir || null),
-                                split_mode: isBioImageArchive ? 'manual' : (trainingParams.split_mode || 'manual'),
-                                train_split_ratio: isBioImageArchive ? 0.8 : Number(trainingParams.train_split_ratio || 0.8),
-                                test_images: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_images || null) : null,
-                                test_annotations: (!isBioImageArchive && trainingParams.split_mode === 'manual') ? (trainingParams.test_annotations || null) : null,
                                 model: trainingParams.model || 'cpsam',
                                 n_samples: null,
                                 n_epochs: trainingParams.n_epochs,
@@ -1885,9 +2562,30 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                                 weight_decay: 1e-4,
                                 min_train_masks: trainingParams.min_train_masks,
                                 validation_interval: null
-                            },
-                            { _rkwargs: true }
-                        );
+                            }
+                            : {
+                                artifact: artifactValue,
+                                train_images: trainingParams.train_images || null,
+                                train_annotations: trainingParams.train_annotations || null,
+                                metadata_dir: trainingParams.metadata_dir || null,
+                                split_mode: trainingParams.split_mode || 'manual',
+                                train_split_ratio: Number(trainingParams.train_split_ratio || 0.8),
+                                test_images: trainingParams.split_mode === 'manual' ? (trainingParams.test_images || null) : null,
+                                test_annotations: trainingParams.split_mode === 'manual' ? (trainingParams.test_annotations || null) : null,
+                                model: trainingParams.model || 'cpsam',
+                                n_samples: null,
+                                n_epochs: trainingParams.n_epochs,
+                                learning_rate: trainingParams.learning_rate,
+                                weight_decay: 1e-4,
+                                min_train_masks: trainingParams.min_train_masks,
+                                validation_interval: null
+                            };
+
+                        console.log("Invoking service.start_training with:", payload);
+                        
+                        console.log("Calling start_training with keyword args via _rkwargs.");
+                        
+                        const result = await service.value.start_training(payload, { _rkwargs: true });
 
                         console.log("start_training result:", result);
                         
@@ -1919,10 +2617,17 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const handleFolderSelect = (event) => {
                      const files = event.target.files;
                      if (files && files.length > 0) {
+                         selectedFileCount.value = files.length;
+                         const rootFolder = String(files[0].webkitRelativePath || '').split('/')[0] || '';
+                         selectedFolderName.value = rootFolder || 'Selected folder';
+
                          // Generate a readable alias
                          const date = new Date().toISOString().slice(0, 10).replace(/-/g, '');
                          const rand = Math.floor(Math.random() * 1000).toString().padStart(3, '0');
                          uploadParams.alias = `dataset-${date}-${rand}`;
+                     } else {
+                         selectedFileCount.value = 0;
+                         selectedFolderName.value = '';
                      }
                 };
 
@@ -1958,6 +2663,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             alias: alias,
                             workspace: getUserWorkspace(),
                             manifest: manifest,
+                            config: {
+                                permissions: {
+                                    "*": ["read", "list", "list_files", "get_file"]
+                                }
+                            },
                             type: "dataset",
                             overwrite: true,
                             stage: true,
@@ -2162,7 +2872,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             alias: tmpAlias,
                             manifest: manifest,
                             workspace: getServiceWorkspace() || getUserWorkspace(),
-                            config: { permissions: { "*": "r" } },
+                            config: {
+                                permissions: {
+                                    "*": ["read", "list", "list_files", "get_file"]
+                                }
+                            },
                             type: "temp",
                             stage: true,
                             overwrite: true,
@@ -2630,8 +3344,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     sessions, refreshSessions, refreshLoading, monitoringSessionId, monitoredSession, selectedSessionId, viewSession, currentSession, hasMetrics, hasValidationSet, stopTraining, restartTraining,
                     runningTrainingSessionsCount,
                     trainingParams, startTraining, isStartingTraining,
-                    applyTrainingGlobExample, applyTrainingFolderExample,
-                    showUploadModal, uploadParams, uploadDataset, isUploading, uploadProgress, uploadStatus, folderInput,
+                    applyTrainingGlobExample, applyTrainingFolderExample, runPathAutoDetection, validateTrainingPaths,
+                    showUploadModal, uploadParams, uploadDataset, isUploading, uploadProgress, uploadStatus, folderInput, selectedFolderName, selectedFileCount,
                     inferenceParams, pretrainedModels, inferenceFile, handleInferenceFile, inferencePreview, isInferring, runInference, inferenceCanvas, previewImg, onImageLoad, canvasStyle, inferenceObjectCount, inferenceResult,
                     exportModelName, isExporting, exportResult, exportModel,
                     
@@ -2640,7 +3354,10 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     browserVisibleItems, browserNavigate, browserNavigateUp, selectFolderFor, browserFeedback,
 
                     // Autocomplete
-                    suggestions, showSuggestions, artifactInputError, artifactIdExists, artifactLookupPending, handleArtifactInput, selectSuggestion, acceptSuggestion, closeSuggestions,
+                    suggestions, showSuggestions, artifactInputError, artifactIdExists, artifactLookupPending, handleArtifactInput, handleArtifactFocus, selectSuggestion, acceptSuggestion, closeSuggestions,
+
+                    // Path auto-detect + validation
+                    isPathDetecting, isPathValidating, pathAutoHint, pathAutoError, showAutoDetectButton, pathValidationMessage, pathValidationOk, markPathFieldEdited, isTrainingStartDisabled,
 
                     // Helpers
                     statusTextColor, formatDuration, handleFolderSelect, getLatestLoss, formatDate, getSelectedArtifactId,
diff --git a/bioengine_apps/cellpose_finetuning/main.py b/bioengine_apps/cellpose_finetuning/main.py
index c264d973..e5c807e4 100644
--- a/bioengine_apps/cellpose_finetuning/main.py
+++ b/bioengine_apps/cellpose_finetuning/main.py
@@ -8,6 +8,7 @@
 
 import asyncio
 import base64
+import csv
 import fnmatch
 import hashlib
 import io
@@ -639,6 +640,9 @@ def get_url_and_artifact_id(artifact_id: str | Any) -> tuple[str, str]:
 
     parsed = urlparse(artifact_id)
     if parsed.scheme in ("http", "https"):
+        if _is_bioimage_archive_url(artifact_id):
+            return DEFAULT_SERVER_URL, artifact_id
+
         path_parts = parsed.path.lstrip("/").split("/")
         if path_parts[1] != "artifacts":
             msg = (
@@ -1856,7 +1860,7 @@ def load_model(
 
     If `identifier` points to an existing file, it is treated as a path to a
     finetuned model. Otherwise, it is treated as a builtin model name
-    (e.g., "cyto3").
+    (e.g., "cpsam").
     """
     from cellpose import core, models  # type: ignore
 
@@ -2021,6 +2025,55 @@ def _extract_bia_accession(url: str) -> str | None:
     return match.group(1).upper() if match else None
 
 
+def _bia_ftp_base_url(accession: str) -> str:
+    accession_number = int(accession.split("S-BIAD", 1)[1])
+    suffix = accession_number % 1000
+    return f"https://ftp.ebi.ac.uk/biostudies/fire/S-BIAD/{suffix}/{accession}"
+
+
+def _extract_tsv_pairs(tsv_content: str) -> list[tuple[str, str]]:
+    pairs: list[tuple[str, str]] = []
+    reader = csv.DictReader(io.StringIO(tsv_content), delimiter="\t")
+    for row in reader:
+        label_path = str(row.get("Files") or "").strip()
+        source_image = str(row.get("Source image") or "").strip()
+        if not label_path or not source_image:
+            continue
+        pairs.append((source_image, label_path))
+    return pairs
+
+
+async def _fetch_bia_tsv_pairs(client: Any, accession: str) -> list[tuple[str, str]]:
+    base_url = _bia_ftp_base_url(accession)
+    candidate_paths = [
+        f"{base_url}/Files/ps_ovule_labels.tsv",
+        f"{base_url}/Files/labels.tsv",
+        f"{base_url}/Files/{accession}_labels.tsv",
+    ]
+
+    for tsv_url in candidate_paths:
+        try:
+            response = await client.get(tsv_url)
+            if response.status_code // 100 != 2 or not response.text.strip():
+                continue
+
+            pairs = _extract_tsv_pairs(response.text)
+            if not pairs:
+                continue
+
+            return [
+                (
+                    f"{base_url}/Files/{image_rel.lstrip('/')}",
+                    f"{base_url}/Files/{label_rel.lstrip('/')}",
+                )
+                for image_rel, label_rel in pairs
+            ]
+        except Exception:
+            continue
+
+    return []
+
+
 def _iter_string_values(payload: Any) -> list[str]:
     out: list[str] = []
     stack = [payload]
@@ -2143,50 +2196,55 @@ async def make_training_pairs_from_bioimage_archive_url(
         )
 
     async with httpx.AsyncClient(timeout=120) as client:
-        payloads = [
-            await _fetch_bia_payload(client, BIA_FTS_ENDPOINT, accession),
-            await _fetch_bia_payload(client, BIA_IMAGE_ENDPOINT, accession),
-        ]
-
-        candidates: set[str] = set()
-        for payload in payloads:
-            if payload is None:
-                continue
-            for value in _iter_string_values(payload):
-                if isinstance(value, str) and value.startswith(("http://", "https://")):
-                    if _looks_like_image_url(value):
-                        candidates.add(value)
+        paired_urls = await _fetch_bia_tsv_pairs(client, accession)
 
-        if not candidates:
-            raise ValueError(
-                "No downloadable image assets were found from BioImage Archive APIs for this accession."
-            )
+        if not paired_urls:
+            payloads = [
+                await _fetch_bia_payload(client, BIA_FTS_ENDPOINT, accession),
+                await _fetch_bia_payload(client, BIA_IMAGE_ENDPOINT, accession),
+            ]
+
+            candidates: set[str] = set()
+            for payload in payloads:
+                if payload is None:
+                    continue
+                for value in _iter_string_values(payload):
+                    if isinstance(value, str) and value.startswith(("http://", "https://")):
+                        if _looks_like_image_url(value) and accession in value.upper():
+                            candidates.add(value)
+
+            if not candidates:
+                raise ValueError(
+                    "No downloadable image assets were found from BioImage Archive for this accession."
+                )
 
-        image_urls = sorted([url for url in candidates if not _looks_like_mask_url(url)])
-        mask_urls = sorted([url for url in candidates if _looks_like_mask_url(url)])
+            image_urls = sorted([
+                url for url in candidates if not _looks_like_mask_url(url)
+            ])
+            mask_urls = sorted([url for url in candidates if _looks_like_mask_url(url)])
 
-        if not image_urls or not mask_urls:
-            raise ValueError(
-                "Found assets in BioImage Archive response, but could not identify both image and mask files."
-            )
+            if not image_urls or not mask_urls:
+                raise ValueError(
+                    "Found BioImage Archive assets, but could not identify both image and mask files."
+                )
 
-        mask_map: dict[str, str] = {}
-        for mask_url in mask_urls:
-            key = _pair_key_from_url(mask_url, is_mask=True)
-            if key and key not in mask_map:
-                mask_map[key] = mask_url
+            mask_map: dict[str, str] = {}
+            for mask_url in mask_urls:
+                key = _pair_key_from_url(mask_url, is_mask=True)
+                if key and key not in mask_map:
+                    mask_map[key] = mask_url
 
-        paired_urls: list[tuple[str, str]] = []
-        for image_url in image_urls:
-            key = _pair_key_from_url(image_url, is_mask=False)
-            mask_url = mask_map.get(key)
-            if mask_url:
-                paired_urls.append((image_url, mask_url))
+            paired_urls = []
+            for image_url in image_urls:
+                key = _pair_key_from_url(image_url, is_mask=False)
+                mask_url = mask_map.get(key)
+                if mask_url:
+                    paired_urls.append((image_url, mask_url))
 
-        if not paired_urls:
-            raise ValueError(
-                "No image/mask pairs could be inferred from BioImage Archive assets."
-            )
+            if not paired_urls:
+                raise ValueError(
+                    "No image/mask pairs could be inferred from BioImage Archive assets."
+                )
 
         if config["n_samples"] is not None and config["n_samples"] < len(paired_urls):
             subset_idx = np.random.default_rng().permutation(len(paired_urls))[ : config["n_samples"] ]
@@ -2300,9 +2358,23 @@ async def list_matching_artifact_paths(
     normalized_pattern = _normalize_artifact_relpath(path_pattern)
 
     if normalized_pattern.endswith("/"):
-        folder = normalized_pattern
-        names = await list_artifact_files(artifact, folder)
-        return [_normalize_artifact_relpath(f"{folder}{name}") for name in names]
+        # Folder-like patterns should match files recursively to stay consistent
+        # with UI validation/path suggestions.
+        if not _contains_glob(normalized_pattern):
+            return await list_artifact_files_recursive(artifact, normalized_pattern)
+
+        base_folder = _glob_base_folder(normalized_pattern)
+        candidates = await list_artifact_files_recursive(artifact, base_folder)
+
+        matched = []
+        for candidate in candidates:
+            parent = _normalize_artifact_relpath(str(Path(candidate).parent))
+            if parent and not parent.endswith("/"):
+                parent += "/"
+            if fnmatch.fnmatch(parent, normalized_pattern):
+                matched.append(candidate)
+
+        return sorted(set(matched))
 
     if not _contains_glob(normalized_pattern):
         # If caller passed a directory without trailing slash (common in UI/manual input),
@@ -3679,14 +3751,16 @@ async def start_training(
         validation_interval = normalize_optional_param(validation_interval)
 
         if metadata_dir is None:
-            if not isinstance(train_images, str) or not train_images:
-                raise ValueError(
-                    "train_images must be a non-empty string when metadata_dir is not provided"
-                )
-            if not isinstance(train_annotations, str) or not train_annotations:
-                raise ValueError(
-                    "train_annotations must be a non-empty string when metadata_dir is not provided"
-                )
+            is_bia = isinstance(artifact, str) and _is_bioimage_archive_url(artifact)
+            if not is_bia:
+                if not isinstance(train_images, str) or not train_images:
+                    raise ValueError(
+                        "train_images must be a non-empty string when metadata_dir is not provided"
+                    )
+                if not isinstance(train_annotations, str) or not train_annotations:
+                    raise ValueError(
+                        "train_annotations must be a non-empty string when metadata_dir is not provided"
+                    )
         else:
             if not isinstance(metadata_dir, str) or not metadata_dir:
                 raise ValueError("metadata_dir must be a non-empty string")
@@ -4834,7 +4908,7 @@ async def test_cellpose_finetune() -> None:
     # Reverting to original call style but casting:
     cellpose_tuner = CellposeFinetune.func_or_class()
 
-    await infer(cellpose_tuner, model="cyto3")
+    await infer(cellpose_tuner, model="cpsam")
 
     session_status = await cellpose_tuner.start_training(
         artifact="ri-scale/zarr-demo",
diff --git a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py b/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
index f10b5b86..0a9e2a44 100644
--- a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
+++ b/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
@@ -79,8 +79,8 @@ def test_match_image_annotation_pairs_with_many_nested_folders() -> None:
     pairs = match_image_annotation_pairs(
         image_files,
         annotation_files,
-        "images/**/*.tif",
-        "annotations/**/*_mask.ome.tif",
+        "images/*/*.tif",
+        "annotations/*/*_mask.ome.tif",
     )
 
     assert pairs == [
@@ -129,22 +129,22 @@ async def get(self, remote_paths, local_paths, on_error="ignore"):
 
 
 class _FakeArtifactNestedMetadata:
-        async def ls(self, folder_path: str):
-                await asyncio.sleep(0)
-                folder = folder_path.rstrip("/") + "/"
-                if folder == "metadata/":
-                        return [{"path": "metadata/records.json", "type": "file"}]
-                return []
-
-        async def get(self, remote_paths, local_paths, on_error="ignore"):
-                await asyncio.sleep(0)
-                for remote, local in zip(remote_paths, local_paths):
-                        local_path = Path(local)
-                        local_path.parent.mkdir(parents=True, exist_ok=True)
-
-                        if str(remote).endswith("metadata/records.json"):
-                                local_path.write_text(
-                                        """
+    async def ls(self, folder_path: str):
+        await asyncio.sleep(0)
+        folder = folder_path.rstrip("/") + "/"
+        if folder == "metadata/":
+            return [{"path": "metadata/records.json", "type": "file"}]
+        return []
+
+    async def get(self, remote_paths, local_paths, on_error="ignore"):
+        await asyncio.sleep(0)
+        for remote, local in zip(remote_paths, local_paths):
+            local_path = Path(local)
+            local_path.parent.mkdir(parents=True, exist_ok=True)
+
+            if str(remote).endswith("metadata/records.json"):
+                local_path.write_text(
+                    """
 {
     "payload": {
         "items": [
@@ -162,10 +162,10 @@ async def get(self, remote_paths, local_paths, on_error="ignore"):
     }
 }
                                         """.strip(),
-                                        encoding="utf-8",
-                                )
-                        else:
-                                local_path.write_bytes(b"dummy")
+                    encoding="utf-8",
+                )
+            else:
+                local_path.write_bytes(b"dummy")
 
 
 class _FakeArtifactAmbiguousDirs:
@@ -246,9 +246,17 @@ def test_make_training_pairs_from_nested_metadata_formats(tmp_path: Path) -> Non
     assert len(train_pairs) == 1
     assert len(test_pairs) == 1
     assert train_pairs[0]["image"].as_posix().endswith("images/train/t0010.tif")
-    assert train_pairs[0]["annotation"].as_posix().endswith("annotations/train/t0010_mask.ome.tif")
+    assert (
+        train_pairs[0]["annotation"]
+        .as_posix()
+        .endswith("annotations/train/t0010_mask.ome.tif")
+    )
     assert test_pairs[0]["image"].as_posix().endswith("images/test/t0011.tif")
-    assert test_pairs[0]["annotation"].as_posix().endswith("annotations/test/t0011_mask.ome.tif")
+    assert (
+        test_pairs[0]["annotation"]
+        .as_posix()
+        .endswith("annotations/test/t0011_mask.ome.tif")
+    )
 
 
 def test_extract_bia_accession_from_gallery_url() -> None:
@@ -259,4 +267,6 @@ def test_extract_bia_accession_from_gallery_url() -> None:
 def test_pair_key_from_url_handles_mask_suffix() -> None:
     img = "https://example.org/data/images/folder1/t0001.ome.tif"
     ann = "https://example.org/data/annotations/folder1/t0001_mask.ome.tif"
-    assert _pair_key_from_url(img, is_mask=False) == _pair_key_from_url(ann, is_mask=True)
+    assert _pair_key_from_url(img, is_mask=False) == _pair_key_from_url(
+        ann, is_mask=True
+    )

From f5c6f2f6db4959b84360e1b7937176a0ab78277c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Sun, 22 Feb 2026 13:51:10 +0100
Subject: [PATCH 10/13] Add sample-percentage controls and ratio-based
 subsampling for Cellpose finetuning

---
 bioengine_apps/cellpose_finetuning/index.html | 214 ++++++++++++++++--
 bioengine_apps/cellpose_finetuning/main.py    | 202 +++++++++++++----
 2 files changed, 353 insertions(+), 63 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 216856dd..07401987 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -463,10 +463,40 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                         <input v-model="trainingParams.metadata_dir" @input="markPathFieldEdited('metadata_dir')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. metadata/">
                                         <p class="text-xs text-gray-500 mt-1">Use this to load image/annotation pairs from metadata JSON files; then explicit annotation masks are optional.</p>
                                     </div>
-                                    <div class="md:col-span-2">
+                                </div>
+                                <div v-else class="md:col-span-2 rounded-lg border border-blue-200 bg-blue-50 px-4 py-3 text-sm text-blue-900">
+                                    BioImage Archive mode discovers image/mask pairs from the archive URL automatically.
+                                </div>
+
+                                <div class="md:col-span-2 rounded-lg border border-gray-200 bg-gray-50 p-4 space-y-4">
+                                    <div>
+                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-2">Total Samples Used (Train + Validation)</label>
+                                        <div class="grid grid-cols-1 md:grid-cols-3 gap-3 items-center">
+                                            <input v-model.number="sampleUsagePercent" type="range" min="1" max="100" step="1" class="w-full md:col-span-2 accent-blue-600">
+                                            <div class="flex gap-2">
+                                                <div class="flex-1">
+                                                    <label class="block text-[11px] text-gray-500 mb-1">Percent</label>
+                                                    <div class="relative">
+                                                        <input v-model.number="sampleUsagePercent" type="number" min="1" max="100" step="1" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 pr-7 text-sm">
+                                                        <span class="absolute right-2 top-1/2 -translate-y-1/2 text-xs text-gray-400">%</span>
+                                                    </div>
+                                                </div>
+                                                <div class="flex-1">
+                                                    <label class="block text-[11px] text-gray-500 mb-1">Decimal</label>
+                                                    <input v-model.number="sampleUsageDecimal" type="number" min="0.01" max="1" step="0.01" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 text-sm">
+                                                </div>
+                                            </div>
+                                        </div>
+                                        <p class="mt-2 text-xs text-gray-600">
+                                            Selected samples: <span class="font-semibold">[[ selectedTotalSampleCount ?? '-' ]]</span>
+                                            <span class="text-gray-400"> / [[ availableTotalSampleCount ?? '-' ]] total</span>
+                                        </p>
+                                    </div>
+
+                                    <div>
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-2">Train/Test Split</label>
-                                        <div class="flex items-center gap-5 text-sm text-gray-700">
-                                            <label class="inline-flex items-center gap-2 cursor-pointer">
+                                        <div class="flex flex-wrap items-center gap-5 text-sm text-gray-700">
+                                            <label v-if="trainingParams.dataset_source !== 'BioImage Archive'" class="inline-flex items-center gap-2 cursor-pointer">
                                                 <input type="radio" value="manual" v-model="trainingParams.split_mode" class="text-blue-600 focus:ring-blue-500">
                                                 <span>Manual test paths</span>
                                             </label>
@@ -474,25 +504,42 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                                 <input type="radio" value="auto" v-model="trainingParams.split_mode" class="text-blue-600 focus:ring-blue-500">
                                                 <span>Auto split</span>
                                             </label>
-                                            <div v-if="trainingParams.split_mode === 'auto'" class="inline-flex items-center gap-2">
-                                                <span class="text-xs text-gray-500">Train ratio</span>
-                                                <input v-model.number="trainingParams.train_split_ratio" type="number" min="0.1" max="0.95" step="0.05" class="w-24 bg-white border border-gray-300 rounded-md px-2 py-1 text-sm">
+                                            <span v-if="trainingParams.dataset_source === 'BioImage Archive'" class="text-xs text-blue-700 bg-blue-100 border border-blue-200 px-2 py-0.5 rounded">BioImage Archive uses auto split</span>
+                                        </div>
+
+                                        <div v-if="trainingParams.split_mode === 'auto'" class="mt-3 space-y-2">
+                                            <div class="grid grid-cols-1 md:grid-cols-3 gap-3 items-center">
+                                                <input v-model.number="trainSplitPercent" type="range" min="5" max="95" step="1" class="w-full md:col-span-2 accent-blue-600">
+                                                <div class="flex gap-2">
+                                                    <div class="flex-1">
+                                                        <label class="block text-[11px] text-gray-500 mb-1">Train %</label>
+                                                        <div class="relative">
+                                                            <input v-model.number="trainSplitPercent" type="number" min="5" max="95" step="1" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 pr-7 text-sm">
+                                                            <span class="absolute right-2 top-1/2 -translate-y-1/2 text-xs text-gray-400">%</span>
+                                                        </div>
+                                                    </div>
+                                                    <div class="flex-1">
+                                                        <label class="block text-[11px] text-gray-500 mb-1">Train ratio</label>
+                                                        <input v-model.number="trainingParams.train_split_ratio" type="number" min="0.05" max="0.95" step="0.01" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 text-sm">
+                                                    </div>
+                                                </div>
                                             </div>
+                                            <p class="text-xs text-gray-600">Estimated split from selected samples: train <span class="font-semibold">[[ selectedTrainSampleCount ?? '-' ]]</span>, validation <span class="font-semibold">[[ selectedValidationSampleCount ?? '-' ]]</span>.</p>
+                                        </div>
+
+                                        <div v-else-if="trainingParams.dataset_source !== 'BioImage Archive'" class="mt-3 grid grid-cols-1 md:grid-cols-2 gap-4">
+                                            <div>
+                                                <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Images (Opt)</label>
+                                                <input v-model="trainingParams.test_images" @input="markPathFieldEdited('test_images')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
+                                            </div>
+                                            <div>
+                                                <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Annotations (Opt)</label>
+                                                <input v-model="trainingParams.test_annotations" @input="markPathFieldEdited('test_annotations')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
+                                            </div>
+                                            <p class="md:col-span-2 text-xs text-gray-600">Estimated split from selected samples: train <span class="font-semibold">[[ selectedTrainSampleCount ?? '-' ]]</span>, validation <span class="font-semibold">[[ selectedValidationSampleCount ?? '-' ]]</span>.</p>
                                         </div>
-                                        <p class="mt-1 text-xs text-gray-500" v-if="trainingParams.split_mode === 'auto'">Uses train paths only and splits into train/test (default 0.8).</p>
-                                    </div>
-                                    <div v-if="trainingParams.split_mode === 'manual'">
-                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Images (Opt)</label>
-                                        <input v-model="trainingParams.test_images" @input="markPathFieldEdited('test_images')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
-                                    </div>
-                                    <div v-if="trainingParams.split_mode === 'manual'">
-                                        <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Annotations (Opt)</label>
-                                        <input v-model="trainingParams.test_annotations" @input="markPathFieldEdited('test_annotations')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
                                     </div>
                                 </div>
-                                <div v-else class="md:col-span-2 rounded-lg border border-blue-200 bg-blue-50 px-4 py-3 text-sm text-blue-900">
-                                    BioImage Archive mode discovers image/mask pairs from the archive URL automatically, so path and split fields are not needed.
-                                </div>
                             </div>
                          </div>
                           </div>
@@ -908,6 +955,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     train_annotations: 'annotations/',
                     metadata_dir: '',
                     split_mode: 'manual',
+                    sample_usage_percent: 100,
                     train_split_ratio: 0.8,
                     test_images: '',
                     test_annotations: '',
@@ -917,6 +965,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 });
                 const isStartingTraining = ref(false);
                 const trainingError = ref('');
+                const sampleInventory = reactive({
+                    trainPairs: 0,
+                    testPairs: 0,
+                    totalPairs: 0,
+                    known: false,
+                });
 
                 // Upload Data
                 const showUploadModal = ref(false);
@@ -1456,6 +1510,87 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     Object.values(sessions.value).filter((s) => s && s.status_type === 'running').length
                 );
 
+                const clamp = (value, min, max) => Math.max(min, Math.min(max, value));
+
+                const sampleUsagePercent = computed({
+                    get: () => clamp(Number(trainingParams.sample_usage_percent || 100), 1, 100),
+                    set: (value) => {
+                        const numeric = Number(value);
+                        if (!Number.isFinite(numeric)) return;
+                        trainingParams.sample_usage_percent = clamp(Math.round(numeric), 1, 100);
+                    },
+                });
+
+                const sampleUsageDecimal = computed({
+                    get: () => Number((sampleUsagePercent.value / 100).toFixed(2)),
+                    set: (value) => {
+                        const numeric = Number(value);
+                        if (!Number.isFinite(numeric)) return;
+                        sampleUsagePercent.value = clamp(Math.round(numeric * 100), 1, 100);
+                    },
+                });
+
+                const trainSplitPercent = computed({
+                    get: () => clamp(Math.round(Number(trainingParams.train_split_ratio || 0.8) * 100), 5, 95),
+                    set: (value) => {
+                        const numeric = Number(value);
+                        if (!Number.isFinite(numeric)) return;
+                        trainingParams.train_split_ratio = clamp(Math.round(numeric), 5, 95) / 100;
+                    },
+                });
+
+                const availableTotalSampleCount = computed(() => sampleInventory.known ? sampleInventory.totalPairs : null);
+
+                const selectedTotalSampleCount = computed(() => {
+                    if (!sampleInventory.known) return null;
+                    const total = sampleInventory.totalPairs;
+                    if (total <= 0) return 0;
+                    const target = Math.round(total * (sampleUsagePercent.value / 100));
+                    return clamp(target, 1, total);
+                });
+
+                const selectedTrainSampleCount = computed(() => {
+                    if (!sampleInventory.known) return null;
+                    const selectedTotal = selectedTotalSampleCount.value;
+                    if (!selectedTotal || selectedTotal <= 0) return 0;
+
+                    if (trainingParams.split_mode === 'auto' || trainingParams.dataset_source === 'BioImage Archive') {
+                        if (selectedTotal === 1) return 1;
+                        const ratio = clamp(Number(trainingParams.train_split_ratio || 0.8), 0.05, 0.95);
+                        return clamp(Math.round(selectedTotal * ratio), 1, selectedTotal - 1);
+                    }
+
+                    const trainAvailable = sampleInventory.trainPairs;
+                    const testAvailable = sampleInventory.testPairs;
+                    const totalAvailable = trainAvailable + testAvailable;
+                    if (totalAvailable <= 0) return 0;
+                    if (testAvailable <= 0) return selectedTotal;
+
+                    let trainTarget = Math.round(selectedTotal * (trainAvailable / totalAvailable));
+                    trainTarget = clamp(trainTarget, 0, Math.min(trainAvailable, selectedTotal));
+                    let testTarget = selectedTotal - trainTarget;
+
+                    if (testTarget > testAvailable) {
+                        const overflow = testTarget - testAvailable;
+                        testTarget = testAvailable;
+                        trainTarget = clamp(trainTarget + overflow, 0, trainAvailable);
+                    }
+                    if (trainTarget > trainAvailable) {
+                        const overflow = trainTarget - trainAvailable;
+                        trainTarget = trainAvailable;
+                        testTarget = clamp(testTarget + overflow, 0, testAvailable);
+                    }
+                    return trainTarget;
+                });
+
+                const selectedValidationSampleCount = computed(() => {
+                    if (!sampleInventory.known) return null;
+                    const selectedTotal = selectedTotalSampleCount.value;
+                    const selectedTrain = selectedTrainSampleCount.value;
+                    if (selectedTotal == null || selectedTrain == null) return null;
+                    return Math.max(0, selectedTotal - selectedTrain);
+                });
+
                 const to2DMaskArray = (maskValue) => {
                     if (!maskValue) return null;
                     if (Array.isArray(maskValue) && Array.isArray(maskValue[0])) return maskValue;
@@ -1924,6 +2059,10 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     if (isBioImageArchive) {
                         pathValidationOk.value = true;
                         pathValidationMessage.value = '';
+                        sampleInventory.trainPairs = 0;
+                        sampleInventory.testPairs = 0;
+                        sampleInventory.totalPairs = 0;
+                        sampleInventory.known = false;
                         return true;
                     }
 
@@ -1970,6 +2109,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         if (!hasMetadataDir) {
                             const trainResult = validatePairSpecs(files, trainingParams.train_images, trainingParams.train_annotations, 'Train set');
                             let message = `Train pairs found: ${trainResult.pairedCount} (images: ${trainResult.imageCount}, masks: ${trainResult.annotationCount}).`;
+                            let testPairCount = 0;
 
                             const hasManualTestPaths = trainingParams.split_mode === 'manual' &&
                                 String(trainingParams.test_images || '').trim() &&
@@ -1978,20 +2118,34 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             if (hasManualTestPaths) {
                                 const testResult = validatePairSpecs(files, trainingParams.test_images, trainingParams.test_annotations, 'Test set');
                                 message += ` Test pairs found: ${testResult.pairedCount}.`;
+                                testPairCount = testResult.pairedCount;
                             }
 
+                            sampleInventory.trainPairs = trainResult.pairedCount;
+                            sampleInventory.testPairs = testPairCount;
+                            sampleInventory.totalPairs = trainResult.pairedCount + testPairCount;
+                            sampleInventory.known = true;
+
                             pathValidationOk.value = true;
                             pathValidationMessage.value = `${message} Supported files in artifact: ${supportedFileCount}.`;
                         } else {
                             if (supportedFileCount === 0) {
                                 throw new Error('No supported image/mask files found in artifact (supported: tif/tiff/png/jpg/jpeg).');
                             }
+                            sampleInventory.trainPairs = 0;
+                            sampleInventory.testPairs = 0;
+                            sampleInventory.totalPairs = 0;
+                            sampleInventory.known = false;
                             pathValidationOk.value = true;
                             pathValidationMessage.value = `Metadata mode enabled. Found ${supportedFileCount} supported files in artifact.`;
                         }
 
                         return true;
                     } catch (error) {
+                        sampleInventory.trainPairs = 0;
+                        sampleInventory.testPairs = 0;
+                        sampleInventory.totalPairs = 0;
+                        sampleInventory.known = false;
                         pathValidationOk.value = false;
                         pathValidationMessage.value = String(error?.message || error);
                         return false;
@@ -2339,6 +2493,10 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 watch(
                     () => trainingParams.split_mode,
                     (mode) => {
+                        if (trainingParams.dataset_source === 'BioImage Archive' && mode !== 'auto') {
+                            trainingParams.split_mode = 'auto';
+                            return;
+                        }
                         if (mode === 'auto') {
                             trainingParams.test_images = '';
                             trainingParams.test_annotations = '';
@@ -2366,6 +2524,14 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 watch(
                     () => [trainingParams.dataset_source, trainingParams.artifact_input],
                     async () => {
+                        if (trainingParams.dataset_source === 'BioImage Archive') {
+                            trainingParams.split_mode = 'auto';
+                            sampleInventory.trainPairs = 0;
+                            sampleInventory.testPairs = 0;
+                            sampleInventory.totalPairs = 0;
+                            sampleInventory.known = false;
+                        }
+
                         if (trainingParams.dataset_source !== 'Artifact') return;
 
                         const artifactId = String(trainingParams.artifact_input || '').trim();
@@ -2399,6 +2565,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         trainingParams.test_images,
                         trainingParams.test_annotations,
                         trainingParams.split_mode,
+                        trainingParams.train_split_ratio,
+                        trainingParams.sample_usage_percent,
                         artifactIdExists.value,
                         artifactLookupPending.value,
                     ],
@@ -2552,11 +2720,15 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             if (!artifactValue) throw new Error("Please specify an Artifact ID.");
                         }
 
+                        const sampleUsageRatio = Number((sampleUsagePercent.value / 100).toFixed(4));
+
                         const payload = isBioImageArchive
                             ? {
                                 artifact: artifactValue,
+                                split_mode: 'auto',
+                                train_split_ratio: Number(trainingParams.train_split_ratio || 0.8),
                                 model: trainingParams.model || 'cpsam',
-                                n_samples: null,
+                                n_samples: sampleUsageRatio,
                                 n_epochs: trainingParams.n_epochs,
                                 learning_rate: trainingParams.learning_rate,
                                 weight_decay: 1e-4,
@@ -2573,7 +2745,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                                 test_images: trainingParams.split_mode === 'manual' ? (trainingParams.test_images || null) : null,
                                 test_annotations: trainingParams.split_mode === 'manual' ? (trainingParams.test_annotations || null) : null,
                                 model: trainingParams.model || 'cpsam',
-                                n_samples: null,
+                                n_samples: sampleUsageRatio,
                                 n_epochs: trainingParams.n_epochs,
                                 learning_rate: trainingParams.learning_rate,
                                 weight_decay: 1e-4,
@@ -3344,6 +3516,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     sessions, refreshSessions, refreshLoading, monitoringSessionId, monitoredSession, selectedSessionId, viewSession, currentSession, hasMetrics, hasValidationSet, stopTraining, restartTraining,
                     runningTrainingSessionsCount,
                     trainingParams, startTraining, isStartingTraining,
+                    sampleUsagePercent, sampleUsageDecimal, trainSplitPercent,
+                    availableTotalSampleCount, selectedTotalSampleCount, selectedTrainSampleCount, selectedValidationSampleCount,
                     applyTrainingGlobExample, applyTrainingFolderExample, runPathAutoDetection, validateTrainingPaths,
                     showUploadModal, uploadParams, uploadDataset, isUploading, uploadProgress, uploadStatus, folderInput, selectedFolderName, selectedFileCount,
                     inferenceParams, pretrainedModels, inferenceFile, handleInferenceFile, inferencePreview, isInferring, runInference, inferenceCanvas, previewImg, onImageLoad, canvasStyle, inferenceObjectCount, inferenceResult,
diff --git a/bioengine_apps/cellpose_finetuning/main.py b/bioengine_apps/cellpose_finetuning/main.py
index e5c807e4..8d8852ec 100644
--- a/bioengine_apps/cellpose_finetuning/main.py
+++ b/bioengine_apps/cellpose_finetuning/main.py
@@ -267,7 +267,7 @@ class TrainingParams(TypedDict):
     learning_rate: float
     weight_decay: float
     server_url: str
-    n_samples: int | None
+    n_samples: int | float | None
     session_id: str
     min_train_masks: int
     validation_interval: int | None
@@ -548,7 +548,7 @@ class SessionStatus(TypedDict, total=False):
     exported_artifact_id: str  # Artifact ID if model has been exported
     model_modified: bool  # Flag indicating if model was modified since last export
     model: str
-    n_samples: int | None
+    n_samples: int | float | None
     n_epochs: int
     learning_rate: float
     weight_decay: float
@@ -582,7 +582,7 @@ class SessionStatusWithId(TypedDict, total=False):
     exported_artifact_id: str
     model_modified: bool
     model: str
-    n_samples: int | None
+    n_samples: int | float | None
     n_epochs: int
     learning_rate: float
     weight_decay: float
@@ -769,7 +769,7 @@ def update_status(
     current_batch: int | None = None,
     total_batches: int | None = None,
     model: str | None = None,
-    n_samples: int | None = None,
+    n_samples: int | float | None = None,
     n_epochs: int | None = None,
     learning_rate: float | None = None,
     weight_decay: float | None = None,
@@ -2101,12 +2101,24 @@ def _iter_string_values(payload: Any) -> list[str]:
 
 def _looks_like_image_url(url: str) -> bool:
     path = urlparse(url).path.lower()
-    return path.endswith((".tif", ".tiff", ".ome.tif", ".ome.tiff", ".png", ".jpg", ".jpeg"))
+    return path.endswith(
+        (".tif", ".tiff", ".ome.tif", ".ome.tiff", ".png", ".jpg", ".jpeg")
+    )
 
 
 def _looks_like_mask_url(url: str) -> bool:
     path = urlparse(url).path.lower()
-    return any(token in path for token in ("_mask", "-mask", "_label", "-label", "annotation", "segmentation"))
+    return any(
+        token in path
+        for token in (
+            "_mask",
+            "-mask",
+            "_label",
+            "-label",
+            "annotation",
+            "segmentation",
+        )
+    )
 
 
 def _pair_key_from_url(url: str, *, is_mask: bool) -> str:
@@ -2209,7 +2221,9 @@ async def make_training_pairs_from_bioimage_archive_url(
                 if payload is None:
                     continue
                 for value in _iter_string_values(payload):
-                    if isinstance(value, str) and value.startswith(("http://", "https://")):
+                    if isinstance(value, str) and value.startswith(
+                        ("http://", "https://")
+                    ):
                         if _looks_like_image_url(value) and accession in value.upper():
                             candidates.add(value)
 
@@ -2218,9 +2232,9 @@ async def make_training_pairs_from_bioimage_archive_url(
                     "No downloadable image assets were found from BioImage Archive for this accession."
                 )
 
-            image_urls = sorted([
-                url for url in candidates if not _looks_like_mask_url(url)
-            ])
+            image_urls = sorted(
+                [url for url in candidates if not _looks_like_mask_url(url)]
+            )
             mask_urls = sorted([url for url in candidates if _looks_like_mask_url(url)])
 
             if not image_urls or not mask_urls:
@@ -2246,17 +2260,17 @@ async def make_training_pairs_from_bioimage_archive_url(
                     "No image/mask pairs could be inferred from BioImage Archive assets."
                 )
 
-        if config["n_samples"] is not None and config["n_samples"] < len(paired_urls):
-            subset_idx = np.random.default_rng().permutation(len(paired_urls))[ : config["n_samples"] ]
-            paired_urls = [paired_urls[i] for i in subset_idx]
-
         bia_cache_root = save_path / "bia_download"
         train_pairs: list[TrainingPair] = []
         test_pairs: list[TrainingPair] = []
 
         for image_url, mask_url in paired_urls:
-            local_image = _local_path_for_remote_url(bia_cache_root, "images", image_url)
-            local_mask = _local_path_for_remote_url(bia_cache_root, "annotations", mask_url)
+            local_image = _local_path_for_remote_url(
+                bia_cache_root, "images", image_url
+            )
+            local_mask = _local_path_for_remote_url(
+                bia_cache_root, "annotations", mask_url
+            )
             local_image.parent.mkdir(parents=True, exist_ok=True)
             local_mask.parent.mkdir(parents=True, exist_ok=True)
 
@@ -2281,7 +2295,9 @@ async def make_training_pairs_from_bioimage_archive_url(
             test_pairs = []
 
         if not train_pairs:
-            raise ValueError("No training pairs found after downloading BioImage Archive assets")
+            raise ValueError(
+                "No training pairs found after downloading BioImage Archive assets"
+            )
 
         return train_pairs, test_pairs
 
@@ -2519,13 +2535,22 @@ def _image_key(path: str) -> str:
     def _annotation_key(path: str) -> str:
         basename = Path(path).name
         key = _strip_known_suffixes(basename).lower()
-        for marker in ("_mask", "-mask", "_label", "-label", "_annotation", "-annotation"):
+        for marker in (
+            "_mask",
+            "-mask",
+            "_label",
+            "-label",
+            "_annotation",
+            "-annotation",
+        ):
             if key.endswith(marker):
                 key = key[: -len(marker)]
                 break
         return key
 
-    image_root = _normalize_artifact_relpath(_glob_base_folder(normalized_image_pattern))
+    image_root = _normalize_artifact_relpath(
+        _glob_base_folder(normalized_image_pattern)
+    )
     annotation_root = _normalize_artifact_relpath(
         _glob_base_folder(normalized_annotation_pattern)
     )
@@ -2766,7 +2791,7 @@ async def make_training_pairs_from_metadata(
     artifact: AsyncHyphaArtifact,
     metadata_dir: str,
     save_path: Path,
-    n_samples: int | None,
+    n_samples: int | float | None,
 ) -> tuple[list[TrainingPair], list[TrainingPair]]:
     metadata_root = _normalize_artifact_relpath(metadata_dir)
     if metadata_root and not metadata_root.endswith("/"):
@@ -2813,12 +2838,6 @@ async def make_training_pairs_from_metadata(
             "Expected keys like image_path/mask_path (also supports camelCase variants)."
         )
 
-    if n_samples is not None and n_samples < len(train_pairs_raw):
-        subset_idx = np.random.default_rng().permutation(len(train_pairs_raw))[
-            :n_samples
-        ]
-        train_pairs_raw = [train_pairs_raw[i] for i in subset_idx]
-
     train_pairs = await download_pairs_from_artifact(
         artifact,
         save_path,
@@ -2945,6 +2964,87 @@ def get_training_subset(
     return image_paths, annotation_paths
 
 
+def resolve_requested_sample_count(
+    n_samples: int | float | None,
+    total_count: int,
+) -> int | None:
+    """Resolve requested sample usage into an absolute count.
+
+    If ``0 < n_samples <= 1``, it is treated as a decimal fraction of
+    ``total_count``. Values greater than 1 are treated as absolute counts.
+    """
+    if n_samples is None:
+        return None
+    if total_count <= 0:
+        return 0
+
+    requested = float(n_samples)
+    if requested <= 0:
+        raise ValueError("n_samples must be > 0")
+
+    if requested <= 1.0:
+        resolved = int(round(total_count * requested))
+    else:
+        resolved = int(round(requested))
+
+    return max(1, min(total_count, resolved))
+
+
+def random_subset_pairs(pairs: list[TrainingPair], count: int) -> list[TrainingPair]:
+    """Return a random subset of ``pairs`` of size ``count``."""
+    if count >= len(pairs):
+        return pairs
+    idx = np.random.default_rng().permutation(len(pairs))[:count]
+    return [pairs[i] for i in idx]
+
+
+def proportional_manual_sample_counts(
+    train_available: int,
+    test_available: int,
+    requested_total: int,
+) -> tuple[int, int]:
+    """Allocate requested manual-split samples across train/test pools."""
+    total_available = train_available + test_available
+    if total_available <= 0:
+        return 0, 0
+
+    requested_total = max(0, min(requested_total, total_available))
+    if requested_total == 0:
+        return 0, 0
+
+    if test_available <= 0:
+        return min(train_available, requested_total), 0
+    if train_available <= 0:
+        return 0, min(test_available, requested_total)
+
+    train_target = int(round(requested_total * (train_available / total_available)))
+    train_target = max(0, min(train_available, train_target))
+    test_target = requested_total - train_target
+
+    if test_target > test_available:
+        overflow = test_target - test_available
+        test_target = test_available
+        train_target = min(train_available, train_target + overflow)
+    if train_target > train_available:
+        overflow = train_target - train_available
+        train_target = train_available
+        test_target = min(test_available, test_target + overflow)
+
+    assigned = train_target + test_target
+    if assigned < requested_total:
+        remaining = requested_total - assigned
+        train_room = train_available - train_target
+        add_train = min(train_room, remaining)
+        train_target += add_train
+        remaining -= add_train
+        if remaining > 0:
+            test_room = test_available - test_target
+            add_test = min(test_room, remaining)
+            test_target += add_test
+
+    return train_target, test_target
+
+
 def split_training_pairs(
     train_pairs: list[TrainingPair],
     train_ratio: float,
@@ -3039,14 +3139,6 @@ async def make_training_pairs(
     train_image_paths = [Path(img) for img, _ in train_matched]
     train_annotation_paths = [Path(ann) for _, ann in train_matched]
 
-    # Apply n_samples if specified
-    if config["n_samples"] is not None and config["n_samples"] < len(train_image_paths):
-        train_image_paths, train_annotation_paths = get_training_subset(
-            train_image_paths,
-            train_annotation_paths,
-            config["n_samples"],
-        )
-
     # Download training pairs
     train_pairs = await download_pairs_from_artifact(
         artifact,
@@ -3088,11 +3180,27 @@ async def make_training_pairs(
         )
 
     split_mode = str(config.get("split_mode", "manual") or "manual").lower()
-    if split_mode == "auto" and not test_pairs:
-        train_pairs, test_pairs = split_training_pairs(
-            train_pairs,
-            float(config.get("train_split_ratio", 0.8)),
-        )
+    train_split_ratio = float(config.get("train_split_ratio", 0.8))
+
+    requested_total = resolve_requested_sample_count(
+        config.get("n_samples"),
+        len(train_pairs) + len(test_pairs),
+    )
+
+    if split_mode == "auto":
+        combined_pairs = [*train_pairs, *test_pairs]
+        if requested_total is not None and requested_total < len(combined_pairs):
+            combined_pairs = random_subset_pairs(combined_pairs, requested_total)
+        train_pairs, test_pairs = split_training_pairs(combined_pairs, train_split_ratio)
+    elif requested_total is not None:
+        if requested_total < len(train_pairs) + len(test_pairs):
+            train_target, test_target = proportional_manual_sample_counts(
+                len(train_pairs),
+                len(test_pairs),
+                requested_total,
+            )
+            train_pairs = random_subset_pairs(train_pairs, train_target)
+            test_pairs = random_subset_pairs(test_pairs, test_target)
 
     return train_pairs, test_pairs
 
@@ -3676,11 +3784,13 @@ async def start_training(
                 *PretrainedModel.values(),
             ],
         ),
-        n_samples: int | None = Field(
+        n_samples: int | float | None = Field(
             None,
             description=(
-                "Optional number of samples to use from the dataset. If None, "
-                "all available samples are used."
+                "Optional number of samples to use from the dataset. "
+                "If 0 < value <= 1, treated as decimal fraction of total samples "
+                "(e.g., 0.6 means 60%). If value > 1, treated as absolute count. "
+                "If None, all available samples are used."
             ),
         ),
         n_epochs: int = Field(10, description="Number of training epochs"),
@@ -3765,8 +3875,12 @@ async def start_training(
             if not isinstance(metadata_dir, str) or not metadata_dir:
                 raise ValueError("metadata_dir must be a non-empty string")
 
+        is_bia = isinstance(artifact, str) and _is_bioimage_archive_url(artifact)
+
         if not isinstance(split_mode, str) or split_mode not in {"manual", "auto"}:
             split_mode = "manual"
+        if is_bia:
+            split_mode = "auto"
 
         if train_split_ratio is None:
             train_split_ratio = 0.8
@@ -3795,7 +3909,9 @@ async def start_training(
         if min_train_masks is None:
             min_train_masks = 5
         if n_samples is not None:
-            n_samples = int(n_samples)
+            n_samples = float(n_samples)
+            if n_samples <= 0:
+                raise ValueError("n_samples must be > 0")
         if validation_interval is not None:
             validation_interval = int(validation_interval)
 

From b7e6f131d9389904b0a3c0a6ceb7e31c6dd9d6d1 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Sun, 22 Feb 2026 23:52:40 +0100
Subject: [PATCH 11/13] Add tests and docs for fractional sample selection in
 cellpose fine-tuning

---
 bioengine_apps/cellpose_finetuning/README.md  |  6 ++-
 .../tests/test_metadata_and_glob.py           | 49 +++++++++++++++++++
 2 files changed, 53 insertions(+), 2 deletions(-)

diff --git a/bioengine_apps/cellpose_finetuning/README.md b/bioengine_apps/cellpose_finetuning/README.md
index 1bc52da7..0be5b342 100644
--- a/bioengine_apps/cellpose_finetuning/README.md
+++ b/bioengine_apps/cellpose_finetuning/README.md
@@ -191,9 +191,11 @@ Start asynchronous model fine-tuning.
 - `metadata_dir` (str, optional): Folder containing metadata JSON files with image/annotation paths (e.g. `image_path` + `mask_path`). If provided, explicit `train_annotations` can be omitted.
 - `test_images` (str, optional): Optional test images path (same format as train_images). Providing test data enables per-epoch pixel-level validation metrics and end-of-training instance segmentation metrics (AP@0.5/0.75/0.9).
 - `test_annotations` (str, optional): Optional test annotations (label masks). Required together with `test_images`.
+- `split_mode` (str): Dataset split mode (`"manual"` or `"auto"`, default: `"manual"`). In `"auto"`, the service creates train/validation split from the selected sample pool.
+- `train_split_ratio` (float): Train ratio used when `split_mode="auto"` (default: `0.8`, i.e. 80% train / 20% validation).
 - `model` (str): Pretrained model to start from (default: "cpsam")
 - `n_epochs` (int): Number of training epochs (default: 10)
-- `n_samples` (int, optional): Limit number of samples to use
+- `n_samples` (int | float, optional): Limit sample usage. If `n_samples > 1`, treated as absolute count. If `0 < n_samples <= 1`, treated as decimal fraction of all available samples (e.g. `0.6` = 60%).
 - `learning_rate` (float): Learning rate (default: 1e-6)
 - `weight_decay` (float): Weight decay (default: 0.0001)
 - `min_train_masks` (int): Minimum number of masks per training batch (default: 5). Lower values speed up training.
@@ -430,7 +432,7 @@ The recommended way to deploy (whether for the first time or to update) is to us
 source .env
 
 # Deploy to production
-python tests/cellpose_legacy_scripts/redeploy_cellpose.py \
+python tests/cellpose/redeploy_cellpose.py \
     --artifact-id ri-scale/cellpose-finetuning \
     --application-id cellpose-finetuning
 ```
diff --git a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py b/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
index 0a9e2a44..094a246b 100644
--- a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
+++ b/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
@@ -9,6 +9,8 @@
     list_matching_artifact_paths,
     make_training_pairs_from_metadata,
     match_image_annotation_pairs,
+    proportional_manual_sample_counts,
+    resolve_requested_sample_count,
 )
 
 
@@ -270,3 +272,50 @@ def test_pair_key_from_url_handles_mask_suffix() -> None:
     assert _pair_key_from_url(img, is_mask=False) == _pair_key_from_url(
         ann, is_mask=True
     )
+
+
+def test_resolve_requested_sample_count_supports_decimal_fraction() -> None:
+    assert resolve_requested_sample_count(0.5, 10) == 5
+    assert resolve_requested_sample_count(0.01, 10) == 1
+    assert resolve_requested_sample_count(1.0, 10) == 10
+
+
+def test_resolve_requested_sample_count_supports_absolute_count() -> None:
+    assert resolve_requested_sample_count(4, 10) == 4
+    assert resolve_requested_sample_count(20, 10) == 10
+
+
+def test_resolve_requested_sample_count_rejects_non_positive_values() -> None:
+    try:
+        resolve_requested_sample_count(0, 10)
+        assert False, "Expected ValueError for n_samples=0"
+    except ValueError as e:
+        assert "n_samples must be > 0" in str(e)
+
+    try:
+        resolve_requested_sample_count(-0.2, 10)
+        assert False, "Expected ValueError for n_samples<0"
+    except ValueError as e:
+        assert "n_samples must be > 0" in str(e)
+
+
+def test_proportional_manual_sample_counts_basic_allocation() -> None:
+    train_count, test_count = proportional_manual_sample_counts(
+        train_available=8,
+        test_available=2,
+        requested_total=5,
+    )
+    assert train_count + test_count == 5
+    assert train_count == 4
+    assert test_count == 1
+
+
+def test_proportional_manual_sample_counts_respects_capacity_limits() -> None:
+    train_count, test_count = proportional_manual_sample_counts(
+        train_available=1,
+        test_available=9,
+        requested_total=6,
+    )
+    assert train_count + test_count == 6
+    assert train_count == 1
+    assert test_count == 5

From 6c2f5e61c890d9d42e96f2df5a2f6f8807cf9240 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Mon, 23 Feb 2026 00:23:19 +0100
Subject: [PATCH 12/13] Improve path validation UX and add pyodide startup
 script for cellpose agent

---
 ...ellpose_training_agent_startup.pyodide.txt | 211 ++++++++++++++++++
 bioengine_apps/cellpose_finetuning/index.html |  40 +++-
 2 files changed, 247 insertions(+), 4 deletions(-)
 create mode 100644 bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt

diff --git a/bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt b/bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt
new file mode 100644
index 00000000..6ef25e24
--- /dev/null
+++ b/bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt
@@ -0,0 +1,211 @@
+import micropip
+await micropip.install(["hypha-rpc", "httpx"])
+
+import os
+from typing import Any
+from hypha_rpc import connect_to_server, login
+
+SERVER_URL = os.environ.get("HYPHA_SERVER_URL", "https://hypha.aicell.io")
+DEFAULT_SERVICE_CANDIDATES = [
+    "bioimage-io/cellpose-finetuning",
+    "ri-scale/cellpose-finetuning",
+    "bioimage-io/cellpose-finetuning-test",
+    "ri-scale/cellpose-finetuning-test",
+]
+
+_state: dict[str, Any] = {
+    "server": None,
+    "workspace": None,
+    "service": None,
+    "service_id": None,
+}
+
+
+async def connect_cellpose(server_url: str | None = None, workspace: str | None = None, token: str | None = None):
+    target_server = server_url or SERVER_URL
+    target_token = token or os.environ.get("HYPHA_TOKEN")
+    if not target_token:
+        target_token = await login({"server_url": target_server})
+
+    config: dict[str, Any] = {"server_url": target_server, "token": target_token}
+    if workspace:
+        config["workspace"] = workspace
+
+    server = await connect_to_server(config)
+    _state["server"] = server
+    _state["workspace"] = server.config.workspace
+    return {
+        "server_url": target_server,
+        "workspace": server.config.workspace,
+        "user_id": server.config.user.get("id"),
+    }
+
+
+async def list_cellpose_services():
+    if not _state["server"]:
+        await connect_cellpose()
+    server = _state["server"]
+
+    try:
+        listed = await server.listServices({"workspace": _state["workspace"]})
+    except Exception:
+        listed = await server.listServices()
+
+    service_ids: list[str] = []
+    for item in (listed if isinstance(listed, list) else []):
+        sid = item if isinstance(item, str) else item.get("id")
+        if isinstance(sid, str) and "cellpose-finetuning" in sid:
+            service_ids.append(sid)
+
+    for fallback in DEFAULT_SERVICE_CANDIDATES:
+        if fallback not in service_ids:
+            try:
+                await server.getService(fallback)
+                service_ids.append(fallback)
+            except Exception:
+                pass
+
+    return sorted(set(service_ids))
+
+
+async def select_cellpose_service(service_id: str | None = None):
+    if not _state["server"]:
+        await connect_cellpose()
+    server = _state["server"]
+
+    chosen = service_id
+    if not chosen:
+        options = await list_cellpose_services()
+        if not options:
+            raise RuntimeError("No cellpose-finetuning service found")
+        chosen = options[0]
+
+    service = await server.getService(chosen)
+    _state["service"] = service
+    _state["service_id"] = chosen
+    return {"service_id": chosen}
+
+
+async def start_cellpose_training(**kwargs):
+    if not _state["service"]:
+        await select_cellpose_service()
+    return await _state["service"].start_training(kwargs, {"_rkwargs": True})
+
+
+async def get_training_status(session_id: str):
+    if not _state["service"]:
+        await select_cellpose_service()
+    return await _state["service"].get_training_status({"session_id": session_id}, {"_rkwargs": True})
+
+
+async def list_training_sessions(status_types: list[str] | None = None):
+    if not _state["service"]:
+        await select_cellpose_service()
+    if status_types:
+        return await _state["service"].list_training_sessions(
+            {"status_types": status_types},
+            {"_rkwargs": True},
+        )
+    return await _state["service"].list_training_sessions()
+
+
+async def stop_training(session_id: str):
+    if not _state["service"]:
+        await select_cellpose_service()
+    return await _state["service"].stop_training({"session_id": session_id}, {"_rkwargs": True})
+
+
+async def restart_training(session_id: str, n_epochs: int | None = None):
+    if not _state["service"]:
+        await select_cellpose_service()
+    payload: dict[str, Any] = {"session_id": session_id}
+    if n_epochs is not None:
+        payload["n_epochs"] = int(n_epochs)
+    return await _state["service"].restart_training(payload, {"_rkwargs": True})
+
+
+async def infer_cellpose(**kwargs):
+    if not _state["service"]:
+        await select_cellpose_service()
+    return await _state["service"].infer(kwargs, {"_rkwargs": True})
+
+
+async def export_model(session_id: str, model_name: str | None = None, collection: str = "bioimage-io/colab-annotations"):
+    if not _state["service"]:
+        await select_cellpose_service()
+    payload: dict[str, Any] = {
+        "session_id": session_id,
+        "collection": collection,
+    }
+    if model_name:
+        payload["model_name"] = model_name
+    return await _state["service"].export_model(payload, {"_rkwargs": True})
+
+
+print("""
+You are a Cellpose training operations assistant for BioEngine.
+
+Your job is to chat with users and execute Cellpose fine-tuning workflows via the utility functions in this runtime.
+
+Available async functions:
+- await connect_cellpose(server_url=None, workspace=None, token=None)
+- await list_cellpose_services()
+- await select_cellpose_service(service_id=None)
+- await start_cellpose_training(**kwargs)
+- await get_training_status(session_id)
+- await list_training_sessions(status_types=None)
+- await stop_training(session_id)
+- await restart_training(session_id, n_epochs=None)
+- await infer_cellpose(**kwargs)
+- await export_model(session_id, model_name=None, collection='bioimage-io/colab-annotations')
+
+Behavior requirements:
+1) Always ensure connectivity first (connect_cellpose, then select_cellpose_service).
+2) For training requests, collect missing required fields before calling start:
+   - artifact
+   - one of:
+     a) train_images + train_annotations
+     b) metadata_dir
+     c) BioImage Archive URL in artifact
+3) Respect split behavior:
+   - For BioImage Archive URLs, use split_mode='auto'.
+   - For auto split, include train_split_ratio if user specifies it.
+   - If user gives sample percentage, convert to decimal and pass as n_samples (0 < n_samples <= 1).
+4) After starting training, report session_id and offer to monitor progress.
+5) Monitoring loop: poll get_training_status(session_id) and summarize key fields:
+   - status_type, message, current_epoch/total_epochs, elapsed_seconds, latest train/test losses.
+6) On errors, provide actionable remediation (artifact path mismatch, missing annotations, token/workspace mismatch, resource limits).
+7) Keep outputs concise and user-friendly; include exact IDs and parameter values used.
+
+Examples:
+- Start from artifact paths:
+  await start_cellpose_training(
+      artifact='workspace/dataset',
+      train_images='images/*/*.tif',
+      train_annotations='annotations/*/*_mask.ome.tif',
+      split_mode='auto',
+      train_split_ratio=0.8,
+      n_samples=0.6,
+      model='cpsam',
+      n_epochs=20,
+      learning_rate=1e-5,
+  )
+
+- Start from metadata:
+  await start_cellpose_training(
+      artifact='workspace/dataset',
+      metadata_dir='metadata/',
+      n_epochs=10,
+      model='cpsam'
+  )
+
+- BioImage Archive:
+  await start_cellpose_training(
+      artifact='https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD1392',
+      split_mode='auto',
+      train_split_ratio=0.8,
+      n_samples=1.0,
+      n_epochs=10,
+      model='cpsam'
+  )
+""")
diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 07401987..3c17daea 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -1006,6 +1006,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const browserPath = ref(''); 
                 const browserLoading = ref(false);
                 const browserError = ref('');
+                const artifactFilesCache = reactive({});
                 const isPathDetecting = ref(false);
                 const isPathValidating = ref(false);
                 const pathAutoHint = ref('');
@@ -1711,7 +1712,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     if (trainingParams.dataset_source === 'BioImage Archive') {
                         return !String(trainingParams.bioimage_archive_url || '').trim();
                     }
-                    return !pathValidationOk.value || isPathValidating.value;
+                    return !pathValidationOk.value;
                 });
 
                 const applyTrainingFolderExample = () => {
@@ -1831,6 +1832,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 };
 
                 const loadArtifactFiles = async (artifactId) => {
+                    const cacheKey = String(artifactId || '').trim();
+                    if (cacheKey && artifactFilesCache[cacheKey]?.files?.length > 0) {
+                        return Array.from(artifactFilesCache[cacheKey].files);
+                    }
+
                     const am = await getArtifactManager();
                     const fileSet = new Set();
                     const folderQueue = [];
@@ -1874,7 +1880,14 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         paths = (Array.isArray(fallback) ? fallback : []).map(normalizeListedPath).filter(Boolean);
                     }
 
-                    return Array.from(new Set(paths.filter((p) => !p.endsWith('/'))));
+                    const uniquePaths = Array.from(new Set(paths.filter((p) => !p.endsWith('/'))));
+                    if (cacheKey) {
+                        artifactFilesCache[cacheKey] = {
+                            files: uniquePaths,
+                            updatedAt: Date.now(),
+                        };
+                    }
+                    return uniquePaths;
                 };
 
                 const baseName = (pathStr) => String(pathStr || '').split('/').pop() || '';
@@ -2036,6 +2049,21 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         trainingParams.train_images = preferredTrainImages || (trainImageDir ? `${trainImageDir}/` : '');
                         trainingParams.train_annotations = preferredTrainAnnotations || (trainMaskDir ? `${trainMaskDir}/` : '');
 
+                        const quickTrainResult = validatePairSpecs(
+                            files,
+                            trainingParams.train_images,
+                            trainingParams.train_annotations,
+                            'Train set'
+                        );
+                        sampleInventory.trainPairs = quickTrainResult.pairedCount;
+                        sampleInventory.testPairs = 0;
+                        sampleInventory.totalPairs = quickTrainResult.pairedCount;
+                        sampleInventory.known = true;
+                        if (quickTrainResult.pairedCount > 0) {
+                            pathValidationOk.value = true;
+                            pathValidationMessage.value = `Train pairs found: ${quickTrainResult.pairedCount}. Finishing full validation…`;
+                        }
+
                         if (trainingParams.split_mode === 'manual' && testImageDir && testMaskDir) {
                             trainingParams.test_images = `${testImageDir}/`;
                             trainingParams.test_annotations = `${testMaskDir}/`;
@@ -2100,6 +2128,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     }
 
                     isPathValidating.value = true;
+                    pathValidationOk.value = false;
                     pathValidationMessage.value = '';
 
                     try {
@@ -2111,6 +2140,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             let message = `Train pairs found: ${trainResult.pairedCount} (images: ${trainResult.imageCount}, masks: ${trainResult.annotationCount}).`;
                             let testPairCount = 0;
 
+                            if (trainResult.pairedCount > 0) {
+                                pathValidationOk.value = true;
+                                pathValidationMessage.value = `Train pairs found: ${trainResult.pairedCount}. Finishing full validation…`;
+                            }
+
                             const hasManualTestPaths = trainingParams.split_mode === 'manual' &&
                                 String(trainingParams.test_images || '').trim() &&
                                 String(trainingParams.test_annotations || '').trim();
@@ -2565,8 +2599,6 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         trainingParams.test_images,
                         trainingParams.test_annotations,
                         trainingParams.split_mode,
-                        trainingParams.train_split_ratio,
-                        trainingParams.sample_usage_percent,
                         artifactIdExists.value,
                         artifactLookupPending.value,
                     ],

From f1ec14f3a58958aaf30573ec47981742ff65df42 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Hugo=20Dettner=20K=C3=A4llander?= <hugokallander@gmail.com>
Date: Tue, 24 Feb 2026 20:29:59 +0100
Subject: [PATCH 13/13] Improve Cellpose preflight, BIA handling, and live e2e
 tooling

---
 .gitignore                                    |    3 +-
 .../bia_resolve_url_proxy/README.md           |    5 +
 bioengine_apps/bia_resolve_url_proxy/main.py  |  136 ++
 .../bia_resolve_url_proxy/manifest.yaml       |   16 +
 bioengine_apps/cellpose_finetuning/README.md  |   27 +-
 ...ellpose_training_agent_startup.pyodide.txt |  211 ---
 bioengine_apps/cellpose_finetuning/index.html |  398 +++++-
 bioengine_apps/cellpose_finetuning/main.py    | 1141 +++++++++++++++--
 scripts/tmp_probe_bia_fetch_only.py           |  140 ++
 ...cellpose_training_agent_startup.pyodide.py |  391 ++++++
 tests/cellpose/live_test_utils.py             |  153 +++
 .../cellpose}/test_metadata_and_glob.py       |  154 +++
 tests/cellpose/test_service.py                |   46 +-
 .../test_training_e2e_small_percentage.py     |  106 ++
 .../test_training_e2e_three_sources.py        |  409 ++++++
 15 files changed, 2962 insertions(+), 374 deletions(-)
 create mode 100644 bioengine_apps/bia_resolve_url_proxy/README.md
 create mode 100644 bioengine_apps/bia_resolve_url_proxy/main.py
 create mode 100644 bioengine_apps/bia_resolve_url_proxy/manifest.yaml
 delete mode 100644 bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt
 create mode 100644 scripts/tmp_probe_bia_fetch_only.py
 create mode 100644 tests/cellpose/cellpose_training_agent_startup.pyodide.py
 create mode 100644 tests/cellpose/live_test_utils.py
 rename {bioengine_apps/cellpose_finetuning/tests => tests/cellpose}/test_metadata_and_glob.py (68%)
 create mode 100644 tests/cellpose/test_training_e2e_small_percentage.py
 create mode 100644 tests/cellpose/test_training_e2e_three_sources.py

diff --git a/.gitignore b/.gitignore
index 81f08936..25aa99bd 100644
--- a/.gitignore
+++ b/.gitignore
@@ -173,4 +173,5 @@ data/
 bioimageio_test_reports/
 
 # MacOS
-.DS_STORE
\ No newline at end of file
+.DS_STORE
+demo-dataset/
\ No newline at end of file
diff --git a/bioengine_apps/bia_resolve_url_proxy/README.md b/bioengine_apps/bia_resolve_url_proxy/README.md
new file mode 100644
index 00000000..54e25671
--- /dev/null
+++ b/bioengine_apps/bia_resolve_url_proxy/README.md
@@ -0,0 +1,5 @@
+# BIA Resolve URL Proxy
+
+Minimal ray-serve app that exposes a single `resolve_url` RPC method.
+
+It is used by `cellpose_finetuning` to fetch BioImage Archive assets through Hypha when direct fetches are blocked.
diff --git a/bioengine_apps/bia_resolve_url_proxy/main.py b/bioengine_apps/bia_resolve_url_proxy/main.py
new file mode 100644
index 00000000..b1537666
--- /dev/null
+++ b/bioengine_apps/bia_resolve_url_proxy/main.py
@@ -0,0 +1,136 @@
+import base64
+import os
+from typing import Any
+from urllib.parse import urlparse
+
+import httpx
+from hypha_rpc.utils.schema import schema_method
+from ray import serve
+
+_DEFAULT_ALLOWED_HOSTS = "beta.bioimagearchive.org,www.ebi.ac.uk,ftp.ebi.ac.uk"
+
+
+def _allowed_hosts() -> set[str]:
+    raw_value = os.environ.get("RESOLVE_URL_ALLOWED_HOSTS", _DEFAULT_ALLOWED_HOSTS)
+    return {part.strip().lower() for part in raw_value.split(",") if part.strip()}
+
+
+def _normalize_headers(headers: dict[str, Any] | None) -> dict[str, str]:
+    normalized: dict[str, str] = {}
+    if not isinstance(headers, dict):
+        return normalized
+    for key, value in headers.items():
+        if not isinstance(key, str):
+            continue
+        if isinstance(value, str):
+            normalized[key] = value
+        elif value is not None:
+            normalized[key] = str(value)
+    return normalized
+
+
+@serve.deployment(
+    ray_actor_options={
+        "num_cpus": 0.5,
+        "num_gpus": 0,
+        "runtime_env": {
+            "pip": [
+                "httpx",
+            ],
+        },
+    }
+)
+class ResolveUrlProxy:
+    @schema_method
+    async def resolve_url(
+        self,
+        url: str,
+        method: str = "GET",
+        headers: dict[str, Any] | None = None,
+        timeout: float = 60.0,
+        body: str | dict[str, Any] | list[Any] | None = None,
+        context: dict[str, Any] | None = None,
+    ) -> dict[str, Any]:
+        try:
+            parsed = urlparse(str(url))
+        except Exception as exp:
+            return {
+                "ok": False,
+                "status_code": 400,
+                "url": str(url),
+                "error": f"Invalid URL: {exp}",
+            }
+
+        if parsed.scheme.lower() != "https":
+            return {
+                "ok": False,
+                "status_code": 400,
+                "url": str(url),
+                "error": "Only https URLs are allowed",
+            }
+
+        host = (parsed.hostname or "").lower()
+        if host not in _allowed_hosts():
+            return {
+                "ok": False,
+                "status_code": 403,
+                "url": str(url),
+                "error": f"Host '{host}' is not allowed",
+            }
+
+        method_value = str(method or "GET").upper()
+        if method_value not in {
+            "GET",
+            "POST",
+            "PUT",
+            "PATCH",
+            "DELETE",
+            "HEAD",
+            "OPTIONS",
+        }:
+            return {
+                "ok": False,
+                "status_code": 400,
+                "url": str(url),
+                "error": f"Unsupported method: {method_value}",
+            }
+
+        request_kwargs: dict[str, Any] = {
+            "method": method_value,
+            "url": str(url),
+            "headers": _normalize_headers(headers),
+        }
+        request_kwargs["headers"].setdefault(
+            "User-Agent", "bioengine-bia-resolve-url-proxy/1.0"
+        )
+
+        if body is not None:
+            if isinstance(body, (dict, list)):
+                request_kwargs["json"] = body
+            else:
+                request_kwargs["content"] = str(body)
+
+        try:
+            async with httpx.AsyncClient(
+                timeout=max(1.0, float(timeout)),
+                follow_redirects=True,
+            ) as client:
+                response = await client.request(**request_kwargs)
+        except Exception as exp:
+            return {
+                "ok": False,
+                "status_code": 502,
+                "url": str(url),
+                "error": str(exp),
+            }
+
+        status_code = int(response.status_code)
+        ok = 200 <= status_code < 300
+        return {
+            "ok": ok,
+            "status_code": status_code,
+            "url": str(response.url),
+            "headers": dict(response.headers),
+            "content_base64": base64.b64encode(response.content).decode("ascii"),
+            "error": "" if ok else f"Upstream returned HTTP {status_code}",
+        }
diff --git a/bioengine_apps/bia_resolve_url_proxy/manifest.yaml b/bioengine_apps/bia_resolve_url_proxy/manifest.yaml
new file mode 100644
index 00000000..9dc5269e
--- /dev/null
+++ b/bioengine_apps/bia_resolve_url_proxy/manifest.yaml
@@ -0,0 +1,16 @@
+name: BIA Resolve URL Proxy
+id: bia-resolve-url-proxy
+id_emoji: '🌐'
+covers: []
+description: Resolve URL proxy for BioImage Archive downloads (resolve_url only).
+documentation: README.md
+format_version: 0.5.0
+git_repo: https://github.com/aicell-lab/bioengine-worker
+links: []
+tags: [bioimage-archive, proxy, resolve-url]
+type: ray-serve
+version: 0.0.1
+deployments:
+  - main:ResolveUrlProxy
+authorized_users:
+  - "*"
diff --git a/bioengine_apps/cellpose_finetuning/README.md b/bioengine_apps/cellpose_finetuning/README.md
index 0be5b342..05b36f12 100644
--- a/bioengine_apps/cellpose_finetuning/README.md
+++ b/bioengine_apps/cellpose_finetuning/README.md
@@ -195,7 +195,7 @@ Start asynchronous model fine-tuning.
 - `train_split_ratio` (float): Train ratio used when `split_mode="auto"` (default: `0.8`, i.e. 80% train / 20% validation).
 - `model` (str): Pretrained model to start from (default: "cpsam")
 - `n_epochs` (int): Number of training epochs (default: 10)
-- `n_samples` (int | float, optional): Limit sample usage. If `n_samples > 1`, treated as absolute count. If `0 < n_samples <= 1`, treated as decimal fraction of all available samples (e.g. `0.6` = 60%).
+- `n_samples` (int | float, optional): Limit sample usage. If `n_samples > 1`, treated as absolute count. If `0 < n_samples <= 1`, treated as decimal fraction of available samples (e.g. `0.05` = 5%).
 - `learning_rate` (float): Learning rate (default: 1e-6)
 - `weight_decay` (float): Weight decay (default: 0.0001)
 - `min_train_masks` (int): Minimum number of masks per training batch (default: 5). Lower values speed up training.
@@ -378,6 +378,31 @@ Notes:
 - Sessions with stale `waiting`/`preparing`/`running` states after a redeploy are normalized to `stopped` (interrupted).
 - Running-session indicators in the UI only track sessions that are currently active.
 
+### `delete_training_session()`
+
+Delete a non-running training session and its local artifacts.
+
+**Parameters:**
+- `session_id` (str): Session ID to delete
+- `force_stop_if_blocked` (bool, optional): If true, stop a blocked `running`/`preparing` session first, then delete
+
+**Returns:** Dict with `deleted` (bool) and `session_id` (str)
+
+### `preflight_training_dataset()`
+
+Run a lightweight backend preflight over training path configuration without starting training.
+
+**Parameters:**
+- `artifact` (str): Artifact ID or BioImage Archive URL
+- `train_images`, `train_annotations` (str, optional): Training path specs
+- `metadata_dir` (str, optional): Metadata mode directory
+- `test_images`, `test_annotations` (str, optional): Optional test path specs
+- `split_mode` (str): `manual` or `auto`
+- `n_samples` (int | float, optional): Optional sample limit/count
+- `max_candidates` (int): Cap for path matching candidates during preflight
+
+**Returns:** Dict including `ok`, pair counts (`train_pair_count`, `test_pair_count`), and an explanatory `message`.
+
 Validation metrics are only produced when both `test_images` and `test_annotations` are provided.
 
 ## Inference Parameters Guide
diff --git a/bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt b/bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt
deleted file mode 100644
index 6ef25e24..00000000
--- a/bioengine_apps/cellpose_finetuning/cellpose_training_agent_startup.pyodide.txt
+++ /dev/null
@@ -1,211 +0,0 @@
-import micropip
-await micropip.install(["hypha-rpc", "httpx"])
-
-import os
-from typing import Any
-from hypha_rpc import connect_to_server, login
-
-SERVER_URL = os.environ.get("HYPHA_SERVER_URL", "https://hypha.aicell.io")
-DEFAULT_SERVICE_CANDIDATES = [
-    "bioimage-io/cellpose-finetuning",
-    "ri-scale/cellpose-finetuning",
-    "bioimage-io/cellpose-finetuning-test",
-    "ri-scale/cellpose-finetuning-test",
-]
-
-_state: dict[str, Any] = {
-    "server": None,
-    "workspace": None,
-    "service": None,
-    "service_id": None,
-}
-
-
-async def connect_cellpose(server_url: str | None = None, workspace: str | None = None, token: str | None = None):
-    target_server = server_url or SERVER_URL
-    target_token = token or os.environ.get("HYPHA_TOKEN")
-    if not target_token:
-        target_token = await login({"server_url": target_server})
-
-    config: dict[str, Any] = {"server_url": target_server, "token": target_token}
-    if workspace:
-        config["workspace"] = workspace
-
-    server = await connect_to_server(config)
-    _state["server"] = server
-    _state["workspace"] = server.config.workspace
-    return {
-        "server_url": target_server,
-        "workspace": server.config.workspace,
-        "user_id": server.config.user.get("id"),
-    }
-
-
-async def list_cellpose_services():
-    if not _state["server"]:
-        await connect_cellpose()
-    server = _state["server"]
-
-    try:
-        listed = await server.listServices({"workspace": _state["workspace"]})
-    except Exception:
-        listed = await server.listServices()
-
-    service_ids: list[str] = []
-    for item in (listed if isinstance(listed, list) else []):
-        sid = item if isinstance(item, str) else item.get("id")
-        if isinstance(sid, str) and "cellpose-finetuning" in sid:
-            service_ids.append(sid)
-
-    for fallback in DEFAULT_SERVICE_CANDIDATES:
-        if fallback not in service_ids:
-            try:
-                await server.getService(fallback)
-                service_ids.append(fallback)
-            except Exception:
-                pass
-
-    return sorted(set(service_ids))
-
-
-async def select_cellpose_service(service_id: str | None = None):
-    if not _state["server"]:
-        await connect_cellpose()
-    server = _state["server"]
-
-    chosen = service_id
-    if not chosen:
-        options = await list_cellpose_services()
-        if not options:
-            raise RuntimeError("No cellpose-finetuning service found")
-        chosen = options[0]
-
-    service = await server.getService(chosen)
-    _state["service"] = service
-    _state["service_id"] = chosen
-    return {"service_id": chosen}
-
-
-async def start_cellpose_training(**kwargs):
-    if not _state["service"]:
-        await select_cellpose_service()
-    return await _state["service"].start_training(kwargs, {"_rkwargs": True})
-
-
-async def get_training_status(session_id: str):
-    if not _state["service"]:
-        await select_cellpose_service()
-    return await _state["service"].get_training_status({"session_id": session_id}, {"_rkwargs": True})
-
-
-async def list_training_sessions(status_types: list[str] | None = None):
-    if not _state["service"]:
-        await select_cellpose_service()
-    if status_types:
-        return await _state["service"].list_training_sessions(
-            {"status_types": status_types},
-            {"_rkwargs": True},
-        )
-    return await _state["service"].list_training_sessions()
-
-
-async def stop_training(session_id: str):
-    if not _state["service"]:
-        await select_cellpose_service()
-    return await _state["service"].stop_training({"session_id": session_id}, {"_rkwargs": True})
-
-
-async def restart_training(session_id: str, n_epochs: int | None = None):
-    if not _state["service"]:
-        await select_cellpose_service()
-    payload: dict[str, Any] = {"session_id": session_id}
-    if n_epochs is not None:
-        payload["n_epochs"] = int(n_epochs)
-    return await _state["service"].restart_training(payload, {"_rkwargs": True})
-
-
-async def infer_cellpose(**kwargs):
-    if not _state["service"]:
-        await select_cellpose_service()
-    return await _state["service"].infer(kwargs, {"_rkwargs": True})
-
-
-async def export_model(session_id: str, model_name: str | None = None, collection: str = "bioimage-io/colab-annotations"):
-    if not _state["service"]:
-        await select_cellpose_service()
-    payload: dict[str, Any] = {
-        "session_id": session_id,
-        "collection": collection,
-    }
-    if model_name:
-        payload["model_name"] = model_name
-    return await _state["service"].export_model(payload, {"_rkwargs": True})
-
-
-print("""
-You are a Cellpose training operations assistant for BioEngine.
-
-Your job is to chat with users and execute Cellpose fine-tuning workflows via the utility functions in this runtime.
-
-Available async functions:
-- await connect_cellpose(server_url=None, workspace=None, token=None)
-- await list_cellpose_services()
-- await select_cellpose_service(service_id=None)
-- await start_cellpose_training(**kwargs)
-- await get_training_status(session_id)
-- await list_training_sessions(status_types=None)
-- await stop_training(session_id)
-- await restart_training(session_id, n_epochs=None)
-- await infer_cellpose(**kwargs)
-- await export_model(session_id, model_name=None, collection='bioimage-io/colab-annotations')
-
-Behavior requirements:
-1) Always ensure connectivity first (connect_cellpose, then select_cellpose_service).
-2) For training requests, collect missing required fields before calling start:
-   - artifact
-   - one of:
-     a) train_images + train_annotations
-     b) metadata_dir
-     c) BioImage Archive URL in artifact
-3) Respect split behavior:
-   - For BioImage Archive URLs, use split_mode='auto'.
-   - For auto split, include train_split_ratio if user specifies it.
-   - If user gives sample percentage, convert to decimal and pass as n_samples (0 < n_samples <= 1).
-4) After starting training, report session_id and offer to monitor progress.
-5) Monitoring loop: poll get_training_status(session_id) and summarize key fields:
-   - status_type, message, current_epoch/total_epochs, elapsed_seconds, latest train/test losses.
-6) On errors, provide actionable remediation (artifact path mismatch, missing annotations, token/workspace mismatch, resource limits).
-7) Keep outputs concise and user-friendly; include exact IDs and parameter values used.
-
-Examples:
-- Start from artifact paths:
-  await start_cellpose_training(
-      artifact='workspace/dataset',
-      train_images='images/*/*.tif',
-      train_annotations='annotations/*/*_mask.ome.tif',
-      split_mode='auto',
-      train_split_ratio=0.8,
-      n_samples=0.6,
-      model='cpsam',
-      n_epochs=20,
-      learning_rate=1e-5,
-  )
-
-- Start from metadata:
-  await start_cellpose_training(
-      artifact='workspace/dataset',
-      metadata_dir='metadata/',
-      n_epochs=10,
-      model='cpsam'
-  )
-
-- BioImage Archive:
-  await start_cellpose_training(
-      artifact='https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD1392',
-      split_mode='auto',
-      train_split_ratio=0.8,
-      n_samples=1.0,
-      n_epochs=10,
-      model='cpsam'
-  )
-""")
diff --git a/bioengine_apps/cellpose_finetuning/index.html b/bioengine_apps/cellpose_finetuning/index.html
index 3c17daea..b6e4ccf0 100644
--- a/bioengine_apps/cellpose_finetuning/index.html
+++ b/bioengine_apps/cellpose_finetuning/index.html
@@ -216,15 +216,35 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                         <div v-if="syncError" class="text-red-500 text-sm px-4 py-1 bg-red-50 rounded-full border border-red-200 flex items-center" title="Connection to worker is unstable">
                             <i class="fas fa-exclamation-triangle mr-2"></i> [[ syncError ]]
                         </div>
-                        <button @click="refreshSessions(false)" class="text-gray-500 hover:text-blue-600 transition text-sm font-medium bg-white px-3 py-1.5 rounded-lg border border-gray-200 shadow-sm">
-                            <i class="fas fa-sync-alt mr-1" :class="{'fa-spin': refreshLoading}"></i> Refresh
-                        </button>
+                        <div class="flex items-center gap-2">
+                            <button
+                                v-if="selectedSessionIds.length > 0"
+                                @click="deleteSelectedSessions"
+                                :disabled="isDeletingSession"
+                                class="text-slate-700 hover:text-slate-900 transition text-sm font-medium bg-white px-3 py-1.5 rounded-lg border border-slate-300 shadow-sm disabled:opacity-50"
+                            >
+                                <i class="fas fa-trash mr-1"></i> Delete Selected ([[ selectedSessionIds.length ]])
+                            </button>
+                            <button @click="refreshSessions(false)" class="text-gray-500 hover:text-blue-600 transition text-sm font-medium bg-white px-3 py-1.5 rounded-lg border border-gray-200 shadow-sm">
+                                <i class="fas fa-sync-alt mr-1" :class="{'fa-spin': refreshLoading}"></i> Refresh
+                            </button>
+                        </div>
                     </div>
 
                     <div class="bg-white rounded-xl shadow-sm border border-gray-200 overflow-hidden">
                          <table class="min-w-full divide-y divide-gray-200">
                             <thead class="bg-gray-50/50">
                                 <tr>
+                                    <th class="px-4 py-3 text-left text-xs font-semibold text-gray-500 uppercase tracking-wider">
+                                        <input
+                                            type="checkbox"
+                                            :checked="allNonRunningSelected"
+                                            @change="toggleSelectAllNonRunning($event.target.checked)"
+                                            title="Select all non-running sessions"
+                                            aria-label="Select all non-running sessions"
+                                            class="rounded border-gray-300 text-blue-600 focus:ring-blue-500"
+                                        >
+                                    </th>
                                     <th class="px-6 py-3 text-left text-xs font-semibold text-gray-500 uppercase tracking-wider">Session ID</th>
                                     <th class="px-6 py-3 text-left text-xs font-semibold text-gray-500 uppercase tracking-wider">Status</th>
                                     <th class="px-6 py-3 text-left text-xs font-semibold text-gray-500 uppercase tracking-wider">Progress</th>
@@ -235,6 +255,17 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                             </thead>
                             <tbody class="bg-white divide-y divide-gray-200">
                                 <tr v-for="(session, id) in sessions" :key="id" class="hover:bg-slate-50 transition-colors cursor-pointer group" @click="viewSession(id)">
+                                    <td class="px-4 py-4 whitespace-nowrap" @click.stop>
+                                        <input
+                                            type="checkbox"
+                                            :checked="selectedSessionIds.includes(id)"
+                                            :disabled="session.status_type === 'running'"
+                                            @change="toggleSessionSelection(id, $event.target.checked)"
+                                            title="Select for batch delete"
+                                            aria-label="Select for batch delete"
+                                            class="rounded border-gray-300 text-blue-600 focus:ring-blue-500 disabled:opacity-40"
+                                        >
+                                    </td>
                                     <td class="px-6 py-4 whitespace-nowrap">
                                         <div class="flex items-center">
                                             <span class="font-mono text-sm text-blue-600 font-medium group-hover:underline">[[ id.slice(0, 8) ]]</span>
@@ -267,12 +298,36 @@ <h2 class="text-2xl font-bold text-gray-900">Dashboard</h2>
                                         [[ formatDate(session.start_time) ]]
                                     </td>
                                     <td class="px-6 py-4 whitespace-nowrap text-right text-sm font-medium" @click.stop>
-                                         <button @click="viewSession(id)" class="text-blue-600 hover:text-blue-900 mr-3">Details</button>
-                                         <button v-if="session.status_type === 'running'" @click="stopTraining(id)" class="text-red-600 hover:text-red-900">Stop</button>
+                                                      <button
+                                                          @click="viewSession(id)"
+                                                          class="inline-flex items-center justify-center w-8 h-8 rounded text-blue-600 hover:text-blue-900 hover:bg-blue-50 mr-1"
+                                                          title="View session details"
+                                                          aria-label="View session details"
+                                                      >
+                                                          <i class="fas fa-chart-area"></i>
+                                                      </button>
+                                                      <button
+                                                          v-if="session.status_type === 'running'"
+                                                          @click="stopTraining(id)"
+                                                          class="inline-flex items-center justify-center w-8 h-8 rounded text-red-600 hover:text-red-900 hover:bg-red-50 mr-1"
+                                                          title="Stop session"
+                                                          aria-label="Stop session"
+                                                      >
+                                                          <i class="fas fa-stop"></i>
+                                                      </button>
+                                                      <button
+                                                          v-if="session.status_type !== 'running'"
+                                                          @click="deleteTrainingSession(id)"
+                                                          class="inline-flex items-center justify-center w-8 h-8 rounded text-slate-700 hover:text-slate-900 hover:bg-slate-100"
+                                                          title="Delete session"
+                                                          aria-label="Delete session"
+                                                      >
+                                                          <i class="fas fa-trash"></i>
+                                                      </button>
                                     </td>
                                 </tr>
                                 <tr v-if="Object.keys(sessions).length === 0">
-                                    <td colspan="6" class="px-6 py-12 text-center text-gray-500 bg-gray-50 rounded-b-xl">
+                                    <td colspan="7" class="px-6 py-12 text-center text-gray-500 bg-gray-50 rounded-b-xl">
                                         <div class="text-4xl mb-2">📭</div>
                                         <p>No training sessions found.</p>
                                         <button @click="currentTab = 'training'" class="mt-4 text-blue-600 font-medium hover:underline">Start a new training session</button>
@@ -463,6 +518,10 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                         <input v-model="trainingParams.metadata_dir" @input="markPathFieldEdited('metadata_dir')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="e.g. metadata/">
                                         <p class="text-xs text-gray-500 mt-1">Use this to load image/annotation pairs from metadata JSON files; then explicit annotation masks are optional.</p>
                                     </div>
+                                    <div v-if="trainingParams.dataset_source !== 'BioImage Archive' && configControlsLocked" class="md:col-span-2 rounded-md border border-amber-200 bg-amber-50 px-3 py-2 text-xs text-amber-800">
+                                        <i class="fas fa-info-circle mr-1"></i>
+                                        Path settings changed. Click <span class="font-semibold">Validate paths</span> to unlock split and sample controls.
+                                    </div>
                                 </div>
                                 <div v-else class="md:col-span-2 rounded-lg border border-blue-200 bg-blue-50 px-4 py-3 text-sm text-blue-900">
                                     BioImage Archive mode discovers image/mask pairs from the archive URL automatically.
@@ -472,18 +531,18 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                     <div>
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-2">Total Samples Used (Train + Validation)</label>
                                         <div class="grid grid-cols-1 md:grid-cols-3 gap-3 items-center">
-                                            <input v-model.number="sampleUsagePercent" type="range" min="1" max="100" step="1" class="w-full md:col-span-2 accent-blue-600">
+                                            <input v-model.number="sampleUsagePercent" :disabled="configControlsLocked" type="range" min="1" max="100" step="1" class="w-full md:col-span-2 accent-blue-600 disabled:opacity-50 disabled:cursor-not-allowed">
                                             <div class="flex gap-2">
                                                 <div class="flex-1">
                                                     <label class="block text-[11px] text-gray-500 mb-1">Percent</label>
                                                     <div class="relative">
-                                                        <input v-model.number="sampleUsagePercent" type="number" min="1" max="100" step="1" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 pr-7 text-sm">
+                                                        <input v-model.number="sampleUsagePercent" :disabled="configControlsLocked" type="number" min="1" max="100" step="1" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 pr-7 text-sm disabled:opacity-50 disabled:cursor-not-allowed">
                                                         <span class="absolute right-2 top-1/2 -translate-y-1/2 text-xs text-gray-400">%</span>
                                                     </div>
                                                 </div>
                                                 <div class="flex-1">
                                                     <label class="block text-[11px] text-gray-500 mb-1">Decimal</label>
-                                                    <input v-model.number="sampleUsageDecimal" type="number" min="0.01" max="1" step="0.01" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 text-sm">
+                                                    <input v-model.number="sampleUsageDecimal" :disabled="configControlsLocked" type="number" min="0.01" max="1" step="0.01" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 text-sm disabled:opacity-50 disabled:cursor-not-allowed">
                                                 </div>
                                             </div>
                                         </div>
@@ -497,11 +556,11 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                         <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-2">Train/Test Split</label>
                                         <div class="flex flex-wrap items-center gap-5 text-sm text-gray-700">
                                             <label v-if="trainingParams.dataset_source !== 'BioImage Archive'" class="inline-flex items-center gap-2 cursor-pointer">
-                                                <input type="radio" value="manual" v-model="trainingParams.split_mode" class="text-blue-600 focus:ring-blue-500">
+                                                <input type="radio" value="manual" v-model="trainingParams.split_mode" :disabled="configControlsLocked" class="text-blue-600 focus:ring-blue-500 disabled:opacity-50 disabled:cursor-not-allowed">
                                                 <span>Manual test paths</span>
                                             </label>
                                             <label class="inline-flex items-center gap-2 cursor-pointer">
-                                                <input type="radio" value="auto" v-model="trainingParams.split_mode" class="text-blue-600 focus:ring-blue-500">
+                                                <input type="radio" value="auto" v-model="trainingParams.split_mode" :disabled="configControlsLocked" class="text-blue-600 focus:ring-blue-500 disabled:opacity-50 disabled:cursor-not-allowed">
                                                 <span>Auto split</span>
                                             </label>
                                             <span v-if="trainingParams.dataset_source === 'BioImage Archive'" class="text-xs text-blue-700 bg-blue-100 border border-blue-200 px-2 py-0.5 rounded">BioImage Archive uses auto split</span>
@@ -509,18 +568,18 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
 
                                         <div v-if="trainingParams.split_mode === 'auto'" class="mt-3 space-y-2">
                                             <div class="grid grid-cols-1 md:grid-cols-3 gap-3 items-center">
-                                                <input v-model.number="trainSplitPercent" type="range" min="5" max="95" step="1" class="w-full md:col-span-2 accent-blue-600">
+                                                <input v-model.number="trainSplitPercent" :disabled="configControlsLocked" type="range" min="5" max="95" step="1" class="w-full md:col-span-2 accent-blue-600 disabled:opacity-50 disabled:cursor-not-allowed">
                                                 <div class="flex gap-2">
                                                     <div class="flex-1">
                                                         <label class="block text-[11px] text-gray-500 mb-1">Train %</label>
                                                         <div class="relative">
-                                                            <input v-model.number="trainSplitPercent" type="number" min="5" max="95" step="1" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 pr-7 text-sm">
+                                                            <input v-model.number="trainSplitPercent" :disabled="configControlsLocked" type="number" min="5" max="95" step="1" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 pr-7 text-sm disabled:opacity-50 disabled:cursor-not-allowed">
                                                             <span class="absolute right-2 top-1/2 -translate-y-1/2 text-xs text-gray-400">%</span>
                                                         </div>
                                                     </div>
                                                     <div class="flex-1">
                                                         <label class="block text-[11px] text-gray-500 mb-1">Train ratio</label>
-                                                        <input v-model.number="trainingParams.train_split_ratio" type="number" min="0.05" max="0.95" step="0.01" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 text-sm">
+                                                        <input v-model.number="trainingParams.train_split_ratio" :disabled="configControlsLocked" type="number" min="0.05" max="0.95" step="0.01" class="w-full bg-white border border-gray-300 rounded-md px-2 py-1.5 text-sm disabled:opacity-50 disabled:cursor-not-allowed">
                                                     </div>
                                                 </div>
                                             </div>
@@ -530,11 +589,11 @@ <h3 class="text-lg font-semibold text-gray-800 flex items-center gap-2"><i class
                                         <div v-else-if="trainingParams.dataset_source !== 'BioImage Archive'" class="mt-3 grid grid-cols-1 md:grid-cols-2 gap-4">
                                             <div>
                                                 <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Images (Opt)</label>
-                                                <input v-model="trainingParams.test_images" @input="markPathFieldEdited('test_images')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
+                                                <input v-model="trainingParams.test_images" :disabled="configControlsLocked" @input="markPathFieldEdited('test_images')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm disabled:opacity-50 disabled:cursor-not-allowed" placeholder="Optional">
                                             </div>
                                             <div>
                                                 <label class="block text-xs font-medium text-gray-500 uppercase tracking-wide mb-1">Test Annotations (Opt)</label>
-                                                <input v-model="trainingParams.test_annotations" @input="markPathFieldEdited('test_annotations')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm" placeholder="Optional">
+                                                <input v-model="trainingParams.test_annotations" :disabled="configControlsLocked" @input="markPathFieldEdited('test_annotations')" class="w-full bg-white border border-gray-300 rounded-md px-3 py-2 text-sm disabled:opacity-50 disabled:cursor-not-allowed" placeholder="Optional">
                                             </div>
                                             <p class="md:col-span-2 text-xs text-gray-600">Estimated split from selected samples: train <span class="font-semibold">[[ selectedTrainSampleCount ?? '-' ]]</span>, validation <span class="font-semibold">[[ selectedValidationSampleCount ?? '-' ]]</span>.</p>
                                         </div>
@@ -565,6 +624,10 @@ <h3 class="text-lg font-semibold text-gray-800 mb-4 flex items-center gap-2"><i
 
                          <div class="flex justify-end gap-4">
                              <button type="button" @click="currentTab = 'dashboard'" class="px-6 py-2.5 text-gray-600 font-medium hover:bg-gray-100 rounded-lg transition">Cancel</button>
+                                      <button type="button" @click="copyTrainingParams" class="px-6 py-2.5 text-blue-700 bg-blue-50 border border-blue-200 font-medium hover:bg-blue-100 rounded-lg transition flex items-center gap-2">
+                                          <i class="fas fa-copy"></i>
+                                          Copy Params
+                                     </button>
                              <button type="submit" :disabled="isTrainingStartDisabled" class="bg-blue-600 text-white px-8 py-2.5 rounded-lg shadow-sm hover:bg-blue-700 hover:shadow-md transition-all font-medium disabled:opacity-50 flex items-center gap-2">
                                 <i v-if="isStartingTraining" class="fas fa-circle-notch fa-spin"></i>
                                 [[ isStartingTraining ? 'Starting...' : 'Start Training' ]]
@@ -600,12 +663,19 @@ <h2 class="text-2xl font-bold text-gray-900 flex items-center gap-3">
                                 <i class="fas fa-stop mr-1"></i> Stop
                             </button>
                             <button
-                                v-if="currentSession && ['stopped','unknown','failed'].includes(currentSession.status_type)"
+                                v-if="currentSession && ['stopped','unknown','failed','completed'].includes(currentSession.status_type)"
                                 @click="restartTraining(selectedSessionId)"
                                 class="px-4 py-2 rounded-lg bg-blue-600 text-white text-sm font-medium hover:bg-blue-700 transition"
                             >
                                 <i class="fas fa-rotate-right mr-1"></i> Restart
                             </button>
+                            <button
+                                v-if="currentSession && currentSession.status_type !== 'running'"
+                                @click="deleteTrainingSession(selectedSessionId)"
+                                class="px-4 py-2 rounded-lg bg-slate-700 text-white text-sm font-medium hover:bg-slate-800 transition"
+                            >
+                                <i class="fas fa-trash mr-1"></i> Delete
+                            </button>
                         </div>
                     </div>
 
@@ -941,6 +1011,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 // Training Data
                 const syncError = ref(null);
                 const sessions = ref({});
+                const selectedSessionIds = ref([]);
                 const selectedSessionId = ref(null);
                 const refreshLoading = ref(false);
                 
@@ -964,6 +1035,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     min_train_masks: 5,
                 });
                 const isStartingTraining = ref(false);
+                const isDeletingSession = ref(false);
                 const trainingError = ref('');
                 const sampleInventory = reactive({
                     trainPairs: 0,
@@ -1021,7 +1093,6 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 });
                 const lastAutoPathArtifact = ref('');
                 const autoDetectedPathsPristine = ref(false);
-                let continuousValidationTimer = null;
 
                 const SUPPORTED_IMAGE_EXT_RE = /\.(ome\.)?(tif|tiff|png|jpg|jpeg)$/i;
 
@@ -1070,7 +1141,6 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     if (workspace.value) workspaces.add(workspace.value);
                     const userWs = getUserWorkspace();
                     if (userWs) workspaces.add(userWs);
-                    workspaces.add('bioimage-io');
 
                     if (!window.hyphaClient) return Array.from(workspaces);
 
@@ -1507,6 +1577,14 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const monitoredSession = computed(() => monitoringSessionId.value ? sessions.value[monitoringSessionId.value] : null);
                 const hasMetrics = computed(() => currentSession.value?.test_metrics?.length > 0);
                 const hasValidationSet = computed(() => (currentSession.value?.n_test || 0) > 0);
+                const nonRunningSessionIds = computed(() => Object.entries(sessions.value)
+                    .filter(([_, session]) => session && session.status_type !== 'running')
+                    .map(([id]) => id));
+                const allNonRunningSelected = computed(() => {
+                    const candidates = nonRunningSessionIds.value;
+                    if (candidates.length === 0) return false;
+                    return candidates.every((id) => selectedSessionIds.value.includes(id));
+                });
                 const runningTrainingSessionsCount = computed(() =>
                     Object.values(sessions.value).filter((s) => s && s.status_type === 'running').length
                 );
@@ -1715,6 +1793,11 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     return !pathValidationOk.value;
                 });
 
+                const configControlsLocked = computed(() => {
+                    if (trainingParams.dataset_source === 'BioImage Archive') return false;
+                    return !pathValidationOk.value || isPathValidating.value;
+                });
+
                 const applyTrainingFolderExample = () => {
                     trainingParams.train_images = 'images/';
                     trainingParams.train_annotations = 'annotations/';
@@ -1725,6 +1808,12 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 const markPathFieldEdited = (field) => {
                     autoDetectedPathsPristine.value = false;
                     if (pathFieldsDirty[field] === false) pathFieldsDirty[field] = true;
+                    pathValidationOk.value = false;
+                    pathValidationMessage.value = 'Path settings changed. Click "Validate paths" to continue.';
+                    sampleInventory.trainPairs = 0;
+                    sampleInventory.testPairs = 0;
+                    sampleInventory.totalPairs = 0;
+                    sampleInventory.known = false;
                 };
 
                 const resetPathDirtyFlags = () => {
@@ -1990,10 +2079,10 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         throw new Error(`${phaseLabel}: no supported mask/annotation files found for '${annotationSpec}'.`);
                     }
 
-                    const annotationStemSet = new Set(annotationFiles.map((pathStr) => normalizePairStem(pathStr)));
-                    const pairedCount = imageFiles.filter((pathStr) => annotationStemSet.has(normalizePairStem(pathStr))).length;
+                    const pairedCount = Math.min(imageFiles.length, annotationFiles.length);
+
                     if (pairedCount === 0) {
-                        throw new Error(`${phaseLabel}: found files, but no image/mask pairs matched by filename stem.`);
+                        throw new Error(`${phaseLabel}: found files, but no trainable pairs can be assumed from file counts.`);
                     }
 
                     return {
@@ -2003,6 +2092,44 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     };
                 };
 
+                const callBackendTrainingPreflight = async (artifactId) => {
+                    if (!service.value || typeof service.value.preflight_training_dataset !== 'function') {
+                        return null;
+                    }
+
+                    const payload = {
+                        artifact: artifactId,
+                        train_images: String(trainingParams.train_images || '').trim() || null,
+                        train_annotations: String(trainingParams.train_annotations || '').trim() || null,
+                        metadata_dir: String(trainingParams.metadata_dir || '').trim() || null,
+                        test_images: String(trainingParams.test_images || '').trim() || null,
+                        test_annotations: String(trainingParams.test_annotations || '').trim() || null,
+                        split_mode: trainingParams.dataset_source === 'BioImage Archive'
+                            ? 'auto'
+                            : String(trainingParams.split_mode || 'manual').toLowerCase(),
+                        n_samples: Number(sampleUsageDecimal.value || 0) > 0
+                            ? Number(sampleUsageDecimal.value)
+                            : null,
+                        max_candidates: 500,
+                    };
+
+                    const attempts = [
+                        () => service.value.preflight_training_dataset(payload, { _rkwargs: true }),
+                        () => service.value.preflight_training_dataset(payload),
+                    ];
+
+                    for (const attempt of attempts) {
+                        try {
+                            const result = await attempt();
+                            if (result && typeof result === 'object') return result;
+                        } catch (_error) {
+                            // Try fallback call signature.
+                        }
+                    }
+
+                    return null;
+                };
+
                 const runPathAutoDetection = async (manualTrigger = false) => {
                     const artifactId = String(getSelectedArtifactId() || '').trim();
                     if (!isCompleteArtifactId(artifactId)) {
@@ -2060,8 +2187,8 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         sampleInventory.totalPairs = quickTrainResult.pairedCount;
                         sampleInventory.known = true;
                         if (quickTrainResult.pairedCount > 0) {
-                            pathValidationOk.value = true;
-                            pathValidationMessage.value = `Train pairs found: ${quickTrainResult.pairedCount}. Finishing full validation…`;
+                            pathValidationOk.value = false;
+                            pathValidationMessage.value = `Auto-detect hint: train files images=${quickTrainResult.imageCount}, masks=${quickTrainResult.annotationCount}, assumed pairs=${quickTrainResult.pairedCount}. Click Validate for backend confirmation.`;
                         }
 
                         if (trainingParams.split_mode === 'manual' && testImageDir && testMaskDir) {
@@ -2132,12 +2259,35 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     pathValidationMessage.value = '';
 
                     try {
+                        const backendPreflight = await callBackendTrainingPreflight(artifactId);
+                        if (backendPreflight) {
+                            const trainPairs = Number(backendPreflight.train_pair_count || 0);
+                            const testPairs = Number(backendPreflight.test_pair_count || 0);
+                            const totalPairs = Number(backendPreflight.sampled_total_count ?? (trainPairs + testPairs));
+
+                            sampleInventory.trainPairs = Math.max(0, trainPairs);
+                            sampleInventory.testPairs = Math.max(0, testPairs);
+                            sampleInventory.totalPairs = Math.max(0, totalPairs);
+                            sampleInventory.known = true;
+
+                            if (!backendPreflight.ok) {
+                                throw new Error(String(backendPreflight.message || 'Dataset preflight failed.'));
+                            }
+
+                            pathValidationOk.value = true;
+                            pathValidationMessage.value = String(
+                                backendPreflight.message ||
+                                `Dataset preflight passed. Train pairs=${trainPairs}, test pairs=${testPairs}.`
+                            );
+                            return true;
+                        }
+
                         const files = await loadArtifactFiles(artifactId);
                         const supportedFileCount = files.filter((pathStr) => isSupportedImagePath(pathStr)).length;
 
                         if (!hasMetadataDir) {
                             const trainResult = validatePairSpecs(files, trainingParams.train_images, trainingParams.train_annotations, 'Train set');
-                            let message = `Train pairs found: ${trainResult.pairedCount} (images: ${trainResult.imageCount}, masks: ${trainResult.annotationCount}).`;
+                            let message = `Train files: images=${trainResult.imageCount}, masks=${trainResult.annotationCount}, assumed pairs=${trainResult.pairedCount}.`;
                             let testPairCount = 0;
 
                             if (trainResult.pairedCount > 0) {
@@ -2151,7 +2301,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
 
                             if (hasManualTestPaths) {
                                 const testResult = validatePairSpecs(files, trainingParams.test_images, trainingParams.test_annotations, 'Test set');
-                                message += ` Test pairs found: ${testResult.pairedCount}.`;
+                                message += ` Test files: images=${testResult.imageCount}, masks=${testResult.annotationCount}, assumed pairs=${testResult.pairedCount}.`;
                                 testPairCount = testResult.pairedCount;
                             }
 
@@ -2189,17 +2339,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                 };
 
                 const scheduleContinuousPathValidation = () => {
-                    if (continuousValidationTimer) {
-                        clearTimeout(continuousValidationTimer);
-                    }
-                    continuousValidationTimer = setTimeout(async () => {
-                        if (trainingParams.dataset_source === 'BioImage Archive') {
-                            pathValidationOk.value = true;
-                            pathValidationMessage.value = '';
-                            return;
-                        }
-                        await validateTrainingPaths(true);
-                    }, 250);
+                    return;
                 };
 
                 const getUserWorkspace = () => {
@@ -2455,6 +2595,13 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                                 sessions.value = result;
                                 syncError.value = null;
 
+                                const existingNonRunning = new Set(
+                                    Object.entries(sessions.value)
+                                        .filter(([_, session]) => session && session.status_type !== 'running')
+                                        .map(([id]) => id)
+                                );
+                                selectedSessionIds.value = selectedSessionIds.value.filter((id) => existingNonRunning.has(id));
+
                                 // Find active session
                                 const running = Object.entries(sessions.value).find(([_, s]) => s && s.status_type === 'running');
                                 if (running) {
@@ -2536,7 +2683,6 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                             trainingParams.test_annotations = '';
                         }
                         autoDetectedPathsPristine.value = false;
-                        scheduleContinuousPathValidation();
                     }
                 );
 
@@ -2546,6 +2692,14 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                         const artifactId = String(getSelectedArtifactId() || '').trim();
                         if (artifactId !== lastAutoPathArtifact.value) {
                             autoDetectedPathsPristine.value = false;
+                            if (trainingParams.dataset_source !== 'BioImage Archive') {
+                                pathValidationOk.value = false;
+                                pathValidationMessage.value = 'Artifact changed. Click "Validate paths" to continue.';
+                                sampleInventory.trainPairs = 0;
+                                sampleInventory.testPairs = 0;
+                                sampleInventory.totalPairs = 0;
+                                sampleInventory.known = false;
+                            }
                         }
                         if (trainingParams.dataset_source === 'BioImage Archive') return;
                         if (!isCompleteArtifactId(artifactId)) return;
@@ -2588,25 +2742,6 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     }
                 );
 
-                watch(
-                    () => [
-                        trainingParams.dataset_source,
-                        trainingParams.artifact_input,
-                        trainingParams.uploaded_artifact,
-                        trainingParams.train_images,
-                        trainingParams.train_annotations,
-                        trainingParams.metadata_dir,
-                        trainingParams.test_images,
-                        trainingParams.test_annotations,
-                        trainingParams.split_mode,
-                        artifactIdExists.value,
-                        artifactLookupPending.value,
-                    ],
-                    () => {
-                        scheduleContinuousPathValidation();
-                    }
-                );
-
                 const viewSession = (id) => {
                     if (!sessions.value[id]) return;
                     selectedSessionId.value = id;
@@ -2696,6 +2831,25 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     }
                 };
 
+                const toggleSessionSelection = (id, checked) => {
+                    if (!id) return;
+                    const session = sessions.value[id];
+                    if (!session || session.status_type === 'running') return;
+
+                    const set = new Set(selectedSessionIds.value);
+                    if (checked) set.add(id);
+                    else set.delete(id);
+                    selectedSessionIds.value = Array.from(set);
+                };
+
+                const toggleSelectAllNonRunning = (checked) => {
+                    if (checked) {
+                        selectedSessionIds.value = [...nonRunningSessionIds.value];
+                    } else {
+                        selectedSessionIds.value = [];
+                    }
+                };
+
                 const restartTraining = async (id) => {
                     try {
                         const result = await service.value.restart_training(
@@ -2719,6 +2873,130 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     }
                 };
 
+                const deleteTrainingSession = async (id) => {
+                    if (isDeletingSession.value) return;
+                    if (!id) return;
+                    if (!confirm(`Delete session ${id}? This removes local session history, status, and checkpoints for this service instance.`)) return;
+
+                    isDeletingSession.value = true;
+                    try {
+                        await service.value.delete_training_session(
+                            { session_id: id, force_stop_if_blocked: true },
+                            { _rkwargs: true }
+                        );
+
+                        if (monitoringSessionId.value === id) {
+                            monitoringSessionId.value = null;
+                        }
+                        if (selectedSessionId.value === id) {
+                            selectedSessionId.value = null;
+                            currentTab.value = 'dashboard';
+                        }
+                        if (sessions.value[id]) {
+                            delete sessions.value[id];
+                        }
+                        selectedSessionIds.value = selectedSessionIds.value.filter((value) => value !== id);
+
+                        await refreshSessions(true);
+                    } catch (e) {
+                        const errMsg = formatServiceError(e, 'Delete session');
+                        trainingError.value = "Delete failed: " + errMsg;
+                        alert("Delete failed: " + errMsg);
+                    } finally {
+                        isDeletingSession.value = false;
+                    }
+                };
+
+                const deleteSelectedSessions = async () => {
+                    if (isDeletingSession.value) return;
+                    const targets = selectedSessionIds.value.filter((id) => {
+                        const session = sessions.value[id];
+                        return session && session.status_type !== 'running';
+                    });
+                    if (targets.length === 0) return;
+
+                    if (!confirm(`Delete ${targets.length} selected session(s)? This removes local session history, status, and checkpoints for this service instance.`)) {
+                        return;
+                    }
+
+                    isDeletingSession.value = true;
+                    const failures = [];
+                    try {
+                        for (const id of targets) {
+                            try {
+                                await service.value.delete_training_session(
+                                    { session_id: id, force_stop_if_blocked: true },
+                                    { _rkwargs: true }
+                                );
+                                if (monitoringSessionId.value === id) monitoringSessionId.value = null;
+                                if (selectedSessionId.value === id) {
+                                    selectedSessionId.value = null;
+                                    currentTab.value = 'dashboard';
+                                }
+                                if (sessions.value[id]) delete sessions.value[id];
+                            } catch (error) {
+                                failures.push({ id, error: formatServiceError(error, 'Delete session') });
+                            }
+                        }
+
+                        selectedSessionIds.value = [];
+                        await refreshSessions(true);
+
+                        if (failures.length > 0) {
+                            const lines = failures.slice(0, 5).map((f) => `- ${f.id}: ${f.error}`);
+                            alert(`Deleted ${targets.length - failures.length}/${targets.length} sessions.\n\nFailures:\n${lines.join('\n')}`);
+                        }
+                    } finally {
+                        isDeletingSession.value = false;
+                    }
+                };
+
+                const getTrainingCopyPayload = () => {
+                    const source = trainingParams.dataset_source;
+                    const artifactOrUrl = source === 'BioImage Archive'
+                        ? String(trainingParams.bioimage_archive_url || '').trim()
+                        : String(getSelectedArtifactId() || '').trim();
+                    const sampleUsageRatio = Number((sampleUsagePercent.value / 100).toFixed(4));
+
+                    return {
+                        dataset_source: source,
+                        artifact_or_bia_url: artifactOrUrl,
+                        artifact_id: source === 'BioImage Archive' ? null : artifactOrUrl,
+                        bioimage_archive_url: source === 'BioImage Archive' ? artifactOrUrl : null,
+                        model: trainingParams.model || 'cpsam',
+                        train_images: source === 'BioImage Archive' ? null : (trainingParams.train_images || null),
+                        train_annotations: source === 'BioImage Archive' ? null : (trainingParams.train_annotations || null),
+                        metadata_dir: source === 'BioImage Archive' ? null : (trainingParams.metadata_dir || null),
+                        split_mode: source === 'BioImage Archive' ? 'auto' : (trainingParams.split_mode || 'manual'),
+                        train_split_ratio: Number(trainingParams.train_split_ratio || 0.8),
+                        test_images: (source !== 'BioImage Archive' && trainingParams.split_mode === 'manual') ? (trainingParams.test_images || null) : null,
+                        test_annotations: (source !== 'BioImage Archive' && trainingParams.split_mode === 'manual') ? (trainingParams.test_annotations || null) : null,
+                        n_samples_ratio: sampleUsageRatio,
+                        n_samples_percent: Number(sampleUsagePercent.value || 0),
+                        selected_samples_estimate: selectedTotalSampleCount.value,
+                        selected_train_samples_estimate: selectedTrainSampleCount.value,
+                        selected_validation_samples_estimate: selectedValidationSampleCount.value,
+                        available_pairs_total: availableTotalSampleCount.value,
+                        available_pairs_train: sampleInventory.known ? sampleInventory.trainPairs : null,
+                        available_pairs_test: sampleInventory.known ? sampleInventory.testPairs : null,
+                        n_epochs: trainingParams.n_epochs,
+                        learning_rate: trainingParams.learning_rate,
+                        weight_decay: 1e-4,
+                        min_train_masks: trainingParams.min_train_masks,
+                        validation_interval: null,
+                    };
+                };
+
+                const copyTrainingParams = async () => {
+                    const payload = getTrainingCopyPayload();
+                    const text = JSON.stringify(payload, null, 2);
+                    try {
+                        await navigator.clipboard.writeText(text);
+                    } catch (_e) {
+                        window.prompt('Copy training parameters:', text);
+                    }
+                };
+
                 const startTraining = async () => {
                     console.log("startTraining called");
                     isStartingTraining.value = true;
@@ -3545,9 +3823,9 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     serverUrl, isConnected, isConnecting, workspace, showTokenInput, authToken, autoConnectError,
                     availableServices, selectedServiceId, service, isScanning, listServices, disconnect, connect, login,
                     currentTab, tabs,
-                    sessions, refreshSessions, refreshLoading, monitoringSessionId, monitoredSession, selectedSessionId, viewSession, currentSession, hasMetrics, hasValidationSet, stopTraining, restartTraining,
+                    sessions, selectedSessionIds, allNonRunningSelected, toggleSessionSelection, toggleSelectAllNonRunning, deleteSelectedSessions, refreshSessions, refreshLoading, monitoringSessionId, monitoredSession, selectedSessionId, viewSession, currentSession, hasMetrics, hasValidationSet, stopTraining, restartTraining, deleteTrainingSession,
                     runningTrainingSessionsCount,
-                    trainingParams, startTraining, isStartingTraining,
+                    trainingParams, startTraining, isStartingTraining, copyTrainingParams,
                     sampleUsagePercent, sampleUsageDecimal, trainSplitPercent,
                     availableTotalSampleCount, selectedTotalSampleCount, selectedTrainSampleCount, selectedValidationSampleCount,
                     applyTrainingGlobExample, applyTrainingFolderExample, runPathAutoDetection, validateTrainingPaths,
@@ -3563,7 +3841,7 @@ <h3 class="text-xl font-bold mb-1 text-gray-900">Upload Dataset</h3>
                     suggestions, showSuggestions, artifactInputError, artifactIdExists, artifactLookupPending, handleArtifactInput, handleArtifactFocus, selectSuggestion, acceptSuggestion, closeSuggestions,
 
                     // Path auto-detect + validation
-                    isPathDetecting, isPathValidating, pathAutoHint, pathAutoError, showAutoDetectButton, pathValidationMessage, pathValidationOk, markPathFieldEdited, isTrainingStartDisabled,
+                    isPathDetecting, isPathValidating, pathAutoHint, pathAutoError, showAutoDetectButton, pathValidationMessage, pathValidationOk, markPathFieldEdited, isTrainingStartDisabled, configControlsLocked,
 
                     // Helpers
                     statusTextColor, formatDuration, handleFolderSelect, getLatestLoss, formatDate, getSelectedArtifactId,
diff --git a/bioengine_apps/cellpose_finetuning/main.py b/bioengine_apps/cellpose_finetuning/main.py
index 8d8852ec..b1832b8f 100644
--- a/bioengine_apps/cellpose_finetuning/main.py
+++ b/bioengine_apps/cellpose_finetuning/main.py
@@ -51,6 +51,11 @@
 STATUS_STALE_SECONDS = 300
 BIA_FTS_ENDPOINT = "https://beta.bioimagearchive.org/search/search/fts"
 BIA_IMAGE_ENDPOINT = "https://beta.bioimagearchive.org/search/search/fts/image"
+MIN_FREE_GPU_MEMORY_TO_START_BYTES = 2 * GB
+BIA_RESOLVE_URL_SERVICE_ID = os.environ.get("BIA_RESOLVE_URL_SERVICE_ID")
+BIA_RESOLVE_URL_APPLICATION_ID = os.environ.get(
+    "BIA_RESOLVE_URL_APPLICATION_ID", "bia-resolve-url-proxy"
+)
 
 # Model template for BioImage.io export
 MODEL_TEMPLATE_PY = '''"""BioImage.io Model Wrapper for Cellpose 4.0.7 (Cellpose-SAM).
@@ -1112,6 +1117,12 @@ def train_seg_with_callbacks(
         nimg_test_val if nimg_test_per_epoch is None else nimg_test_per_epoch
     )
 
+    if nimg_per_epoch_val <= 0:
+        raise ValueError(
+            "No training samples available after dataset filtering. "
+            "Try increasing n_samples or lowering min_train_masks."
+        )
+
     # learning rate schedule
     learning_rate_schedule = np.linspace(0, learning_rate, 10)
     learning_rate_schedule = np.append(
@@ -1408,8 +1419,87 @@ def run_blocking_task(
     """Run the blocking training task."""
     import time
 
+    from cellpose import io as cp_io  # type: ignore
+
     session_id = training_params["session_id"]
 
+    def _is_readable_pair(
+        image_path: Path, label_path: Path
+    ) -> tuple[bool, str | None]:
+        try:
+            image_array = cp_io.imread(str(image_path))
+            label_array = cp_io.imread(str(label_path))
+        except Exception as exc:
+            return False, str(exc)
+
+        if image_array is None:
+            return False, "image could not be decoded"
+        if label_array is None:
+            return False, "annotation could not be decoded"
+
+        if not hasattr(image_array, "ndim") or int(image_array.ndim) < 2:
+            return False, "image has invalid ndim"
+        if not hasattr(label_array, "ndim") or int(label_array.ndim) < 2:
+            return False, "annotation has invalid ndim"
+
+        return True, None
+
+    train_files = list(dataset_split["train_files"])
+    train_labels_files = list(dataset_split["train_labels_files"])
+    test_files = list(dataset_split["test_files"] or [])
+    test_labels_files = list(dataset_split["test_labels_files"] or [])
+
+    filtered_train_files: list[Path] = []
+    filtered_train_labels: list[Path] = []
+    dropped_train: list[str] = []
+    for image_path, label_path in zip(train_files, train_labels_files):
+        ok, reason = _is_readable_pair(image_path, label_path)
+        if ok:
+            filtered_train_files.append(image_path)
+            filtered_train_labels.append(label_path)
+        else:
+            dropped_train.append(f"{image_path} <-> {label_path} ({reason})")
+
+    filtered_test_files: list[Path] = []
+    filtered_test_labels: list[Path] = []
+    dropped_test: list[str] = []
+    for image_path, label_path in zip(test_files, test_labels_files):
+        ok, reason = _is_readable_pair(image_path, label_path)
+        if ok:
+            filtered_test_files.append(image_path)
+            filtered_test_labels.append(label_path)
+        else:
+            dropped_test.append(f"{image_path} <-> {label_path} ({reason})")
+
+    if dropped_train or dropped_test:
+        append_info(
+            session_id,
+            (
+                f"Skipped unreadable pairs: train={len(dropped_train)}, "
+                f"test={len(dropped_test)}"
+            ),
+            with_time=True,
+        )
+        logger.warning(
+            "Session %s: skipped unreadable pairs (train=%d, test=%d)",
+            session_id,
+            len(dropped_train),
+            len(dropped_test),
+        )
+
+    if len(filtered_train_files) == 0:
+        raise ValueError(
+            "No readable training pairs remain after file validation. "
+            "Please verify that selected image/annotation files are valid TIFF/PNG/JPEG masks and images."
+        )
+
+    dataset_split = DatasetSplit(
+        train_files=filtered_train_files,
+        train_labels_files=filtered_train_labels,
+        test_files=filtered_test_files if filtered_test_files else None,
+        test_labels_files=filtered_test_labels if filtered_test_labels else None,
+    )
+
     # Calculate dataset sizes
     n_train = len(dataset_split["train_files"])
     n_test = (
@@ -1996,6 +2086,26 @@ def _contains_glob(path_pattern: str) -> bool:
     return "*" in path_pattern
 
 
+SUPPORTED_DATASET_IMAGE_SUFFIXES = (
+    ".ome.tiff",
+    ".ome.tif",
+    ".tiff",
+    ".tif",
+    ".png",
+    ".jpg",
+    ".jpeg",
+)
+
+
+def _is_supported_dataset_image_path(path_value: str) -> bool:
+    normalized = _normalize_artifact_relpath(path_value).lower().rstrip("/")
+    return normalized.endswith(SUPPORTED_DATASET_IMAGE_SUFFIXES)
+
+
+def _filter_supported_dataset_image_paths(paths: list[str]) -> list[str]:
+    return [path for path in paths if _is_supported_dataset_image_path(path)]
+
+
 def _glob_base_folder(path_pattern: str) -> str:
     normalized = _normalize_artifact_relpath(path_pattern)
     wildcard_index = normalized.find("*")
@@ -2014,11 +2124,97 @@ def _glob_base_folder(path_pattern: str) -> str:
 
 def _is_bioimage_archive_url(candidate: str) -> bool:
     parsed = urlparse(str(candidate or ""))
-    return (
-        parsed.scheme in {"http", "https"}
-        and "bioimagearchive.org" in parsed.netloc.lower()
+    host = parsed.netloc.lower()
+    path = parsed.path.lower()
+    return parsed.scheme in {"http", "https"} and (
+        "bioimagearchive.org" in host
+        or ("ebi.ac.uk" in host and "/biostudies/bioimages" in path)
+    )
+
+
+async def _get_bia_resolver_service() -> tuple[Any, Any]:
+    """Connect to Hypha and resolve the BIA URL resolver service."""
+    from hypha_rpc import connect_to_server
+
+    server_url = os.environ.get("HYPHA_SERVER_URL", DEFAULT_SERVER_URL)
+    workspace = os.environ.get("BIA_RESOLVE_URL_WORKSPACE", "ri-scale")
+    token = (
+        os.environ.get("RI_SCALE_TOKEN")
+        or os.environ.get("HYPHA_TOKEN")
+        or os.environ.get("BIOENGINE_HYPHA_TOKEN")
+    )
+    if not token:
+        raise RuntimeError(
+            "Missing token for resolver service connection. "
+            "Set RI_SCALE_TOKEN or HYPHA_TOKEN in environment."
+        )
+
+    server: Any = await connect_to_server(
+        {
+            "server_url": server_url,
+            "token": token,
+            "workspace": workspace,
+        }
     )
 
+    resolved_service_id = BIA_RESOLVE_URL_SERVICE_ID
+    if not resolved_service_id:
+        worker_id = os.environ.get("HYPHA_WORKER_SERVICE_ID", "bioimage-io/bioengine-worker")
+        worker: Any = await server.get_service(worker_id)
+        app_status = await worker.get_application_status(
+            application_ids=[BIA_RESOLVE_URL_APPLICATION_ID]
+        )
+        app = (app_status or {}).get(BIA_RESOLVE_URL_APPLICATION_ID, {})
+        service_ids = app.get("service_ids") or []
+        for item in service_ids:
+            if isinstance(item, dict) and item.get("websocket_service_id"):
+                resolved_service_id = item.get("websocket_service_id")
+                break
+            if isinstance(item, str) and item:
+                resolved_service_id = item
+                break
+
+    if not isinstance(resolved_service_id, str) or not resolved_service_id:
+        await server.disconnect()
+        raise RuntimeError(
+            "Could not resolve BIA resolver service ID. "
+            "Set BIA_RESOLVE_URL_SERVICE_ID or ensure application "
+            f"'{BIA_RESOLVE_URL_APPLICATION_ID}' is running on the worker."
+        )
+
+    service: Any = await server.get_service(resolved_service_id)
+    return server, service
+
+
+async def _download_remote_bytes_via_resolver(
+    resolver_service: Any,
+    url: str,
+    *,
+    timeout_seconds: float = 120.0,
+) -> bytes:
+    """Download binary content through resolver service."""
+    payload = await resolver_service.resolve_url(
+        url=url,
+        method="GET",
+        timeout=timeout_seconds,
+    )
+    if not isinstance(payload, dict):
+        raise RuntimeError(f"Resolver returned invalid payload type for {url}")
+
+    if not payload.get("ok"):
+        error = payload.get("error") or f"HTTP {payload.get('status_code', 'unknown')}"
+        raise RuntimeError(f"Resolver request failed for {url}: {error}")
+
+    content_base64 = payload.get("content_base64")
+    if isinstance(content_base64, str) and content_base64:
+        return base64.b64decode(content_base64)
+
+    text_value = payload.get("text")
+    if isinstance(text_value, str):
+        return text_value.encode("utf-8")
+
+    raise RuntimeError(f"Resolver returned no content for {url}")
+
 
 def _extract_bia_accession(url: str) -> str | None:
     match = re.search(r"(S-BIAD\d+)", str(url or ""), flags=re.IGNORECASE)
@@ -2043,7 +2239,22 @@ def _extract_tsv_pairs(tsv_content: str) -> list[tuple[str, str]]:
     return pairs
 
 
-async def _fetch_bia_tsv_pairs(client: Any, accession: str) -> list[tuple[str, str]]:
+def _bia_pair_cap_from_n_samples(n_samples: Any) -> int | None:
+    if isinstance(n_samples, bool):
+        return None
+    if isinstance(n_samples, int) and n_samples > 0:
+        return max(20, n_samples * 8)
+    if isinstance(n_samples, float) and float(n_samples).is_integer() and n_samples > 0:
+        as_int = int(n_samples)
+        return max(20, as_int * 8)
+    return None
+
+
+async def _fetch_bia_tsv_pairs(
+    client: Any,
+    accession: str,
+    max_pairs: int | None = None,
+) -> list[tuple[str, str]]:
     base_url = _bia_ftp_base_url(accession)
     candidate_paths = [
         f"{base_url}/Files/ps_ovule_labels.tsv",
@@ -2053,11 +2264,48 @@ async def _fetch_bia_tsv_pairs(client: Any, accession: str) -> list[tuple[str, s
 
     for tsv_url in candidate_paths:
         try:
-            response = await client.get(tsv_url)
-            if response.status_code // 100 != 2 or not response.text.strip():
-                continue
+            pairs: list[tuple[str, str]] = []
+            async with client.stream("GET", tsv_url) as response:
+                if response.status_code // 100 != 2:
+                    continue
+
+                header: list[str] | None = None
+                label_idx: int | None = None
+                source_idx: int | None = None
+
+                async for line in response.aiter_lines():
+                    if not line:
+                        continue
+
+                    row = next(csv.reader([line], delimiter="\t"), [])
+                    if not row:
+                        continue
+
+                    if header is None:
+                        header = [str(col).strip().lower() for col in row]
+                        try:
+                            label_idx = header.index("files")
+                            source_idx = header.index("source image")
+                        except ValueError:
+                            label_idx = None
+                            source_idx = None
+                            break
+                        continue
+
+                    if label_idx is None or source_idx is None:
+                        continue
+                    if len(row) <= max(label_idx, source_idx):
+                        continue
+
+                    label_path = str(row[label_idx] or "").strip()
+                    source_image = str(row[source_idx] or "").strip()
+                    if not label_path or not source_image:
+                        continue
+
+                    pairs.append((source_image, label_path))
+                    if max_pairs is not None and len(pairs) >= max_pairs:
+                        break
 
-            pairs = _extract_tsv_pairs(response.text)
             if not pairs:
                 continue
 
@@ -2208,7 +2456,8 @@ async def make_training_pairs_from_bioimage_archive_url(
         )
 
     async with httpx.AsyncClient(timeout=120) as client:
-        paired_urls = await _fetch_bia_tsv_pairs(client, accession)
+        pair_cap = _bia_pair_cap_from_n_samples(config.get("n_samples"))
+        paired_urls = await _fetch_bia_tsv_pairs(client, accession, max_pairs=pair_cap)
 
         if not paired_urls:
             payloads = [
@@ -2260,39 +2509,124 @@ async def make_training_pairs_from_bioimage_archive_url(
                     "No image/mask pairs could be inferred from BioImage Archive assets."
                 )
 
+        train_url_pairs: list[tuple[str, str]] = []
+        test_url_pairs: list[tuple[str, str]] = []
+        for image_url, mask_url in paired_urls:
+            if _is_test_url(image_url) or _is_test_url(mask_url):
+                test_url_pairs.append((image_url, mask_url))
+            else:
+                train_url_pairs.append((image_url, mask_url))
+
+        if not train_url_pairs and test_url_pairs:
+            train_url_pairs = test_url_pairs
+            test_url_pairs = []
+
+        if not train_url_pairs:
+            raise ValueError(
+                "No training pairs found after downloading BioImage Archive assets"
+            )
+
+        split_mode = str(config.get("split_mode", "manual") or "manual").lower()
+        train_split_ratio = float(config.get("train_split_ratio", 0.8))
+        requested_total = resolve_requested_sample_count(
+            config.get("n_samples"),
+            len(train_url_pairs) + len(test_url_pairs),
+        )
+
+        selected_train_url_pairs = list(train_url_pairs)
+        selected_test_url_pairs = list(test_url_pairs)
+
+        if requested_total is not None:
+            if split_mode == "auto":
+                combined_pairs = [*selected_train_url_pairs, *selected_test_url_pairs]
+                if requested_total < len(combined_pairs):
+                    indices = np.random.default_rng().permutation(len(combined_pairs))[  # type: ignore[index]
+                        :requested_total
+                    ]
+                    combined_pairs = [combined_pairs[i] for i in indices]
+
+                if len(combined_pairs) <= 1:
+                    selected_train_url_pairs = combined_pairs
+                    selected_test_url_pairs = []
+                else:
+                    ratio = max(0.05, min(0.95, float(train_split_ratio)))
+                    n_train = int(round(len(combined_pairs) * ratio))
+                    n_train = max(1, min(len(combined_pairs) - 1, n_train))
+                    shuffled = np.random.default_rng().permutation(len(combined_pairs))
+                    train_idx = set(shuffled[:n_train])
+                    selected_train_url_pairs = [
+                        pair
+                        for idx, pair in enumerate(combined_pairs)
+                        if idx in train_idx
+                    ]
+                    selected_test_url_pairs = [
+                        pair
+                        for idx, pair in enumerate(combined_pairs)
+                        if idx not in train_idx
+                    ]
+            else:
+                train_target, test_target = proportional_manual_sample_counts(
+                    len(selected_train_url_pairs),
+                    len(selected_test_url_pairs),
+                    requested_total,
+                )
+                if train_target < len(selected_train_url_pairs):
+                    train_idx = np.random.default_rng().permutation(
+                        len(selected_train_url_pairs)
+                    )[:train_target]
+                    selected_train_url_pairs = [
+                        selected_train_url_pairs[i] for i in train_idx
+                    ]
+                if test_target < len(selected_test_url_pairs):
+                    test_idx = np.random.default_rng().permutation(
+                        len(selected_test_url_pairs)
+                    )[:test_target]
+                    selected_test_url_pairs = [
+                        selected_test_url_pairs[i] for i in test_idx
+                    ]
+
+        selected_train_lookup = set(selected_train_url_pairs)
+
         bia_cache_root = save_path / "bia_download"
         train_pairs: list[TrainingPair] = []
         test_pairs: list[TrainingPair] = []
 
-        for image_url, mask_url in paired_urls:
-            local_image = _local_path_for_remote_url(
-                bia_cache_root, "images", image_url
-            )
-            local_mask = _local_path_for_remote_url(
-                bia_cache_root, "annotations", mask_url
-            )
-            local_image.parent.mkdir(parents=True, exist_ok=True)
-            local_mask.parent.mkdir(parents=True, exist_ok=True)
-
-            if not local_image.exists() or local_image.stat().st_size <= 0:
-                resp = await client.get(image_url)
-                resp.raise_for_status()
-                local_image.write_bytes(resp.content)
+        resolver_server, resolver_service = await _get_bia_resolver_service()
+        try:
+            for image_url, mask_url in [
+                *selected_train_url_pairs,
+                *selected_test_url_pairs,
+            ]:
+                local_image = _local_path_for_remote_url(
+                    bia_cache_root, "images", image_url
+                )
+                local_mask = _local_path_for_remote_url(
+                    bia_cache_root, "annotations", mask_url
+                )
+                local_image.parent.mkdir(parents=True, exist_ok=True)
+                local_mask.parent.mkdir(parents=True, exist_ok=True)
 
-            if not local_mask.exists() or local_mask.stat().st_size <= 0:
-                resp = await client.get(mask_url)
-                resp.raise_for_status()
-                local_mask.write_bytes(resp.content)
+                if not local_image.exists() or local_image.stat().st_size <= 0:
+                    image_bytes = await _download_remote_bytes_via_resolver(
+                        resolver_service,
+                        image_url,
+                    )
+                    local_image.write_bytes(image_bytes)
 
-            pair = TrainingPair(image=local_image, annotation=local_mask)
-            if _is_test_url(image_url) or _is_test_url(mask_url):
-                test_pairs.append(pair)
-            else:
-                train_pairs.append(pair)
+                if not local_mask.exists() or local_mask.stat().st_size <= 0:
+                    mask_bytes = await _download_remote_bytes_via_resolver(
+                        resolver_service,
+                        mask_url,
+                    )
+                    local_mask.write_bytes(mask_bytes)
 
-        if not train_pairs and test_pairs:
-            train_pairs = test_pairs
-            test_pairs = []
+                pair = TrainingPair(image=local_image, annotation=local_mask)
+                if (image_url, mask_url) in selected_train_lookup:
+                    train_pairs.append(pair)
+                else:
+                    test_pairs.append(pair)
+        finally:
+            await resolver_server.disconnect()
 
         if not train_pairs:
             raise ValueError(
@@ -2305,15 +2639,20 @@ async def make_training_pairs_from_bioimage_archive_url(
 async def list_artifact_files_recursive(
     artifact: AsyncHyphaArtifact,
     folder_path: str,
+    max_results: int | None = None,
 ) -> list[str]:
     """Recursively list artifact file paths (relative paths, not folders)."""
+    limit = None
+    if isinstance(max_results, int) and max_results > 0:
+        limit = max_results
+
     root = _normalize_artifact_relpath(folder_path)
     if root and not root.endswith("/"):
         root += "/"
 
     queue = [root]
     visited: set[str] = set()
-    collected: list[str] = []
+    collected: set[str] = set()
 
     while queue:
         current = queue.pop(0)
@@ -2361,36 +2700,153 @@ async def list_artifact_files_recursive(
                 continue
 
             if normalized:
-                collected.append(normalized)
+                collected.add(normalized)
+                if limit is not None and len(collected) >= limit:
+                    return sorted(collected)
 
-    return sorted(set(collected))
+    return sorted(collected)
 
 
 async def list_matching_artifact_paths(
     artifact: AsyncHyphaArtifact,
     path_pattern: str,
+    max_results: int | None = None,
 ) -> list[str]:
     """List artifact files that match a folder path or glob path pattern."""
     normalized_pattern = _normalize_artifact_relpath(path_pattern)
+    limit = max_results if isinstance(max_results, int) and max_results > 0 else None
+
+    def _cap(values: list[str]) -> list[str]:
+        if limit is None:
+            return values
+        return values[:limit]
+
+    def _path_depth(path: str) -> int:
+        stripped = path.strip("/")
+        if not stripped:
+            return 0
+        return len([segment for segment in stripped.split("/") if segment])
+
+    async def _list_matching_artifact_dirs(dir_pattern: str) -> list[str]:
+        base_folder = _glob_base_folder(dir_pattern)
+        target_depth = _path_depth(dir_pattern)
+        queue = [base_folder]
+        visited: set[str] = set()
+        matched: set[str] = set()
+
+        while queue:
+            current_dir = queue.pop(0)
+            if current_dir in visited:
+                continue
+            visited.add(current_dir)
+
+            try:
+                entries = await artifact.ls(current_dir)
+            except Exception:
+                continue
+
+            for entry in entries:
+                entry_type = ""
+                if isinstance(entry, dict):
+                    entry_type = str(entry.get("type") or "").lower()
+                    raw_path = str(entry.get("path") or entry.get("name") or "")
+                else:
+                    raw_path = str(entry)
+
+                if raw_path and "/" not in raw_path and current_dir:
+                    raw_path = f"{current_dir.rstrip('/')}/{raw_path}"
+
+                normalized = _normalize_artifact_relpath(raw_path)
+                if not normalized:
+                    continue
+
+                candidate_dir = (
+                    normalized if normalized.endswith("/") else normalized + "/"
+                )
+                is_dir = entry_type in {"directory", "dir"} or normalized.endswith("/")
+                if not is_dir:
+                    try:
+                        probe_entries = await artifact.ls(candidate_dir)
+                        is_dir = (
+                            isinstance(probe_entries, list) and len(probe_entries) > 0
+                        )
+                    except Exception:
+                        is_dir = False
+
+                if not is_dir:
+                    continue
+
+                if fnmatch.fnmatch(candidate_dir, dir_pattern):
+                    matched.add(candidate_dir)
+
+                if (
+                    _path_depth(candidate_dir) < target_depth
+                    and candidate_dir not in visited
+                ):
+                    queue.append(candidate_dir)
+
+        return sorted(matched)
 
     if normalized_pattern.endswith("/"):
         # Folder-like patterns should match files recursively to stay consistent
         # with UI validation/path suggestions.
         if not _contains_glob(normalized_pattern):
-            return await list_artifact_files_recursive(artifact, normalized_pattern)
+            scan_limit = limit * 5 if limit is not None else None
+            candidates = await list_artifact_files_recursive(
+                artifact,
+                normalized_pattern,
+                max_results=scan_limit,
+            )
+            return _cap(_filter_supported_dataset_image_paths(candidates))
 
-        base_folder = _glob_base_folder(normalized_pattern)
-        candidates = await list_artifact_files_recursive(artifact, base_folder)
+        matched_dirs = await _list_matching_artifact_dirs(normalized_pattern)
+        if matched_dirs:
+            matched_files: list[str] = []
+            seen: set[str] = set()
+            for folder in matched_dirs:
+                try:
+                    names = await list_artifact_files(artifact, folder)
+                except Exception:
+                    continue
+
+                folder_prefix = folder if folder.endswith("/") else f"{folder}/"
+                for name in names:
+                    candidate = _normalize_artifact_relpath(f"{folder_prefix}{name}")
+                    if not candidate or not _is_supported_dataset_image_path(candidate):
+                        continue
+                    if candidate in seen:
+                        continue
+                    seen.add(candidate)
+                    matched_files.append(candidate)
+                    if limit is not None and len(matched_files) >= limit:
+                        return _cap(sorted(matched_files))
+
+            if matched_files:
+                return _cap(sorted(matched_files))
+
+            return _cap(matched_dirs)
 
+        base_folder = _glob_base_folder(normalized_pattern)
+        scan_limit = limit * 5 if limit is not None else None
+        candidates = await list_artifact_files_recursive(
+            artifact,
+            base_folder,
+            max_results=scan_limit,
+        )
         matched = []
         for candidate in candidates:
-            parent = _normalize_artifact_relpath(str(Path(candidate).parent))
-            if parent and not parent.endswith("/"):
-                parent += "/"
-            if fnmatch.fnmatch(parent, normalized_pattern):
+            parent = Path(candidate).parent.as_posix()
+            parent_pattern_value = parent + "/" if parent and parent != "." else ""
+            if fnmatch.fnmatch(parent_pattern_value, normalized_pattern):
                 matched.append(candidate)
+                if limit is not None and len(matched) >= limit:
+                    break
+
+        filtered_files = sorted(set(_filter_supported_dataset_image_paths(matched)))
+        if filtered_files:
+            return _cap(filtered_files)
 
-        return sorted(set(matched))
+        return []
 
     if not _contains_glob(normalized_pattern):
         # If caller passed a directory without trailing slash (common in UI/manual input),
@@ -2404,21 +2860,29 @@ async def list_matching_artifact_paths(
                     if normalized_pattern.endswith("/")
                     else normalized_pattern + "/"
                 )
-                return [
+                candidates = [
                     _normalize_artifact_relpath(f"{folder}{name}") for name in names
                 ]
+                return _cap(_filter_supported_dataset_image_paths(candidates))
         except Exception:
             pass
-        return [normalized_pattern]
+        if _is_supported_dataset_image_path(normalized_pattern):
+            return [normalized_pattern]
+        return []
 
     base_folder = _glob_base_folder(normalized_pattern)
-    candidates = await list_artifact_files_recursive(artifact, base_folder)
+    scan_limit = limit * 5 if limit is not None else None
+    candidates = await list_artifact_files_recursive(
+        artifact,
+        base_folder,
+        max_results=scan_limit,
+    )
     matched = [
         candidate
         for candidate in candidates
         if fnmatch.fnmatch(candidate, normalized_pattern)
     ]
-    return sorted(set(matched))
+    return _cap(sorted(set(_filter_supported_dataset_image_paths(matched))))
 
 
 async def list_artifact_files(
@@ -2838,6 +3302,18 @@ async def make_training_pairs_from_metadata(
             "Expected keys like image_path/mask_path (also supports camelCase variants)."
         )
 
+    requested_total = resolve_requested_sample_count(
+        n_samples,
+        len(train_pairs_raw) + len(test_pairs_raw),
+    )
+    train_pairs_raw, test_pairs_raw = sample_pair_lists(
+        train_pairs_raw,
+        test_pairs_raw,
+        requested_total,
+        split_mode="manual",
+        train_split_ratio=0.8,
+    )
+
     train_pairs = await download_pairs_from_artifact(
         artifact,
         save_path,
@@ -2876,10 +3352,31 @@ async def download_pairs_from_artifact(
         logger.info("Downloading %d files from artifact", len(missing_rpaths))
         try:
             # Add a timeout to prevent indefinite hanging (10 minutes)
-            await asyncio.wait_for(
-                artifact.get(missing_rpaths, missing_lpaths, on_error="ignore"),
-                timeout=600,
-            )
+            try:
+                await asyncio.wait_for(
+                    artifact.get(missing_rpaths, missing_lpaths, on_error="ignore"),
+                    timeout=600,
+                )
+            except Exception as first_error:
+                err_msg = str(first_error).lower()
+                needs_recursive = (
+                    "path is a directory" in err_msg or "recursive" in err_msg
+                )
+                if needs_recursive:
+                    logger.info(
+                        "Retrying artifact download with recursive=True for directory paths"
+                    )
+                    await asyncio.wait_for(
+                        artifact.get(
+                            missing_rpaths,
+                            missing_lpaths,
+                            on_error="ignore",
+                            recursive=True,
+                        ),
+                        timeout=600,
+                    )
+                else:
+                    raise
             logger.info("Download completed successfully")
         except asyncio.TimeoutError:
             logger.error("Download timed out after 600 seconds")
@@ -3045,6 +3542,38 @@ def proportional_manual_sample_counts(
     return train_target, test_target
 
 
+def sample_pair_lists(
+    train_pairs: list[TrainingPair],
+    test_pairs: list[TrainingPair],
+    requested_total: int | None,
+    *,
+    split_mode: str,
+    train_split_ratio: float,
+) -> tuple[list[TrainingPair], list[TrainingPair]]:
+    """Apply random sample limits to pair lists."""
+    if requested_total is None:
+        return train_pairs, test_pairs
+
+    total_pairs = len(train_pairs) + len(test_pairs)
+    if requested_total >= total_pairs:
+        return train_pairs, test_pairs
+
+    if split_mode == "auto":
+        combined_pairs = [*train_pairs, *test_pairs]
+        combined_pairs = random_subset_pairs(combined_pairs, requested_total)
+        return split_training_pairs(combined_pairs, train_split_ratio)
+
+    train_target, test_target = proportional_manual_sample_counts(
+        len(train_pairs),
+        len(test_pairs),
+        requested_total,
+    )
+    return (
+        random_subset_pairs(train_pairs, train_target),
+        random_subset_pairs(test_pairs, test_target),
+    )
+
+
 def split_training_pairs(
     train_pairs: list[TrainingPair],
     train_ratio: float,
@@ -3121,12 +3650,30 @@ async def make_training_pairs(
             "train_annotations must be provided."
         )
 
-    train_image_files = await list_matching_artifact_paths(artifact, train_images)
+    requested_by_user = config.get("n_samples")
+    scan_limit: int | None = None
+    if isinstance(requested_by_user, (int, float)):
+        requested_float = float(requested_by_user)
+        if requested_float >= 1.0:
+            requested_count = max(1, int(round(requested_float)))
+            # For low-sample interactive runs, avoid scanning entire artifacts.
+            # Keep a generous buffer so wildcard matching still finds enough pairs.
+            scan_limit = max(100, requested_count * 40)
+
+    train_image_files = await list_matching_artifact_paths(
+        artifact,
+        train_images,
+        max_results=scan_limit,
+    )
     train_annotation_files = await list_matching_artifact_paths(
         artifact,
         train_annotations,
+        max_results=scan_limit,
     )
 
+    split_mode = str(config.get("split_mode", "manual") or "manual").lower()
+    train_split_ratio = float(config.get("train_split_ratio", 0.8))
+
     # Match training pairs
     train_matched = match_image_annotation_pairs(
         train_image_files,
@@ -3134,29 +3681,23 @@ async def make_training_pairs(
         train_images,
         train_annotations,
     )
-
-    # Build full paths for training files
-    train_image_paths = [Path(img) for img, _ in train_matched]
-    train_annotation_paths = [Path(ann) for _, ann in train_matched]
-
-    # Download training pairs
-    train_pairs = await download_pairs_from_artifact(
-        artifact,
-        save_path,
-        train_image_paths,
-        train_annotation_paths,
-    )
+    train_pairs_raw = [
+        TrainingPair(image=Path(img), annotation=Path(ann))
+        for img, ann in train_matched
+    ]
 
     # Handle test pairs if specified
-    test_pairs: list[TrainingPair] = []
+    test_pairs_raw: list[TrainingPair] = []
     if config["test_images"] and config["test_annotations"]:
         test_image_files = await list_matching_artifact_paths(
             artifact,
             config["test_images"],
+            max_results=scan_limit,
         )
         test_annotation_files = await list_matching_artifact_paths(
             artifact,
             config["test_annotations"],
+            max_results=scan_limit,
         )
 
         # Match test pairs
@@ -3167,40 +3708,43 @@ async def make_training_pairs(
             config["test_annotations"],
         )
 
-        # Build full paths for test files
-        test_image_paths = [Path(img) for img, _ in test_matched]
-        test_annotation_paths = [Path(ann) for _, ann in test_matched]
+        test_pairs_raw = [
+            TrainingPair(image=Path(img), annotation=Path(ann))
+            for img, ann in test_matched
+        ]
+
+    requested_total = resolve_requested_sample_count(
+        config.get("n_samples"),
+        len(train_pairs_raw) + len(test_pairs_raw),
+    )
+    train_pairs_raw, test_pairs_raw = sample_pair_lists(
+        train_pairs_raw,
+        test_pairs_raw,
+        requested_total,
+        split_mode=split_mode,
+        train_split_ratio=train_split_ratio,
+    )
 
-        # Download test pairs
+    # Download selected training pairs only
+    train_pairs = await download_pairs_from_artifact(
+        artifact,
+        save_path,
+        [pair["image"] for pair in train_pairs_raw],
+        [pair["annotation"] for pair in train_pairs_raw],
+    )
+
+    # Download selected test pairs only
+    test_pairs: list[TrainingPair] = []
+    if test_pairs_raw:
         test_pairs = await download_pairs_from_artifact(
             artifact,
             save_path,
-            test_image_paths,
-            test_annotation_paths,
+            [pair["image"] for pair in test_pairs_raw],
+            [pair["annotation"] for pair in test_pairs_raw],
         )
 
-    split_mode = str(config.get("split_mode", "manual") or "manual").lower()
-    train_split_ratio = float(config.get("train_split_ratio", 0.8))
-
-    requested_total = resolve_requested_sample_count(
-        config.get("n_samples"),
-        len(train_pairs) + len(test_pairs),
-    )
-
-    if split_mode == "auto":
-        combined_pairs = [*train_pairs, *test_pairs]
-        if requested_total is not None and requested_total < len(combined_pairs):
-            combined_pairs = random_subset_pairs(combined_pairs, requested_total)
-        train_pairs, test_pairs = split_training_pairs(combined_pairs, train_split_ratio)
-    elif requested_total is not None:
-        if requested_total < len(train_pairs) + len(test_pairs):
-            train_target, test_target = proportional_manual_sample_counts(
-                len(train_pairs),
-                len(test_pairs),
-                requested_total,
-            )
-            train_pairs = random_subset_pairs(train_pairs, train_target)
-            test_pairs = random_subset_pairs(test_pairs, test_target)
+    if split_mode == "auto" and not test_pairs:
+        train_pairs, test_pairs = split_training_pairs(train_pairs, train_split_ratio)
 
     return train_pairs, test_pairs
 
@@ -3681,6 +4225,51 @@ def __init__(self) -> None:
         self.tasks = {}
         self._session_lock = asyncio.Lock()
 
+    def _has_active_training_task(self) -> bool:
+        """Return True when any known training task is still running."""
+        for task in self.tasks.values():
+            if task is not None and not task.done():
+                return True
+        return False
+
+    def _get_free_gpu_memory_bytes(self) -> int | None:
+        """Return free GPU memory in bytes, or None if unavailable."""
+        try:
+            import torch
+        except Exception:
+            return None
+
+        try:
+            if not torch.cuda.is_available():
+                return None
+            free_bytes, _total_bytes = torch.cuda.mem_get_info()
+            return int(free_bytes)
+        except Exception:
+            return None
+
+    def _admission_denial_message(self) -> str | None:
+        """Return a human-readable reason why a new training run should not start."""
+        if self._has_active_training_task():
+            return (
+                "Deferred start to avoid GPU contention: another training session is already active. "
+                "Use restart_training(session_id=...) when resources are available."
+            )
+
+        free_gpu_bytes = self._get_free_gpu_memory_bytes()
+        if free_gpu_bytes is None:
+            return None
+
+        if free_gpu_bytes < MIN_FREE_GPU_MEMORY_TO_START_BYTES:
+            free_gb = free_gpu_bytes / GB
+            required_gb = MIN_FREE_GPU_MEMORY_TO_START_BYTES / GB
+            return (
+                "Deferred start due to low free GPU memory "
+                f"({free_gb:.2f} GB available, need at least {required_gb:.2f} GB). "
+                "Use restart_training(session_id=...) after active workloads finish."
+            )
+
+        return None
+
     def get_model_id(self, model: str) -> str | Path:
         """Return a model identifier suitable for loading by Cellpose."""
         # print all contents of get_session_path(model):
@@ -3694,6 +4283,277 @@ def get_model_id(self, model: str) -> str | Path:
         )
         raise ValueError(msg)
 
+    @schema_method(arbitrary_types_allowed=True)  # type: ignore
+    async def preflight_training_dataset(
+        self,
+        artifact: str = Field(
+            description="Artifact identifier or BioImage Archive URL",
+            examples=["ri-scale/zarr-demo"],
+        ),
+        train_images: str | None = Field(
+            None,
+            description="Training images path or glob pattern",
+        ),
+        train_annotations: str | None = Field(
+            None,
+            description="Training annotations path or glob pattern",
+        ),
+        metadata_dir: str | None = Field(
+            None,
+            description="Optional metadata directory path",
+        ),
+        test_images: str | None = Field(
+            None,
+            description="Optional test images path or glob pattern",
+        ),
+        test_annotations: str | None = Field(
+            None,
+            description="Optional test annotations path or glob pattern",
+        ),
+        split_mode: str = Field(
+            "manual",
+            description="Dataset split mode: manual or auto",
+        ),
+        n_samples: int | float | None = Field(
+            None,
+            description="Optional sample count or fraction",
+        ),
+        max_candidates: int = Field(
+            500,
+            description=(
+                "Maximum matching candidates to inspect per path spec. "
+                "Keeps preflight fast for large artifacts."
+            ),
+        ),
+    ) -> dict[str, Any]:
+        """Preflight-check dataset specs with lightweight backend matching.
+
+        Returns counts and readiness information without starting training.
+        """
+        if isinstance(artifact, dict):
+            wrapped = artifact
+            artifact = wrapped.get(
+                "artifact", wrapped.get("artifact_id", wrapped.get("id", artifact))
+            )
+            train_images = wrapped.get("train_images", train_images)
+            train_annotations = wrapped.get("train_annotations", train_annotations)
+            metadata_dir = wrapped.get("metadata_dir", metadata_dir)
+            test_images = wrapped.get("test_images", test_images)
+            test_annotations = wrapped.get("test_annotations", test_annotations)
+            split_mode = wrapped.get("split_mode", split_mode)
+            n_samples = wrapped.get("n_samples", n_samples)
+            max_candidates = wrapped.get("max_candidates", max_candidates)
+
+        artifact = normalize_optional_param(artifact)
+        train_images = normalize_optional_param(train_images)
+        train_annotations = normalize_optional_param(train_annotations)
+        metadata_dir = normalize_optional_param(metadata_dir)
+        test_images = normalize_optional_param(test_images)
+        test_annotations = normalize_optional_param(test_annotations)
+        split_mode = normalize_optional_param(split_mode)
+        n_samples = normalize_optional_param(n_samples)
+        max_candidates = int(max_candidates) if max_candidates is not None else 500
+        max_candidates = max(10, min(max_candidates, 5000))
+
+        split_mode_value = str(split_mode or "manual").lower()
+        if split_mode_value not in {"manual", "auto"}:
+            split_mode_value = "manual"
+
+        result: dict[str, Any] = {
+            "ok": False,
+            "artifact_id": None,
+            "mode": "artifact",
+            "split_mode": split_mode_value,
+            "train_image_count": 0,
+            "train_annotation_count": 0,
+            "train_pair_count": 0,
+            "test_image_count": 0,
+            "test_annotation_count": 0,
+            "test_pair_count": 0,
+            "sampled_total_count": None,
+            "requested_total_count": None,
+            "message": "",
+        }
+
+        if not isinstance(artifact, str) or not artifact:
+            result["message"] = "artifact must be a non-empty string"
+            return result
+
+        is_bia = _is_bioimage_archive_url(artifact)
+        if is_bia:
+            result["ok"] = True
+            result["mode"] = "bioimage-archive"
+            result["artifact_id"] = artifact
+            result["split_mode"] = "auto"
+            result["message"] = (
+                "BioImage Archive mode detected. Pair matching happens during preparation."
+            )
+            return result
+
+        server_url, artifact_id = get_url_and_artifact_id(artifact)
+        result["artifact_id"] = artifact_id
+
+        try:
+            artifact_client = await make_artifact_client(artifact_id, server_url)
+        except Exception as e:
+            result["message"] = f"Failed to access artifact '{artifact_id}': {e}"
+            return result
+
+        if metadata_dir:
+            metadata_dir_value = str(metadata_dir)
+            if not metadata_dir_value.endswith("/"):
+                metadata_dir_value += "/"
+            try:
+                entries = await artifact_client.ls(metadata_dir_value)
+            except Exception as e:
+                result["message"] = (
+                    f"Metadata directory '{metadata_dir_value}' is not readable: {e}"
+                )
+                return result
+
+            metadata_file_count = 0
+            for entry in entries or []:
+                raw_path = ""
+                if isinstance(entry, dict):
+                    raw_path = str(entry.get("path") or entry.get("name") or "")
+                else:
+                    raw_path = str(entry)
+                if raw_path.lower().endswith(".json"):
+                    metadata_file_count += 1
+
+            if metadata_file_count <= 0:
+                result["message"] = (
+                    f"No metadata JSON files found under '{metadata_dir_value}'."
+                )
+                return result
+
+            result["ok"] = True
+            result["mode"] = "metadata"
+            result["metadata_file_count"] = metadata_file_count
+            result["message"] = (
+                f"Metadata mode ready. Found {metadata_file_count} metadata JSON files."
+            )
+            return result
+
+        if not isinstance(train_images, str) or not train_images:
+            result["message"] = (
+                "train_images must be provided when metadata_dir is not used"
+            )
+            return result
+        if not isinstance(train_annotations, str) or not train_annotations:
+            result["message"] = (
+                "train_annotations must be provided when metadata_dir is not used"
+            )
+            return result
+
+        try:
+            train_image_paths = await list_matching_artifact_paths(
+                artifact_client,
+                train_images,
+                max_results=max_candidates,
+            )
+            train_annotation_paths = await list_matching_artifact_paths(
+                artifact_client,
+                train_annotations,
+                max_results=max_candidates,
+            )
+        except Exception as e:
+            result["message"] = f"Failed to list training paths: {e}"
+            return result
+
+        train_image_count = len(
+            [
+                path
+                for path in train_image_paths
+                if _is_supported_dataset_image_path(path)
+            ]
+        )
+        train_annotation_count = len(
+            [
+                path
+                for path in train_annotation_paths
+                if _is_supported_dataset_image_path(path)
+            ]
+        )
+        train_pair_count = min(train_image_count, train_annotation_count)
+
+        result["train_image_count"] = train_image_count
+        result["train_annotation_count"] = train_annotation_count
+        result["train_pair_count"] = train_pair_count
+
+        if train_pair_count <= 0:
+            result["message"] = (
+                "No training pairs found from current training path specs. "
+                f"images={train_image_count}, annotations={train_annotation_count}."
+            )
+            return result
+
+        test_pair_count = 0
+        if split_mode_value == "manual" and test_images and test_annotations:
+            try:
+                test_image_paths = await list_matching_artifact_paths(
+                    artifact_client,
+                    test_images,
+                    max_results=max_candidates,
+                )
+                test_annotation_paths = await list_matching_artifact_paths(
+                    artifact_client,
+                    test_annotations,
+                    max_results=max_candidates,
+                )
+            except Exception as e:
+                result["message"] = f"Failed to list test paths: {e}"
+                return result
+
+            test_image_count = len(
+                [
+                    path
+                    for path in test_image_paths
+                    if _is_supported_dataset_image_path(path)
+                ]
+            )
+            test_annotation_count = len(
+                [
+                    path
+                    for path in test_annotation_paths
+                    if _is_supported_dataset_image_path(path)
+                ]
+            )
+            test_pair_count = min(test_image_count, test_annotation_count)
+
+            result["test_image_count"] = test_image_count
+            result["test_annotation_count"] = test_annotation_count
+            result["test_pair_count"] = test_pair_count
+
+            if test_pair_count <= 0:
+                result["message"] = (
+                    "Manual split selected but no test pairs were found from test path specs. "
+                    f"images={test_image_count}, annotations={test_annotation_count}."
+                )
+                return result
+
+        total_pairs = train_pair_count + test_pair_count
+        requested_total = None
+        if n_samples is not None:
+            try:
+                requested_total = resolve_requested_sample_count(
+                    float(n_samples), total_pairs
+                )
+            except Exception as e:
+                result["message"] = f"Invalid n_samples value: {e}"
+                return result
+
+        result["requested_total_count"] = requested_total
+        result["sampled_total_count"] = (
+            requested_total if requested_total is not None else total_pairs
+        )
+        result["ok"] = True
+        result["message"] = (
+            "Dataset preflight passed. "
+            f"train_pairs={train_pair_count}, test_pairs={test_pair_count}, total={total_pairs}."
+        )
+        return result
+
     @schema_method(arbitrary_types_allowed=True)  # type: ignore
     async def start_training(
         self,
@@ -3966,6 +4826,29 @@ async def start_training(
         training_params_path.write_text(json.dumps(params_dict, indent=2))
 
         async with self._session_lock:
+            denial_message = self._admission_denial_message()
+            if denial_message is not None:
+                update_status(
+                    session_id=session_id,
+                    status_type=StatusType.STOPPED,
+                    message=denial_message,
+                    dataset_artifact_id=artifact_id,
+                    model=model_id,
+                    n_samples=n_samples,
+                    n_epochs=n_epochs,
+                    learning_rate=learning_rate,
+                    weight_decay=weight_decay,
+                    min_train_masks=min_train_masks,
+                    validation_interval=validation_interval,
+                )
+                append_info(
+                    session_id,
+                    denial_message,
+                    with_time=True,
+                )
+                status = get_status(session_id)
+                return SessionStatusWithId(**status, session_id=session_id)
+
             executor = ThreadPoolExecutor(max_workers=1)
             self.executors[session_id] = executor
             task = asyncio.create_task(launch_training_task(training_params, executor))
@@ -4194,6 +5077,72 @@ async def restart_training(
         restarted["restarted_from"] = session_id
         return restarted
 
+    @schema_method(arbitrary_types_allowed=True)  # type: ignore
+    async def delete_training_session(
+        self,
+        session_id: str = Field(description="Session ID to delete"),
+        force_stop_if_blocked: bool = Field(
+            False,
+            description="If true, stop blocked running/preparing session first, then delete.",
+        ),
+    ) -> dict[str, Any]:
+        """Delete a non-running training session and its local artifacts."""
+        import shutil
+
+        if isinstance(session_id, dict):
+            wrapped = session_id
+            session_id = wrapped.get("session_id", wrapped.get("id", session_id))
+
+        session_id = str(session_id).strip().replace("\\", "/")
+        if session_id.endswith("/status.json"):
+            session_id = session_id[: -len("/status.json")]
+        session_id = Path(session_id).name
+
+        status_path = get_status_path(session_id)
+        if not status_path.exists():
+            raise ValueError(f"Session {session_id} does not exist")
+
+        try:
+            status_mtime = status_path.stat().st_mtime
+        except OSError:
+            status_mtime = time.time()
+
+        status = self._normalize_session_status(
+            session_id=session_id,
+            session_data=dict(get_status(session_id)),
+            status_mtime=status_mtime,
+        )
+        status_type = str(status.get("status_type") or "").lower()
+        blocked = {StatusType.RUNNING.value, StatusType.PREPARING.value}
+        if status_type in blocked:
+            if force_stop_if_blocked:
+                await self.stop_training(session_id=session_id)
+            else:
+                raise ValueError(
+                    f"Session {session_id} has status '{status_type}'. "
+                    "Running or preparing sessions cannot be deleted. Stop the session first."
+                )
+
+        task = self.tasks.get(session_id)
+        if task and not task.done():
+            task.cancel()
+        if session_id in self.tasks:
+            del self.tasks[session_id]
+
+        executor = self.executors.get(session_id)
+        if executor is not None:
+            executor.shutdown(wait=False, cancel_futures=True)
+            del self.executors[session_id]
+
+        session_path = get_session_path(session_id)
+        if session_path.exists():
+            shutil.rmtree(session_path, ignore_errors=False)
+
+        return {
+            "deleted": True,
+            "session_id": session_id,
+        }
+
     def _list_sessions_sync(
         self,
         status_types: list[str] | None,
diff --git a/scripts/tmp_probe_bia_fetch_only.py b/scripts/tmp_probe_bia_fetch_only.py
new file mode 100644
index 00000000..ee1ac181
--- /dev/null
+++ b/scripts/tmp_probe_bia_fetch_only.py
@@ -0,0 +1,140 @@
+import argparse
+import asyncio
+import csv
+import io
+import time
+from dataclasses import dataclass
+
+import httpx
+
+
+@dataclass
+class FetchResult:
+    label: str
+    url: str
+    ok: bool
+    status_code: int | None
+    elapsed_s: float
+    size_bytes: int | None = None
+    error: str | None = None
+
+
+def _bia_ftp_base_url(accession: str) -> str:
+    accession_number = int(accession.split("S-BIAD", 1)[1])
+    suffix = accession_number % 1000
+    return f"https://ftp.ebi.ac.uk/biostudies/fire/S-BIAD/{suffix}/{accession}"
+
+
+def _extract_tsv_pairs(tsv_content: str) -> list[tuple[str, str]]:
+    pairs: list[tuple[str, str]] = []
+    reader = csv.DictReader(io.StringIO(tsv_content), delimiter="\t")
+    for row in reader:
+        label_path = str(row.get("Files") or "").strip()
+        source_image = str(row.get("Source image") or "").strip()
+        if not label_path or not source_image:
+            continue
+        pairs.append((source_image, label_path))
+    return pairs
+
+
+async def _timed_get(client: httpx.AsyncClient, label: str, url: str) -> FetchResult:
+    started = time.perf_counter()
+    try:
+        resp = await client.get(url)
+        elapsed = time.perf_counter() - started
+        return FetchResult(
+            label=label,
+            url=url,
+            ok=200 <= resp.status_code < 300,
+            status_code=resp.status_code,
+            elapsed_s=elapsed,
+            size_bytes=len(resp.content),
+            error=None if 200 <= resp.status_code < 300 else f"HTTP {resp.status_code}",
+        )
+    except Exception as exp:
+        elapsed = time.perf_counter() - started
+        return FetchResult(
+            label=label,
+            url=url,
+            ok=False,
+            status_code=None,
+            elapsed_s=elapsed,
+            size_bytes=None,
+            error=str(exp),
+        )
+
+
+def _print_result(result: FetchResult) -> None:
+    status = result.status_code if result.status_code is not None else "ERR"
+    size = result.size_bytes if result.size_bytes is not None else "-"
+    msg = f"[{result.label}] ok={result.ok} status={status} time={result.elapsed_s:.2f}s size={size}"
+    if result.error:
+        msg += f" error={result.error}"
+    print(msg, flush=True)
+
+
+async def main() -> None:
+    parser = argparse.ArgumentParser(description="BIA-only fetch probe (no training calls).")
+    parser.add_argument("--accession", default="S-BIAD1392")
+    parser.add_argument("--sample-pairs", type=int, default=3)
+    args = parser.parse_args()
+
+    accession = str(args.accession).upper()
+    base_url = _bia_ftp_base_url(accession)
+
+    study_url = f"https://www.ebi.ac.uk/biostudies/bioimages/studies/{accession}"
+    tsv_candidates = [
+        f"{base_url}/Files/ps_ovule_labels.tsv",
+        f"{base_url}/Files/labels.tsv",
+        f"{base_url}/Files/{accession}_labels.tsv",
+    ]
+
+    timeout = httpx.Timeout(connect=20.0, read=120.0, write=20.0, pool=20.0)
+    async with httpx.AsyncClient(timeout=timeout, follow_redirects=True) as client:
+        print(f"Probing accession: {accession}")
+
+        study_result = await _timed_get(client, "study_page", study_url)
+        _print_result(study_result)
+
+        tsv_result: FetchResult | None = None
+        tsv_text: str | None = None
+        for tsv_url in tsv_candidates:
+            result = await _timed_get(client, "tsv", tsv_url)
+            _print_result(result)
+            if result.ok:
+                resp = await client.get(tsv_url)
+                tsv_text = resp.text
+                tsv_result = result
+                break
+
+        if not tsv_result or not tsv_text:
+            print("No TSV source succeeded. BIA pair listing is blocked/slow at source.")
+            return
+
+        pairs = _extract_tsv_pairs(tsv_text)
+        print(f"Parsed TSV pairs: {len(pairs)}")
+        if not pairs:
+            print("TSV loaded but contained no usable image/mask pairs.")
+            return
+
+        sample_count = max(1, min(int(args.sample_pairs), len(pairs)))
+        sample_pairs = pairs[:sample_count]
+
+        urls: list[tuple[str, str]] = []
+        for idx, (image_rel, mask_rel) in enumerate(sample_pairs, start=1):
+            image_url = f"{base_url}/Files/{image_rel.lstrip('/')}"
+            mask_url = f"{base_url}/Files/{mask_rel.lstrip('/')}"
+            urls.append((f"pair{idx}_image", image_url))
+            urls.append((f"pair{idx}_mask", mask_url))
+
+        print(f"Fetching {len(urls)} sample files from FTP...")
+        results = await asyncio.gather(*[_timed_get(client, label, url) for label, url in urls])
+        for result in results:
+            _print_result(result)
+
+        ok_count = sum(1 for item in results if item.ok)
+        print(f"Sample fetch success: {ok_count}/{len(results)}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/tests/cellpose/cellpose_training_agent_startup.pyodide.py b/tests/cellpose/cellpose_training_agent_startup.pyodide.py
new file mode 100644
index 00000000..77eeb88f
--- /dev/null
+++ b/tests/cellpose/cellpose_training_agent_startup.pyodide.py
@@ -0,0 +1,391 @@
+import micropip
+
+await micropip.install(["hypha-rpc", "httpx"])
+
+import json
+import os
+from typing import Any
+
+from hypha_rpc import connect_to_server, login
+
+SERVER_URL = os.environ.get("HYPHA_SERVER_URL", "https://hypha.aicell.io")
+DEFAULT_SERVICE = "bioimage-io/cellpose-finetuning"
+
+_state: dict[str, Any] = {
+    "server": None,
+    "workspace": None,
+    "service": None,
+    "service_id": None,
+    "schemas": {},
+}
+
+
+async def connect_cellpose(
+    server_url: str | None = None,
+    workspace: str | None = None,
+    token: str | None = None,
+):
+    target_server = server_url or SERVER_URL
+    target_token = token or os.environ.get("HYPHA_TOKEN")
+    if not target_token:
+        target_token = await login({"server_url": target_server})
+
+    config: dict[str, Any] = {"server_url": target_server, "token": target_token}
+    if workspace:
+        config["workspace"] = workspace
+
+    server = await connect_to_server(config)
+    _state["server"] = server
+    _state["workspace"] = server.config.workspace
+    return {
+        "server_url": target_server,
+        "workspace": server.config.workspace,
+        "user_id": server.config.user.get("id"),
+    }
+
+
+async def connect_cellpose_service():
+    return await select_cellpose_service(DEFAULT_SERVICE)
+
+
+async def list_cellpose_services():
+    if not _state["server"]:
+        await connect_cellpose()
+
+    server = _state["server"]
+    workspace = _state.get("workspace")
+    candidates: list[str] = []
+
+    env_service = os.environ.get("HYPHA_TEST_SERVICE_ID")
+    if env_service:
+        candidates.append(env_service)
+
+    if workspace:
+        candidates.append(f"{workspace}/cellpose-finetuning")
+
+    candidates.extend(
+        [
+            DEFAULT_SERVICE,
+            "bioimage-io/cellpose-finetuning",
+        ]
+    )
+
+    # Worker-derived service ids (best source after redeploys)
+    worker_candidates = ["bioimage-io/bioengine-worker"]
+    if workspace:
+        worker_candidates.append(f"{workspace}/bioengine-worker")
+
+    for worker_id in worker_candidates:
+        try:
+            worker = await server.get_service(worker_id)
+            app_status = await worker.get_application_status(
+                application_ids=["cellpose-finetuning"]
+            )
+            app_data = (
+                app_status.get("cellpose-finetuning", {})
+                if isinstance(app_status, dict)
+                else {}
+            )
+            for entry in app_data.get("service_ids", []):
+                if isinstance(entry, dict):
+                    service_id = entry.get("websocket_service_id")
+                else:
+                    service_id = entry
+                if isinstance(service_id, str) and service_id:
+                    candidates.append(service_id)
+        except Exception:
+            continue
+
+    deduped: list[str] = []
+    seen: set[str] = set()
+    for service_id in candidates:
+        if service_id and service_id not in seen:
+            seen.add(service_id)
+            deduped.append(service_id)
+
+    services: list[dict[str, Any]] = []
+    for service_id in deduped:
+        available = False
+        error = None
+        try:
+            await server.get_service(service_id)
+            available = True
+        except Exception as e:
+            error = str(e)
+        services.append(
+            {"service_id": service_id, "available": available, "error": error}
+        )
+
+    return services
+
+
+async def select_cellpose_service(service_id: str | None = None):
+    if not _state["server"]:
+        await connect_cellpose()
+
+    server = _state["server"]
+    target = service_id
+    if not target:
+        services = await list_cellpose_services()
+        for service_info in services:
+            if service_info.get("available"):
+                target = service_info.get("service_id")
+                break
+
+    if not target:
+        raise RuntimeError(
+            "Could not resolve an available Cellpose service. "
+            "Call list_cellpose_services() and pass an explicit service_id."
+        )
+
+    service = await server.get_service(target)
+    _state["service"] = service
+    _state["service_id"] = target
+    await _refresh_service_schemas()
+    return {"service_id": target, "workspace": _state.get("workspace")}
+
+
+async def _refresh_service_schemas():
+    if not _state["service"]:
+        await connect_cellpose_service()
+
+    service = _state["service"]
+    method_names = [
+        "start_training",
+        "get_training_status",
+        "list_training_sessions",
+        "stop_training",
+        "restart_training",
+        "infer",
+        "export_model",
+    ]
+
+    schemas: dict[str, Any] = {}
+    for name in method_names:
+        method = getattr(service, name, None)
+        if method is None:
+            continue
+        schema = getattr(method, "__schema__", None)
+        if schema is not None:
+            schemas[name] = schema
+
+    _state["schemas"] = schemas
+    return schemas
+
+
+async def start_cellpose_training(**kwargs):
+    if not _state["service"]:
+        await connect_cellpose_service()
+    return await _state["service"].start_training(**kwargs)
+
+
+async def get_training_status(session_id: str):
+    if not _state["service"]:
+        await connect_cellpose_service()
+    return await _state["service"].get_training_status(session_id=session_id)
+
+
+async def list_training_sessions(status_types: list[str] | None = None):
+    if not _state["service"]:
+        await connect_cellpose_service()
+    if status_types:
+        return await _state["service"].list_training_sessions(
+            status_types=status_types,
+        )
+    return await _state["service"].list_training_sessions()
+
+
+async def stop_training(session_id: str):
+    if not _state["service"]:
+        await connect_cellpose_service()
+    return await _state["service"].stop_training(session_id=session_id)
+
+
+async def restart_training(session_id: str, n_epochs: int | None = None):
+    if not _state["service"]:
+        await connect_cellpose_service()
+    payload: dict[str, Any] = {"session_id": session_id}
+    if n_epochs is not None:
+        payload["n_epochs"] = int(n_epochs)
+    return await _state["service"].restart_training(**payload)
+
+
+async def infer_cellpose(**kwargs):
+    if not _state["service"]:
+        await connect_cellpose_service()
+    return await _state["service"].infer(**kwargs)
+
+
+async def export_model(
+    session_id: str,
+    model_name: str | None = None,
+    collection: str = "bioimage-io/colab-annotations",
+):
+    if not _state["service"]:
+        await connect_cellpose_service()
+    payload: dict[str, Any] = {
+        "session_id": session_id,
+        "collection": collection,
+    }
+    if model_name:
+        payload["model_name"] = model_name
+    return await _state["service"].export_model(**payload)
+
+
+async def get_cellpose_schemas(refresh: bool = False):
+    if refresh or not _state.get("schemas"):
+        await _refresh_service_schemas()
+    return _state.get("schemas", {})
+
+
+async def print_cellpose_runtime_docs(include_schemas: bool = True):
+    docs = RUNTIME_GUIDE
+    if include_schemas:
+        schemas = await get_cellpose_schemas(refresh=True)
+        if schemas:
+            docs += "\n\nLive service schemas (from service.<function>.__schema__):\n"
+            docs += json.dumps(schemas, indent=2, sort_keys=True)
+        else:
+            docs += (
+                "\n\nLive service schemas are unavailable. "
+                "Connect/select a service first, then call print_cellpose_runtime_docs()."
+            )
+    print(docs)
+
+
+RUNTIME_GUIDE = """
+You are a Cellpose training operations assistant for BioEngine.
+
+Primary goal:
+- Help users run reliable Cellpose fine-tuning workflows end-to-end: connect, select service, start training, monitor, troubleshoot, stop/restart, infer, export.
+
+Recommended execution order:
+1) Connect:
+     - await connect_cellpose(server_url=None, workspace=None, token=None)
+2) Discover and select service (important after redeploys):
+     - await list_cellpose_services()
+     - await select_cellpose_service(service_id=None)
+3) Start training:
+     - await start_cellpose_training(...)
+4) Monitor:
+     - await get_training_status(session_id)
+     - optional: await list_training_sessions(status_types=[...])
+5) Control:
+     - await stop_training(session_id)
+     - await restart_training(session_id, n_epochs=None)
+6) Post-training:
+     - await infer_cellpose(...)
+     - await export_model(session_id, model_name=None, collection='bioimage-io/colab-annotations')
+
+Function reference (wrapper-level):
+- connect_cellpose(server_url=None, workspace=None, token=None)
+    - Connects to Hypha server and caches server/workspace in _state.
+    - If token missing, invokes interactive login().
+    - Returns {server_url, workspace, user_id}.
+
+- list_cellpose_services()
+    - Lists candidate Cellpose service IDs from defaults + worker app status.
+    - Returns [{service_id, available, error}].
+
+- select_cellpose_service(service_id=None)
+    - Selects explicit service_id or first available from list_cellpose_services().
+    - Caches service + service_id in _state.
+    - Refreshes live __schema__ docs.
+    - Returns {service_id, workspace}.
+
+- start_cellpose_training(**kwargs)
+    - Forwards kwargs to service.start_training.
+    - Required: artifact and either
+        a) train_images + train_annotations, or
+        b) metadata_dir, or
+        c) BioImage Archive URL in artifact.
+
+- get_training_status(session_id)
+    - Returns latest status including status_type/message/progress/losses/metrics.
+
+- list_training_sessions(status_types=None)
+    - Lists known sessions, optionally filtered by status.
+
+- stop_training(session_id)
+    - Requests graceful stop for running session.
+
+- restart_training(session_id, n_epochs=None)
+    - Starts a new run from previous session checkpoint/params.
+
+- infer_cellpose(**kwargs)
+    - Forwards infer parameters to service.infer.
+
+- export_model(session_id, model_name=None, collection='bioimage-io/colab-annotations')
+    - Exports completed training to BioImage.IO artifact.
+
+- get_cellpose_schemas(refresh=False)
+    - Returns cached or live-refreshed method schemas from service.<function>.__schema__.
+
+- print_cellpose_runtime_docs(include_schemas=True)
+    - Prints this guide plus live schemas.
+
+Training parameter guidance:
+- split_mode: use 'auto' for BioImage Archive URL artifacts.
+- train_split_ratio: used only in auto split (e.g. 0.8 for 80/20).
+- n_samples:
+    - if user gives percentage p%, pass n_samples=p/100.
+    - valid decimal range: 0 < n_samples <= 1.
+    - examples: 0.005 = 0.5%, 0.02 = 2%, 1.0 = 100%.
+- validation metrics require test data (manual split) or auto split that creates test data.
+
+Monitoring loop checklist:
+- Poll every 2-5 seconds.
+- Report: status_type, message, current_epoch/total_epochs, elapsed_seconds.
+- Include latest available train/test losses and validation metrics when present.
+- Stop polling when status_type is completed/failed/stopped.
+
+Error triage checklist:
+- Service resolution failure:
+    - call list_cellpose_services(); pick available websocket service ID.
+- Artifact/path mismatch:
+    - verify artifact workspace and glob patterns.
+- Token/workspace mismatch:
+    - reconnect with correct token/workspace.
+- Resource pressure (VRAM/OOM):
+    - stop competing sessions, retry with smaller n_samples and/or lower load.
+
+Examples:
+1) Start from artifact paths:
+     await start_cellpose_training(
+             artifact='workspace/dataset',
+             train_images='images/*/*.tif',
+             train_annotations='annotations/*/*_mask.ome.tif',
+             split_mode='auto',
+             train_split_ratio=0.8,
+             n_samples=0.005,
+             model='cpsam',
+             n_epochs=1,
+             learning_rate=1e-5,
+             min_train_masks=1,
+             validation_interval=1,
+     )
+
+2) Start from metadata:
+     await start_cellpose_training(
+             artifact='workspace/dataset',
+             metadata_dir='metadata/',
+             model='cpsam',
+             n_epochs=1,
+     )
+
+3) BioImage Archive URL:
+     await start_cellpose_training(
+             artifact='https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD1392',
+             split_mode='auto',
+             train_split_ratio=0.8,
+             n_samples=0.005,
+             n_epochs=1,
+             model='cpsam',
+     )
+
+Schema usage note:
+- To inspect exact callable schemas at runtime, run:
+    await print_cellpose_runtime_docs(include_schemas=True)
+"""
+
+print(RUNTIME_GUIDE)
diff --git a/tests/cellpose/live_test_utils.py b/tests/cellpose/live_test_utils.py
new file mode 100644
index 00000000..e8daca80
--- /dev/null
+++ b/tests/cellpose/live_test_utils.py
@@ -0,0 +1,153 @@
+from __future__ import annotations
+
+import asyncio
+import os
+from typing import Any
+
+import pytest
+from hypha_rpc import connect_to_server
+
+SERVER_URL = os.environ.get("HYPHA_SERVER_URL", "https://hypha.aicell.io")
+
+
+def is_transient_connection_error(err: Exception) -> bool:
+    text = str(err).lower()
+    return (
+        "websocket" in text
+        or "http 404" in text
+        or "http 502" in text
+        or "server rejected websocket connection" in text
+        or "method call timed out" in text
+    )
+
+
+def is_unavailable_workspace_error(err: Exception) -> bool:
+    text = str(err).lower()
+    return (
+        "workspace" in text
+        and ("does not exist" in text or "not accessible" in text)
+    )
+
+
+async def with_live_server(
+    token: str,
+    workspace: str,
+    callback: Any,
+    *,
+    retries: int = 6,
+    method_timeout: int = 240000,
+) -> Any:
+    last_error: Exception | None = None
+    for attempt in range(retries):
+        try:
+            async with connect_to_server(
+                {
+                    "server_url": SERVER_URL,
+                    "token": token,
+                    "workspace": workspace,
+                    "method_timeout": method_timeout,
+                }
+            ) as server:
+                return await callback(server)
+        except Exception as e:
+            last_error = e
+            if attempt < retries - 1 and is_transient_connection_error(e):
+                await asyncio.sleep(10)
+                continue
+            if is_transient_connection_error(e):
+                pytest.skip(
+                    "Live Hypha websocket endpoint is temporarily unavailable "
+                    "(HTTP 404/502/timeouts during connection)."
+                )
+            if is_unavailable_workspace_error(e):
+                pytest.skip(
+                    "Live workspace is currently not accessible for this token/session."
+                )
+            raise
+
+    if last_error is not None:
+        raise last_error
+    return None
+
+
+def _supports_split_mode(service: Any) -> bool:
+    try:
+        schema = getattr(service.start_training, "__schema__", None)
+        if isinstance(schema, dict):
+            params = schema.get("parameters")
+            if isinstance(params, dict):
+                return "split_mode" in params.get("properties", {})
+            if isinstance(params, list):
+                return any(
+                    isinstance(p, dict) and p.get("name") == "split_mode"
+                    for p in params
+                )
+    except Exception:
+        return False
+    return False
+
+
+async def resolve_cellpose_service(
+    server: Any,
+    workspace: str,
+    *,
+    application_id: str = "cellpose-finetuning",
+    requested_service_id: str | None = None,
+) -> Any:
+    if requested_service_id:
+        service = await server.get_service(requested_service_id)
+        if _supports_split_mode(service):
+            return service
+
+    worker_candidates = [
+        "bioimage-io/bioengine-worker",
+        f"{workspace}/bioengine-worker",
+    ]
+    for worker_id in worker_candidates:
+        try:
+            worker = await server.get_service(worker_id)
+            status = await worker.get_application_status(
+                application_ids=[application_id]
+            )
+        except Exception:
+            continue
+
+        app_data = status.get(application_id, {}) if isinstance(status, dict) else {}
+        service_entries = app_data.get("service_ids", [])
+        for entry in reversed(service_entries):
+            if isinstance(entry, dict):
+                websocket_id = (
+                    entry.get("websocket_service_id")
+                    or entry.get("service_id")
+                    or entry.get("id")
+                )
+            else:
+                websocket_id = entry
+
+            if isinstance(websocket_id, str) and websocket_id:
+                try:
+                    service = await server.get_service(websocket_id)
+                    if _supports_split_mode(service):
+                        return service
+                except Exception:
+                    continue
+
+    candidate_ids = [
+        os.environ.get("HYPHA_TEST_SERVICE_ID"),
+        f"{workspace}/{application_id}",
+        "bioimage-io/cellpose-finetuning",
+    ]
+    for candidate in candidate_ids:
+        if not candidate:
+            continue
+        try:
+            service = await server.get_service(candidate)
+            if _supports_split_mode(service):
+                return service
+        except Exception:
+            continue
+
+    raise RuntimeError(
+        "Could not resolve cellpose service id from workspace or worker status. "
+        "Verify app deployment and websocket service ids in worker status."
+    )
diff --git a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py b/tests/cellpose/test_metadata_and_glob.py
similarity index 68%
rename from bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
rename to tests/cellpose/test_metadata_and_glob.py
index 094a246b..835c5083 100644
--- a/bioengine_apps/cellpose_finetuning/tests/test_metadata_and_glob.py
+++ b/tests/cellpose/test_metadata_and_glob.py
@@ -4,6 +4,7 @@
 from pathlib import Path
 
 from bioengine_apps.cellpose_finetuning.main import (
+    CellposeFinetune,
     _extract_bia_accession,
     _pair_key_from_url,
     list_matching_artifact_paths,
@@ -11,6 +12,7 @@
     match_image_annotation_pairs,
     proportional_manual_sample_counts,
     resolve_requested_sample_count,
+    sample_pair_lists,
 )
 
 
@@ -197,6 +199,31 @@ async def get(self, remote_paths, local_paths, on_error="ignore"):
             local_path.write_bytes(b"dummy")
 
 
+class _FakeArtifactFolderGlobDirs:
+    async def ls(self, folder_path: str):
+        await asyncio.sleep(0)
+        folder = folder_path.rstrip("/") + "/"
+        if folder == "images/":
+            return [
+                {"path": "images/sample-a"},
+                {"path": "images/sample-b"},
+                {"path": "images/readme.txt", "type": "file"},
+            ]
+        if folder in {"images/sample-a/", "images/sample-b/"}:
+            return [
+                {"path": f"{folder}0", "type": "file"},
+                {"path": f"{folder}1", "type": "file"},
+            ]
+        return []
+
+    async def get(self, remote_paths, local_paths, on_error="ignore"):
+        await asyncio.sleep(0)
+        for local in local_paths:
+            local_path = Path(local)
+            local_path.parent.mkdir(parents=True, exist_ok=True)
+            local_path.write_bytes(b"dummy")
+
+
 def test_make_training_pairs_from_metadata(tmp_path: Path) -> None:
     artifact = _FakeArtifact()
     train_pairs, test_pairs = asyncio.run(
@@ -234,6 +261,15 @@ def test_list_matching_artifact_paths_with_ambiguous_directory_entries() -> None
     ]
 
 
+def test_list_matching_artifact_paths_folder_glob_returns_directories() -> None:
+    artifact = _FakeArtifactFolderGlobDirs()
+    matches = asyncio.run(list_matching_artifact_paths(artifact, "images/*/"))
+    assert matches == [
+        "images/sample-a/",
+        "images/sample-b/",
+    ]
+
+
 def test_make_training_pairs_from_nested_metadata_formats(tmp_path: Path) -> None:
     artifact = _FakeArtifactNestedMetadata()
     train_pairs, test_pairs = asyncio.run(
@@ -319,3 +355,121 @@ def test_proportional_manual_sample_counts_respects_capacity_limits() -> None:
     assert train_count + test_count == 6
     assert train_count == 1
     assert test_count == 5
+
+
+def test_sample_pair_lists_auto_mode_subsets_and_splits() -> None:
+    train_pairs = [
+        {"image": Path(f"images/t{i}.tif"), "annotation": Path(f"annotations/t{i}.tif")}
+        for i in range(10)
+    ]
+    test_pairs = []
+
+    sampled_train, sampled_test = sample_pair_lists(
+        train_pairs,
+        test_pairs,
+        requested_total=4,
+        split_mode="auto",
+        train_split_ratio=0.75,
+    )
+
+    assert len(sampled_train) + len(sampled_test) == 4
+    assert len(sampled_train) == 3
+    assert len(sampled_test) == 1
+
+
+def test_sample_pair_lists_manual_mode_proportional() -> None:
+    train_pairs = [
+        {"image": Path(f"images/t{i}.tif"), "annotation": Path(f"annotations/t{i}.tif")}
+        for i in range(8)
+    ]
+    test_pairs = [
+        {"image": Path(f"images/v{i}.tif"), "annotation": Path(f"annotations/v{i}.tif")}
+        for i in range(2)
+    ]
+
+    sampled_train, sampled_test = sample_pair_lists(
+        train_pairs,
+        test_pairs,
+        requested_total=5,
+        split_mode="manual",
+        train_split_ratio=0.8,
+    )
+
+    assert len(sampled_train) + len(sampled_test) == 5
+    assert len(sampled_train) == 4
+    assert len(sampled_test) == 1
+
+
+def test_preflight_training_dataset_reports_pair_counts(monkeypatch) -> None:
+    class _FakeArtifactForPreflight:
+        async def ls(self, _folder_path: str):
+            await asyncio.sleep(0)
+            return []
+
+    async def _fake_make_artifact_client(_artifact_id: str, _server_url: str):
+        return _FakeArtifactForPreflight()
+
+    async def _fake_list_matching_paths(_artifact, path_pattern: str, max_results=None):
+        if "images" in path_pattern:
+            return [f"images/t{i}.tif" for i in range(6)]
+        if "annotations" in path_pattern:
+            return [f"annotations/t{i}_mask.tif" for i in range(5)]
+        return []
+
+    monkeypatch.setattr(
+        "bioengine_apps.cellpose_finetuning.main.make_artifact_client",
+        _fake_make_artifact_client,
+    )
+    monkeypatch.setattr(
+        "bioengine_apps.cellpose_finetuning.main.list_matching_artifact_paths",
+        _fake_list_matching_paths,
+    )
+
+    service = CellposeFinetune.func_or_class()
+    result = asyncio.run(
+        service.preflight_training_dataset(
+            artifact="ri-scale/zarr-demo",
+            train_images="images/*/",
+            train_annotations="annotations/*/",
+            split_mode="auto",
+            n_samples=5,
+        )
+    )
+
+    assert result["ok"] is True
+    assert result["train_image_count"] == 6
+    assert result["train_annotation_count"] == 5
+    assert result["train_pair_count"] == 5
+    assert result["sampled_total_count"] == 5
+
+
+def test_preflight_training_dataset_metadata_mode(monkeypatch) -> None:
+    class _FakeArtifactWithMetadata:
+        async def ls(self, folder_path: str):
+            await asyncio.sleep(0)
+            if folder_path == "metadata/":
+                return [
+                    {"path": "metadata/records.json", "type": "file"},
+                    {"path": "metadata/readme.txt", "type": "file"},
+                ]
+            return []
+
+    async def _fake_make_artifact_client(_artifact_id: str, _server_url: str):
+        return _FakeArtifactWithMetadata()
+
+    monkeypatch.setattr(
+        "bioengine_apps.cellpose_finetuning.main.make_artifact_client",
+        _fake_make_artifact_client,
+    )
+
+    service = CellposeFinetune.func_or_class()
+    result = asyncio.run(
+        service.preflight_training_dataset(
+            artifact="ri-scale/zarr-demo",
+            metadata_dir="metadata/",
+        )
+    )
+
+    assert result["ok"] is True
+    assert result["mode"] == "metadata"
+    assert result["metadata_file_count"] == 1
diff --git a/tests/cellpose/test_service.py b/tests/cellpose/test_service.py
index 8df44ced..a0dc84e7 100644
--- a/tests/cellpose/test_service.py
+++ b/tests/cellpose/test_service.py
@@ -39,12 +39,17 @@
 import asyncio
 import logging
 import os
+import sys
 from datetime import datetime, timezone
+from pathlib import Path
 from typing import TYPE_CHECKING
 
 # from dotenv import load_dotenv
 from hypha_rpc import connect_to_server, login
 
+sys.path.append(str(Path(__file__).resolve().parent))
+from live_test_utils import resolve_cellpose_service
+
 if TYPE_CHECKING:
     from hypha_rpc.rpc import RemoteService
 
@@ -67,9 +72,22 @@ async def get_cellpose_service(
     resolved_service_id = service_id or f"{workspace}/{application_id}"
     logger.info("Using cellpose service id: %s", resolved_service_id)
 
-    cellpose_service = await server.get_service(resolved_service_id)
-    logger.info("Obtained Cellpose Fine-Tuning service")
-    return cellpose_service
+    try:
+        cellpose_service = await resolve_cellpose_service(
+            server,
+            workspace,
+            application_id=application_id,
+            requested_service_id=resolved_service_id,
+        )
+        logger.info("Obtained Cellpose Fine-Tuning service")
+        return cellpose_service
+    except Exception as original_error:
+        msg = (
+            f"Could not resolve Cellpose service '{resolved_service_id}'. "
+            "Verify deployment and websocket service ids in "
+            "worker.get_application_status(...)[application_id]['service_ids']."
+        )
+        raise RuntimeError(msg) from original_error
 
 
 async def infer(
@@ -318,7 +336,9 @@ async def start_training(
     test_images: str | None,
     test_annotations: str | None,
     n_epochs: int,
-    n_samples: int | None,
+    n_samples: float | None,
+    split_mode: str,
+    train_split_ratio: float,
     learning_rate: float,
     min_train_masks: int,
     validation_interval: int | None,
@@ -333,6 +353,8 @@ async def start_training(
         "learning_rate": learning_rate,
         "n_samples": n_samples,
         "min_train_masks": min_train_masks,
+        "split_mode": split_mode,
+        "train_split_ratio": train_split_ratio,
     }
     if validation_interval is not None:
         kwargs["validation_interval"] = validation_interval
@@ -379,7 +401,19 @@ def _parse_args() -> argparse.Namespace:
     p.add_argument("--test-annotations", default=None)
 
     p.add_argument("--n-epochs", type=int, default=10)
-    p.add_argument("--n-samples", type=int, default=None)
+    p.add_argument("--n-samples", type=float, default=None)
+    p.add_argument(
+        "--split-mode",
+        default="manual",
+        choices=["manual", "auto"],
+        help="Dataset split mode for training.",
+    )
+    p.add_argument(
+        "--train-split-ratio",
+        type=float,
+        default=0.8,
+        help="Train split ratio when --split-mode auto.",
+    )
     p.add_argument("--learning-rate", type=float, default=1e-5)
     p.add_argument("--min-train-masks", type=int, default=5)
     p.add_argument(
@@ -471,6 +505,8 @@ async def main() -> None:
                 test_annotations=args.test_annotations,
                 n_epochs=args.n_epochs,
                 n_samples=args.n_samples,
+                split_mode=args.split_mode,
+                train_split_ratio=args.train_split_ratio,
                 learning_rate=args.learning_rate,
                 min_train_masks=args.min_train_masks,
                 validation_interval=args.validation_interval,
diff --git a/tests/cellpose/test_training_e2e_small_percentage.py b/tests/cellpose/test_training_e2e_small_percentage.py
new file mode 100644
index 00000000..fddf9654
--- /dev/null
+++ b/tests/cellpose/test_training_e2e_small_percentage.py
@@ -0,0 +1,106 @@
+from __future__ import annotations
+
+import asyncio
+import os
+import sys
+import time
+from pathlib import Path
+
+import pytest
+from hypha_rpc import connect_to_server
+
+sys.path.append(str(Path(__file__).resolve().parent))
+from live_test_utils import resolve_cellpose_service
+
+SERVER_URL = os.environ.get("HYPHA_SERVER_URL", "https://hypha.aicell.io")
+RUN_LIVE_REGRESSION_TESTS = os.environ.get("RUN_LIVE_REGRESSION_TESTS") == "1"
+
+
+def _require_live_env() -> tuple[str, str, str]:
+    if not RUN_LIVE_REGRESSION_TESTS:
+        pytest.skip("Set RUN_LIVE_REGRESSION_TESTS=1 to run live e2e tests")
+
+    token = os.environ.get("HYPHA_TOKEN")
+    if not token:
+        pytest.skip("HYPHA_TOKEN is required for live e2e tests")
+
+    workspace = os.environ.get("HYPHA_TEST_WORKSPACE", "ri-scale")
+    service_id = os.environ.get("HYPHA_TEST_SERVICE_ID", "bioimage-io/cellpose-finetuning")
+    return token, workspace, service_id
+
+
+@pytest.mark.asyncio
+async def test_small_percentage_auto_split_progresses() -> None:
+    token, workspace, service_id = _require_live_env()
+
+    async with connect_to_server(
+        {"server_url": SERVER_URL, "token": token, "workspace": workspace}
+    ) as server:
+        service = await resolve_cellpose_service(
+            server,
+            workspace,
+            requested_service_id=service_id,
+        )
+
+        started = await service.start_training(
+            artifact="ri-scale/zarr-demo",
+            train_images="images/*/",
+            train_annotations="annotations/*/",
+            split_mode="auto",
+            train_split_ratio=0.8,
+            n_samples=0.02,
+            n_epochs=2,
+            min_train_masks=1,
+        )
+
+        session_id = started["session_id"]
+        assert session_id
+
+        reached_non_preparing = False
+        deadline = time.time() + 240
+
+        try:
+            while time.time() < deadline:
+                status = await service.get_training_status(session_id)
+                status_type = str(status.get("status_type", "")).lower()
+
+                if status_type != "preparing":
+                    reached_non_preparing = True
+                    break
+
+                await asyncio.sleep(3)
+        finally:
+            latest = await service.get_training_status(session_id)
+            latest_type = str(latest.get("status_type", "")).lower()
+            if latest_type in {"preparing", "running", "waiting"}:
+                await service.stop_training(session_id=session_id)
+
+        assert reached_non_preparing, "Training remained stuck in 'preparing' state"
+
+
+@pytest.mark.asyncio
+async def test_infer_pretrained_model_e2e() -> None:
+    token, workspace, service_id = _require_live_env()
+
+    async with connect_to_server(
+        {"server_url": SERVER_URL, "token": token, "workspace": workspace}
+    ) as server:
+        service = await resolve_cellpose_service(
+            server,
+            workspace,
+            requested_service_id=service_id,
+        )
+
+        result = await service.infer(
+            artifact="ri-scale/zarr-demo",
+            image_paths=["images/108bb69d-2e52-4382-8100-e96173db24ee/t0000.ome.tif"],
+            model="cpsam",
+            diameter=40,
+            json_safe=True,
+        )
+
+        assert isinstance(result, list)
+        assert len(result) == 1
+        output = result[0].get("output")
+        assert output is not None
+        assert output.get("encoding") in {"ndarray_base64", "mask_png_base64"}
diff --git a/tests/cellpose/test_training_e2e_three_sources.py b/tests/cellpose/test_training_e2e_three_sources.py
new file mode 100644
index 00000000..4b55fff5
--- /dev/null
+++ b/tests/cellpose/test_training_e2e_three_sources.py
@@ -0,0 +1,409 @@
+from __future__ import annotations
+
+import asyncio
+import os
+import sys
+import uuid
+from pathlib import Path
+from typing import Any
+
+import httpx
+import pytest
+
+sys.path.append(str(Path(__file__).resolve().parent))
+from live_test_utils import resolve_cellpose_service, with_live_server
+
+RUN_LIVE_REGRESSION_TESTS = os.environ.get("RUN_LIVE_REGRESSION_TESTS") == "1"
+ROOT_DIR = Path(__file__).resolve().parents[2]
+DEMO_DATASET_DIR = ROOT_DIR / "demo-dataset"
+
+
+def _require_live_env() -> tuple[str, str]:
+    if not RUN_LIVE_REGRESSION_TESTS:
+        pytest.skip("Set RUN_LIVE_REGRESSION_TESTS=1 to run live e2e tests")
+
+    token = os.environ.get("HYPHA_TOKEN")
+    if not token:
+        pytest.skip("HYPHA_TOKEN is required for live e2e tests")
+
+    workspace = os.environ.get("HYPHA_TEST_WORKSPACE", "ri-scale")
+    return token, workspace
+
+
+async def _upload_demo_dataset(server: Any, workspace: str, local_dir: Path) -> str:
+    artifact_manager = await server.get_service("public/artifact-manager")
+    alias = f"demo-dataset-e2e-{uuid.uuid4().hex[:8]}"
+
+    artifact = await artifact_manager.create(
+        type="dataset",
+        alias=alias,
+        manifest={
+            "name": alias,
+            "description": "Temporary demo dataset for cellpose e2e tests",
+            "type": "dataset",
+        },
+        config={"permissions": {"@": "*"}},
+        stage=True,
+    )
+
+    all_files = [p for p in local_dir.rglob("*") if p.is_file()]
+    timeout = httpx.Timeout(120.0)
+    async with httpx.AsyncClient(timeout=timeout) as client:
+        for file_path in all_files:
+            rel_path = file_path.relative_to(local_dir).as_posix()
+            put_url = await artifact_manager.put_file(artifact.id, file_path=rel_path)
+            with open(file_path, "rb") as f:
+                response = await client.put(put_url, content=f.read())
+            response.raise_for_status()
+
+    await artifact_manager.commit(artifact.id)
+    return artifact.id
+
+
+async def _run_training_to_finish(
+    service: Any,
+    *,
+    artifact: str,
+    train_images: str | None,
+    train_annotations: str | None,
+    n_samples: float = 5,
+    retry_sample_fractions: tuple[float, ...] = (8, 10),
+    train_split_ratio: float = 0.8,
+    n_epochs: int = 2,
+    timeout_seconds: int = 360,
+) -> dict[str, Any]:
+    print(
+        f"[TRAIN] artifact={artifact} n_samples={n_samples} n_epochs={n_epochs}",
+        flush=True,
+    )
+
+    async def _get_status_with_retries(
+        session_id: str,
+        attempts: int = 10,
+        delay_seconds: float = 5.0,
+    ) -> dict[str, Any]:
+        last_error: Exception | None = None
+        for _ in range(attempts):
+            try:
+                return await service.get_training_status(session_id=session_id)
+            except Exception as e:
+                last_error = e
+                text = str(e).lower()
+                is_timeout = "timed out" in text or isinstance(e, TimeoutError)
+                if not is_timeout:
+                    raise
+                await asyncio.sleep(delay_seconds)
+        if last_error is not None:
+            return {
+                "status_type": "preparing",
+                "message": f"Transient status polling timeout: {last_error}",
+            }
+        return {}
+
+    async def _stop_active_sessions() -> None:
+        try:
+            sessions = await service.list_training_sessions(
+                status_types=["waiting", "preparing", "running"]
+            )
+        except Exception:
+            return
+
+        if not isinstance(sessions, dict):
+            return
+
+        for existing_session_id, session_status in sessions.items():
+            status_type = str((session_status or {}).get("status_type", "")).lower()
+            if status_type in {"waiting", "preparing", "running"}:
+                try:
+                    await service.stop_training(session_id=existing_session_id)
+                except Exception:
+                    continue
+
+        await asyncio.sleep(2)
+
+    candidate_fractions = (n_samples, *retry_sample_fractions)
+    last_status: dict[str, Any] = {}
+
+    for sample_fraction in candidate_fractions:
+        for _network_retry in range(2):
+            print(
+                f"[TRAIN] attempt sample_fraction={sample_fraction} network_retry={_network_retry}",
+                flush=True,
+            )
+            await _stop_active_sessions()
+
+            kwargs: dict[str, Any] = {
+                "artifact": artifact,
+                "split_mode": "auto",
+                "train_split_ratio": train_split_ratio,
+                "n_samples": sample_fraction,
+                "n_epochs": n_epochs,
+                "validation_interval": 1,
+                "min_train_masks": 0,
+            }
+            if train_images is not None:
+                kwargs["train_images"] = train_images
+            if train_annotations is not None:
+                kwargs["train_annotations"] = train_annotations
+
+            try:
+                started = await service.start_training(**kwargs)
+                session_id = started["session_id"]
+                print(
+                    f"[TRAIN] started session_id={session_id} status={started.get('status_type')} msg={started.get('message')}",
+                    flush=True,
+                )
+
+                started_type = str(started.get("status_type", "")).lower()
+                started_message = str(started.get("message", "")).lower()
+                deferred_start = (
+                    started_type == "stopped"
+                    and "deferred start" in started_message
+                    and "gpu contention" in started_message
+                )
+                if deferred_start:
+                    restarted = await service.restart_training(
+                        session_id=session_id, n_epochs=n_epochs
+                    )
+                    session_id = restarted["session_id"]
+                    print(
+                        f"[TRAIN] restarted deferred session -> {session_id}",
+                        flush=True,
+                    )
+
+                assert session_id
+
+                deadline = asyncio.get_running_loop().time() + timeout_seconds
+                latest_status: dict[str, Any] = {}
+
+                try:
+                    while asyncio.get_running_loop().time() < deadline:
+                        latest_status = await _get_status_with_retries(
+                            session_id=session_id
+                        )
+                        status_type = str(latest_status.get("status_type", "")).lower()
+                        print(
+                            f"[TRAIN] session_id={session_id} status={status_type} msg={latest_status.get('message')}",
+                            flush=True,
+                        )
+                        if status_type in {"completed", "failed", "stopped"}:
+                            break
+                        await asyncio.sleep(3)
+                finally:
+                    status_type = str(latest_status.get("status_type", "")).lower()
+                    if status_type in {"waiting", "preparing", "running"}:
+                        try:
+                            await service.stop_training(session_id=session_id)
+                        except Exception:
+                            pass
+
+                latest_status = await _get_status_with_retries(session_id=session_id)
+                last_status = latest_status
+                final_type = str(latest_status.get("status_type", "")).lower()
+                print(
+                    f"[TRAIN] final session_id={session_id} status={final_type} msg={latest_status.get('message')}",
+                    flush=True,
+                )
+                if final_type == "completed":
+                    latest_status["n_samples_used"] = sample_fraction
+                    latest_status["session_id"] = session_id
+                    return latest_status
+
+                message = str(latest_status.get("message") or "").lower()
+                retriable = (
+                    "no training samples available" in message
+                    or "no training pairs found" in message
+                    or "float division by zero" in message
+                )
+                if not retriable:
+                    break
+            except Exception as e:
+                err = str(e).lower()
+                transient = (
+                    "websocket" in err
+                    or "http 404" in err
+                    or "http 502" in err
+                    or "method call timed out" in err
+                )
+                if transient and _network_retry == 0:
+                    await asyncio.sleep(5)
+                    continue
+                raise
+
+    final_type = str(last_status.get("status_type", "")).lower()
+    assert final_type == "completed", (
+        f"Training did not complete for artifact={artifact}. "
+        f"final status={final_type}, message={last_status.get('message')}"
+    )
+    return last_status
+
+
+async def _infer_with_session_model(
+    service: Any,
+    *,
+    session_id: str,
+    artifact: str,
+    image_path: str,
+) -> dict[str, Any]:
+    print(
+        f"[INFER] model_session={session_id} artifact={artifact} image={image_path}",
+        flush=True,
+    )
+    result = await service.infer(
+        model=session_id,
+        artifact=artifact,
+        image_paths=[image_path],
+        diameter=40,
+        json_safe=True,
+    )
+    assert isinstance(result, list)
+    assert len(result) == 1
+    output = result[0].get("output")
+    assert output is not None
+    assert output.get("encoding") in {"ndarray_base64", "mask_png_base64"}
+    print(
+        f"[INFER] ok model_session={session_id} encoding={output.get('encoding')}",
+        flush=True,
+    )
+    return result[0]
+
+
+@pytest.mark.asyncio
+async def test_training_bia_url_small_percentage_auto_split() -> None:
+    token, workspace = _require_live_env()
+
+    async def _run(server: Any) -> None:
+        service = await resolve_cellpose_service(server, workspace)
+        await _run_training_to_finish(
+            service,
+            artifact="https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD1392",
+            train_images="images/*/",
+            train_annotations="annotations/*/",
+            n_samples=0.005,
+            train_split_ratio=0.8,
+            timeout_seconds=420,
+        )
+
+    await with_live_server(token, workspace, _run)
+
+
+@pytest.mark.asyncio
+async def test_training_ri_scale_zarr_demo_small_percentage_auto_split() -> None:
+    token, workspace = _require_live_env()
+
+    async def _run(server: Any) -> None:
+        service = await resolve_cellpose_service(server, workspace)
+        await _run_training_to_finish(
+            service,
+            artifact="ri-scale/zarr-demo",
+            train_images="images/108bb69d-2e52-4382-8100-e96173db24ee/*.ome.tif",
+            train_annotations="annotations/108bb69d-2e52-4382-8100-e96173db24ee/*_mask.ome.tif",
+            n_samples=0.005,
+            train_split_ratio=0.8,
+            timeout_seconds=1200,
+        )
+
+    await with_live_server(token, workspace, _run)
+
+
+@pytest.mark.asyncio
+async def test_training_local_demo_dataset_small_percentage_auto_split() -> None:
+    token, workspace = _require_live_env()
+
+    if not DEMO_DATASET_DIR.exists():
+        pytest.skip(f"demo dataset not found at {DEMO_DATASET_DIR}")
+
+    async def _run(server: Any) -> None:
+        service = await resolve_cellpose_service(server, workspace)
+        artifact_id = await _upload_demo_dataset(server, workspace, DEMO_DATASET_DIR)
+
+        await _run_training_to_finish(
+            service,
+            artifact=artifact_id,
+            train_images="images/*/*.ome.tif",
+            train_annotations="annotations/*/*_mask.ome.tif",
+            n_samples=0.005,
+            train_split_ratio=0.8,
+            timeout_seconds=1200,
+        )
+
+    await with_live_server(token, workspace, _run)
+
+
+@pytest.mark.asyncio
+async def test_three_sources_sequential_with_inference() -> None:
+    token, workspace = _require_live_env()
+
+    async def _run(server: Any) -> None:
+        service = await resolve_cellpose_service(server, workspace)
+
+        if not DEMO_DATASET_DIR.exists():
+            pytest.skip(f"demo dataset not found at {DEMO_DATASET_DIR}")
+
+        shared_infer_artifact = "ri-scale/zarr-demo"
+        shared_infer_image = "images/108bb69d-2e52-4382-8100-e96173db24ee/t0000.ome.tif"
+
+        # 1) Artifact source (pre-existing artifact id) using uploaded demo dataset
+        print("[E2E] Source 1/3: artifact-id source", flush=True)
+        artifact_source_id = await _upload_demo_dataset(server, workspace, DEMO_DATASET_DIR)
+        artifact_status = await _run_training_to_finish(
+            service,
+            artifact=artifact_source_id,
+            train_images="images/*/*.ome.tif",
+            train_annotations="annotations/*/*_mask.ome.tif",
+            n_samples=5,
+            n_epochs=2,
+            timeout_seconds=1200,
+        )
+        artifact_session_id = str(artifact_status.get("session_id") or "")
+        assert artifact_session_id
+        await _infer_with_session_model(
+            service,
+            session_id=artifact_session_id,
+            artifact=shared_infer_artifact,
+            image_path=shared_infer_image,
+        )
+
+        # 2) BioImage Archive URL source
+        print("[E2E] Source 2/3: BioImage Archive URL source", flush=True)
+        bia_status = await _run_training_to_finish(
+            service,
+            artifact="https://www.ebi.ac.uk/biostudies/bioimages/studies/S-BIAD1392",
+            train_images="images/*/",
+            train_annotations="annotations/*/",
+            n_samples=5,
+            n_epochs=2,
+            timeout_seconds=1200,
+        )
+        bia_session_id = str(bia_status.get("session_id") or "")
+        assert bia_session_id
+        await _infer_with_session_model(
+            service,
+            session_id=bia_session_id,
+            artifact=shared_infer_artifact,
+            image_path=shared_infer_image,
+        )
+
+        # 3) Uploaded local demo dataset source
+        print("[E2E] Source 3/3: uploaded local dataset source", flush=True)
+        uploaded_artifact_id = await _upload_demo_dataset(server, workspace, DEMO_DATASET_DIR)
+        uploaded_status = await _run_training_to_finish(
+            service,
+            artifact=uploaded_artifact_id,
+            train_images="images/*/*.ome.tif",
+            train_annotations="annotations/*/*_mask.ome.tif",
+            n_samples=5,
+            n_epochs=2,
+            timeout_seconds=1200,
+        )
+        uploaded_session_id = str(uploaded_status.get("session_id") or "")
+        assert uploaded_session_id
+        await _infer_with_session_model(
+            service,
+            session_id=uploaded_session_id,
+            artifact=shared_infer_artifact,
+            image_path=shared_infer_image,
+        )
+        print("[E2E] Completed all three sources with inference", flush=True)
+
+    await with_live_server(token, workspace, _run)