Background jobs and polling

app.py

@@ -210,7 +210,8 @@ def start_evaluation(
                 f"❌ {error}",
                 f"❌ Invalid model ID: {error}",
                 format_leaderboard_df(),
-                gr.update(active=False),  # timer
+                gr.update(visible=False),  # check_status_btn
+                gr.update(visible=False, value=""),  # job_id_display
             )
     else:
         model_id = model_choice
@@ -227,7 +228,8 @@ def start_evaluation(
                 "❌ API key required",
                 f"❌ API key required for {model_id}. Please enter your {key_type.upper()} API key or set the {env_var} environment variable.",
                 format_leaderboard_df(),
-                gr.update(active=False),
+                gr.update(visible=False),
+                gr.update(visible=False, value=""),
             )
 
     # Get model display name
@@ -248,27 +250,28 @@ def start_evaluation(
 
     return (
         job.job_id,
-        "⏳ Starting evaluation...",
+        "⏳ Evaluation started! Click 'Check Status' to see progress (auto-refreshes every 5 seconds).",
         "",
         gr.update(),  # Don't update leaderboard yet
-        gr.update(active=True),  # Activate timer
+        gr.update(visible=True),  # Show check_status_btn
+        gr.update(visible=True, value=f"Job ID: {job.job_id[:8]}..."),  # Show job_id_display
     )
 
 
-def poll_evaluation(job_id: str):
+def check_job_status(job_id: str):
     """
-    Poll for evaluation status by reading the job file.
+    Check job status by reading the job file.
 
-    This is a stateless operation - each poll reads fresh data from disk.
-    This allows the UI to track progress even if the original HTTP connection
-    timed out or the page was refreshed.
+    This is a stateless operation - each check reads fresh data from disk.
+    Uses regular HTTP POST (not SSE) so it survives HF Spaces proxy timeouts.
     """
     if not job_id:
         return (
             "",
             "",
             gr.update(),
-            gr.update(active=False),
+            gr.update(visible=False),  # Hide check button
+            gr.update(visible=False, value=""),  # Hide job ID
         )
 
     job = get_job(job_id)
@@ -276,10 +279,11 @@ def poll_evaluation(job_id: str):
     if job is None:
         # Job not found - might have been cleaned up or never existed
         return (
-            "Job not found. It may have expired or been cleaned up.",
+            "⚠️ Job not found. It may have expired or been cleaned up.",
             "",
             gr.update(),
-            gr.update(active=False),
+            gr.update(visible=False),
+            gr.update(visible=False, value=""),
         )
 
     if job.status == "completed":
@@ -291,7 +295,8 @@ def poll_evaluation(job_id: str):
             job.progress,
             job.result or "",
             format_leaderboard_df(),
-            gr.update(active=False),  # Stop timer
+            gr.update(visible=False),  # Hide check button
+            gr.update(visible=False, value=""),  # Hide job ID
         )
 
     elif job.status == "failed":
@@ -304,7 +309,8 @@ def poll_evaluation(job_id: str):
             job.progress,
             f"❌ Error: {error_msg}",
             format_leaderboard_df(),
-            gr.update(active=False),  # Stop timer
+            gr.update(visible=False),  # Hide check button
+            gr.update(visible=False, value=""),  # Hide job ID
         )
 
     else:
@@ -317,7 +323,8 @@ def poll_evaluation(job_id: str):
             progress_with_pct,
             "",
             gr.update(),  # Don't update leaderboard yet
-            gr.update(active=True),  # Keep timer running
+            gr.update(visible=True),  # Keep check button visible
+            gr.update(visible=True),  # Keep job ID visible
         )
 
 
@@ -524,13 +531,27 @@ def create_app():
                             variant="primary",
                             size="lg",
                         )
-                        
+
+                        # Manual refresh button - visible when a job is running
+                        # This uses regular HTTP POST (not SSE) so it survives proxy timeouts
+                        with gr.Row():
+                            check_status_btn = gr.Button(
+                                "🔄 Check Status",
+                                variant="secondary",
+                                size="sm",
+                                visible=False,
+                            )
+                            job_id_display = gr.Textbox(
+                                label="",
+                                visible=False,
+                                interactive=False,
+                                container=False,
+                                scale=2,
+                            )
+
                         status_text = gr.Markdown("")
-                        
+
                         results_markdown = gr.Markdown("")
-                        
-                        # Timer for polling (initially inactive)
-                        poll_timer = gr.Timer(value=2, active=False)
         
         gr.Markdown("""
         ---
@@ -548,21 +569,45 @@ def create_app():
             inputs=[model_dropdown],
             outputs=[custom_model_input, api_key_input],
         )
-        
+
         # Start evaluation: creates persistent job file and spawns background thread.
         # Returns immediately with job_id so UI doesn't timeout waiting.
         evaluate_btn.click(
             fn=start_evaluation,
             inputs=[model_dropdown, custom_model_input, api_key_input, max_pairs_slider],
-            outputs=[job_id_state, status_text, results_markdown, leaderboard_table, poll_timer],
+            outputs=[job_id_state, status_text, results_markdown, leaderboard_table, check_status_btn, job_id_display],
+        ).then(
+            # Start JavaScript auto-polling after evaluation begins
+            fn=None,
+            inputs=None,
+            outputs=None,
+            js="""
+            () => {
+                // Clear any existing interval
+                if (window.jobPollInterval) {
+                    clearInterval(window.jobPollInterval);
+                }
+                // Auto-click the check status button every 5 seconds
+                window.jobPollInterval = setInterval(() => {
+                    const checkBtn = document.querySelector('button:has(span:contains("Check Status"))') ||
+                                     Array.from(document.querySelectorAll('button')).find(b => b.textContent.includes('Check Status'));
+                    if (checkBtn && checkBtn.offsetParent !== null) {
+                        checkBtn.click();
+                    } else {
+                        // Button is hidden (job done), stop polling
+                        clearInterval(window.jobPollInterval);
+                        window.jobPollInterval = null;
+                    }
+                }, 5000);
+            }
+            """,
         )
 
-        # Poll for updates by reading job file (stateless - survives connection timeouts).
-        # Each poll is a fresh HTTP request that reads the job status from disk.
-        poll_timer.tick(
-            fn=poll_evaluation,
+        # Check status button - uses regular HTTP POST (not SSE) so it survives proxy timeouts
+        check_status_btn.click(
+            fn=check_job_status,
             inputs=[job_id_state],
-            outputs=[status_text, results_markdown, leaderboard_table, poll_timer],
+            outputs=[status_text, results_markdown, leaderboard_table, check_status_btn, job_id_display],
         )
         
         def refresh_leaderboard():