Add manual RTX benchmark action with Pages dashboard

README.md

@@ -179,6 +179,8 @@ On macOS, update by hand: `git pull` then `make restart`.
 - [docs/updating.md](docs/updating.md) — how updates and the weekly timer work.
 - [docs/codex.md](docs/codex.md) — using Codex CLI with this server as the model.
 - [docs/claude-code.md](docs/claude-code.md) — same for Claude Code, which also gets MCP.
+- [docs/benchmarks.md](docs/benchmarks.md) — browser-agent benchmark findings, all in one place.
+- [docs/benchmark-action.md](docs/benchmark-action.md) — the manual GitHub Actions job that runs the benchmark on the RTX worker.
 
 ## License
 

bench-site/data/index.json

@@ -0,0 +1,101 @@
+{
+  "schema": 1,
+  "runs": [
+    {
+      "schema": 1,
+      "run_id": "2026092501",
+      "created_utc": "2026-09-25T20:01:34+00:00",
+      "commit": "9f1c2ab",
+      "model": "qwen3.5-9b-q4_k_m",
+      "ctx": 32768,
+      "gpu": "RTX 3070 Ti",
+      "suites": {
+        "fixture": {
+          "passed": 1,
+          "total": 1,
+          "pass_rate": 100.0,
+          "median_seconds": 15.7,
+          "peak_vram_mib": 6269,
+          "peak_prompt_tokens": 4874,
+          "ctx": 32768,
+          "parallel": 1
+        },
+        "live_web": {
+          "passed": 1,
+          "total": 2,
+          "pass_rate": 50.0,
+          "median_seconds": 49.5,
+          "peak_vram_mib": 6120,
+          "peak_prompt_tokens": 11124,
+          "ctx": 32768,
+          "parallel": 1
+        }
+      },
+      "sample": true
+    },
+    {
+      "schema": 1,
+      "run_id": "2026092502",
+      "created_utc": "2026-09-25T21:10:00+00:00",
+      "commit": "b2ac901",
+      "model": "qwen3.5-9b-q4_k_m",
+      "ctx": 49152,
+      "gpu": "RTX 3070 Ti",
+      "suites": {
+        "concurrency_serial": {
+          "passed": 3,
+          "total": 4,
+          "pass_rate": 75.0,
+          "median_seconds": 16.3,
+          "peak_vram_mib": 6823,
+          "peak_prompt_tokens": 6801,
+          "ctx": 49152,
+          "parallel": 2
+        },
+        "concurrency_parallel": {
+          "passed": 2,
+          "total": 4,
+          "pass_rate": 50.0,
+          "median_seconds": 29.7,
+          "peak_vram_mib": 6823,
+          "peak_prompt_tokens": 6659,
+          "ctx": 49152,
+          "parallel": 2
+        }
+      },
+      "sample": true
+    },
+    {
+      "schema": 1,
+      "run_id": "2026092503",
+      "created_utc": "2026-09-26T05:50:24+00:00",
+      "commit": "c77e410",
+      "model": "qwen3.5-9b-q4_k_m",
+      "ctx": 65536,
+      "gpu": "RTX 3070 Ti",
+      "suites": {
+        "long_context": {
+          "passed": 1,
+          "total": 1,
+          "pass_rate": 100.0,
+          "median_seconds": 26.1,
+          "peak_vram_mib": 7323,
+          "peak_prompt_tokens": 59906,
+          "ctx": 65536,
+          "parallel": 1
+        },
+        "live_web": {
+          "passed": 2,
+          "total": 3,
+          "pass_rate": 66.7,
+          "median_seconds": 69.5,
+          "peak_vram_mib": 7100,
+          "peak_prompt_tokens": 11124,
+          "ctx": 65536,
+          "parallel": 1
+        }
+      },
+      "sample": true
+    }
+  ]
+}

bench-site/index.html

@@ -0,0 +1,38 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>RTX benchmark dashboard</title>
+<meta name="description" content="Manual GitHub Actions benchmark results for the local LLM browser agent, on RTX hardware.">
+<link rel="stylesheet" href="style.css">
+</head>
+<body>
+<div class="page">
+
+  <header class="site-header">
+    <h1>RTX benchmark dashboard</h1>
+    <p id="header-note">Loading run data&hellip;</p>
+  </header>
+
+  <div id="app" role="main">
+    <!-- Filled in by app.js. Kept empty here so the page never ships
+         stale numbers baked into the HTML. -->
+  </div>
+
+  <footer class="site-footer">
+    Data from <code>data/index.json</code>, written by <code>tests/bench/collect.py</code>
+    after each manual run of the benchmark workflow.
+  </footer>
+
+</div>
+
+<noscript>
+  <p style="padding:16px;">This dashboard needs JavaScript to read
+  <code>data/index.json</code> and draw the charts. Enable JavaScript, or
+  open that file directly to see the raw numbers.</p>
+</noscript>
+
+<script src="app.js"></script>
+</body>
+</html>