Limit execute concurrency by default

browser_use/browser/watchdogs/local_browser_watchdog.py

@@ -24,6 +24,23 @@
 	pass
 
 
+CDP_READY_TIMEOUT_ENV = 'TIMEOUT_BrowserCDPReady'
+DEFAULT_CDP_READY_TIMEOUT = 180.0
+
+
+def cdp_ready_timeout() -> float:
+	"""Return the timeout for local Chromium to expose its CDP endpoint."""
+	value = os.getenv(CDP_READY_TIMEOUT_ENV)
+	if value:
+		try:
+			parsed = float(value)
+			if parsed >= 0:
+				return parsed
+		except ValueError:
+			pass
+	return DEFAULT_CDP_READY_TIMEOUT
+
+
 class LocalBrowserWatchdog(BaseWatchdog):
 	"""Manages local browser subprocess lifecycle."""
 
@@ -155,7 +172,7 @@ async def _launch_browser(self, max_retries: int = 3) -> tuple[psutil.Process, s
 				process = psutil.Process(subprocess.pid)
 
 				# Wait for CDP to be ready and get the URL
-				cdp_url = await self._wait_for_cdp_url(debug_port)
+				cdp_url = await self._wait_for_cdp_url(debug_port, timeout=cdp_ready_timeout())
 
 				# Success! Clean up only the temp dirs we created but didn't use
 				currently_used_dir = str(profile.user_data_dir)

docker/Dockerfile.api

@@ -7,12 +7,13 @@
 # Required env vars: MURPHY_API_KEY, OPENAI_API_KEY
 # Optional env vars: MURPHY_MAX_CONCURRENT_JOBS (default: 2), MURPHY_API_PORT (default: 8000)
 
-FROM python:3.11-slim
+FROM python:3.12-slim
 
 ENV DEBIAN_FRONTEND=noninteractive \
     PYTHONUNBUFFERED=1 \
     UV_CACHE_DIR=/root/.cache/uv \
     UV_LINK_MODE=copy \
+    UV_PYTHON_PREFERENCE=only-system \
     CHROME_PATH=/usr/bin/chromium
 
 # Install Chromium + minimal deps
@@ -33,12 +34,22 @@ WORKDIR /app
 
 # Install dependencies first (layer caching)
 COPY pyproject.toml uv.lock* /app/
-RUN uv venv --python 3.11 && uv sync --no-dev --no-install-project
+RUN uv venv --python 3.12 && uv sync --no-dev --no-install-project
 
 # Copy source
 COPY . /app
 RUN uv sync --no-dev
 
+# Create non-root user 1000:1000 with a home directory so browser-use's
+# config/profile dirs (~/.config/browseruse, ~/.cache, etc.) have a writable
+# location when ECS runs the container as user 1000.
+RUN groupadd --gid 1000 murphy && \
+    useradd --uid 1000 --gid 1000 --create-home --home-dir /home/murphy murphy && \
+    chown -R murphy:murphy /app /home/murphy
+
+ENV HOME=/home/murphy
+USER murphy:murphy
+
 EXPOSE 8000
 
-CMD ["uv", "run", "uvicorn", "murphy.api:app", "--host", "0.0.0.0", "--port", "8000", "--timeout-keep-alive", "1800"]
+CMD ["/app/.venv/bin/uvicorn", "murphy.api.rest:app", "--host", "0.0.0.0", "--port", "8000", "--timeout-keep-alive", "1800"]

docs/BROWSER_USE_MODIFICATIONS.md

@@ -32,6 +32,8 @@
 
 13. **Broader empty-response detection** (`browser_use/llm/openai/chat.py`) — Changed empty-content check from `is None` to falsy (`not content`) to also catch empty strings; updated error message and status code to 502 (provider-side issue).
 
+14. **Configurable local CDP readiness timeout** (`browser_use/browser/watchdogs/local_browser_watchdog.py`) — Added `TIMEOUT_BrowserCDPReady` with a 180-second default so slower containerized Chromium startups can wait longer for `/json/version` without modifying the event-level browser start timeouts.
+
 ## Syncing with Upstream
 
 ```bash

murphy/__init__.py

@@ -1,6 +1,6 @@
 """Murphy — AI-driven website evaluation powered by browser-use."""
 
-__version__ = '1.1.0'
+__version__ = '1.1.3'
 
 from murphy.core.analysis import analyze_website as analyze_website
 from murphy.core.execution import execute_tests as execute_tests

murphy/api/jobs.py

@@ -107,14 +107,22 @@ async def _acquire_semaphore() -> bool:
 
 async def _execute_with_semaphore(job: Job, core_fn: Any, req: Any, timeout: int | float) -> None:
 	"""Run core_fn under semaphore, update job status on completion/failure."""
+	task: asyncio.Task[Any] | None = None
 	try:
 		effective = _effective_timeout(timeout)
-		job.result = await asyncio.wait_for(core_fn(req), timeout=effective)
+		task = asyncio.create_task(core_fn(req))
+		job.result = await asyncio.wait_for(task, timeout=effective)
 		job.status = 'completed'
-	except TimeoutError:
-		logger.error('Job %s timed out after %ds', job.id, _effective_timeout(timeout))
-		job.status = 'failed'
-		job.error = f'Job timed out after {_effective_timeout(timeout)}s'
+	except TimeoutError as exc:
+		if task is not None and task.done() and not task.cancelled():
+			tb = traceback.format_exc()
+			logger.error('Job %s failed: %s\n%s', job.id, exc, tb)
+			job.status = 'failed'
+			job.error = f'{type(exc).__name__}: {exc}'
+		else:
+			logger.error('Job %s timed out after %ds', job.id, _effective_timeout(timeout))
+			job.status = 'failed'
+			job.error = f'Job timed out after {_effective_timeout(timeout)}s'
 	except Exception as exc:
 		tb = traceback.format_exc()
 		logger.error('Job %s failed: %s\n%s', job.id, exc, tb)

murphy/api/request_models.py

@@ -54,7 +54,7 @@ class ExecuteRequest(BaseModel):
 	judge_provider: str | None = None
 	judge_model: str | None = None
 	max_steps: int = 15
-	max_concurrent: int = 3
+	max_concurrent: int = 1
 	webhook_url: str | None = None
 	async_mode: bool = Field(False, alias='async')
 

pyproject.toml

@@ -2,7 +2,7 @@
 name = "murphy"
 description = "AI-driven website evaluation powered by browser-use"
 authors = [{ name = "MIH AI B.V." }]
-version = "1.1.0"
+version = "1.1.3"
 readme = "README.md"
 requires-python = ">=3.11,<4.0"
 classifiers = [
@@ -210,3 +210,15 @@ dev-dependencies = [
     "pytest-timeout>=2.4.0",
     "pytest-cov>=7.0.0",
 ]
+
+# Pull torch from the PyTorch CPU-only index. The default PyPI wheel bundles
+# ~2.5 GB of unused NVIDIA CUDA libraries — Fargate has no GPU and torch falls
+# back to CPU mode either way, so the CUDA payload is dead weight that bloats
+# the container image and slows cold starts.
+[tool.uv.sources]
+torch = { index = "pytorch-cpu" }
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true

tests/browser_use/browser/test_local_browser_watchdog.py

@@ -0,0 +1,19 @@
+from browser_use.browser.watchdogs.local_browser_watchdog import cdp_ready_timeout
+
+
+def test_cdp_ready_timeout_defaults_to_180(monkeypatch):
+	monkeypatch.delenv('TIMEOUT_BrowserCDPReady', raising=False)
+
+	assert cdp_ready_timeout() == 180.0
+
+
+def test_cdp_ready_timeout_uses_env_override(monkeypatch):
+	monkeypatch.setenv('TIMEOUT_BrowserCDPReady', '45.5')
+
+	assert cdp_ready_timeout() == 45.5
+
+
+def test_cdp_ready_timeout_ignores_invalid_env_override(monkeypatch):
+	monkeypatch.setenv('TIMEOUT_BrowserCDPReady', 'not-a-number')
+
+	assert cdp_ready_timeout() == 180.0

tests/murphy/api/test_jobs.py

@@ -124,6 +124,19 @@ async def slow_fn(req):
 	assert job.error is not None and 'timed out' in job.error
 
 
+@pytest.mark.asyncio
+async def test_execute_with_semaphore_preserves_inner_timeout(monkeypatch):
+	monkeypatch.setattr('murphy.api.jobs.MURPHY_JOB_TIMEOUT_OVERRIDE', None)
+
+	async def browser_start_fn(req):
+		raise TimeoutError('Browser did not start within 180 seconds')
+
+	job = Job(id='browser-timeout')
+	await _execute_with_semaphore(job, browser_start_fn, {}, timeout=1800)
+	assert job.status == 'failed'
+	assert job.error == 'TimeoutError: Browser did not start within 180 seconds'
+
+
 @pytest.mark.asyncio
 async def test_execute_with_semaphore_exception(monkeypatch):
 	monkeypatch.setattr('murphy.api.jobs.MURPHY_JOB_TIMEOUT_OVERRIDE', None)

tests/murphy/api/test_request_models.py

@@ -100,7 +100,7 @@ def test_execute_request_defaults():
 	assert r.test_plan is None
 	assert r.evaluate_job_id is None
 	assert r.max_steps == 15
-	assert r.max_concurrent == 3
+	assert r.max_concurrent == 1
 
 
 def test_execute_request_with_json_string_test_plan():