From 4bee0ff18f1d5ac9b2dee854be26aa2fa6dfe1d7 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Sat, 15 Aug 2026 23:24:01 -0700
Subject: [PATCH] [eric] apps: a first boot's pip install gets a real deadline,
 so a loaded machine stops killing new apps at 60s

---
 backend/apps/outputs/webapp_template/run.sh  | 10 ++++-
 backend/tests/test_template_boot_deadline.py | 41 ++++++++++++++++++++
 2 files changed, 50 insertions(+), 1 deletion(-)
 create mode 100644 backend/tests/test_template_boot_deadline.py

diff --git a/backend/apps/outputs/webapp_template/run.sh b/backend/apps/outputs/webapp_template/run.sh
index 35376b48..7310cf54 100755
--- a/backend/apps/outputs/webapp_template/run.sh
+++ b/backend/apps/outputs/webapp_template/run.sh
@@ -70,7 +70,15 @@ if [[ "${BACKEND_PORT}" == "NONE" || -z "${BACKEND_PORT}" || ! -f "$ROOT_DIR/bac
     done
 else
     BACKEND_URL="http://localhost:${BACKEND_PORT:-8324}/api/health/check"
-    MAX_WAIT=60
+    # First boot runs a full venv create + pip install, which takes minutes on a loaded machine;
+    # a 60s wall clock killed exactly those boots mid-install (measured live: "Installing build
+    # dependencies" -> "failed to start within 60s" with 10 agents running). The sentinel is only
+    # written after a successful install, so its absence means the slow path is ahead of us.
+    if [ -f "$ROOT_DIR/backend/.venv/.openswarm_installed" ]; then
+        MAX_WAIT=60
+    else
+        MAX_WAIT=600
+    fi
 
     echo "Starting backend..."
     echo ""
diff --git a/backend/tests/test_template_boot_deadline.py b/backend/tests/test_template_boot_deadline.py
new file mode 100644
index 00000000..414232e0
--- /dev/null
+++ b/backend/tests/test_template_boot_deadline.py
@@ -0,0 +1,41 @@
+"""The template's backend boot deadline must survive a first-boot pip install (found live).
+
+Measured on a packaged build under real load (10 agents + ~20 runtimes): a fresh app's backend was
+still at "Installing build dependencies" when run.sh's flat 60s wall clock fired, printed "Backend
+failed to start within 60s. Aborting.", and the app card sat on "Starting preview" forever. The
+install sentinel is written only after a successful pip install, so its absence IS the signal that
+the slow path lies ahead; the deadline now keys off it. These pin that shape so a refactor cannot
+quietly return to one flat number.
+"""
+import os
+import re
+
+P_TEMPLATE = os.path.join(os.path.dirname(__file__), "..", "apps", "outputs", "webapp_template")
+
+
+def p_read(name: str) -> str:
+    with open(os.path.join(P_TEMPLATE, name), encoding="utf-8") as fh:
+        return fh.read()
+
+
+def test_first_boot_gets_a_pip_sized_deadline():
+    src = p_read("run.sh")
+    assert ".openswarm_installed" in src, "the deadline must key off the install sentinel"
+    waits = [int(m) for m in re.findall(r"MAX_WAIT=(\d+)", src)]
+    assert len(waits) >= 2, "one flat MAX_WAIT is the bug: first boot and warm boot are different animals"
+    assert max(waits) >= 300, f"first-boot budget {max(waits)}s cannot cover a pip install on a loaded machine"
+    assert min(waits) <= 90, "the warm-boot budget must stay tight so a genuinely dead backend still fails fast"
+
+
+def test_the_sentinel_paths_agree_between_the_two_scripts():
+    # run.sh checks the sentinel that backend/run.sh writes; if either side renames it, the check
+    # silently always takes the short deadline, which is the original bug wearing a new hat.
+    outer = p_read("run.sh")
+    inner = p_read(os.path.join("backend", "run.sh"))
+    assert 'SENTINEL="$VENV_DIR/.openswarm_installed"' in inner
+    assert 'backend/.venv/.openswarm_installed' in outer
+
+
+def test_dead_process_still_fails_fast_regardless_of_deadline():
+    src = p_read("run.sh")
+    assert "Backend process died before becoming ready" in src, "a crashed backend must not wait out the long budget"