[eric] webapp-cache: validate-or-nuke half-installed warm cache, heal off the hot path

backend/apps/outputs/view_builder_templates.py

@@ -319,13 +319,21 @@ def _ensure_warm_cache() -> str | None:
     with _warm_cache_lock:
         if _warm_cache_is_complete(cache_modules):
             return cache_modules
+        # A node_modules that exists but flunks the completeness check is a
+        # half-finished install; wipe it so the rebuild below starts on clean
+        # ground instead of layering onto a broken tree.
+        if os.path.isdir(cache_modules):
+            shutil.rmtree(cache_modules, ignore_errors=True)
         # Fast path: pre-built archive shipped inside the release. The
         # build script generates this so users hitting OpenSwarm for the
         # first time skip the ~22 s live `npm install`. Falls through on
         # any failure so dev installs (no archive) keep working.
         if _try_extract_bundled_archive(cache_dir, _warm_cache_digest()):
-            logger.info("webapp-template: warm cache ready from bundled archive")
-            return cache_modules
+            if _warm_cache_is_complete(cache_modules):
+                logger.info("webapp-template: warm cache ready from bundled archive")
+                return cache_modules
+            # Archive unpacked a tree without the launch bin; don't trust it.
+            shutil.rmtree(cache_modules, ignore_errors=True)
         try:
             os.makedirs(cache_dir, exist_ok=True)
             # Copy package.json + lockfile (if it exists) into the cache
@@ -370,6 +378,10 @@ def _ensure_warm_cache() -> str | None:
                     (result.stderr or "")[-1500:],
                 )
                 return None
+            # Never hand back a tree the workspace can't actually launch from.
+            if not _warm_cache_is_complete(cache_modules):
+                logger.warning("webapp-template: warm-cache install left no .bin/vite; not caching")
+                return None
             return cache_modules
         except Exception as exc:
             logger.warning("webapp-template warm-cache failed: %s", exc)
@@ -549,7 +561,7 @@ def warm_cache_in_background() -> None:
     global _warm_cache_thread
     if _warm_cache_thread is not None and _warm_cache_thread.is_alive():
         return
-    node_done = os.path.isdir(os.path.join(_warm_cache_dir(), "node_modules"))
+    node_done = _warm_cache_is_complete(os.path.join(_warm_cache_dir(), "node_modules"))
     venv_done = os.path.isfile(os.path.join(_warm_venv_dir(), ".populated"))
     if node_done and venv_done:
         return

backend/apps/outputs/webapp_template/frontend/run.sh

@@ -40,17 +40,18 @@ fi
 # Fast path: the seeder usually symlinks node_modules to a shared warm
 # cache (~/.openswarm/cache/webapp_template_node_modules/<hash>), so the
 # dependency install has already been done once and we can skip straight
-# to vite. Only run npm install when node_modules is genuinely missing
-# or empty — e.g. a workspace seeded before the warm-cache existed, or
-# the user's cache was cleared.
-# A non-empty node_modules is NOT proof of a finished install. 
-# non-empty -> skip" check then trusted that, and `npm run dev` died with
-# `vite: command not found`. Gate on the bin we actually launch with so a
-# broken/partial tree self-heals via install instead of being skipped.
-#So thats why i  explicitly have "/.bin/vite"
+# to vite. A non-empty node_modules is NOT proof of a finished install
+# (npm links .bin/* last, so a killed install leaves trees but no bin and
+# vite dies with "command not found"); gate on the bin we actually launch.
 if [ -e node_modules/.bin/vite ]; then
     echo "Dependencies already present - skipping install."
 else
+    # Incomplete tree. If node_modules is a SYMLINK to the shared warm cache,
+    # never install through it: that writes into the cache every other app
+    # shares (corruption) and stampedes when several apps boot at once. Drop
+    # the link and install a private tree so this app heals alone while the
+    # backend's background warmer rebuilds the shared cache for everyone else.
+    [ -L node_modules ] && rm -f node_modules
     echo "Installing dependencies..."
     "$NPM" install --prefer-offline --no-audit --no-fund
 fi

backend/tests/test_bundled_extracted_modules.py

@@ -31,3 +31,38 @@ def test_no_bundled_tree_returns_none(monkeypatch, tmp_path):
     # caller falls through to the .tar.gz extract or live npm.
     monkeypatch.setattr(vt, "_BUNDLED_ARCHIVE_DIR", str(tmp_path / "empty"))
     assert vt._bundled_extracted_modules() is None
+
+
+def test_warm_cache_is_complete_requires_launch_bin(tmp_path):
+    # A package tree on disk is NOT a finished install; the .bin/vite launch
+    # shim is what proves npm finished its bin-linking phase.
+    nm = tmp_path / "node_modules"
+    (nm / "vite" / "bin").mkdir(parents=True)
+    (nm / "vite" / "bin" / "vite.js").write_text("// vite")
+    assert vt._warm_cache_is_complete(str(nm)) is False
+    bindir = nm / ".bin"
+    bindir.mkdir()
+    (bindir / "vite").symlink_to("../vite/bin/vite.js")
+    assert vt._warm_cache_is_complete(str(nm)) is True
+
+
+def test_ensure_warm_cache_wipes_partial_and_never_returns_incomplete(monkeypatch, tmp_path):
+    # A half-finished cache (package tree present, .bin/vite missing) must be
+    # WIPED and never handed back, so no workspace symlinks to an unlaunchable
+    # tree and run.sh is never pushed into installing through the shared cache.
+    digest = vt._warm_cache_digest()
+    home = tmp_path / "home"
+    monkeypatch.setenv("OPENSWARM_WEBAPP_CACHE_DIR", str(home))
+    cache_modules = home / digest / "node_modules"
+    (cache_modules / "vite" / "bin").mkdir(parents=True)
+    (cache_modules / "vite" / "bin" / "vite.js").write_text("// vite")
+    assert vt._warm_cache_is_complete(str(cache_modules)) is False
+
+    # No bundled tree, no archive, no npm: the only honest answer is "not ready"
+    # (None), and the broken tree must be gone, not cached for the next caller.
+    monkeypatch.setattr(vt, "_BUNDLED_ARCHIVE_DIR", str(tmp_path / "noresources"))
+    monkeypatch.setattr(vt, "_try_extract_bundled_archive", lambda *a, **k: False)
+    monkeypatch.setattr(vt, "_resolve_npm", lambda: None)
+
+    assert vt._ensure_warm_cache() is None
+    assert not cache_modules.exists()