diff --git a/AGENTS.md b/AGENTS.md
index fba1e17f..53d938ee 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -304,14 +304,18 @@ hatch, the explicit `getLiveUsers` limit and the current sidebar labels.
 - `place.stream.live.getLiveUsers` with no `limit` answers `{}` no matter
   how live the stream is: the handler truncates the streamer list to `limit`,
   so zero means zero. Pass `?limit=50`.
-- A killed harness leaves orphaned node/PDS processes behind. They hold
-  ports and their data dirs under `/tmp` look like the run you are
+- A harness you kill by hand leaves orphaned node/PDS processes behind.
+  They hold ports and their data dirs under `/tmp` look like the run you are
   debugging, so sweep them before re-running:
   `pkill -f 'libstreamplace e2e'; pkill -f 'js/dev-env/run.mjs'`. If you run
   that from a wrapper whose own command line contains the pattern (e.g.
   `docker exec … bash -c "pkill -f 'libstreamplace e2e'"`) it matches and
-  kills itself; bracket a character (`'libstreamplace e2[o]e'`) to exclude
-  the wrapper.
+  kills itself; bracket a character that is in the target
+  (`'libstreamplace e2[e]'`) to exclude the wrapper. The harness's ingest
+  worker re-`setsid`s itself, so it outlives a plain top-level kill — kill it
+  by name too (`pkill -f 'libstreamplace ingest-worker'`), or just let
+  `hack/e2e-web-local.sh` clean up after itself (it sweeps the binaries that
+  appeared during the run, leaving any pre-existing scratch node alone).
 - `js/dev-env` needs Node 22 (better-sqlite3 pin).
 
 ## 6. Git and GitHub _(unverified)_
diff --git a/hack/e2e-web-local.sh b/hack/e2e-web-local.sh
index 6931b71e..1923abf8 100755
--- a/hack/e2e-web-local.sh
+++ b/hack/e2e-web-local.sh
@@ -40,14 +40,30 @@ fi
 ENVFILE="$(mktemp)"
 LOGFILE="$(mktemp)"
 echo "starting e2e harness…"
-"$BUILDDIR/streamplace" e2e > "$ENVFILE" 2> "$LOGFILE" &
+# `streamplace e2e` forks a dev-env node, the server node and an ingest
+# worker; run it in its own process group (setsid) so cleanup can take the
+# whole tree down instead of orphaning the grandchildren.
+#
+# PIDs matching this checkout's binary before we start — never kill a scratch
+# node someone else is running from the same build dir.
+PREEXISTING=" $( { pgrep -f "$BUILDDIR/libstreamplace" 2>/dev/null || true; } | tr '\n' ' ') "
+if command -v setsid >/dev/null 2>&1; then
+  setsid "$BUILDDIR/streamplace" e2e > "$ENVFILE" 2> "$LOGFILE" &
+else
+  "$BUILDDIR/streamplace" e2e > "$ENVFILE" 2> "$LOGFILE" &
+fi
 HARNESS_PID=$!
 cleanup() {
-  kill "$HARNESS_PID" 2>/dev/null || true
-  pkill -P "$HARNESS_PID" 2>/dev/null || true
+  kill -- -"$HARNESS_PID" 2>/dev/null || kill "$HARNESS_PID" 2>/dev/null || true
+  wait "$HARNESS_PID" 2>/dev/null || true
+  # The ingest worker re-setsid's itself, escaping the group kill; sweep the
+  # binaries that appeared since we started.
+  for pid in $(pgrep -f "$BUILDDIR/libstreamplace" 2>/dev/null || true); do
+    case "$PREEXISTING" in *" $pid "*) ;; *) kill "$pid" 2>/dev/null || true ;; esac
+  done
   rm -f "$ENVFILE" "$LOGFILE"
 }
-trap cleanup EXIT
+trap cleanup EXIT INT TERM
 
 for _ in $(seq 1 90); do grep -q SERVER_URL "$ENVFILE" && break; sleep 2; done
 if ! grep -q SERVER_URL "$ENVFILE"; then
@@ -59,5 +75,7 @@ export SERVER_URL ACCOUNT_HANDLE ACCOUNT_DID
 echo "harness up: SERVER_URL=$SERVER_URL ACCOUNT_HANDLE=$ACCOUNT_HANDLE"
 
 # --- run the flows ---------------------------------------------------------
+# No `exec` here: that would replace this shell and discard the EXIT trap,
+# leaving the harness and its children running after a successful run.
 cd "$REPO/js/e2e-web"
-exec pnpm exec playwright test "$@"
+pnpm exec playwright test "$@"