Lightpanda: the headless browser designed for AI and automation
test: surface agent stderr on live-layer failures
Adrià Arrufate2195817911b
diff --git a/test/agent/run.sh b/test/agent/run.shindex 727de70ba..1734b64e6 100755--- a/test/agent/run.sh+++ b/test/agent/run.sh@@ -94,6 +94,13 @@ run_deterministic() {fi}+# Surface the agent's own error output on failure — an instant empty answer+# usually means an API error (bad key, quota, restriction) that only appears+# on stderr.+show_err() {+ sed -n '/^\$usage /!p' "$1" | tail -3 | sed 's/^/ /'+}+# Grep the stable "$usage total=N" line lightpanda prints to stderr and fail# if a single task blew past the token ceiling (runaway agent loop).check_usage() {@@ -117,6 +124,7 @@ run_live_qa() {elsefail "Q&A: expected \"$expected\" not found in answer"info " answer: $(tr '\n' ' ' <"$TMP/out" | cut -c1-200)"+ show_err "$TMP/err"ficheck_usage "$TMP/err" "Q&A \"$expected\""done <"$HERE/cases/static-qa.tsv"@@ -134,7 +142,9 @@ run_live_hn() {check_usage "$TMP/err" "HN save"if [ ! -s "$script" ]; then- fail "HN save produced no script"; return+ fail "HN save produced no script"+ show_err "$TMP/err"+ returnfi# A real replay script drives the page — it must navigate and extract.if grep -q 'goto(' "$script" && grep -q 'extract(' "$script"; then