Local code intelligence for developers and AI agents. Connect code, docs and infrastructure in a persistent graph, with CBM Lens, MCP tools, agent integrations and change-aware test selection.
C aiderastclaude-codecode-analysiscode-intelligencecodexcursorcypherdeveloper-toolsgemini-cligraph-visualizationkilocodeknowledge-graphmcpmcp-servermodel-context-protocolopencodesqlitetree-sitterwindsurf
fix(vm): require the completion marker of the full leg only
Martin Vogel9db5217bdbb7
diff --git a/test-infrastructure/vm/vm-run-tests.sh b/test-infrastructure/vm/vm-run-tests.shindex 5e5760741..ec17f771d 100644--- a/test-infrastructure/vm/vm-run-tests.sh+++ b/test-infrastructure/vm/vm-run-tests.sh@@ -49,9 +49,17 @@ esac# logic with synthetic logs instead of a copy of it that can drift# (tests/test_vm_verdict_contract.sh). See the two-channel note at the call# site for why the log outranks the exit status.+# `mode` is "full" for the whole venue leg and "iteration" for a named subset of+# suites. Only the full leg prints the completion marker, because only the full+# leg HAS an end to reach: scripts/test.sh --suites finishes after the suites it+# was given and says nothing more. Requiring the marker in both modes made every+# `win.sh test <suites>` run report failure while passing -- 23 passed, 1+# skipped, rc=0, called red. A guard against false greens is not allowed to+# invent false reds.vm_verdict() {local log="$1"local rc="$2"+ local mode="${3:-full}"if ! grep -Eq '[0-9]+ passed' "$log"; thenecho "GUARD: test runner produced no completion summary — the suites did" \"not validly run; treating as failure (runner rc=$rc)" >&2@@ -66,11 +74,19 @@ vm_verdict() {echo "GUARD: the log reports $failed_total failed test(s) (runner rc=$rc)" >&2return 1fi- if [ "$complete" -eq 0 ]; then+ if [ "$mode" = "full" ] && [ "$complete" -eq 0 ]; thenecho "GUARD: the log has a summary but no completion marker — the leg" \"stopped before the end (runner rc=$rc)" >&2return 1fi+ if [ "$mode" != "full" ] && [ "${rc:-1}" -ne 0 ]; then+ # No marker to lean on here, so a non-zero status is the only evidence+ # that the run ended badly after its last summary line. Iteration mode+ # is a developer tool, not a gate, so it obeys rc rather than overriding+ # it the way the full leg does.+ echo "GUARD: suites reported no failures but the run exited $rc" >&2+ return 1+ fiif [ "${rc:-1}" -ne 0 ]; thenecho "GUARD: every suite passed and the run completed, but the exit status" \"came back as $rc — the ssh/msys2_shell chain lost it. Reporting the" \@@ -81,11 +97,11 @@ vm_verdict() {# Verdict-only mode for the contract test: decide a log WITHOUT a VM.if [ "${1:-}" = "--verdict" ]; then- if [ $# -ne 3 ]; then- echo "usage: vm-run-tests.sh --verdict <log> <rc>" >&2+ if [ $# -lt 3 ] || [ $# -gt 4 ]; then+ echo "usage: vm-run-tests.sh --verdict <log> <rc> [full|iteration]" >&2exit 2fi- vm_verdict "$2" "$3"+ vm_verdict "$2" "$3" "${4:-full}"exit $?fi@@ -227,7 +243,9 @@ fi## So the LOG decides the test outcome (the runner writes it locally; it cannot# be mangled in transit) and rc decides what the log cannot see. Green requires-# zero reported failures AND the completion marker that scripts/test.sh prints-# as its last statement — a leg that stopped early has a summary but no marker.-vm_verdict "$LOG" "$rc"+# zero reported failures AND, for the full leg, the completion marker that+# scripts/test.sh prints as its last statement — a leg that stopped early has a+# summary but no marker. A named subset of suites prints no marker at all, so it+# is judged on its summaries and its exit status instead.+vm_verdict "$LOG" "$rc" "$([ "$1" = "--par" ] && echo full || echo iteration)"exit $?diff --git a/tests/test_vm_verdict_contract.sh b/tests/test_vm_verdict_contract.shindex 3076eb820..275f865e8 100755--- a/tests/test_vm_verdict_contract.sh+++ b/tests/test_vm_verdict_contract.sh@@ -36,9 +36,9 @@ printf ' 7925 passed, 0 failed, 74 skipped\n' > "$WORK/no_marker"printf 'configure: something exploded\n' > "$WORK/no_summary"failures=0-expect() { # description log rc want- local description="$1" log="$2" rc="$3" want="$4" got- bash "$RUNNER" --verdict "$WORK/$log" "$rc" >/dev/null 2>&1+expect() { # description log rc want [mode]+ local description="$1" log="$2" rc="$3" want="$4" mode="${5:-full}" got+ bash "$RUNNER" --verdict "$WORK/$log" "$rc" "$mode" >/dev/null 2>&1got=$?if [ "$got" -ne "$want" ]; thenecho "FAIL: $description — verdict $got, expected $want" >&2@@ -65,8 +65,21 @@ expect "a summary without the completion marker is red" no_marker 0 1# distinguishable from "ran and failed".expect "a log with no summary is the no-summary guard" no_summary 0 90+# Iteration mode: `test.sh --suites` finishes after the suites it was named and+# prints no completion marker, because it has no end to reach. Demanding one+# turned a passing subset run into a failure — 23 passed, 1 skipped, rc=0,+# reported red (win.sh test artifact, 2026-09-19). A guard against false greens+# must not invent false reds, so the marker is required of the full leg only.+expect "a subset run with no marker is green in iteration mode" no_marker 0 0 iteration+expect "the same log is still red for the full leg" no_marker 0 1 full+# Without a marker to lean on, rc is the only evidence left that a subset run+# ended badly after its last summary, so iteration mode obeys it.+expect "a non-zero status is red in iteration mode" no_marker 1 1 iteration+# Failures still outrank everything, in either mode.+expect "failures are red in iteration mode too" red 0 1 iteration+if [ "$failures" -gt 0 ]; thenecho "VM verdict contract VIOLATED: $failures case(s)" >&2exit 1fi-echo "VM verdict contract passed (8 cases: lost exit status, false green, partial runs)"+echo "VM verdict contract passed (12 cases: lost exit status, false green, partial runs, iteration mode)"