tests/e2e/test_local_provision_lifecycle_e2e.sh

@@ -614,6 +614,21 @@ else
 fi
 echo ""
 
+# ----------------------------------------------------------------------------
+# Advisory-lane infra-skip helper (core#2917 follow-on). The mandatory stub
+# lane must keep asserting the real proxy round-trip; only the advisory
+# real-LLM lane may go green-with-skip when the platform A2A layer is under
+# a known transient degradation (queue never drains / gateway errors).
+infra_skip_advisory() {
+  local reason="$1"
+  local detail="${2:-}"
+  if [ "$LIFECYCLE_LLM" = "minimax" ]; then
+    echo "[$(date +%H:%M:%S)] ⚠️  scan_status: infra-skip:${reason}${detail:+ $detail}"
+    echo "=== Results: advisory infra-skip (${reason}) ==="
+    exit 0
+  fi
+}
+
 # ----------------------------------------------------------------------------
 # Step 5 — proxy reach (ws-<id>:8000 Docker-DNS rewrite, end to end).
 # ----------------------------------------------------------------------------
@@ -640,9 +655,98 @@ print(json.dumps({'method':'message/send','params':{'message':{'role':'user','pa
 # slower than the stub; give the real-LLM call a longer ceiling.
 A2A_CEIL="$A2A_TIMEOUT"
 [ "$LIFECYCLE_LLM" = "minimax" ] && A2A_CEIL="${A2A_MINIMAX_TIMEOUT:-120}"
-A2A=$(curl -s --max-time "$A2A_CEIL" -X POST "$BASE/workspaces/$WSID/a2a" \
+
+# Capture both body and HTTP code so we can detect gateway/queued responses.
+A2A_TMP=$(mktemp)
+set +e
+A2A_CODE=$(curl -s -o "$A2A_TMP" -w '%{http_code}' --max-time "$A2A_CEIL" \
+  -X POST "$BASE/workspaces/$WSID/a2a" \
   -H "Content-Type: application/json" \
   -d "$A2A_BODY")
+A2A_RC=$?
+set -e
+A2A=$(cat "$A2A_TMP" 2>/dev/null || echo "")
+rm -f "$A2A_TMP"
+
+# Gateway/transport failure on the initial POST is an A2A-layer infra issue,
+# not a local-provision code regression. Only skip the advisory lane.
+if [ "$A2A_RC" -ne 0 ] || [ "$A2A_CODE" = "000" ] || [[ "$A2A_CODE" == 5* ]]; then
+  infra_skip_advisory "a2a-gateway-error" "curl_rc=$A2A_RC http=$A2A_CODE"
+fi
+
+# core#2917: the A2A proxy can return a 202-queued envelope instead of a
+# synchronous result. Poll the durable queue result; if the queue never drains,
+# infra-skip the advisory lane rather than falsely blaming local-provision code.
+A2A_QUEUED=$(printf '%s' "$A2A" | python3 -c "
+import sys,json
+try:
+  d=json.load(sys.stdin)
+  print('true' if d.get('queued') is True or (d.get('status') or '').lower() == 'queued' else 'false')
+except Exception:
+  print('false')" 2>/dev/null || echo "false")
+if [ "$A2A_QUEUED" = "true" ]; then
+  A2A_QID=$(printf '%s' "$A2A" | python3 -c "
+import sys,json
+try:
+  print(json.load(sys.stdin).get('queue_id',''))
+except Exception:
+  print('')" 2>/dev/null || echo "")
+  if [ -z "$A2A_QID" ]; then
+    infra_skip_advisory "a2a-queued-no-queue-id" "initial POST was queued but returned no queue_id"
+  fi
+  echo "  A2A queued (queue_id=$A2A_QID); polling durable result..."
+  A2A_POLL_TMP=$(mktemp)
+  A2A_LAST_STATUS=""
+  A2A_POLL_COUNT=0
+  for poll_attempt in $(seq 1 30); do
+    : >"$A2A_POLL_TMP"
+    set +e
+    curl -s -o "$A2A_POLL_TMP" -w '%{http_code}' --max-time 30 \
+      -H "X-Workspace-ID: $WSID" \
+      "$BASE/workspaces/$WSID/a2a/queue/$A2A_QID" >/dev/null 2>&1
+    set -e
+    A2A_POLL_RESP=$(cat "$A2A_POLL_TMP" 2>/dev/null || echo "")
+    A2A_POLL_STATUS=$(printf '%s' "$A2A_POLL_RESP" | python3 -c "
+import sys,json
+try:
+  print(json.load(sys.stdin).get('status',''))
+except Exception:
+  print('')" 2>/dev/null || echo "")
+    A2A_LAST_STATUS="$A2A_POLL_STATUS"
+    A2A_POLL_COUNT=$poll_attempt
+    case "$A2A_POLL_STATUS" in
+      completed)
+        A2A=$(printf '%s' "$A2A_POLL_RESP" | python3 -c "
+import sys,json
+try:
+  rb=json.load(sys.stdin).get('response_body')
+  print(json.dumps(rb) if rb is not None else '')
+except Exception:
+  print('')" 2>/dev/null || echo "")
+        if [ -n "$A2A" ]; then
+          break
+        fi
+        ;;
+      failed|dropped)
+        rm -f "$A2A_POLL_TMP"
+        infra_skip_advisory "a2a-queue-terminal" "queue_id=$A2A_QID status=$A2A_POLL_STATUS"
+        ;;
+      queued|dispatched|in_progress|"")
+        echo "    queue poll $poll_attempt/30 status=$A2A_POLL_STATUS — backing off 2s"
+        sleep 2
+        ;;
+      *)
+        rm -f "$A2A_POLL_TMP"
+        infra_skip_advisory "a2a-queue-unexpected" "queue_id=$A2A_QID status=$A2A_POLL_STATUS"
+        ;;
+    esac
+  done
+  rm -f "$A2A_POLL_TMP"
+  if [ -z "$A2A" ]; then
+    infra_skip_advisory "a2a-queue-timeout" "queue_id=$A2A_QID poll_count=${A2A_POLL_COUNT}/30 last_status=${A2A_LAST_STATUS:-<empty>}"
+  fi
+fi
+
 # Extract the assistant text part once (shared by the minimax assertion +
 # diagnostics). Tolerates result.parts[].text and result.message.parts[].text.
 a2a_text() {