#!/usr/bin/env bash
set -uo pipefail

# Smoke test: /skill:propagate loads and the model responds with test generation methodology
# Expects: the response discusses test obligations, assertions, or test categories

cd "$(dirname "$0")/.."

MODEL="Qwen3.6-35B-A3B-MXFP4_MOE.gguf"
TIMEOUT=180

echo "=== propagate: invoking pi ==="

output=$(timeout "$TIMEOUT" pi -p --no-session --model "$MODEL" \
    "/skill:propagate Given an Allium entity Order with status: pending | confirmed | shipped and a rule ConfirmOrder that transitions pending to confirmed, what test obligations would you derive? Keep it under 100 words." 2>&1)
rc=$?

if [ $rc -ne 0 ]; then
    echo "  FAIL: pi exited with code $rc"
    echo "$output" | tail -20
    exit 1
fi

if [ -z "$output" ]; then
    echo "  FAIL: empty response"
    exit 1
fi

echo "$output"
echo ""

# Check the response is relevant to test generation
if echo "$output" | grep -iqE 'test|assert|obligation|transition|valid|invalid|state|verif|propert'; then
    echo "  PASS: response contains propagation-relevant content"
    exit 0
else
    echo "  FAIL: response does not appear to follow propagation methodology"
    exit 1
fi