diff --git a/.gitignore b/.gitignore
index 6b2d7e4..15eac4f 100644
--- a/.gitignore
+++ b/.gitignore
@@ -9,3 +9,12 @@ archived
 *.bak
 *.clean
 session*
+llm.env
+
+# System test artifacts
+tests/results/
+tests/__pycache__/
+tests/**/__pycache__/
+tests/.pytest_cache/
+tests/llm/.cache/
+*.pyc
diff --git a/runner.sh b/runner.sh
new file mode 100755
index 0000000..a849795
--- /dev/null
+++ b/runner.sh
@@ -0,0 +1,112 @@
+#!/bin/bash
+set -euo pipefail
+
+# =============================================================================
+# HoneyBiscuitWorkshop System Test Runner
+# =============================================================================
+# Usage: ./runner.sh [auto|manual] [-- pytest-args...]
+#
+# Modes:
+#   auto   - Build, run all tests, collect reports, cleanup (default)
+#   manual - Build, start container shell for interactive testing
+#
+# Examples:
+#   ./runner.sh                          # Run all tests automatically
+#   ./runner.sh auto -k test_prebake     # Run only prebake tests
+#   ./runner.sh manual                   # Drop into container shell
+# =============================================================================
+
+MODE="${1:-auto}"
+shift 2>/dev/null || true
+PYTEST_ARGS="$@"
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+STAGING_DIR="/tmp/hbw-system-test"
+IMAGE_NAME="hbw-system-test:latest"
+
+# ---- Cleanup handler ----
+cleanup() {
+    echo "[runner] Cleaning up..."
+    docker rm -f hbw-system-test-container 2>/dev/null || true
+    rm -rf "$STAGING_DIR"
+}
+trap cleanup EXIT
+
+# ---- Step 1: Build the binary ----
+build_binary() {
+    echo "[runner] Building workshop-baker (release)..."
+    cd "$SCRIPT_DIR/workshop-baker"
+    cargo build --release 2>&1
+    echo "[runner] Build complete."
+}
+
+# ---- Step 2: Prepare staging directory ----
+prepare_staging() {
+    echo "[runner] Preparing staging directory..."
+    rm -rf "$STAGING_DIR"
+    mkdir -p "$STAGING_DIR"
+
+    # Copy binary
+    cp "$SCRIPT_DIR/workshop-baker/target/release/workshop-baker" "$STAGING_DIR/"
+    chmod +x "$STAGING_DIR/workshop-baker"
+
+    # Copy examples if they exist
+    if [ -d "$SCRIPT_DIR/workshop-baker/examples" ]; then
+        cp -r "$SCRIPT_DIR/workshop-baker/examples" "$STAGING_DIR/examples"
+    else
+        mkdir -p "$STAGING_DIR/examples"
+    fi
+
+    # Copy llm.env if it exists
+    if [ -f "$SCRIPT_DIR/llm.env" ]; then
+        cp "$SCRIPT_DIR/llm.env" "$STAGING_DIR/llm.env"
+    fi
+
+    echo "[runner] Staging ready at $STAGING_DIR"
+}
+
+# ---- Step 3: Build Docker image ----
+build_image() {
+    echo "[runner] Building Docker image: $IMAGE_NAME ..."
+    docker build -t "$IMAGE_NAME" -f "$SCRIPT_DIR/tests/Dockerfile" "$SCRIPT_DIR"
+    echo "[runner] Image built: $IMAGE_NAME"
+}
+
+# ---- Step 4: Run container ----
+run_container() {
+    local mode="$1"
+    echo "[runner] Starting container in $mode mode..."
+
+    mkdir -p "$SCRIPT_DIR/tests/results"
+
+    if [ "$mode" = "manual" ]; then
+        docker run -it --rm \
+            --name hbw-system-test-container \
+            --privileged \
+            -v "$STAGING_DIR:/workshop:ro" \
+            -v "$SCRIPT_DIR/tests/results:/results" \
+            -e TEST_MODE=manual \
+            "$IMAGE_NAME"
+    else
+        docker run --rm \
+            --name hbw-system-test-container \
+            --privileged \
+            -v "$STAGING_DIR:/workshop:ro" \
+            -v "$SCRIPT_DIR/tests/results:/results" \
+            -e TEST_MODE=auto \
+            "$IMAGE_NAME" $PYTEST_ARGS
+    fi
+}
+
+# ---- Main flow ----
+echo "============================================"
+echo "  HoneyBiscuitWorkshop System Test Runner"
+echo "  Mode: $MODE"
+echo "============================================"
+
+build_binary
+prepare_staging
+build_image
+run_container "$MODE"
+
+echo "[runner] Done."
diff --git a/templates/finalize.yml.tmpl b/templates/finalize.yml.tmpl
index e8ee493..cd7b97c 100644
--- a/templates/finalize.yml.tmpl
+++ b/templates/finalize.yml.tmpl
@@ -85,86 +85,67 @@ plugin:
 # 建议将优先级设为 0、1、2... 等连续值
 
 notification:
+  templates:
+    sms-tmpl1:
+      on_success: "Build success!"
+      on_failure: "Build failed!"
+      on_finish_of: "Build stage {{ build.status }} finished."
+    sms-tmpl2:
+      use: "git://https://gitee.com/example_user/example_template_for_sms.git@a1b2c3d4"
+
   # 最高优先级组 - 首选通知方式
   0:
     # Webhook 通知
     webhook:
       url: "https://example.com/hook"
-      # schema 定义负载格式：
-      #   - "default": 使用插件内置的默认格式
-      #   - "custom": 自定义格式，需要提供 on_* 模板
-      # 当设置 on_success/on_failure 时，schema 隐式为 "custom"
-      schema: "default"
+      template:
+        # schema 定义负载格式：
+        #   - "default": 使用系统(ws-base)指定的默认格式，standalone模式不可用，client模式的默认
+        #   - "fallback": 使用插件指定的默认格式，standalone模式的默认
+        #   - "custom": 自定义格式，需要提供全部 on_* 模板
+        #   - "<name>": 使用templates节中定义的名字
+        # 当任何on_*被设置，则等效于replace模式：schema做基底，on_*做替换
+        schema: "default"
 
     # Satori 协议通知（如 Koishi）
     satori:
-      url: "http://127.0.0.1:5140/satori/v1/message.create"
-      token: "Bearer {{ secret.koishi_token }}"
-      schema: "custom"
-
-      # on_success: 构建成功时发送的内容
-      # 支持模板变量：
-      #   {{ build.status }} - 构建状态
-      #   {{ build.duration }} - 构建耗时
-      #   {{ target.xxx }} - 目标相关变量
-      on_success: |
-        {
-          "channel_id": "{{ target.group_id }}",
-          "content": "🍯 HBW 构建完成\n状态: {{ build.status }}\n耗时: {{ build.duration }}"
-        }
-
-      # on_failure: 构建失败时发送的内容
-      on_failure: |
-        {
-          "channel_id": "{{ target.group_id }}",
-          "content": "🍯 HBW 构建失败！\n状态: {{ build.status }}\n耗时: {{ build.duration }}"
-        }
-
-      # on_finish_of: 在指定阶段完成时发送通知
-      # 可用于在构建过程中发送进度通知
-      # 阶段命名规则：
-      #   - prebake/finalize: 第二字段为静态名称，如 "prebake.ready"
-      #   - bake: 第二字段为 bake.sh 中的函数名，如 "bake.build"
-      # 每个条件有独立的 content，会在对应时机发送
-      on_finish_of:
-        stage: "prebake.ready"
-        content: |
-          {
-            "channel_id": "{{ target.group_id }}",
-            "content": "🍯 HBW 准备完毕\n状态: {{ build.status }}\n耗时: {{ build.duration }}"
-          }
+      # TODO: 需要大量重构
 
   # 次优先级组 - 备用通知方式（当优先级 0 中任一方式失败时触发）
   1:
     # 邮件通知（内置插件）
     mail:
       to: "user@gmail.com"
-      from:
-        host: "mail.163.com"
-        mode: "smtp"
-        # 加密模式：
-        #   - "TLS": 使用 TLS 加密（端口 465）
-        #   - "StartTLS": 使用 STARTTLS（端口 587）
-        #   - "no": 不加密（不推荐）
-        secure: "TLS"
-        username: {{ variable.mail.user }}
-        password: {{ secret.mail.password }}
+      # TODO: 替换为实际mail里的字段
       schema: "default"
 
     # 插件也可以直接作为通知方式使用
     # 需要插件支持通知触发器（on_success/on_failure/on_finish_of）
     sms-notify:
       to: "+8610012345678"
-      schema: "custom"
       policy:
-      - on_success: "[HBW Notify] Build Succeeded: {{ build.status }} ({{ build.duration }})"
+      - trigger:
+          - on_success
+        template:
+          schema: "custom"
+          on_success: "[HBW Notify] Build Succeeded: {{ build.status }} ({{ build.duration }})"
         to: "+8610012345678"
-      - on_failure: "[HBW Notify] Build Failed: {{ build.status }} ({{ build.duration }})"
+      - trigger:
+          - on_success
+          - on_failure
+        template:
+          schema: "sms-tmpl1"
+          # on_success: "[HBW Notify] Build Succeeded: {{ build.status }} ({{ build.duration }})"
+          # on_failure: "[HBW Notify] Build Failed: {{ build.status }} ({{ build.duration }})"
         to: "+8610012345679"
-
-      - on_finish_of:
-          stage: "bake.build"
-          content: "[HBW Notify] Build Progress: {{ build.status }} ({{ build.duration }})"
+      - trigger:
+          - on_finish_of:
+            - "bake.build"
+            - "bake.clean.*"
+            - 're:bake\.prepare\.*'
+        template:
+          schema: "custom"
+          on_finish_of: "[HBW Notify] Build Progress: {{ build.status }} ({{ build.duration }})"
         to: "+8610012345677"
       fallible: true
 
diff --git a/tests/Dockerfile b/tests/Dockerfile
new file mode 100644
index 0000000..22357da
--- /dev/null
+++ b/tests/Dockerfile
@@ -0,0 +1,32 @@
+FROM archlinux:latest
+
+RUN sed -i 's|https://geo.mirror.pkgbuild.com|https://mirrors.tuna.tsinghua.edu.cn/archlinux|g' /etc/pacman.d/mirrorlist && \
+    sed -i 's|https://mirror.rackspace.com/archlinux|https://mirrors.tuna.tsinghua.edu.cn/archlinux|g' /etc/pacman.d/mirrorlist && \
+    echo 'Server = https://mirrors.tuna.tsinghua.edu.cn/archlinux/$repo/os/$arch' >> /etc/pacman.d/mirrorlist
+
+RUN pacman -Syu --noconfirm && \
+    pacman -S --noconfirm \
+    python python-pip base-devel git curl wget sudo \
+    which procps-ng shadow && \
+    pacman -Scc --noconfirm
+
+# Create test user
+RUN useradd -m -s /bin/bash testuser
+
+# Python environment
+WORKDIR /app
+COPY tests/requirements.txt ./requirements.txt
+RUN python -m venv /opt/test-venv && \
+    /opt/test-venv/bin/pip install -r requirements.txt
+
+# Copy test suite
+COPY tests/ ./tests/
+COPY tests/entrypoint.sh ./entrypoint.sh
+RUN chmod +x entrypoint.sh
+
+# Environment
+ENV PATH="/opt/test-venv/bin:$PATH"
+ENV PYTHONUNBUFFERED=1
+ENV PYTHONDONTWRITEBYTECODE=1
+
+ENTRYPOINT ["./entrypoint.sh"]
diff --git a/tests/conftest.py b/tests/conftest.py
new file mode 100644
index 0000000..23ebd55
--- /dev/null
+++ b/tests/conftest.py
@@ -0,0 +1,175 @@
+"""
+Root conftest for HoneyBiscuitWorkshop system tests.
+
+Provides core fixtures for:
+- Locating the baker binary and examples
+- Running subprocess commands with output capture
+- Managing temporary working directories
+"""
+import os
+import subprocess
+import shutil
+import tempfile
+import pytest
+
+
+def pytest_configure(config):
+    """Register custom markers."""
+    config.addinivalue_line("markers", "prebake: prebake stage tests")
+    config.addinivalue_line("markers", "bake: bake stage tests")
+    config.addinivalue_line("markers", "finalize: finalize stage tests")
+    config.addinivalue_line("markers", "llm: tests requiring LLM API")
+    config.addinivalue_line("markers", "slow: slow running tests")
+
+
+@pytest.fixture(scope="session")
+def baker_bin():
+    """Path to the workshop-baker binary."""
+    path = os.environ.get("BAKER_BIN", "/workshop/workshop-baker")
+    if not os.path.isfile(path):
+        pytest.skip(f"Baker binary not found at {path}")
+    if not os.access(path, os.X_OK):
+        pytest.skip(f"Baker binary not executable at {path}")
+    return path
+
+
+@pytest.fixture(scope="session")
+def examples_dir():
+    """Path to the examples directory (may be None)."""
+    path = os.environ.get("EXAMPLES_DIR", "/workshop/examples")
+    if not os.path.isdir(path):
+        return None
+    return path
+
+
+@pytest.fixture(scope="session")
+def llm_env():
+    """LLM environment variables dict (from llm.env)."""
+    env = {}
+    for var in [
+        "OPENAI_BASE_URL",
+        "OPENAI_API_KEY",
+        "ANTHROPIC_BASE_URL",
+        "ANTHROPIC_API_KEY",
+        "MODEL",
+        "REASONING_EFFORT",
+    ]:
+        val = os.environ.get(var)
+        if val:
+            env[var] = val
+    return env
+
+
+@pytest.fixture
+def work_dir():
+    """Create a temporary working directory for a test, cleaned up after."""
+    d = tempfile.mkdtemp(prefix="hbw-test-")
+    yield d
+    shutil.rmtree(d, ignore_errors=True)
+
+
+@pytest.fixture
+def run_cmd():
+    """Fixture that provides a command runner function.
+
+    Returns a function that runs a command and returns
+    (stdout, stderr, returncode, success).
+    """
+
+    def _run(args, cwd=None, env=None, timeout=120):
+        """Run a command and capture output.
+
+        Args:
+            args: Command as list of strings.
+            cwd: Working directory.
+            env: Environment variables (merged with os.environ).
+            timeout: Timeout in seconds.
+
+        Returns:
+            dict with keys: stdout, stderr, returncode, success
+        """
+        full_env = {**os.environ}
+        if env:
+            full_env.update(env)
+
+        try:
+            result = subprocess.run(
+                args,
+                cwd=cwd,
+                env=full_env,
+                capture_output=True,
+                text=True,
+                timeout=timeout,
+            )
+            return {
+                "stdout": result.stdout,
+                "stderr": result.stderr,
+                "returncode": result.returncode,
+                "success": result.returncode == 0,
+            }
+        except subprocess.TimeoutExpired:
+            return {
+                "stdout": "",
+                "stderr": f"Command timed out after {timeout}s",
+                "returncode": -1,
+                "success": False,
+            }
+        except Exception as e:
+            return {
+                "stdout": "",
+                "stderr": str(e),
+                "returncode": -1,
+                "success": False,
+            }
+
+    return _run
+
+
+@pytest.fixture
+def run_baker(baker_bin, run_cmd):
+    """Fixture that provides a baker command runner.
+
+    Returns a function that runs workshop-baker with common flags.
+    """
+
+    def _run(subcommand, *args, cwd=None, env=None, timeout=120):
+        """Run workshop-baker with a subcommand.
+
+        Args:
+            subcommand: Baker subcommand (prebake, bake, finalize).
+            *args: Additional arguments to the subcommand.
+            cwd: Working directory.
+            env: Extra environment variables.
+            timeout: Timeout in seconds.
+
+        Returns:
+            dict with keys: stdout, stderr, returncode, success
+        """
+        cmd = [
+            baker_bin,
+            "-u", "testuser",
+            "-p", "test-pipeline",
+            "-b", "test-build-001",
+            subcommand,
+        ]
+        cmd.extend(args)
+        return run_cmd(cmd, cwd=cwd, env=env, timeout=timeout)
+
+    return _run
+
+
+@pytest.fixture
+def fixtures_dir():
+    """Path to the test fixtures directory."""
+    return os.path.join(os.path.dirname(__file__), "fixtures")
+
+
+@pytest.fixture
+def bake_workspace():
+    """Create /workspace for bake tests (baker writes temp scripts there)."""
+    existed = os.path.exists("/workspace")
+    if not existed:
+        os.makedirs("/workspace", exist_ok=True)
+    yield
+    if not existed:
+        shutil.rmtree("/workspace", ignore_errors=True)
diff --git a/tests/entrypoint.sh b/tests/entrypoint.sh
new file mode 100755
index 0000000..42563b4
--- /dev/null
+++ b/tests/entrypoint.sh
@@ -0,0 +1,36 @@
+#!/bin/bash
+set -e
+
+# Load LLM environment if available
+if [ -f /workshop/llm.env ]; then
+    set -a
+    source /workshop/llm.env
+    set +a
+fi
+
+# Export test configuration
+export BAKER_BIN=/workshop/workshop-baker
+export EXAMPLES_DIR=/workshop/examples
+export TEST_MODE=${TEST_MODE:-auto}
+
+if [ "$TEST_MODE" = "manual" ]; then
+    echo "========================================"
+    echo "  HoneyBiscuitWorkshop System Tests"
+    echo "  Mode: MANUAL (interactive shell)"
+    echo "========================================"
+    echo ""
+    echo "Binary:   $BAKER_BIN"
+    echo "Examples: $EXAMPLES_DIR"
+    echo ""
+    echo "Run tests manually:"
+    echo "  cd /app && pytest tests/ -v"
+    echo ""
+    exec /bin/bash
+else
+    echo "========================================"
+    echo "  HoneyBiscuitWorkshop System Tests"
+    echo "  Mode: AUTO (running pytest)"
+    echo "========================================"
+    cd /app
+    exec python -m pytest tests/ -v --tb=short --junitxml=/results/report.xml "$@"
+fi
diff --git a/tests/fixtures/bake/bake_base.sh b/tests/fixtures/bake/bake_base.sh
new file mode 100644
index 0000000..7108dfa
--- /dev/null
+++ b/tests/fixtures/bake/bake_base.sh
@@ -0,0 +1,6 @@
+#!/bin/bash
+[ -f /dev/shm/bakeexport.{{ name }} ] && . /dev/shm/bakeexport.{{ name }} || true
+{{ condition }}
+{{ main }}
+{{ name }}
+{{ export }}
diff --git a/tests/fixtures/bake/pipeline.sh b/tests/fixtures/bake/pipeline.sh
new file mode 100644
index 0000000..8ffc060
--- /dev/null
+++ b/tests/fixtures/bake/pipeline.sh
@@ -0,0 +1,25 @@
+#!/bin/bash
+
+# @pipeline
+step_one() {
+    echo "Step 1: Setup"
+    mkdir -p /tmp/bake-test-output
+    echo "step1-done" > /tmp/bake-test-output/step1.txt
+}
+
+# @pipeline
+step_two() {
+    echo "Step 2: Build"
+    echo "step2-done" > /tmp/bake-test-output/step2.txt
+}
+
+# @pipeline
+step_three() {
+    echo "Step 3: Verify"
+    if [ -f /tmp/bake-test-output/step1.txt ] && [ -f /tmp/bake-test-output/step2.txt ]; then
+        echo "All steps completed successfully"
+    else
+        echo "ERROR: Missing step outputs"
+        exit 1
+    fi
+}
diff --git a/tests/fixtures/bake/simple.sh b/tests/fixtures/bake/simple.sh
new file mode 100644
index 0000000..3f51aa4
--- /dev/null
+++ b/tests/fixtures/bake/simple.sh
@@ -0,0 +1,5 @@
+#!/bin/bash
+echo "Hello from simple bake script"
+echo "Task ID: $HBW_TASKID"
+echo "Pipeline: $HBW_PIPELINE"
+exit 0
diff --git a/tests/fixtures/bake/with_after.sh b/tests/fixtures/bake/with_after.sh
new file mode 100644
index 0000000..b7c70af
--- /dev/null
+++ b/tests/fixtures/bake/with_after.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+# @pipeline
+init_workspace() {
+    echo "Initializing workspace"
+    mkdir -p /tmp/bake-after-test
+    echo "init" > /tmp/bake-after-test/status.txt
+}
+
+# @pipeline
+# @after(init_workspace)
+fetch_source() {
+    echo "Fetching source (depends on init_workspace)"
+    echo "fetched" > /tmp/bake-after-test/source.txt
+}
+
+# @pipeline
+# @after(fetch_source)
+build_project() {
+    echo "Building project (depends on fetch_source)"
+    echo "built" > /tmp/bake-after-test/build.txt
+}
diff --git a/tests/fixtures/finalize/hook_early.sh b/tests/fixtures/finalize/hook_early.sh
new file mode 100644
index 0000000..02d15c5
--- /dev/null
+++ b/tests/fixtures/finalize/hook_early.sh
@@ -0,0 +1,4 @@
+#!/bin/bash
+echo "EARLY_HOOK_EXECUTED"
+echo "Working directory: $(pwd)"
+echo "User: $(whoami)"
diff --git a/tests/fixtures/finalize/hook_late.sh b/tests/fixtures/finalize/hook_late.sh
new file mode 100644
index 0000000..a13290d
--- /dev/null
+++ b/tests/fixtures/finalize/hook_late.sh
@@ -0,0 +1,3 @@
+#!/bin/bash
+echo "LATE_HOOK_EXECUTED"
+echo "Working directory: $(pwd)"
diff --git a/tests/fixtures/finalize/minimal.yml b/tests/fixtures/finalize/minimal.yml
new file mode 100644
index 0000000..e35f838
--- /dev/null
+++ b/tests/fixtures/finalize/minimal.yml
@@ -0,0 +1,6 @@
+version: "0.0.1"
+plugin: {}
+notification: {}
+artifact: []
+cleanup:
+  policy: "auto"
diff --git a/tests/fixtures/finalize/with_hooks.yml b/tests/fixtures/finalize/with_hooks.yml
new file mode 100644
index 0000000..8cace4d
--- /dev/null
+++ b/tests/fixtures/finalize/with_hooks.yml
@@ -0,0 +1,13 @@
+version: "0.0.1"
+plugin: {}
+notification: {}
+artifact: []
+cleanup:
+  policy: "auto"
+hooks:
+  early:
+    - name: "finalize-early"
+      command: "echo 'finalize early hook'"
+  late:
+    - name: "finalize-late"
+      command: "echo 'finalize late hook'"
diff --git a/tests/fixtures/prebake/bootstrap_only.yml b/tests/fixtures/prebake/bootstrap_only.yml
new file mode 100644
index 0000000..34a1b43
--- /dev/null
+++ b/tests/fixtures/prebake/bootstrap_only.yml
@@ -0,0 +1,8 @@
+version: "1.0"
+environment:
+  builder: "baremetal"
+bootstrap:
+  user: "vulcan"
+  workspace:
+    path: "/home/vulcan/workspace"
+    fallback: true
diff --git a/tests/fixtures/prebake/minimal.yml b/tests/fixtures/prebake/minimal.yml
new file mode 100644
index 0000000..34a1b43
--- /dev/null
+++ b/tests/fixtures/prebake/minimal.yml
@@ -0,0 +1,8 @@
+version: "1.0"
+environment:
+  builder: "baremetal"
+bootstrap:
+  user: "vulcan"
+  workspace:
+    path: "/home/vulcan/workspace"
+    fallback: true
diff --git a/tests/fixtures/prebake/with_hooks.yml b/tests/fixtures/prebake/with_hooks.yml
new file mode 100644
index 0000000..3e105da
--- /dev/null
+++ b/tests/fixtures/prebake/with_hooks.yml
@@ -0,0 +1,15 @@
+version: "1.0"
+environment:
+  builder: "baremetal"
+bootstrap:
+  user: "vulcan"
+  workspace:
+    path: "/home/vulcan/workspace"
+    fallback: true
+hooks:
+  early:
+    - name: "test-early-hook"
+      command: "echo 'early hook executed'"
+  late:
+    - name: "test-late-hook"
+      command: "echo 'late hook executed'"
diff --git a/tests/fixtures/prebake/with_pacman.yml b/tests/fixtures/prebake/with_pacman.yml
new file mode 100644
index 0000000..36da80b
--- /dev/null
+++ b/tests/fixtures/prebake/with_pacman.yml
@@ -0,0 +1,14 @@
+version: "1.0"
+environment:
+  builder: "baremetal"
+bootstrap:
+  user: "vulcan"
+  workspace:
+    path: "/home/vulcan/workspace"
+    fallback: true
+dependencies:
+  system:
+    pacman:
+      packages:
+        - "curl"
+        - "wget"
diff --git a/tests/llm/__init__.py b/tests/llm/__init__.py
new file mode 100644
index 0000000..15f1768
--- /dev/null
+++ b/tests/llm/__init__.py
@@ -0,0 +1,12 @@
+"""
+LLM-based log judgment system for system tests.
+
+Uses LLM APIs to evaluate non-deterministic test outputs (logs, stderr, etc.)
+where simple string matching is insufficient.
+"""
+
+from .client import LLMClient
+from .cache import LLMCache
+from .judge import judge_log, JudgmentResult
+
+__all__ = ["LLMClient", "LLMCache", "judge_log", "JudgmentResult"]
diff --git a/tests/llm/cache.py b/tests/llm/cache.py
new file mode 100644
index 0000000..baf12a5
--- /dev/null
+++ b/tests/llm/cache.py
@@ -0,0 +1,56 @@
+"""
+Disk-based cache for LLM API responses.
+
+Uses SHA256 of (prompt + model + system_prompt) as cache key.
+Cache files stored as JSON in ``.cache/`` directory.
+"""
+
+import hashlib
+import json
+from pathlib import Path
+
+
+class LLMCache:
+    """SHA256-based disk cache for LLM API responses."""
+
+    def __init__(self, cache_dir: str | None = None):
+        if cache_dir is None:
+            import os
+
+            cache_dir = os.path.join(os.path.dirname(__file__), ".cache")
+        self.cache_dir = Path(cache_dir)
+        self.cache_dir.mkdir(parents=True, exist_ok=True)
+
+    def _make_key(self, prompt: str, model: str, system_prompt: str = "") -> str:
+        """Create deterministic cache key from request parameters."""
+        data = json.dumps(
+            {"prompt": prompt, "model": model, "system": system_prompt},
+            sort_keys=True,
+        )
+        return hashlib.sha256(data.encode()).hexdigest()
+
+    def get(self, prompt: str, model: str, system_prompt: str = "") -> dict | None:
+        """Retrieve cached response if exists."""
+        key = self._make_key(prompt, model, system_prompt)
+        path = self.cache_dir / f"{key}.json"
+        if path.exists():
+            try:
+                with open(path) as f:
+                    return json.load(f)
+            except (json.JSONDecodeError, IOError):
+                return None
+        return None
+
+    def set(
+        self, prompt: str, model: str, system_prompt: str, response: dict
+    ) -> None:
+        """Store response in cache."""
+        key = self._make_key(prompt, model, system_prompt)
+        path = self.cache_dir / f"{key}.json"
+        with open(path, "w") as f:
+            json.dump(response, f, ensure_ascii=False)
+
+    def clear(self) -> None:
+        """Clear all cached responses."""
+        for f in self.cache_dir.glob("*.json"):
+            f.unlink()
diff --git a/tests/llm/client.py b/tests/llm/client.py
new file mode 100644
index 0000000..72400fa
--- /dev/null
+++ b/tests/llm/client.py
@@ -0,0 +1,187 @@
+"""
+LLM client supporting both Anthropic and OpenAI APIs.
+
+Prefers Anthropic format (per project convention).
+Falls back to OpenAI format if Anthropic is unavailable.
+"""
+
+import os
+import logging
+from .cache import LLMCache
+
+logger = logging.getLogger(__name__)
+
+
+class LLMClient:
+    """LLM client with caching support.  Prefers Anthropic API."""
+
+    def __init__(self, cache: LLMCache | None = None):
+        self.cache = cache or LLMCache()
+        self._anthropic_client = None
+        self._openai_client = None
+        self._init_clients()
+
+    def _init_clients(self):
+        """Initialize API clients based on available environment variables."""
+        # Anthropic (preferred)
+        anthropic_key = os.environ.get("ANTHROPIC_API_KEY")
+        anthropic_base = os.environ.get("ANTHROPIC_BASE_URL")
+        if anthropic_key:
+            try:
+                import anthropic
+
+                kwargs: dict = {"api_key": anthropic_key}
+                if anthropic_base:
+                    kwargs["base_url"] = anthropic_base
+                self._anthropic_client = anthropic.Anthropic(**kwargs)
+                logger.info("Anthropic client initialized")
+            except ImportError:
+                logger.warning("anthropic package not installed")
+
+        # OpenAI (fallback)
+        openai_key = os.environ.get("OPENAI_API_KEY")
+        openai_base = os.environ.get("OPENAI_BASE_URL")
+        if openai_key:
+            try:
+                import openai
+
+                kwargs = {"api_key": openai_key}
+                if openai_base:
+                    kwargs["base_url"] = openai_base
+                self._openai_client = openai.OpenAI(**kwargs)
+                logger.info("OpenAI client initialized")
+            except ImportError:
+                logger.warning("openai package not installed")
+
+    @property
+    def available(self) -> bool:
+        """Check if any LLM client is available."""
+        return (
+            self._anthropic_client is not None or self._openai_client is not None
+        )
+
+    def complete(
+        self, prompt: str, system_prompt: str = "", max_tokens: int = 2000
+    ) -> str:
+        """Send a completion request, with caching.
+
+        Args:
+            prompt: User message.
+            system_prompt: System message (sets role / persona).
+            max_tokens: Max response tokens.
+
+        Returns:
+            LLM response text.
+
+        Raises:
+            RuntimeError: If no client is available.
+        """
+        model = os.environ.get("MODEL", "deepseek-v4-flash")
+
+        # Check cache
+        cached = self.cache.get(prompt, model, system_prompt)
+        if cached:
+            logger.info("Cache hit for LLM request")
+            return cached["content"]
+
+        # Try Anthropic first
+        if self._anthropic_client:
+            content = self._call_anthropic(
+                prompt, system_prompt, model, max_tokens
+            )
+        elif self._openai_client:
+            content = self._call_openai(prompt, system_prompt, model, max_tokens)
+        else:
+            raise RuntimeError(
+                "No LLM client available. "
+                "Set ANTHROPIC_API_KEY or OPENAI_API_KEY."
+            )
+
+        # Cache the response
+        self.cache.set(prompt, model, system_prompt, {"content": content})
+        return content
+
+    # ------------------------------------------------------------------
+    # Private helpers
+    # ------------------------------------------------------------------
+
+    def _call_anthropic(
+        self, prompt: str, system_prompt: str, model: str, max_tokens: int
+    ) -> str:
+        """Call Anthropic API."""
+        kwargs: dict = {
+            "model": model,
+            "max_tokens": max_tokens,
+            "messages": [{"role": "user", "content": prompt}],
+        }
+        if system_prompt:
+            kwargs["system"] = system_prompt
+
+        # Only add reasoning_effort if the model supports it
+        reasoning_effort = os.environ.get("REASONING_EFFORT")
+        if reasoning_effort:
+            try:
+                kwargs["reasoning_effort"] = reasoning_effort
+                response = self._anthropic_client.messages.create(**kwargs)
+            except Exception as e:
+                if (
+                    "reasoning_effort" in str(e).lower()
+                    or "unexpected" in str(e).lower()
+                ):
+                    del kwargs["reasoning_effort"]
+                    response = self._anthropic_client.messages.create(**kwargs)
+                else:
+                    raise
+        else:
+            response = self._anthropic_client.messages.create(**kwargs)
+
+        # Extract text from response content blocks
+        if hasattr(response, "content") and response.content:
+            block = response.content[0]
+            if hasattr(block, "text"):
+                return block.text
+            if hasattr(block, "type") and block.type == "thinking":
+                for b in response.content:
+                    if hasattr(b, "text"):
+                        return b.text
+            return str(block)
+        return str(response)
+
+    def _call_openai(
+        self, prompt: str, system_prompt: str, model: str, max_tokens: int
+    ) -> str:
+        """Call OpenAI-compatible API."""
+        messages: list[dict] = []
+        if system_prompt:
+            messages.append({"role": "system", "content": system_prompt})
+        messages.append({"role": "user", "content": prompt})
+
+        kwargs: dict = {
+            "model": model,
+            "messages": messages,
+            "max_tokens": max_tokens,
+        }
+
+        # Only add reasoning_effort if supported
+        reasoning_effort = os.environ.get("REASONING_EFFORT")
+        if reasoning_effort:
+            try:
+                kwargs["reasoning_effort"] = reasoning_effort
+                response = self._openai_client.chat.completions.create(
+                    **kwargs
+                )
+            except Exception as e:
+                if (
+                    "reasoning_effort" in str(e).lower()
+                    or "unexpected" in str(e).lower()
+                ):
+                    del kwargs["reasoning_effort"]
+                    response = self._openai_client.chat.completions.create(
+                        **kwargs
+                    )
+                else:
+                    raise
+        else:
+            response = self._openai_client.chat.completions.create(**kwargs)
+
+        return response.choices[0].message.content
diff --git a/tests/llm/judge.py b/tests/llm/judge.py
new file mode 100644
index 0000000..f9703b3
--- /dev/null
+++ b/tests/llm/judge.py
@@ -0,0 +1,149 @@
+"""
+LLM-based log judgment for system tests.
+
+Evaluates logs against expected behavior using LLM semantic analysis.
+Returns structured pass/fail results with reasoning.
+"""
+
+import json
+import re
+import logging
+from dataclasses import dataclass
+from .client import LLMClient
+
+logger = logging.getLogger(__name__)
+
+JUDGE_SYSTEM_PROMPT = """\
+You are a test result judge for a CI/CD system called HoneyBiscuitWorkshop.
+
+Your job: Evaluate whether actual command output/logs match expected behavior.
+
+RULES:
+1. You MUST respond with valid JSON in this exact format:
+   {"verdict": "pass" | "fail", "reason": "brief explanation"}
+2. "pass" = the output confirms the expected behavior occurred
+3. "fail" = the output shows the expected behavior did NOT occur, or shows an error
+4. Ignore irrelevant noise (timestamps, debug lines, unrelated warnings)
+5. Focus ONLY on whether the specific expected behavior is confirmed or contradicted
+6. Be strict: if the evidence is ambiguous, verdict is "fail"
+7. Do NOT include any text outside the JSON object"""
+
+
+@dataclass
+class JudgmentResult:
+    """Result of an LLM log judgment."""
+
+    verdict: str  # "pass" or "fail"
+    reason: str
+    raw_response: str
+
+    @property
+    def passed(self) -> bool:
+        return self.verdict == "pass"
+
+
+def _extract_relevant_logs(full_log: str, max_chars: int = 4000) -> str:
+    """Extract relevant portions from logs, removing noise.
+
+    Keeps the last *max_chars* characters (most recent / relevant).
+    Removes common noise patterns.
+    """
+    lines = full_log.strip().split("\n")
+
+    noise_patterns = [
+        r"^\d{4}-\d{2}-\d{2}T",
+        r"^DEBUG ",
+        r"^TRACE ",
+    ]
+    filtered = [
+        line
+        for line in lines
+        if not any(re.match(p, line) for p in noise_patterns)
+    ]
+
+    text = "\n".join(filtered)
+    if len(text) > max_chars:
+        text = "...\n" + text[-max_chars:]
+    return text
+
+
+def judge_log(
+    operation: str,
+    expected_behavior: str,
+    actual_log: str,
+    client: LLMClient | None = None,
+) -> JudgmentResult:
+    """Judge whether actual logs match expected behavior.
+
+    Args:
+        operation: Description of what was attempted
+            (e.g. ``"prebake bootstrap stage"``).
+        expected_behavior: What should have happened
+            (e.g. ``"user vulcan was created"``).
+        actual_log: The actual stdout / stderr output.
+        client: ``LLMClient`` instance (creates default if ``None``).
+
+    Returns:
+        ``JudgmentResult`` with verdict and reason.
+    """
+    if client is None:
+        client = LLMClient()
+
+    if not client.available:
+        logger.warning("No LLM client available, returning fail verdict")
+        return JudgmentResult(
+            verdict="fail",
+            reason="No LLM client available for judgment",
+            raw_response="",
+        )
+
+    relevant_log = _extract_relevant_logs(actual_log)
+
+    prompt = f"""\
+## Operation
+{operation}
+
+## Expected Behavior
+{expected_behavior}
+
+## Actual Output (relevant excerpts)
+```
+{relevant_log}
+```
+
+Evaluate: Does the actual output confirm the expected behavior occurred?
+Respond with JSON only."""
+
+    try:
+        response = client.complete(
+            prompt=prompt,
+            system_prompt=JUDGE_SYSTEM_PROMPT,
+            max_tokens=500,
+        )
+
+        # Parse JSON response — handle potential markdown code blocks
+        text = response.strip()
+        if text.startswith("```"):
+            lines = text.split("\n")
+            text = "\n".join(lines[1:-1]) if len(lines) > 2 else text
+
+        result = json.loads(text)
+        return JudgmentResult(
+            verdict=result.get("verdict", "fail"),
+            reason=result.get("reason", "No reason provided"),
+            raw_response=response,
+        )
+    except json.JSONDecodeError as e:
+        logger.error(f"Failed to parse LLM response as JSON: {e}")
+        return JudgmentResult(
+            verdict="fail",
+            reason=f"LLM response was not valid JSON: {response[:200]}",
+            raw_response=response,
+        )
+    except Exception as e:
+        logger.error(f"LLM judgment failed: {e}")
+        return JudgmentResult(
+            verdict="fail",
+            reason=f"LLM call failed: {str(e)}",
+            raw_response="",
+        )
diff --git a/tests/probes/__init__.py b/tests/probes/__init__.py
new file mode 100644
index 0000000..4d71a54
--- /dev/null
+++ b/tests/probes/__init__.py
@@ -0,0 +1,34 @@
+"""
+Environment probes for verifying system state in workshop-baker tests.
+
+Each probe is a standalone function that can be called from any test.
+Probes return bool or raise AssertionError with descriptive message.
+"""
+
+from .user import user_exists, user_uid, user_gid, user_in_group, user_shell
+from .file import (
+    file_exists,
+    dir_exists,
+    file_permissions,
+    file_contains,
+    file_owned_by,
+)
+from .package import package_installed, package_version
+from .process import process_running, process_count
+
+__all__ = [
+    "user_exists",
+    "user_uid",
+    "user_gid",
+    "user_in_group",
+    "user_shell",
+    "file_exists",
+    "dir_exists",
+    "file_permissions",
+    "file_contains",
+    "file_owned_by",
+    "package_installed",
+    "package_version",
+    "process_running",
+    "process_count",
+]
diff --git a/tests/probes/file.py b/tests/probes/file.py
new file mode 100644
index 0000000..7f73fb1
--- /dev/null
+++ b/tests/probes/file.py
@@ -0,0 +1,43 @@
+"""File and directory existence / permission probes."""
+
+import os
+import stat
+import grp
+import pwd
+
+
+def file_exists(path: str) -> bool:
+    """Check if *path* exists and is a regular file."""
+    return os.path.isfile(path)
+
+
+def dir_exists(path: str) -> bool:
+    """Check if *path* exists and is a directory."""
+    return os.path.isdir(path)
+
+
+def file_permissions(path: str) -> int:
+    """Return the permission mode of *path* as an octal int (e.g. ``0o755``).
+
+    Raises ``FileNotFoundError`` if *path* doesn't exist.
+    """
+    return stat.S_IMODE(os.stat(path).st_mode)
+
+
+def file_contains(path: str, text: str) -> bool:
+    """Check if a file contains the specified text."""
+    with open(path, "r", errors="replace") as f:
+        return text in f.read()
+
+
+def file_owned_by(path: str, username: str) -> bool:
+    """Check if a file is owned by the specified user."""
+    st = os.stat(path)
+    owner = pwd.getpwuid(st.st_uid).pw_name
+    return owner == username
+
+
+def file_group(path: str) -> str:
+    """Get the group owner of a file."""
+    st = os.stat(path)
+    return grp.getgrgid(st.st_gid).gr_name
diff --git a/tests/probes/package.py b/tests/probes/package.py
new file mode 100644
index 0000000..295fcfa
--- /dev/null
+++ b/tests/probes/package.py
@@ -0,0 +1,34 @@
+"""Package installation probes for Arch Linux (pacman)."""
+
+import subprocess
+
+
+def package_installed(package_name: str) -> bool:
+    """Check if a pacman package is installed."""
+    try:
+        result = subprocess.run(
+            ["pacman", "-Q", package_name],
+            capture_output=True,
+            text=True,
+        )
+        return result.returncode == 0
+    except FileNotFoundError:
+        return False
+
+
+def package_version(package_name: str) -> str | None:
+    """Get installed version of a package, or ``None`` if not installed."""
+    try:
+        result = subprocess.run(
+            ["pacman", "-Q", package_name],
+            capture_output=True,
+            text=True,
+        )
+        if result.returncode == 0:
+            # Output format: "package_name version"
+            parts = result.stdout.strip().split()
+            if len(parts) >= 2:
+                return parts[1]
+        return None
+    except FileNotFoundError:
+        return None
diff --git a/tests/probes/process.py b/tests/probes/process.py
new file mode 100644
index 0000000..ac432da
--- /dev/null
+++ b/tests/probes/process.py
@@ -0,0 +1,44 @@
+"""Process existence probes."""
+
+import os
+import subprocess
+
+
+def process_running(process_name: str) -> bool:
+    """Check if a process with the given name is running."""
+    try:
+        result = subprocess.run(
+            ["pgrep", "-x", process_name],
+            capture_output=True,
+            text=True,
+        )
+        return result.returncode == 0
+    except FileNotFoundError:
+        # Fall back to /proc scan if pgrep not available
+        try:
+            for pid_dir in os.listdir("/proc"):
+                if pid_dir.isdigit():
+                    try:
+                        with open(f"/proc/{pid_dir}/comm") as f:
+                            if f.read().strip() == process_name:
+                                return True
+                    except (IOError, OSError):
+                        continue
+        except OSError:
+            pass
+        return False
+
+
+def process_count(process_name: str) -> int:
+    """Count processes with the given name."""
+    try:
+        result = subprocess.run(
+            ["pgrep", "-c", "-x", process_name],
+            capture_output=True,
+            text=True,
+        )
+        if result.returncode == 0:
+            return int(result.stdout.strip())
+        return 0
+    except (FileNotFoundError, ValueError):
+        return 0
diff --git a/tests/probes/user.py b/tests/probes/user.py
new file mode 100644
index 0000000..94ade6d
--- /dev/null
+++ b/tests/probes/user.py
@@ -0,0 +1,37 @@
+"""User existence and attribute probes."""
+
+import pwd
+import grp
+
+
+def user_exists(username: str) -> bool:
+    """Check if a system user exists."""
+    try:
+        pwd.getpwnam(username)
+        return True
+    except KeyError:
+        return False
+
+
+def user_uid(username: str) -> int:
+    """Get the UID of a user.  Raises KeyError if user doesn't exist."""
+    return pwd.getpwnam(username).pw_uid
+
+
+def user_gid(username: str) -> int:
+    """Get the primary GID of a user.  Raises KeyError if user doesn't exist."""
+    return pwd.getpwnam(username).pw_gid
+
+
+def user_in_group(username: str, groupname: str) -> bool:
+    """Check if a user is a member of a group."""
+    try:
+        group = grp.getgrnam(groupname)
+        return username in group.gr_mem
+    except KeyError:
+        return False
+
+
+def user_shell(username: str) -> str:
+    """Get the login shell of a user.  Raises KeyError if user doesn't exist."""
+    return pwd.getpwnam(username).pw_shell
diff --git a/tests/requirements.txt b/tests/requirements.txt
new file mode 100644
index 0000000..04fca7d
--- /dev/null
+++ b/tests/requirements.txt
@@ -0,0 +1,6 @@
+pytest>=8.0.0
+pytest-timeout>=2.2.0
+pytest-xdist>=3.5.0
+anthropic>=0.40.0,<0.90.0
+openai>=1.50.0,<2.0.0
+pyyaml>=6.0
diff --git a/tests/test_bake.py b/tests/test_bake.py
new file mode 100644
index 0000000..ce95e9a
--- /dev/null
+++ b/tests/test_bake.py
@@ -0,0 +1,129 @@
+import os
+import pytest
+
+
+FIXTURES = os.path.join(os.path.dirname(__file__), "fixtures", "bake")
+
+
+@pytest.mark.bake
+class TestBakeSimpleScript:
+
+    def test_simple_script_runs(self, run_baker, bake_workspace, work_dir):
+        prebake_cfg = os.path.join(
+            os.path.dirname(__file__), "fixtures", "prebake", "minimal.yml"
+        )
+        script = os.path.join(FIXTURES, "simple.sh")
+        bake_base = os.path.join(FIXTURES, "bake_base.sh")
+
+        result = run_baker(
+            "bake", script,
+            "--bake-base", bake_base,
+            "--prebake", prebake_cfg,
+            cwd=work_dir,
+            timeout=60,
+        )
+        assert result["success"], f"Bake failed: {result['stderr']}"
+        assert "Hello from simple bake script" in result["stdout"]
+
+    def test_simple_script_sets_env_vars(self, run_baker, bake_workspace, work_dir):
+        prebake_cfg = os.path.join(
+            os.path.dirname(__file__), "fixtures", "prebake", "minimal.yml"
+        )
+        script = os.path.join(FIXTURES, "simple.sh")
+        bake_base = os.path.join(FIXTURES, "bake_base.sh")
+
+        result = run_baker(
+            "bake", script,
+            "--bake-base", bake_base,
+            "--prebake", prebake_cfg,
+            cwd=work_dir,
+            timeout=60,
+        )
+        assert result["success"]
+        assert "test-pipeline" in result["stdout"]
+
+
+@pytest.mark.bake
+class TestBakePipeline:
+
+    def test_pipeline_functions_execute(self, run_baker, bake_workspace, work_dir):
+        prebake_cfg = os.path.join(
+            os.path.dirname(__file__), "fixtures", "prebake", "minimal.yml"
+        )
+        script = os.path.join(FIXTURES, "pipeline.sh")
+        bake_base = os.path.join(FIXTURES, "bake_base.sh")
+
+        result = run_baker(
+            "bake", script,
+            "--bake-base", bake_base,
+            "--prebake", prebake_cfg,
+            cwd=work_dir,
+            timeout=60,
+        )
+        assert result["success"], f"Pipeline bake failed: {result['stderr']}"
+        assert os.path.exists("/tmp/bake-test-output/step1.txt")
+        assert os.path.exists("/tmp/bake-test-output/step2.txt")
+
+    def test_pipeline_step_outputs_correct(self, run_baker, bake_workspace, work_dir):
+        prebake_cfg = os.path.join(
+            os.path.dirname(__file__), "fixtures", "prebake", "minimal.yml"
+        )
+        script = os.path.join(FIXTURES, "pipeline.sh")
+        bake_base = os.path.join(FIXTURES, "bake_base.sh")
+
+        result = run_baker(
+            "bake", script,
+            "--bake-base", bake_base,
+            "--prebake", prebake_cfg,
+            cwd=work_dir,
+            timeout=60,
+        )
+        assert result["success"]
+        with open("/tmp/bake-test-output/step1.txt") as f:
+            assert f.read().strip() == "step1-done"
+        with open("/tmp/bake-test-output/step2.txt") as f:
+            assert f.read().strip() == "step2-done"
+
+
+@pytest.mark.bake
+class TestBakeAfterDependencies:
+
+    def test_after_ordering(self, run_baker, bake_workspace, work_dir):
+        prebake_cfg = os.path.join(
+            os.path.dirname(__file__), "fixtures", "prebake", "minimal.yml"
+        )
+        script = os.path.join(FIXTURES, "with_after.sh")
+        bake_base = os.path.join(FIXTURES, "bake_base.sh")
+
+        result = run_baker(
+            "bake", script,
+            "--bake-base", bake_base,
+            "--prebake", prebake_cfg,
+            cwd=work_dir,
+            timeout=60,
+        )
+        assert result["success"], f"@after bake failed: {result['stderr']}"
+        assert os.path.exists("/tmp/bake-after-test/status.txt")
+        assert os.path.exists("/tmp/bake-after-test/source.txt")
+        assert os.path.exists("/tmp/bake-after-test/build.txt")
+
+
+@pytest.mark.bake
+class TestBakeStatusFiles:
+
+    def test_status_written_after_bake(self, run_baker, bake_workspace, work_dir):
+        prebake_cfg = os.path.join(
+            os.path.dirname(__file__), "fixtures", "prebake", "minimal.yml"
+        )
+        script = os.path.join(FIXTURES, "simple.sh")
+        bake_base = os.path.join(FIXTURES, "bake_base.sh")
+
+        result = run_baker(
+            "bake", script,
+            "--bake-base", bake_base,
+            "--prebake", prebake_cfg,
+            cwd=work_dir,
+            timeout=60,
+        )
+        assert result["success"]
+        assert os.path.exists("/tmp/.hbwstatus")
diff --git a/tests/test_finalize.py b/tests/test_finalize.py
new file mode 100644
index 0000000..572b8a5
--- /dev/null
+++ b/tests/test_finalize.py
@@ -0,0 +1,77 @@
+import os
+import pytest
+
+from probes import file
+from llm import judge_log, LLMClient
+
+
+FIXTURES = os.path.join(os.path.dirname(__file__), "fixtures", "finalize")
+
+
+@pytest.mark.finalize
+class TestFinalizeMinimal:
+
+    def test_minimal_finalize_succeeds(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "minimal.yml")
+        result = run_baker("finalize", config, cwd=work_dir, timeout=60)
+        assert result["success"], f"Minimal finalize failed: {result['stderr']}"
+
+
+@pytest.mark.finalize
+class TestFinalizeHooks:
+
+    def test_early_hook_executes(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "with_hooks.yml")
+        result = run_baker("finalize", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert "finalize early hook" in result["stdout"] or "finalize early hook" in result["stderr"]
+
+    def test_late_hook_executes(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "with_hooks.yml")
+        result = run_baker("finalize", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert "finalize late hook" in result["stdout"] or "finalize late hook" in result["stderr"]
+
+
+@pytest.mark.finalize
+class TestFinalizeStatus:
+
+    def test_finalize_writes_status(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "minimal.yml")
+        result = run_baker("finalize", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert os.path.exists("/tmp/.hbwstatus")
+
+
+@pytest.mark.finalize
+@pytest.mark.llm
+class TestFinalizeWithLLM:
+
+    def test_finalize_hooks_via_llm(self, run_baker, llm_env, work_dir):
+        if not llm_env:
+            pytest.skip("LLM environment not configured")
+
+        config = os.path.join(FIXTURES, "with_hooks.yml")
+        result = run_baker("finalize", config, cwd=work_dir, timeout=60)
+
+        combined_log = result["stdout"] + "\n" + result["stderr"]
+
+        judgment = judge_log(
+            operation="finalize with early and late hooks",
+            expected_behavior="Both early and late hooks executed successfully, "
+                              "printing 'finalize early hook' and 'finalize late hook'",
+            actual_log=combined_log,
+        )
+        assert judgment.passed, f"LLM judgment failed: {judgment.reason}"
+
+
+@pytest.mark.finalize
+class TestFinalizePlugins:
+
+    @pytest.mark.xfail(reason="Plugin download requires network/git access", strict=False)
+    def test_plugin_download(self, run_baker, work_dir):
+        pass
+
+    @pytest.mark.xfail(reason="Artifact stage not yet implemented", strict=False)
+    def test_artifact_collection(self, run_baker, work_dir):
+        pass
diff --git a/tests/test_prebake.py b/tests/test_prebake.py
new file mode 100644
index 0000000..9e11292
--- /dev/null
+++ b/tests/test_prebake.py
@@ -0,0 +1,94 @@
+import os
+import pytest
+
+from probes import user, file, package
+
+
+FIXTURES = os.path.join(os.path.dirname(__file__), "fixtures", "prebake")
+
+
+@pytest.mark.prebake
+class TestPrebakeBootstrap:
+
+    def test_bootstrap_creates_vulcan_user(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "bootstrap_only.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"], f"prebake failed: {result['stderr']}"
+        assert user.user_exists("vulcan")
+
+    def test_bootstrap_creates_workspace(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "bootstrap_only.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert file.dir_exists("/home/vulcan/workspace") or file.dir_exists("/workspace")
+
+    def test_bootstrap_creates_symlink(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "bootstrap_only.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert os.path.islink("/workspace")
+
+    @pytest.mark.xfail(reason="Sudoers not created without package manager config", strict=False)
+    def test_bootstrap_creates_sudoers(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "bootstrap_only.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert file.file_exists("/etc/sudoers.d/workshop")
+
+    def test_bootstrap_generates_bootstrap_script(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "bootstrap_only.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert file.file_exists("/bootstrap.sh")
+
+    def test_bootstrap_script_permissions(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "bootstrap_only.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        mode = file.file_permissions("/bootstrap.sh")
+        assert mode & 0o111, f"/bootstrap.sh not executable, mode: {oct(mode)}"
+
+
+@pytest.mark.prebake
+class TestPrebakeDeps:
+
+    @pytest.mark.xfail(reason="Package installation may fail in test environment")
+    def test_pacman_packages_installed(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "with_pacman.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=120)
+        assert result["success"], f"prebake failed: {result['stderr']}"
+        assert package.package_installed("curl")
+        assert package.package_installed("wget")
+
+
+@pytest.mark.prebake
+class TestPrebakeHooks:
+
+    @pytest.mark.xfail(reason="Late hook fails after privilege drop to vulcan (Permission denied)", strict=False)
+    def test_early_hook_executes(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "with_hooks.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"], f"prebake with hooks failed: stdout={result['stdout']!r} stderr={result['stderr']!r}"
+        assert "early hook executed" in result["stdout"] or "early hook executed" in result["stderr"]
+
+    @pytest.mark.xfail(reason="Late hook fails after privilege drop to vulcan (Permission denied)", strict=False)
+    def test_late_hook_executes(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "with_hooks.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"], f"prebake with hooks failed: stdout={result['stdout']!r} stderr={result['stderr']!r}"
+        assert "late hook executed" in result["stdout"] or "late hook executed" in result["stderr"]
+
+
+@pytest.mark.prebake
+class TestPrebakeStatus:
+
+    def test_status_files_written(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "bootstrap_only.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"]
+        assert os.path.exists("/tmp/.hbwstatus")
+
+    def test_minimal_config_succeeds(self, run_baker, work_dir):
+        config = os.path.join(FIXTURES, "minimal.yml")
+        result = run_baker("prebake", config, cwd=work_dir, timeout=60)
+        assert result["success"], f"Minimal prebake failed: {result['stderr']}"