chore: bump bundled CLI version to 2.0.74

Replace heredoc with echo statements to fix YAML parsing issue. The
unindented heredoc content was breaking out of the literal block scalar,
causing `---` to be interpreted as a YAML document separator.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude <noreply@anthropic.com>

.claude/agents/test-agent.md

@@ -0,0 +1,9 @@
+---
+name: test-agent
+description: A simple test agent for SDK testing
+tools: Read
+---
+
+# Test Agent
+
+You are a simple test agent. When asked a question, provide a brief, helpful answer.

.dockerignore

@@ -0,0 +1,49 @@
+# Git
+.git
+.gitignore
+
+# Python
+__pycache__
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# Virtual environments
+.env
+.venv
+env/
+venv/
+ENV/
+
+# IDE
+.idea/
+.vscode/
+*.swp
+*.swo
+
+# Testing/Coverage
+.coverage
+.pytest_cache/
+htmlcov/
+.tox/
+.nox/
+
+# Misc
+*.log
+.DS_Store

.github/workflows/create-release-tag.yml

@@ -24,12 +24,6 @@ jobs:
         VERSION="${BRANCH_NAME#release/v}"
         echo "version=$VERSION" >> $GITHUB_OUTPUT
     
-    - name: Get previous release tag
-      id: previous_tag
-      run: |
-        PREVIOUS_TAG=$(git describe --tags --abbrev=0 HEAD~1 2>/dev/null || echo "")
-        echo "previous_tag=$PREVIOUS_TAG" >> $GITHUB_OUTPUT
-
     - name: Create and push tag
       run: |
         git config --local user.email "github-actions[bot]@users.noreply.github.com"
@@ -46,14 +40,34 @@ jobs:
       env:
         GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       run: |
-        # Create release with auto-generated notes
-        gh release create "v${{ steps.extract_version.outputs.version }}" \
-          --title "Release v${{ steps.extract_version.outputs.version }}" \
-          --generate-notes \
-          --notes-start-tag "${{ steps.previous_tag.outputs.previous_tag }}" \
-          --notes "Published to PyPI: https://pypi.org/project/claude-agent-sdk/${{ steps.extract_version.outputs.version }}/
+        VERSION="${{ steps.extract_version.outputs.version }}"
 
-        ### Installation
-        \`\`\`bash
-        pip install claude-agent-sdk==${{ steps.extract_version.outputs.version }}
-        \`\`\`"
+        # Extract changelog section for this version to a temp file
+        awk -v ver="$VERSION" '
+          /^## / {
+            if (found) exit
+            if ($2 == ver) found=1
+            next
+          }
+          found { print }
+        ' CHANGELOG.md > release_notes.md
+
+        # Append install instructions
+        {
+          echo ""
+          echo "---"
+          echo ""
+          echo "**PyPI:** https://pypi.org/project/claude-agent-sdk/VERSION/"
+          echo ""
+          echo '```bash'
+          echo "pip install claude-agent-sdk==VERSION"
+          echo '```'
+        } >> release_notes.md
+
+        # Replace VERSION placeholder
+        sed -i "s/VERSION/$VERSION/g" release_notes.md
+
+        # Create release with notes from file
+        gh release create "v$VERSION" \
+          --title "v$VERSION" \
+          --notes-file release_notes.md

.github/workflows/publish.yml

@@ -68,8 +68,6 @@ jobs:
 
     steps:
     - uses: actions/checkout@v4
-      with:
-        fetch-depth: 0 # Fetch all history including tags (necessary for changelog generation)
 
     - name: Set up Python
       uses: actions/setup-python@v5
@@ -109,6 +107,7 @@ jobs:
     - uses: actions/checkout@v4
       with:
         token: ${{ secrets.GITHUB_TOKEN }}
+        fetch-depth: 0  # Fetch all history including tags for changelog generation
 
     - name: Set up Python
       uses: actions/setup-python@v5
@@ -189,6 +188,7 @@ jobs:
         anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
         github_token: ${{ secrets.GITHUB_TOKEN }}
         claude_args: |
+          --model claude-opus-4-5
           --allowedTools 'Bash(git add:*),Bash(git commit:*),Edit'
 
     - name: Push branch and create PR

.github/workflows/test.yml

@@ -81,6 +81,24 @@ jobs:
         run: |
           python -m pytest e2e-tests/ -v -m e2e
 
+  test-e2e-docker:
+    runs-on: ubuntu-latest
+    needs: test # Run after unit tests pass
+    # Run e2e tests in Docker to catch container-specific issues like #406
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Build Docker test image
+        run: docker build -f Dockerfile.test -t claude-sdk-test .
+
+      - name: Run e2e tests in Docker
+        env:
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+        run: |
+          docker run --rm -e ANTHROPIC_API_KEY \
+            claude-sdk-test python -m pytest e2e-tests/ -v -m e2e
+
   test-examples:
     runs-on: ubuntu-latest
     needs: test-e2e # Run after e2e tests

CHANGELOG.md

@@ -1,5 +1,70 @@
 # Changelog
 
+## 0.1.18
+
+### Internal/Other Changes
+
+- **Docker-based test infrastructure**: Added Docker support for running e2e tests in containerized environments, helping catch Docker-specific issues (#424)
+- Updated bundled Claude CLI to version 2.0.72
+
+## 0.1.17
+
+### New Features
+
+- **UserMessage UUID field**: Added `uuid` field to `UserMessage` response type, making it easier to use the `rewind_files()` method by providing direct access to message identifiers needed for file checkpointing (#418)
+
+### Internal/Other Changes
+
+- Updated bundled Claude CLI to version 2.0.70
+
+## 0.1.16
+
+### Bug Fixes
+
+- **Rate limit detection**: Fixed parsing of the `error` field in `AssistantMessage`, enabling applications to detect and handle API errors like rate limits. Previously, the `error` field was defined but never populated from CLI responses (#405)
+
+### Internal/Other Changes
+
+- Updated bundled Claude CLI to version 2.0.68
+
+## 0.1.15
+
+### New Features
+
+- **File checkpointing and rewind**: Added `enable_file_checkpointing` option to `ClaudeAgentOptions` and `rewind_files(user_message_id)` method to `ClaudeSDKClient` and `Query`. This enables reverting file changes made during a session back to a specific checkpoint, useful for exploring different approaches or recovering from unwanted modifications (#395)
+
+### Documentation
+
+- Added license and terms section to README (#399)
+
+## 0.1.14
+
+### Internal/Other Changes
+
+- Updated bundled Claude CLI to version 2.0.62
+
+## 0.1.13
+
+### Bug Fixes
+
+- **Faster error handling**: CLI errors (e.g., invalid session ID) now propagate to pending requests immediately instead of waiting for the 60-second timeout (#388)
+- **Pydantic 2.12+ compatibility**: Fixed `PydanticUserError` caused by `McpServer` type only being imported under `TYPE_CHECKING` (#385)
+- **Concurrent subagent writes**: Added write lock to prevent `BusyResourceError` when multiple subagents invoke MCP tools in parallel (#391)
+
+### Internal/Other Changes
+
+- Updated bundled Claude CLI to version 2.0.59
+
+## 0.1.12
+
+### New Features
+
+- **Tools option**: Added `tools` option to `ClaudeAgentOptions` for controlling the base set of available tools, matching the TypeScript SDK functionality. Supports three modes:
+  - Array of tool names to specify which tools should be available (e.g., `["Read", "Edit", "Bash"]`)
+  - Empty array `[]` to disable all built-in tools
+  - Preset object `{"type": "preset", "preset": "claude_code"}` to use the default Claude Code toolset
+- **SDK beta support**: Added `betas` option to `ClaudeAgentOptions` for enabling Anthropic API beta features. Currently supports `"context-1m-2025-08-07"` for extended context window
+
 ## 0.1.11
 
 ### Internal/Other Changes

Dockerfile.test

@@ -0,0 +1,29 @@
+# Dockerfile for running SDK tests in a containerized environment
+# This helps catch Docker-specific issues like #406
+
+FROM python:3.12-slim
+
+# Install dependencies for Claude CLI and git (needed for some tests)
+RUN apt-get update && apt-get install -y \
+    curl \
+    git \
+    && rm -rf /var/lib/apt/lists/*
+
+# Install Claude Code CLI
+RUN curl -fsSL https://claude.ai/install.sh | bash
+ENV PATH="/root/.local/bin:$PATH"
+
+# Set up working directory
+WORKDIR /app
+
+# Copy the SDK source
+COPY . .
+
+# Install SDK with dev dependencies
+RUN pip install -e ".[dev]"
+
+# Verify CLI installation
+RUN claude -v
+
+# Default: run unit tests
+CMD ["python", "-m", "pytest", "tests/", "-v"]

README.md

@@ -351,6 +351,6 @@ The package is published to PyPI via the GitHub Actions workflow in `.github/wor
 
 The workflow tracks both the package version and the bundled CLI version separately, allowing you to release a new package version with an updated CLI without code changes.
 
-## License
+## License and terms
 
-MIT
+Use of this SDK is governed by Anthropic's [Commercial Terms of Service](https://www.anthropic.com/legal/commercial-terms), including when you use it to power products and services that you make available to your own customers and end users, except to the extent a specific component or dependency is covered by a different license as indicated in that component's LICENSE file.

e2e-tests/test_agents_and_settings.py

@@ -38,15 +38,88 @@ async def test_agent_definition():
         async for message in client.receive_response():
             if isinstance(message, SystemMessage) and message.subtype == "init":
                 agents = message.data.get("agents", [])
-                assert isinstance(
-                    agents, list
-                ), f"agents should be a list of strings, got: {type(agents)}"
-                assert (
-                    "test-agent" in agents
-                ), f"test-agent should be available, got: {agents}"
+                assert isinstance(agents, list), (
+                    f"agents should be a list of strings, got: {type(agents)}"
+                )
+                assert "test-agent" in agents, (
+                    f"test-agent should be available, got: {agents}"
+                )
                 break
 
 
+@pytest.mark.e2e
+@pytest.mark.asyncio
+async def test_filesystem_agent_loading():
+    """Test that filesystem-based agents load via setting_sources and produce full response.
+
+    This is the core test for issue #406. It verifies that when using
+    setting_sources=["project"] with a .claude/agents/ directory containing
+    agent definitions, the SDK:
+    1. Loads the agents (they appear in init message)
+    2. Produces a full response with AssistantMessage
+    3. Completes with a ResultMessage
+
+    The bug in #406 causes the iterator to complete after only the
+    init SystemMessage, never yielding AssistantMessage or ResultMessage.
+    """
+    with tempfile.TemporaryDirectory() as tmpdir:
+        # Create a temporary project with a filesystem agent
+        project_dir = Path(tmpdir)
+        agents_dir = project_dir / ".claude" / "agents"
+        agents_dir.mkdir(parents=True)
+
+        # Create a test agent file
+        agent_file = agents_dir / "fs-test-agent.md"
+        agent_file.write_text(
+            """---
+name: fs-test-agent
+description: A filesystem test agent for SDK testing
+tools: Read
+---
+
+# Filesystem Test Agent
+
+You are a simple test agent. When asked a question, provide a brief, helpful answer.
+"""
+        )
+
+        options = ClaudeAgentOptions(
+            setting_sources=["project"],
+            cwd=project_dir,
+            max_turns=1,
+        )
+
+        messages = []
+        async with ClaudeSDKClient(options=options) as client:
+            await client.query("Say hello in exactly 3 words")
+            async for msg in client.receive_response():
+                messages.append(msg)
+
+        # Must have at least init, assistant, result
+        message_types = [type(m).__name__ for m in messages]
+
+        assert "SystemMessage" in message_types, "Missing SystemMessage (init)"
+        assert "AssistantMessage" in message_types, (
+            f"Missing AssistantMessage - got only: {message_types}. "
+            "This may indicate issue #406 (silent failure with filesystem agents)."
+        )
+        assert "ResultMessage" in message_types, "Missing ResultMessage"
+
+        # Find the init message and check for the filesystem agent
+        for msg in messages:
+            if isinstance(msg, SystemMessage) and msg.subtype == "init":
+                agents = msg.data.get("agents", [])
+                # Agents are returned as strings (just names)
+                assert "fs-test-agent" in agents, (
+                    f"fs-test-agent not loaded from filesystem. Found: {agents}"
+                )
+                break
+
+        # On Windows, wait for file handles to be released before cleanup
+        if sys.platform == "win32":
+            await asyncio.sleep(0.5)
+
+
 @pytest.mark.e2e
 @pytest.mark.asyncio
 async def test_setting_sources_default():
@@ -74,12 +147,12 @@ async def test_setting_sources_default():
             async for message in client.receive_response():
                 if isinstance(message, SystemMessage) and message.subtype == "init":
                     output_style = message.data.get("output_style")
-                    assert (
-                        output_style != "local-test-style"
-                    ), f"outputStyle should NOT be from local settings (default is no settings), got: {output_style}"
-                    assert (
-                        output_style == "default"
-                    ), f"outputStyle should be 'default', got: {output_style}"
+                    assert output_style != "local-test-style", (
+                        f"outputStyle should NOT be from local settings (default is no settings), got: {output_style}"
+                    )
+                    assert output_style == "default", (
+                        f"outputStyle should be 'default', got: {output_style}"
+                    )
                     break
 
         # On Windows, wait for file handles to be released before cleanup
@@ -121,9 +194,9 @@ This is a test command.
             async for message in client.receive_response():
                 if isinstance(message, SystemMessage) and message.subtype == "init":
                     commands = message.data.get("slash_commands", [])
-                    assert (
-                        "testcmd" not in commands
-                    ), f"testcmd should NOT be available with user-only sources, got: {commands}"
+                    assert "testcmd" not in commands, (
+                        f"testcmd should NOT be available with user-only sources, got: {commands}"
+                    )
                     break
 
         # On Windows, wait for file handles to be released before cleanup
@@ -159,11 +232,11 @@ async def test_setting_sources_project_included():
             async for message in client.receive_response():
                 if isinstance(message, SystemMessage) and message.subtype == "init":
                     output_style = message.data.get("output_style")
-                    assert (
-                        output_style == "local-test-style"
-                    ), f"outputStyle should be from local settings, got: {output_style}"
+                    assert output_style == "local-test-style", (
+                        f"outputStyle should be from local settings, got: {output_style}"
+                    )
                     break
 
         # On Windows, wait for file handles to be released before cleanup
         if sys.platform == "win32":
-            await asyncio.sleep(0.5)
+            await asyncio.sleep(0.5)

examples/filesystem_agents.py

@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""Example of loading filesystem-based agents via setting_sources.
+
+This example demonstrates how to load agents defined in .claude/agents/ files
+using the setting_sources option. This is different from inline AgentDefinition
+objects - these agents are loaded from markdown files on disk.
+
+This example tests the scenario from issue #406 where filesystem-based agents
+loaded via setting_sources=["project"] may silently fail in certain environments.
+
+Usage:
+./examples/filesystem_agents.py
+"""
+
+import asyncio
+from pathlib import Path
+
+from claude_agent_sdk import (
+    AssistantMessage,
+    ClaudeAgentOptions,
+    ClaudeSDKClient,
+    ResultMessage,
+    SystemMessage,
+    TextBlock,
+)
+
+
+def extract_agents(msg: SystemMessage) -> list[str]:
+    """Extract agent names from system message init data."""
+    if msg.subtype == "init":
+        agents = msg.data.get("agents", [])
+        # Agents can be either strings or dicts with a 'name' field
+        result = []
+        for a in agents:
+            if isinstance(a, str):
+                result.append(a)
+            elif isinstance(a, dict):
+                result.append(a.get("name", ""))
+        return result
+    return []
+
+
+async def main():
+    """Test loading filesystem-based agents."""
+    print("=== Filesystem Agents Example ===")
+    print("Testing: setting_sources=['project'] with .claude/agents/test-agent.md")
+    print()
+
+    # Use the SDK repo directory which has .claude/agents/test-agent.md
+    sdk_dir = Path(__file__).parent.parent
+
+    options = ClaudeAgentOptions(
+        setting_sources=["project"],
+        cwd=sdk_dir,
+    )
+
+    message_types: list[str] = []
+    agents_found: list[str] = []
+
+    async with ClaudeSDKClient(options=options) as client:
+        await client.query("Say hello in exactly 3 words")
+
+        async for msg in client.receive_response():
+            message_types.append(type(msg).__name__)
+
+            if isinstance(msg, SystemMessage) and msg.subtype == "init":
+                agents_found = extract_agents(msg)
+                print(f"Init message received. Agents loaded: {agents_found}")
+
+            elif isinstance(msg, AssistantMessage):
+                for block in msg.content:
+                    if isinstance(block, TextBlock):
+                        print(f"Assistant: {block.text}")
+
+            elif isinstance(msg, ResultMessage):
+                print(
+                    f"Result: subtype={msg.subtype}, cost=${msg.total_cost_usd or 0:.4f}"
+                )
+
+    print()
+    print("=== Summary ===")
+    print(f"Message types received: {message_types}")
+    print(f"Total messages: {len(message_types)}")
+
+    # Validate the results
+    has_init = "SystemMessage" in message_types
+    has_assistant = "AssistantMessage" in message_types
+    has_result = "ResultMessage" in message_types
+    has_test_agent = "test-agent" in agents_found
+
+    print()
+    if has_init and has_assistant and has_result:
+        print("SUCCESS: Received full response (init, assistant, result)")
+    else:
+        print("FAILURE: Did not receive full response")
+        print(f"  - Init: {has_init}")
+        print(f"  - Assistant: {has_assistant}")
+        print(f"  - Result: {has_result}")
+
+    if has_test_agent:
+        print("SUCCESS: test-agent was loaded from filesystem")
+    else:
+        print("WARNING: test-agent was NOT loaded (may not exist in .claude/agents/)")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

examples/tools_option.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+"""Example demonstrating the tools option and verifying tools in system message."""
+
+import anyio
+
+from claude_agent_sdk import (
+    AssistantMessage,
+    ClaudeAgentOptions,
+    ResultMessage,
+    SystemMessage,
+    TextBlock,
+    query,
+)
+
+
+async def tools_array_example():
+    """Example with tools as array of specific tool names."""
+    print("=== Tools Array Example ===")
+    print("Setting tools=['Read', 'Glob', 'Grep']")
+    print()
+
+    options = ClaudeAgentOptions(
+        tools=["Read", "Glob", "Grep"],
+        max_turns=1,
+    )
+
+    async for message in query(
+        prompt="What tools do you have available? Just list them briefly.",
+        options=options,
+    ):
+        if isinstance(message, SystemMessage) and message.subtype == "init":
+            tools = message.data.get("tools", [])
+            print(f"Tools from system message: {tools}")
+            print()
+        elif isinstance(message, AssistantMessage):
+            for block in message.content:
+                if isinstance(block, TextBlock):
+                    print(f"Claude: {block.text}")
+        elif isinstance(message, ResultMessage):
+            if message.total_cost_usd:
+                print(f"\nCost: ${message.total_cost_usd:.4f}")
+    print()
+
+
+async def tools_empty_array_example():
+    """Example with tools as empty array (disables all built-in tools)."""
+    print("=== Tools Empty Array Example ===")
+    print("Setting tools=[] (disables all built-in tools)")
+    print()
+
+    options = ClaudeAgentOptions(
+        tools=[],
+        max_turns=1,
+    )
+
+    async for message in query(
+        prompt="What tools do you have available? Just list them briefly.",
+        options=options,
+    ):
+        if isinstance(message, SystemMessage) and message.subtype == "init":
+            tools = message.data.get("tools", [])
+            print(f"Tools from system message: {tools}")
+            print()
+        elif isinstance(message, AssistantMessage):
+            for block in message.content:
+                if isinstance(block, TextBlock):
+                    print(f"Claude: {block.text}")
+        elif isinstance(message, ResultMessage):
+            if message.total_cost_usd:
+                print(f"\nCost: ${message.total_cost_usd:.4f}")
+    print()
+
+
+async def tools_preset_example():
+    """Example with tools preset (all default Claude Code tools)."""
+    print("=== Tools Preset Example ===")
+    print("Setting tools={'type': 'preset', 'preset': 'claude_code'}")
+    print()
+
+    options = ClaudeAgentOptions(
+        tools={"type": "preset", "preset": "claude_code"},
+        max_turns=1,
+    )
+
+    async for message in query(
+        prompt="What tools do you have available? Just list them briefly.",
+        options=options,
+    ):
+        if isinstance(message, SystemMessage) and message.subtype == "init":
+            tools = message.data.get("tools", [])
+            print(f"Tools from system message ({len(tools)} tools): {tools[:5]}...")
+            print()
+        elif isinstance(message, AssistantMessage):
+            for block in message.content:
+                if isinstance(block, TextBlock):
+                    print(f"Claude: {block.text}")
+        elif isinstance(message, ResultMessage):
+            if message.total_cost_usd:
+                print(f"\nCost: ${message.total_cost_usd:.4f}")
+    print()
+
+
+async def main():
+    """Run all examples."""
+    await tools_array_example()
+    await tools_empty_array_example()
+    await tools_preset_example()
+
+
+if __name__ == "__main__":
+    anyio.run(main)

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "claude-agent-sdk"
-version = "0.1.11"
+version = "0.1.18"
 description = "Python SDK for Claude Code"
 readme = "README.md"
 requires-python = ">=3.10"

scripts/test-docker.sh

@@ -0,0 +1,77 @@
+#!/bin/bash
+# Run SDK tests in a Docker container
+# This helps catch Docker-specific issues like #406
+#
+# Usage:
+#   ./scripts/test-docker.sh [unit|e2e|all]
+#
+# Examples:
+#   ./scripts/test-docker.sh unit                    # Run unit tests only
+#   ANTHROPIC_API_KEY=sk-... ./scripts/test-docker.sh e2e   # Run e2e tests
+#   ANTHROPIC_API_KEY=sk-... ./scripts/test-docker.sh all   # Run all tests
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_DIR="$(dirname "$SCRIPT_DIR")"
+
+cd "$PROJECT_DIR"
+
+usage() {
+    echo "Usage: $0 [unit|e2e|all]"
+    echo ""
+    echo "Commands:"
+    echo "  unit  - Run unit tests only (no API key needed)"
+    echo "  e2e   - Run e2e tests (requires ANTHROPIC_API_KEY)"
+    echo "  all   - Run both unit and e2e tests"
+    echo ""
+    echo "Examples:"
+    echo "  $0 unit"
+    echo "  ANTHROPIC_API_KEY=sk-... $0 e2e"
+    exit 1
+}
+
+echo "Building Docker test image..."
+docker build -f Dockerfile.test -t claude-sdk-test .
+
+case "${1:-unit}" in
+    unit)
+        echo ""
+        echo "Running unit tests in Docker..."
+        docker run --rm claude-sdk-test \
+            python -m pytest tests/ -v
+        ;;
+    e2e)
+        if [ -z "$ANTHROPIC_API_KEY" ]; then
+            echo "Error: ANTHROPIC_API_KEY environment variable is required for e2e tests"
+            echo ""
+            echo "Usage: ANTHROPIC_API_KEY=sk-... $0 e2e"
+            exit 1
+        fi
+        echo ""
+        echo "Running e2e tests in Docker..."
+        docker run --rm -e ANTHROPIC_API_KEY \
+            claude-sdk-test python -m pytest e2e-tests/ -v -m e2e
+        ;;
+    all)
+        echo ""
+        echo "Running unit tests in Docker..."
+        docker run --rm claude-sdk-test \
+            python -m pytest tests/ -v
+
+        echo ""
+        if [ -n "$ANTHROPIC_API_KEY" ]; then
+            echo "Running e2e tests in Docker..."
+            docker run --rm -e ANTHROPIC_API_KEY \
+                claude-sdk-test python -m pytest e2e-tests/ -v -m e2e
+        else
+            echo "Skipping e2e tests (ANTHROPIC_API_KEY not set)"
+        fi
+        ;;
+    *)
+        usage
+        ;;
+esac
+
+echo ""
+echo "Done!"

src/claude_agent_sdk/__init__.py

@@ -42,6 +42,7 @@ from .types import (
     SandboxIgnoreViolations,
     SandboxNetworkConfig,
     SandboxSettings,
+    SdkBeta,
     SdkPluginConfig,
     SettingSource,
     StopHookInput,
@@ -345,6 +346,8 @@ __all__ = [
     "SettingSource",
     # Plugin support
     "SdkPluginConfig",
+    # Beta support
+    "SdkBeta",
     # Sandbox support
     "SandboxSettings",
     "SandboxNetworkConfig",

src/claude_agent_sdk/_cli_version.py

@@ -1,3 +1,3 @@
 """Bundled Claude Code CLI version."""
 
-__cli_version__ = "2.0.57"
+__cli_version__ = "2.0.74"

src/claude_agent_sdk/_internal/message_parser.py

@@ -48,6 +48,7 @@ def parse_message(data: dict[str, Any]) -> Message:
         case "user":
             try:
                 parent_tool_use_id = data.get("parent_tool_use_id")
+                uuid = data.get("uuid")
                 if isinstance(data["message"]["content"], list):
                     user_content_blocks: list[ContentBlock] = []
                     for block in data["message"]["content"]:
@@ -74,10 +75,12 @@ def parse_message(data: dict[str, Any]) -> Message:
                                 )
                     return UserMessage(
                         content=user_content_blocks,
+                        uuid=uuid,
                         parent_tool_use_id=parent_tool_use_id,
                     )
                 return UserMessage(
                     content=data["message"]["content"],
+                    uuid=uuid,
                     parent_tool_use_id=parent_tool_use_id,
                 )
             except KeyError as e:
@@ -120,6 +123,7 @@ def parse_message(data: dict[str, Any]) -> Message:
                     content=content_blocks,
                     model=data["message"]["model"],
                     parent_tool_use_id=data.get("parent_tool_use_id"),
+                    error=data["message"].get("error"),
                 )
             except KeyError as e:
                 raise MessageParseError(

src/claude_agent_sdk/_internal/query.py

@@ -214,6 +214,11 @@ class Query:
             raise  # Re-raise to properly handle cancellation
         except Exception as e:
             logger.error(f"Fatal error in message reader: {e}")
+            # Signal all pending control requests so they fail fast instead of timing out
+            for request_id, event in list(self.pending_control_responses.items()):
+                if request_id not in self.pending_control_results:
+                    self.pending_control_results[request_id] = e
+                    event.set()
             # Put error in stream so iterators can handle it
             await self._message_send.send({"type": "error", "error": str(e)})
         finally:
@@ -534,6 +539,21 @@ class Query:
             }
         )
 
+    async def rewind_files(self, user_message_id: str) -> None:
+        """Rewind tracked files to their state at a specific user message.
+
+        Requires file checkpointing to be enabled via the `enable_file_checkpointing` option.
+
+        Args:
+            user_message_id: UUID of the user message to rewind to
+        """
+        await self._send_control_request(
+            {
+                "subtype": "rewind_files",
+                "user_message_id": user_message_id,
+            }
+        )
+
     async def stream_input(self, stream: AsyncIterable[dict[str, Any]]) -> None:
         """Stream input messages to transport.
 

src/claude_agent_sdk/_internal/transport/subprocess_cli.py

@@ -65,6 +65,7 @@ class SubprocessCLITransport(Transport):
             else _DEFAULT_MAX_BUFFER_SIZE
         )
         self._temp_files: list[str] = []  # Track temporary files for cleanup
+        self._write_lock: anyio.Lock = anyio.Lock()
 
     def _find_cli(self) -> str:
         """Find Claude Code CLI binary."""
@@ -185,6 +186,18 @@ class SubprocessCLITransport(Transport):
                     ["--append-system-prompt", self._options.system_prompt["append"]]
                 )
 
+        # Handle tools option (base set of tools)
+        if self._options.tools is not None:
+            tools = self._options.tools
+            if isinstance(tools, list):
+                if len(tools) == 0:
+                    cmd.extend(["--tools", ""])
+                else:
+                    cmd.extend(["--tools", ",".join(tools)])
+            else:
+                # Preset object - 'claude_code' preset maps to 'default'
+                cmd.extend(["--tools", "default"])
+
         if self._options.allowed_tools:
             cmd.extend(["--allowedTools", ",".join(self._options.allowed_tools)])
 
@@ -203,6 +216,9 @@ class SubprocessCLITransport(Transport):
         if self._options.fallback_model:
             cmd.extend(["--fallback-model", self._options.fallback_model])
 
+        if self._options.betas:
+            cmd.extend(["--betas", ",".join(self._options.betas)])
+
         if self._options.permission_prompt_tool_name:
             cmd.extend(
                 ["--permission-prompt-tool", self._options.permission_prompt_tool_name]
@@ -368,6 +384,10 @@ class SubprocessCLITransport(Transport):
                 "CLAUDE_AGENT_SDK_VERSION": __version__,
             }
 
+            # Enable file checkpointing if requested
+            if self._options.enable_file_checkpointing:
+                process_env["CLAUDE_CODE_ENABLE_SDK_FILE_CHECKPOINTING"] = "true"
+
             if self._cwd:
                 process_env["PWD"] = self._cwd
 
@@ -456,8 +476,6 @@ class SubprocessCLITransport(Transport):
 
     async def close(self) -> None:
         """Close the transport and clean up resources."""
-        self._ready = False
-
         # Clean up temporary files first (before early return)
         for temp_file in self._temp_files:
             with suppress(Exception):
@@ -465,6 +483,7 @@ class SubprocessCLITransport(Transport):
         self._temp_files.clear()
 
         if not self._process:
+            self._ready = False
             return
 
         # Close stderr task group if active
@@ -474,21 +493,19 @@ class SubprocessCLITransport(Transport):
                 await self._stderr_task_group.__aexit__(None, None, None)
             self._stderr_task_group = None
 
-        # Close streams
-        if self._stdin_stream:
-            with suppress(Exception):
-                await self._stdin_stream.aclose()
-            self._stdin_stream = None
+        # Close stdin stream (acquire lock to prevent race with concurrent writes)
+        async with self._write_lock:
+            self._ready = False  # Set inside lock to prevent TOCTOU with write()
+            if self._stdin_stream:
+                with suppress(Exception):
+                    await self._stdin_stream.aclose()
+                self._stdin_stream = None
 
         if self._stderr_stream:
             with suppress(Exception):
                 await self._stderr_stream.aclose()
             self._stderr_stream = None
 
-        if self._process.stdin:
-            with suppress(Exception):
-                await self._process.stdin.aclose()
-
         # Terminate and wait for process
         if self._process.returncode is None:
             with suppress(ProcessLookupError):
@@ -506,37 +523,37 @@ class SubprocessCLITransport(Transport):
 
     async def write(self, data: str) -> None:
         """Write raw data to the transport."""
-        # Check if ready (like TypeScript)
-        if not self._ready or not self._stdin_stream:
-            raise CLIConnectionError("ProcessTransport is not ready for writing")
+        async with self._write_lock:
+            # All checks inside lock to prevent TOCTOU races with close()/end_input()
+            if not self._ready or not self._stdin_stream:
+                raise CLIConnectionError("ProcessTransport is not ready for writing")
 
-        # Check if process is still alive (like TypeScript)
-        if self._process and self._process.returncode is not None:
-            raise CLIConnectionError(
-                f"Cannot write to terminated process (exit code: {self._process.returncode})"
-            )
+            if self._process and self._process.returncode is not None:
+                raise CLIConnectionError(
+                    f"Cannot write to terminated process (exit code: {self._process.returncode})"
+                )
 
-        # Check for exit errors (like TypeScript)
-        if self._exit_error:
-            raise CLIConnectionError(
-                f"Cannot write to process that exited with error: {self._exit_error}"
-            ) from self._exit_error
+            if self._exit_error:
+                raise CLIConnectionError(
+                    f"Cannot write to process that exited with error: {self._exit_error}"
+                ) from self._exit_error
 
-        try:
-            await self._stdin_stream.send(data)
-        except Exception as e:
-            self._ready = False  # Mark as not ready (like TypeScript)
-            self._exit_error = CLIConnectionError(
-                f"Failed to write to process stdin: {e}"
-            )
-            raise self._exit_error from e
+            try:
+                await self._stdin_stream.send(data)
+            except Exception as e:
+                self._ready = False
+                self._exit_error = CLIConnectionError(
+                    f"Failed to write to process stdin: {e}"
+                )
+                raise self._exit_error from e
 
     async def end_input(self) -> None:
         """End the input stream (close stdin)."""
-        if self._stdin_stream:
-            with suppress(Exception):
-                await self._stdin_stream.aclose()
-            self._stdin_stream = None
+        async with self._write_lock:
+            if self._stdin_stream:
+                with suppress(Exception):
+                    await self._stdin_stream.aclose()
+                self._stdin_stream = None
 
     def read_messages(self) -> AsyncIterator[dict[str, Any]]:
         """Read and parse messages from the transport."""

src/claude_agent_sdk/_version.py

@@ -1,3 +1,3 @@
 """Version information for claude-agent-sdk."""
 
-__version__ = "0.1.11"
+__version__ = "0.1.18"

src/claude_agent_sdk/client.py

@@ -261,6 +261,38 @@ class ClaudeSDKClient:
             raise CLIConnectionError("Not connected. Call connect() first.")
         await self._query.set_model(model)
 
+    async def rewind_files(self, user_message_id: str) -> None:
+        """Rewind tracked files to their state at a specific user message.
+
+        Requires:
+            - `enable_file_checkpointing=True` to track file changes
+            - `extra_args={"replay-user-messages": None}` to receive UserMessage
+              objects with `uuid` in the response stream
+
+        Args:
+            user_message_id: UUID of the user message to rewind to. This should be
+                the `uuid` field from a `UserMessage` received during the conversation.
+
+        Example:
+            ```python
+            options = ClaudeAgentOptions(
+                enable_file_checkpointing=True,
+                extra_args={"replay-user-messages": None},
+            )
+            async with ClaudeSDKClient(options) as client:
+                await client.query("Make some changes to my files")
+                async for msg in client.receive_response():
+                    if isinstance(msg, UserMessage) and msg.uuid:
+                        checkpoint_id = msg.uuid  # Save this for later
+
+                # Later, rewind to that point
+                await client.rewind_files(checkpoint_id)
+            ```
+        """
+        if not self._query:
+            raise CLIConnectionError("Not connected. Call connect() first.")
+        await self._query.rewind_files(user_message_id)
+
     async def get_server_info(self) -> dict[str, Any] | None:
         """Get server initialization info including available commands and output styles.
 

src/claude_agent_sdk/types.py

@@ -10,10 +10,16 @@ from typing_extensions import NotRequired
 
 if TYPE_CHECKING:
     from mcp.server import Server as McpServer
+else:
+    # Runtime placeholder for forward reference resolution in Pydantic 2.12+
+    McpServer = Any
 
 # Permission modes
 PermissionMode = Literal["default", "acceptEdits", "plan", "bypassPermissions"]
 
+# SDK Beta features - see https://docs.anthropic.com/en/api/beta-headers
+SdkBeta = Literal["context-1m-2025-08-07"]
+
 # Agent definitions
 SettingSource = Literal["user", "project", "local"]
 
@@ -26,6 +32,13 @@ class SystemPromptPreset(TypedDict):
     append: NotRequired[str]
 
 
+class ToolsPreset(TypedDict):
+    """Tools preset configuration."""
+
+    type: Literal["preset"]
+    preset: Literal["claude_code"]
+
+
 @dataclass
 class AgentDefinition:
     """Agent definition configuration."""
@@ -549,6 +562,7 @@ class UserMessage:
     """User message."""
 
     content: str | list[ContentBlock]
+    uuid: str | None = None
     parent_tool_use_id: str | None = None
 
 
@@ -603,6 +617,7 @@ Message = UserMessage | AssistantMessage | SystemMessage | ResultMessage | Strea
 class ClaudeAgentOptions:
     """Query options for Claude SDK."""
 
+    tools: list[str] | ToolsPreset | None = None
     allowed_tools: list[str] = field(default_factory=list)
     system_prompt: str | SystemPromptPreset | None = None
     mcp_servers: dict[str, McpServerConfig] | str | Path = field(default_factory=dict)
@@ -614,6 +629,8 @@ class ClaudeAgentOptions:
     disallowed_tools: list[str] = field(default_factory=list)
     model: str | None = None
     fallback_model: str | None = None
+    # Beta features - see https://docs.anthropic.com/en/api/beta-headers
+    betas: list[SdkBeta] = field(default_factory=list)
     permission_prompt_tool_name: str | None = None
     cwd: str | Path | None = None
     cli_path: str | Path | None = None
@@ -657,6 +674,10 @@ class ClaudeAgentOptions:
     # Output format for structured outputs (matches Messages API structure)
     # Example: {"type": "json_schema", "schema": {"type": "object", "properties": {...}}}
     output_format: dict[str, Any] | None = None
+    # Enable file checkpointing to track file changes during the session.
+    # When enabled, files can be rewound to their state at any user message
+    # using `ClaudeSDKClient.rewind_files()`.
+    enable_file_checkpointing: bool = False
 
 
 # SDK Control Protocol
@@ -697,6 +718,11 @@ class SDKControlMcpMessageRequest(TypedDict):
     message: Any
 
 
+class SDKControlRewindFilesRequest(TypedDict):
+    subtype: Literal["rewind_files"]
+    user_message_id: str
+
+
 class SDKControlRequest(TypedDict):
     type: Literal["control_request"]
     request_id: str
@@ -707,6 +733,7 @@ class SDKControlRequest(TypedDict):
         | SDKControlSetPermissionModeRequest
         | SDKHookCallbackRequest
         | SDKControlMcpMessageRequest
+        | SDKControlRewindFilesRequest
     )
 
 

tests/test_message_parser.py

@@ -31,6 +31,21 @@ class TestMessageParser:
         assert isinstance(message.content[0], TextBlock)
         assert message.content[0].text == "Hello"
 
+    def test_parse_user_message_with_uuid(self):
+        """Test parsing a user message with uuid field (issue #414).
+
+        The uuid field is needed for file checkpointing with rewind_files().
+        """
+        data = {
+            "type": "user",
+            "uuid": "msg-abc123-def456",
+            "message": {"content": [{"type": "text", "text": "Hello"}]},
+        }
+        message = parse_message(data)
+        assert isinstance(message, UserMessage)
+        assert message.uuid == "msg-abc123-def456"
+        assert len(message.content) == 1
+
     def test_parse_user_message_with_tool_use(self):
         """Test parsing a user message with tool_use block."""
         data = {

tests/test_transport.py

@@ -647,3 +647,182 @@ class TestSubprocessCLITransport:
         assert network["allowLocalBinding"] is True
         assert network["httpProxyPort"] == 8080
         assert network["socksProxyPort"] == 8081
+
+    def test_build_command_with_tools_array(self):
+        """Test building CLI command with tools as array of tool names."""
+        transport = SubprocessCLITransport(
+            prompt="test",
+            options=make_options(tools=["Read", "Edit", "Bash"]),
+        )
+
+        cmd = transport._build_command()
+        assert "--tools" in cmd
+        tools_idx = cmd.index("--tools")
+        assert cmd[tools_idx + 1] == "Read,Edit,Bash"
+
+    def test_build_command_with_tools_empty_array(self):
+        """Test building CLI command with tools as empty array (disables all tools)."""
+        transport = SubprocessCLITransport(
+            prompt="test",
+            options=make_options(tools=[]),
+        )
+
+        cmd = transport._build_command()
+        assert "--tools" in cmd
+        tools_idx = cmd.index("--tools")
+        assert cmd[tools_idx + 1] == ""
+
+    def test_build_command_with_tools_preset(self):
+        """Test building CLI command with tools preset."""
+        transport = SubprocessCLITransport(
+            prompt="test",
+            options=make_options(tools={"type": "preset", "preset": "claude_code"}),
+        )
+
+        cmd = transport._build_command()
+        assert "--tools" in cmd
+        tools_idx = cmd.index("--tools")
+        assert cmd[tools_idx + 1] == "default"
+
+    def test_build_command_without_tools(self):
+        """Test building CLI command without tools option (default None)."""
+        transport = SubprocessCLITransport(
+            prompt="test",
+            options=make_options(),
+        )
+
+        cmd = transport._build_command()
+        assert "--tools" not in cmd
+
+    def test_concurrent_writes_are_serialized(self):
+        """Test that concurrent write() calls are serialized by the lock.
+
+        When parallel subagents invoke MCP tools, they trigger concurrent write()
+        calls. Without the _write_lock, trio raises BusyResourceError.
+
+        Uses a real subprocess with the same stream setup as production:
+        process.stdin -> TextSendStream
+        """
+
+        async def _test():
+            import sys
+            from subprocess import PIPE
+
+            from anyio.streams.text import TextSendStream
+
+            # Create a real subprocess that consumes stdin (cross-platform)
+            process = await anyio.open_process(
+                [sys.executable, "-c", "import sys; sys.stdin.read()"],
+                stdin=PIPE,
+                stdout=PIPE,
+                stderr=PIPE,
+            )
+
+            try:
+                transport = SubprocessCLITransport(
+                    prompt="test",
+                    options=ClaudeAgentOptions(cli_path="/usr/bin/claude"),
+                )
+
+                # Same setup as production: TextSendStream wrapping process.stdin
+                transport._ready = True
+                transport._process = MagicMock(returncode=None)
+                transport._stdin_stream = TextSendStream(process.stdin)
+
+                # Spawn concurrent writes - the lock should serialize them
+                num_writes = 10
+                errors: list[Exception] = []
+
+                async def do_write(i: int):
+                    try:
+                        await transport.write(f'{{"msg": {i}}}\n')
+                    except Exception as e:
+                        errors.append(e)
+
+                async with anyio.create_task_group() as tg:
+                    for i in range(num_writes):
+                        tg.start_soon(do_write, i)
+
+                # All writes should succeed - the lock serializes them
+                assert len(errors) == 0, f"Got errors: {errors}"
+            finally:
+                process.terminate()
+                await process.wait()
+
+        anyio.run(_test, backend="trio")
+
+    def test_concurrent_writes_fail_without_lock(self):
+        """Verify that without the lock, concurrent writes cause BusyResourceError.
+
+        Uses a real subprocess with the same stream setup as production.
+        """
+
+        async def _test():
+            import sys
+            from contextlib import asynccontextmanager
+            from subprocess import PIPE
+
+            from anyio.streams.text import TextSendStream
+
+            # Create a real subprocess that consumes stdin (cross-platform)
+            process = await anyio.open_process(
+                [sys.executable, "-c", "import sys; sys.stdin.read()"],
+                stdin=PIPE,
+                stdout=PIPE,
+                stderr=PIPE,
+            )
+
+            try:
+                transport = SubprocessCLITransport(
+                    prompt="test",
+                    options=ClaudeAgentOptions(cli_path="/usr/bin/claude"),
+                )
+
+                # Same setup as production
+                transport._ready = True
+                transport._process = MagicMock(returncode=None)
+                transport._stdin_stream = TextSendStream(process.stdin)
+
+                # Replace lock with no-op to trigger the race condition
+                class NoOpLock:
+                    @asynccontextmanager
+                    async def __call__(self):
+                        yield
+
+                    async def __aenter__(self):
+                        return self
+
+                    async def __aexit__(self, *args):
+                        pass
+
+                transport._write_lock = NoOpLock()
+
+                # Spawn concurrent writes - should fail without lock
+                num_writes = 10
+                errors: list[Exception] = []
+
+                async def do_write(i: int):
+                    try:
+                        await transport.write(f'{{"msg": {i}}}\n')
+                    except Exception as e:
+                        errors.append(e)
+
+                async with anyio.create_task_group() as tg:
+                    for i in range(num_writes):
+                        tg.start_soon(do_write, i)
+
+                # Should have gotten errors due to concurrent access
+                assert len(errors) > 0, (
+                    "Expected errors from concurrent access, but got none"
+                )
+
+                # Check that at least one error mentions the concurrent access
+                error_strs = [str(e) for e in errors]
+                assert any("another task" in s for s in error_strs), (
+                    f"Expected 'another task' error, got: {error_strs}"
+                )
+            finally:
+                process.terminate()
+                await process.wait()
+
+        anyio.run(_test, backend="trio")