Implements the remaining M12 features from AGENTS.md:
**Plugin System (crates/platform/plugins)**
- Plugin manifest schema with plugin.json support
- Plugin loader for commands, agents, skills, hooks, and MCP servers
- Discovers plugins from ~/.config/owlen/plugins and .owlen/plugins
- Includes comprehensive tests (4 passing)
**Session Checkpointing (crates/core/agent)**
- Checkpoint struct capturing session state and file diffs
- CheckpointManager with snapshot, diff, save, load, and rewind capabilities
- File diff tracking with before/after content
- Checkpoint persistence to .owlen/checkpoints/
- Includes comprehensive tests (6 passing)
**REPL Commands (crates/app/cli)**
- /checkpoint - Save current session with file diffs
- /checkpoints - List all saved checkpoints
- /rewind <id> - Restore session and files from checkpoint
- Updated /help documentation
M12 milestone now fully complete:
✅ /permissions, /status, /cost (previously implemented)
✅ Checkpointing and /rewind
✅ Plugin loader with manifest schema
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Milestone M8 implementation adds MCP integration for connecting to external
tool servers and resources.
New crate: crates/integration/mcp-client
- JSON-RPC 2.0 protocol implementation
- Stdio transport for spawning MCP server processes
- Capability negotiation (initialize handshake)
- Tool operations:
* tools/list: List available tools from server
* tools/call: Invoke tools with arguments
- Resource operations:
* resources/list: List available resources
* resources/read: Read resource contents
- Async design using tokio for non-blocking I/O
MCP Client Features:
- McpClient: Main client with subprocess management
- ServerCapabilities: Capability discovery
- McpTool: Tool definitions with JSON schema
- McpResource: Resource definitions with URI/mime-type
- Automatic request ID management
- Error handling with proper JSON-RPC error codes
Permission Integration:
- Added Tool::Mcp to permission system
- Pattern matching support for mcp__server__tool format
* "filesystem__*" matches all filesystem server tools
* "filesystem__read_file" matches specific tool
- MCP requires Ask permission in Plan/AcceptEdits modes
- MCP allowed in Code mode (like Bash)
Tests added (3 new tests with mock Python servers):
1. mcp_server_capability_negotiation - Verifies initialize handshake
2. mcp_tool_invocation - Tests tool listing and calling
3. mcp_resource_reads - Tests resource listing and reading
Permission tests added (2 new tests):
1. mcp_server_pattern_matching - Verifies server-level wildcards
2. mcp_exact_tool_matching - Verifies tool-level exact matching
All 75 tests passing (up from 68).
Note: CLI integration deferred - MCP infrastructure is in place and fully
tested. Future work will add MCP server configuration and CLI commands to
invoke MCP tools.
Protocol: Implements MCP 2024-11-05 specification over stdio transport.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Milestone M6 implementation adds a comprehensive hook system that allows
users to run custom scripts at various lifecycle events.
New crate: crates/platform/hooks
- HookEvent enum with multiple event types:
* PreToolUse: fires before tool execution, can deny operations (exit code 2)
* PostToolUse: fires after tool execution
* SessionStart: fires at session start, can persist env vars
* SessionEnd, UserPromptSubmit, PreCompact (defined for future use)
- HookManager for executing hooks with timeout support
- JSON I/O: hooks receive event data via stdin, can output to stdout
- Hooks located in .owlen/hooks/{EventName}
CLI integration:
- All tool commands (Read, Write, Edit, Glob, Grep, Bash, SlashCommand)
now fire PreToolUse hooks before execution
- Hooks can deny operations by exiting with code 2
- Hooks timeout after 5 seconds by default
Tests added:
- pretooluse_can_deny_call: verifies hooks can block tool execution
- posttooluse_runs_parallel: verifies PostToolUse hooks execute
- sessionstart_persists_env: verifies SessionStart can create env files
- hook_timeout_works: verifies timeout mechanism
- hook_not_found_is_ok: verifies missing hooks don't cause errors
All 63 tests passing.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
This commit implements the complete M1 milestone (Config & Permissions) including:
- New permissions crate with Tool, Action, Mode, and PermissionManager
- Three permission modes: Plan (read-only default), AcceptEdits, Code
- Pattern matching for permission rules (exact match and prefix with *)
- Integration with config-agent for mode-based permission management
- CLI integration with --mode flag to override configured mode
- Permission checks for Read, Glob, and Grep operations
- Comprehensive test suite (10 tests in permissions, 4 in config, 4 in CLI)
Also fixes:
- Fixed failing test in tools-fs (glob pattern issue)
- Improved glob_list() root extraction to handle patterns like "/*.txt"
All 21 workspace tests passing.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
