- Created a comprehensive guide for setting up TypeScript LSP in Claude Code, detailing installation methods, configuration, and troubleshooting.
- Added multiple debugging test scripts to validate LSP communication with pyright, including direct communication tests, configuration checks, and document symbol retrieval.
- Implemented error handling and logging for better visibility during LSP interactions.
- Implement comprehensive unit tests for the LspGraphBuilder class to validate its functionality in building code association graphs.
- Tests cover various scenarios including single level graph expansion, max nodes and depth boundaries, concurrent expansion limits, document symbol caching, error handling during node expansion, and edge cases such as empty seed lists and self-referencing nodes.
- Utilize pytest and asyncio for asynchronous testing and mocking of LspBridge methods.
- Introduced a new test script for the CodexLens LSP tool to validate core functionalities including symbol search, find definition, find references, and get hover.
- Created comprehensive documentation for the MCP endpoint design, detailing the architecture, features, and integration with the CCW MCP Manager.
- Developed a detailed implementation plan for transitioning to a real LSP server, outlining phases, architecture, and acceptance criteria.
Introduce dedicated agent_message IR type to distinguish final AI responses
from generic stdout. This enables --final flag to show only agent messages,
filtering out all intermediate content (JSONL events, reasoning, tool calls).
Changes:
- Add agent_message type to CliOutputUnitType
- Update JsonLinesParser to map final responses from all tools (codex,
gemini, claude, opencode) to agent_message type
- Add final_output field to database schema with migration
- Update getCachedOutput and getConversation to return finalOutput
- Prefer finalOutput in outputAction for --final flag
Fixes issue where --final showed raw JSONL instead of filtered content.
- Introduced `simulate-cli-prompt.js` to simulate various prompt formats and display the final content passed to the CLI.
- Added `test-shell-prompt.js` to test actual shell execution of different prompt formats, demonstrating correct vs incorrect multi-line prompt handling.
- Created comprehensive tests in `cli-prompt-parsing.test.ts` to validate prompt parsing, including single-line, multi-line, special characters, and template concatenation.
- Implemented edge case handling for empty lines, long prompts, and Unicode characters.
- Skip template concatenation when using --uncommitted/--base/--commit
- Allow empty prompt for review mode with target flags
- Add hasReviewTarget check in command routing
- Update documentation with validation constraints
codex review constraint: target flags and prompt are mutually exclusive
- Replace $PROTO/$TMPL environment variable injection with systemRules/roles direct concatenation
- Append rules to END of prompt instead of prepending
- Change prompt field name from RULES to CONSTRAINTS in all prompts
- Default to universal-rigorous-style template when --rule not specified
- Update all .claude documentation, agents, commands, and skills
- Add streaming_content type support for Gemini delta messages
Breaking: Prompts now use CONSTRAINTS field instead of RULES
- #54: Add API endpoint configuration documentation to DASHBOARD_GUIDE.md
- #50: Add brainstorm context overflow protection with output size limits
Note: #72 and #53 not changed per user feedback - existing behavior is sufficient
(users can configure envFile themselves; default Python version is appropriate)
