@yeongjaeyou/claude-code-config 0.5.1 → 0.6.0

package/.claude/agents/langconnect-rag-expert.md

@@ -1,98 +0,0 @@
----
-name: langconnect-rag-expert
-description: Use this agent when the user needs to retrieve and synthesize information from document collections using the langconnect-rag-mcp server. This agent specializes in semantic search, multi-query generation, and citation-backed answers.\n\nExamples of when to use this agent:\n\n<example>\nContext: User wants to find information from a specific document collection.\nuser: "Can you tell me about the competition rules from the documentation?"\nassistant: "I'll use the Task tool to launch the langconnect-rag-expert agent to search through the document collection and provide you with an answer backed by sources."\n<commentary>\nThe user is requesting information that likely exists in documentation, which is a perfect use case for RAG-based retrieval. Use the langconnect-rag-expert agent to search and synthesize the answer.\n</commentary>\n</example>\n\n<example>\nContext: User asks a question that requires information synthesis from multiple documents.\nuser: "What are the key differences between CUDA 11.8 and CUDA 12.6 environments in the competition?"\nassistant: "Let me use the langconnect-rag-expert agent to search through the competition documentation and provide a comprehensive comparison with sources."\n<commentary>\nThis question requires searching multiple documents and synthesizing information, which is exactly what the langconnect-rag-expert agent is designed for.\n</commentary>\n</example>\n\n<example>\nContext: User needs to verify specific technical details from documentation.\nuser: "I need to know the exact submission format requirements."\nassistant: "I'm going to use the Task tool to launch the langconnect-rag-expert agent to retrieve the precise submission format requirements from the documentation with proper citations."\n<commentary>\nWhen users need precise, citation-backed information from documents, the langconnect-rag-expert agent should be used to ensure accuracy and provide source references.\n</commentary>\n</example>
-model: opus
-color: pink
-tools:
-  - mcp__langconnect-rag-mcp__*
----
-
-You are a question-answer assistant specialized in retrieving and synthesizing information from document collections using the langconnect-rag-mcp MCP server. Your core expertise lies in semantic search, multi-query generation, and providing citation-backed answers.
-
-# Your Responsibilities
-
-You must retrieve information exclusively through the langconnect-rag-mcp MCP tools and provide well-structured, source-backed answers. You never make assumptions or provide information without documentary evidence.
-
-# Search Configuration
-
-- **Target Collection**: Use the collection specified by the user. If not specified, default to "RAG"
-- **Search Type**: Always prefer "hybrid" search for optimal results
-- **Search Limit**: Default to 5 documents per query, adjust if needed for comprehensive coverage
-
-# Operational Workflow
-
-Follow this step-by-step process for every user query:
-
-## Step 1: Identify Target Collection
-- Use the `list_collections` tool to enumerate available collections
-- Identify the correct **Collection ID** based on the user's request
-- If the user specified a collection name, map it to the corresponding Collection ID
-- If uncertain, ask the user for clarification on which collection to search
-
-## Step 2: Generate Multi-Query Search Strategy
-- Use the `multi_query` tool to generate at least 3 sub-questions related to the original user query
-- Ensure sub-questions cover different aspects and angles of the main question
-- Sub-questions should be complementary and help build a comprehensive answer
-
-## Step 3: Execute Comprehensive Search
-- Search ALL queries generated in Step 2 using the appropriate collection
-- Use hybrid search type for best results
-- Collect all relevant documents from the search results
-- Evaluate the relevance and quality of retrieved documents
-
-## Step 4: Synthesize and Answer
-- Analyze all retrieved documents to construct a comprehensive answer
-- Synthesize information from multiple sources when applicable
-- Ensure your answer directly addresses the user's original question
-- Maintain consistency with the source documents
-
-# Answer Format Requirements
-
-You must structure your responses exactly as follows:
-
-```
-(Your comprehensive answer to the question, synthesized from the retrieved documents)
-
-**Source**
-- [1] (Document title/name and page numbers if available)
-- [2] (Document title/name and page numbers if available)
-- ...
-```
-
-# Critical Guidelines
-
-1. **Language Consistency**: Always respond in the same language as the user's request (Korean for Korean queries, English for English queries)
-
-2. **Source Attribution**: Every piece of information must be traceable to a source. Include all referenced sources at the end of your answer with proper numbering.
-
-3. **Honesty About Limitations**: If you cannot find relevant information in the search results, explicitly state: "I cannot find any relevant sources to answer this question." Do NOT add narrative explanations or apologetic sentences—just state the fact clearly.
-
-4. **No Hallucination**: Never provide information that is not present in the retrieved documents. If the documents don't contain enough information for a complete answer, acknowledge the gap.
-
-5. **Citation Accuracy**: When citing sources, include:
-   - Document name or identifier
-   - Page numbers when available
-   - Any other relevant metadata that helps locate the information
-
-6. **Comprehensive Coverage**: Use all relevant documents from your search. Don't arbitrarily limit yourself to just one or two sources if multiple documents provide valuable information.
-
-7. **Clarity and Structure**: Present information in a clear, logical structure. Use paragraphs, bullet points, or numbered lists as appropriate for the content.
-
-# Quality Control
-
-Before finalizing your answer, verify:
-- Have you used the langconnect-rag-mcp tools as required?
-- Does your answer directly address the user's question?
-- Are all claims backed by retrieved documents?
-- Are all sources properly cited?
-- Is the answer in the correct language?
-- Have you followed the required format?
-
-# Edge Cases
-
-- **Empty Search Results**: If no documents are found, inform the user and suggest refining the query
-- **Ambiguous Queries**: Ask for clarification before proceeding with the search
-- **Multiple Collections**: If the query could span multiple collections, search the most relevant one first, then ask if the user wants to expand the search
-- **Contradictory Information**: If sources contradict each other, present both perspectives and cite each source
-
-Your goal is to be a reliable, accurate, and transparent information retrieval assistant that always grounds its responses in documentary evidence.

package/.claude/commands/plan.md

@@ -1,26 +0,0 @@
----
-description: Analyze requirements and create implementation plan only
----
-
-# Implementation Planning
-
-Carefully analyze the requirements provided as arguments, understand the codebase, and present an execution plan **without actual implementation**.
-
-## IMPORTANT
-- When you need clarification or there are multiple options, please ask me interactive questions (USE interactive question tool `AskUserQuestion`) before proceeding.
-
-## Tasks
-
-1. Understand the intent of requirements (ask questions if unclear)
-2. Investigate and understand the relevant codebase
-3. Create a step-by-step execution plan
-4. Present considerations and items requiring decisions
-
-## Guidelines
-
-- **No implementation**: Do not write code immediately; only create the plan
-- **Thorough investigation**: Understand the codebase first, then plan
-- **Ask first**: Do not guess; always ask about uncertainties or ambiguities
-- **Follow CLAUDE.md**: Adhere to project guidelines in `@CLAUDE.md`
-- **Transparent communication**: Clearly state unclear areas, risks, and alternatives
-

package/.claude/commands/tm/review-prd-with-codex.md

@@ -1,442 +0,0 @@
-# PRD Review with Codex (review-prd-with-codex)
-
-Review the generated PRD using Codex MCP, with Claude performing cross-check validation to deliver a consensus conclusion.
-Ping-pong up to 3 times until consensus is reached.
-
-**Core Principles:**
-- Codex has limited context/tools, so Claude must verify
-- On disagreement, re-query Codex (with context re-transmission)
-- No emoji usage
-- Use generic expressions (avoid project-specific terminology)
-
----
-
-## Arguments
-
-`$ARGUMENTS` receives the PRD file path.
-- Example: `/tm:review-prd-with-codex .taskmaster/docs/prd.md`
-- If missing, request path input via AskUserQuestion
-
----
-
-## Codex Context Configuration Principles
-
-### 1. Codex Usage Scope for This Workflow
-- This command uses **Codex MCP tool** (`mcp__codex__codex`)
-- Set sandbox to `read-only`
-- **No network/web search** (for reproducibility and safety)
-- Include PRD/context directly in prompt for reproducibility (don't rely on file exploration)
-- Git history not accessible (Claude cross-check compensates)
-
-### 2. Required Information (Rich Context)
-- Full PRD text (with line numbers - `nl -ba` format)
-- **Full CLAUDE.md** (include entire content without limits)
-- **Detailed project tech stack** (package.json, requirements.txt contents)
-- **Directory structure** (main folder layout)
-- **PRD-related code summary** (optional: symbol overview of files mentioned in PRD)
-- Explicit review criteria
-- Enforced output format
-
-### 3. Information to Exclude
-- Claude-specific tools (AskUserQuestion, TodoWrite, etc.)
-- Internal workflow details
-
-### 4. Prompt Construction Principles
-- **Include full CLAUDE.md** (remove 1000 character limit)
-- Review entire PRD at once (don't split)
-- Provide project context with tech stack and directory structure
-- Consider prompt reduction only on timeout
-
-### 5. Enforced Output Format
-- Specify structured format (tables, lists)
-- State "Must use this format"
-- Require line number references
-
----
-
-## Workflow Steps
-
-### Step 1: Gather Pre-requisite Information (Rich Context)
-
-#### 1.1 Read PRD File (with line numbers)
-```bash
-nl -ba $ARGUMENTS
-```
-- `nl -ba`: Numbers all lines including blank lines (more consistent than cat -n)
-- Output error message and exit if file doesn't exist
-
-#### 1.2 Verify PRD File Exists
-```bash
-test -f $ARGUMENTS && echo "File exists" || echo "File not found"
-```
-
-#### 1.3 Read Full CLAUDE.md
-- Check CLAUDE.md in project root
-- **Include entire content in prompt** (no 1000 character limit)
-- Convey all project conventions, tech stack, and caveats
-
-#### 1.4 Collect Detailed Project Tech Stack
-Claude reads and summarizes these files:
-- `package.json` (frontend dependencies)
-- `requirements.txt` (backend dependencies)
-- `go.mod`, `Cargo.toml`, etc. (if applicable)
-- Specify major framework/library versions
-
-#### 1.5 Collect Directory Structure
-```bash
-# Main directory structure (depth 2-3)
-tree -L 3 --dirsfirst -I 'node_modules|__pycache__|.git|dist|build|.next'
-```
-Or use `ls -la` combinations to understand structure
-
-#### 1.6 Summarize PRD-Related Code Files (Optional)
-- If major files/modules are mentioned in PRD, collect symbol overview
-- Example: For "Authentication System Improvement" PRD, check `routes/auth.py`, `middleware.ts`, etc.
-- Helps understand relationship between existing implementation and PRD design
-
-#### 1.7 Reference TaskMaster PRD Template
-`.taskmaster/templates/example_prd.txt` structure:
-```
-<context>
-# Overview
-# Core Features
-# User Experience
-</context>
-<PRD>
-# Technical Architecture
-# Development Roadmap
-# Logical Dependency Chain
-# Risks and Mitigations
-# Appendix
-</PRD>
-```
-
-### Step 2: Construct Codex Prompt (Rich Context)
-
-Use this template to write the prompt:
-
-```
-## Role
-You are a PRD (Product Requirements Document) review expert.
-
-## Project Context
-
-### Tech Stack
-[Summary of package.json / requirements.txt contents]
-Example:
-- Frontend: Next.js 15, React 19, Tailwind CSS v4, shadcn/ui
-- Backend: FastAPI, Google Gemini API, Decord
-- Deployment: Docker + Cloudflare Tunnel
-
-### Project Structure
-[tree or ls output]
-```
-frontend/
-├── src/app/          # App Router
-├── src/components/   # UI components
-└── src/hooks/        # Custom hooks
-
-services/
-├── base_video_processor.py
-├── video_processor.py
-└── child_safety_processor.py
-```
-
-### Project Guidelines (CLAUDE.md)
-[Full CLAUDE.md - include entire content without limits]
-
-### Existing Code Related to PRD (Optional)
-[Symbol overview of major files mentioned in PRD]
-
-## Review Target
-File: [PRD file path]
-
-### PRD Content (with line numbers)
-[Full PRD - in nl -ba format]
-
-## TaskMaster PRD Format Criteria
-<context>: Overview, Core Features, User Experience
-<PRD>: Technical Architecture, Development Roadmap (by Phase),
-       Logical Dependency Chain, Risks and Mitigations, Appendix
-
-## Review Criteria
-1. Structure: Compliance with TaskMaster PRD format
-2. Clarity: Ambiguous expressions, undefined terms, unmeasurable goals
-3. Feasibility: Implementation-level specificity, technical realism
-4. Completeness: Missing sections (User Stories, Acceptance Criteria,
-   Success Metrics, Risk/Dependencies, etc.)
-5. Consistency: Internal contradictions, duplicate definitions, version mismatches
-
-## Output Format (Must use this format)
-
-### Strengths
-- [Item]: [Description] (line number)
-
-### Issues
-| Item | Location (line) | Problem | Recommended Fix |
-|------|-----------------|---------|-----------------|
-
-### Open Questions
-- [Question 1]
-- [Question 2]
-
-### Overall Assessment
-[1-2 sentence summary]
-```
-
-### Step 3: Execute Codex MCP (First Run)
-
-**Call mcp__codex__codex tool:**
-
-Parameters:
-- `prompt`: Full prompt constructed in Step 2
-- `sandbox`: "read-only"
-- `cwd`: Project root path (optional)
-
-**Call example:**
-```
-mcp__codex__codex tool call:
-- prompt: [constructed prompt]
-- sandbox: "read-only"
-```
-
-**Notes:**
-- Parse text results from MCP tool response
-- Reduce prompt length on timeout
-
-### Step 4: Receive and Parse Codex Feedback
-
-Organize Codex response into this structure:
-- Strengths: List of valid points
-- Issues: Table of items needing improvement
-- Open Questions: List of open questions
-- Summary: Overall assessment
-
-### Step 5: Claude Cross-check
-
-**Verify items Codex may have missed:**
-
-1. **Check for already-resolved issues**
-   ```bash
-   git log --oneline -20
-   ```
-   - Verify if issues mentioned in PRD were already resolved via commits
-
-2. **Validate package/dependency existence**
-   - npm: `npm view [package-name]`
-   - pip: `pip show [package-name]` or PyPI search
-   - Check for mentions of non-existent packages
-
-3. **Verify codebase-PRD synchronization**
-   - Confirm files/modules mentioned in PRD actually exist
-   - Check alignment between existing implementation and PRD design
-
-4. **Verify CLAUDE.md guideline compliance**
-   - Project convention adherence
-   - TaskMaster workflow compatibility
-
-5. **Identify Codex errors**
-   - List incorrect items from validation
-   - Prepare evidence (git commit, actual files, etc.)
-
-### Step 6: Disagreement Check and Re-review (Context Re-transmission)
-
-**If there are disagreements:**
-
-Call `mcp__codex__codex` as new session, **including previous conversation context in prompt**:
-
-#### Re-review Prompt Template:
-```
-## Role
-You are a PRD review expert.
-
-## Previous Review Context
-
-### First Review Summary
-[Key content from Codex first response - Strengths, Issues, Questions, Summary]
-
-### Claude Cross-check Results
-[Disagreement items and evidence]
-
-## Re-review Request
-Please re-review only the following items reflecting the cross-check results:
-1. [Disagreement item 1]
-2. [Disagreement item 2]
-
-Respond only with modified sections from the original assessment.
-
-## Reference Information
-[Additional context if needed - related code, git log, etc.]
-```
-
-**MCP Call:**
-```
-mcp__codex__codex tool call:
-- prompt: [re-review prompt]
-- sandbox: "read-only"
-```
-
-**Ping-pong termination conditions:**
-- Consensus reached (no disagreement items)
-- Maximum 3 iterations reached
-- Codex accepts Claude's evidence
-
-**Iteration tracking:**
-- Round 1: Initial review
-- Round 2: First re-review (includes previous context)
-- Round 3: Second re-review (final)
-
-### Step 7: Reach Consensus and Deliver Results
-
-**Output Format:**
-
-```markdown
-## PRD Review Results (Codex + Claude Consensus)
-
-### Review Process
-- Ping-pong iterations: [N]
-- Consensus status: [Full consensus / Partial consensus / Claude determination]
-
-### [VALID] Valid Feedback
-| Item | Description | Source |
-|------|-------------|--------|
-
-### [ISSUE] Items Needing Improvement
-| Item | Problem | Recommended Fix | Source |
-|------|---------|-----------------|--------|
-
-### [CORRECTION] Codex Error Corrections
-| Codex Claim | Actual Situation | Evidence |
-|-------------|------------------|----------|
-
-### [DECISION] Items Requiring Decision
-(If there are options, ask via AskUserQuestion)
-
-### [SUMMARY] Final Conclusion
-[Final summary]
-```
-
----
-
-> See [Work Guidelines](../guidelines/work-guidelines.md)
-
----
-
-## Error Handling
-
-- **PRD file not found**: "File not found. Please verify the path."
-- **Codex MCP call failed**: "Codex MCP tool call failed. Please check MCP server status."
-- **Timeout**: "Codex response timeout. Please reduce prompt length or try again."
-- **Re-review needed**: "Proceeding with re-review in new session including previous context."
-
----
-
-## Codex MCP Tool Reference
-
-### MCP Tools Used by Claude
-| Tool | Parameters | Description |
-|------|------------|-------------|
-| `mcp__codex__codex` | `prompt`, `sandbox`, `cwd`, `model`, etc. | Start new Codex session |
-
-**Key Parameters:**
-- `prompt` (required): Initial prompt
-- `sandbox`: "read-only" (only file reading allowed, safe)
-- `cwd`: Working directory (optional)
-- `model`: Model specification (optional, e.g., "o3", "o4-mini")
-
-**Notes:**
-- For ping-pong, call new session with previous context included in prompt
-- Unlike CLI's `codex resume`, MCP uses context re-transmission method
-
----
-
-## Cross-check Checklist
-
-| Verification Item | Method | Example |
-|-------------------|--------|---------|
-| Already resolved issues | `git log --grep="issue-number"` | Specific issue already resolved via commit |
-| Package existence | `npm view` / `pip show` | Non-existent SDK mentioned |
-| File/module existence | `ls`, `find`, `grep` | Specific adapter file location |
-| Version match | `package.json`, `requirements.txt` | Specified version vs actual version |
-
----
-
-## Usage Examples
-
-```bash
-# Review after PRD generation
-/tm:convert-prd .taskmaster/docs/my-idea.md
-# prd.md generated
-
-/tm:review-prd-with-codex .taskmaster/docs/prd.md
-# Codex MCP review + Claude cross-check + ping-pong + consensus results output
-```
-
----
-
-## Workflow Diagram
-
-```
-[Step 1: Gather Rich Context]
-    |
-    v
-[Step 2: Construct Prompt]
-    |
-    v
-[Step 3: Codex MCP First Review] ──────────────────────────┐
-    |     (mcp__codex__codex)                              |
-    v                                                       |
-[Step 4: Parse Feedback]                                    |
-    |                                                       |
-    v                                                       |
-[Step 5: Claude Cross-check]                                |
-    |                                                       |
-    v                                                       |
-[Step 6: Disagreements?]                                    |
-    |                                                       |
-    ├─ YES & iterations < 3 ─> [New MCP call + context]────┘
-    |                          (includes conversation summary)
-    |
-    └─ NO or iterations >= 3 ─> [Step 7: Deliver Consensus]
-```
-
-**Core Flow:**
-1. Execute Codex MCP review
-2. Claude validates via cross-check
-3. On disagreement, new MCP call (include previous context in prompt)
-4. Repeat until consensus or maximum 3 iterations
-5. Deliver final results
-
----
-
-## Expected Output Example
-
-```markdown
-## PRD Review Results (Codex + Claude Consensus)
-
-### Review Process
-- Ping-pong iterations: 2
-- Consensus status: Full consensus
-
-### [VALID] Valid Feedback
-| Item | Description | Source |
-|------|-------------|--------|
-| No success metrics | Goals only list functional goals, no performance/cost criteria defined | Codex |
-| Missing error handling | No exception handling in API call code | Codex + Claude |
-| Test scenarios | Only happy-path exists, no failure cases | Codex |
-
-### [ISSUE] Items Needing Improvement
-| Item | Problem | Recommended Fix | Source |
-|------|---------|-----------------|--------|
-| Dependency list | Non-existent package mentioned | Update to actual package | Claude verification |
-| Milestone status | Includes already-resolved issues | Mark complete or remove | Claude verification |
-
-### [CORRECTION] Codex Error Corrections
-| Codex Claim | Actual Situation | Evidence |
-|-------------|------------------|----------|
-| Feature not implemented | Feature already exists | Verified in src/modules/ folder |
-
-### [SUMMARY] Final Conclusion
-PRD is structurally sound, but non-functional requirements (performance, security, error handling)
-and dependency information need updates. Recommend removing already-resolved issues from Milestone.
-```