# AGENTS.md - Development Guidelines

## Code Style Guidelines

### General Principles
- Use relative paths only (never absolute paths)
- Operate strictly on explicitly mentioned files
- Generate minimal, single implementations
- Avoid speculation about missing files

## Error Handling
- Use structured error messages
- Avoid verbose stack traces in user-facing output

## Naming Conventions
- Variables/functions: `snake_case`
- Types/interfaces: `PascalCase`
- Constants: `UPPER_SNAKE_CASE`

## CRITICAL Security Rules

### File Operations
- **NEVER** delete files without explicit human authorization
- **NEVER** modify files outside the scope of the requested task
- Always use relative paths (never absolute paths)

### Version Control with `jj`
- Projects are managed with `jj` (Jujutsu), NOT Git
- **Before starting modifications to respond to a user prompt**, run `jj new` to create a commit
- This creates a baseline to revert all changes from that prompt if needed
- Do NOT run `jj new` for every individual file change—only once at the start of a prompt response

## File Generation Rules
- Output only requested files
- Do not modify unrelated files
- Valid JSON without comments when requested
- Assume no knowledge of other files unless provided

## MCP Tools
- Located in `/opt/dev-tools/`
- Use SQLite for graph storage (`.ai_index/graph.db`)
- Embeddings stored in `.ai_index/embeddings/`
- AST data in `.ai_index/ast/`

## LM Studio Integration
- Local LLM API: `http://host.docker.internal:1248`
- Model: `qwen/qwen3-coder-next@4bit`
- Context window: 32768 tokens
- Max output: 65536 tokens