Quick Start

Install the skill for your AI coding agent:

npx skills add AvivK5498/Agentifind

Then ask your agent: "Run /agentifind" or "Set up codebase intelligence"

Works with: Claude Code, Cursor, Windsurf, Cline, and other supported agents.

What It Does

The /agentifind skill:

1. Runs agentifind sync to extract code structure (LSP or tree-sitter)
2. Analyzes the extracted data to identify architecture
3. Synthesizes .claude/CODEBASE.md - a navigation guide for your agent

Output

.claude/
├── codebase.json         # Structured extraction (modules, call graph, imports)
├── CODEBASE.md           # Navigation guide (synthesized by skill)
└── .agentifind-checksum  # Staleness detection

Benchmark Skill

Measure how much CODEBASE.md improves agent efficiency:

npx skills add AvivK5498/Agentifind/skills/agentifind_benchmark

Then run /agentifind-benchmark to:

• Set up hooks that enforce guide restrictions
• Create 7 benchmark tasks (fundamental → expert)
• Run two agents in parallel: one WITH guide, one WITHOUT
• Compare tool calls, accuracy, and token usage

Requires: Run /agentifind first to generate CODEBASE.md

Supported Languages

| Language | Extraction | Validation | |----------|------------|------------| | Python | LSP (pyright) or tree-sitter | ruff, mypy | | TypeScript | LSP (tsserver) or tree-sitter | eslint, tsc | | JavaScript | tree-sitter | eslint |

Extraction Methods

• LSP first (if available): Accurate cross-file resolution, 5-15 min on large codebases
• Tree-sitter fallback: Fast AST parsing, ~30 seconds

Direct CLI Usage

You can also run the CLI directly without the skill:

# Extract code structure
npx agentifind sync

# Check if index is stale
npx agentifind status

Options: | Option | Description | |--------|-------------| | --skip-validate | Skip linting/type checks (faster) | | --if-stale | Only sync if source files changed | | --verbose | Show extraction progress | | --ci | Exit non-zero on validation failures |

Requirements

• Node.js >= 18
• Git repository
• macOS or Linux (Windows via WSL)

License

MIT