tree-sitting

AST-powered code navigation using tree-sitter. Each invocation auto-scans the codebase (~700ms for 250 files), then runs queries at sub-millisecond speed.

Setup

uv pip install --system --break-system-packages tree-sitter

Grammars are loaded from bundled parsers/*.so files — no network fetch, no tree-sitter-language-pack dependency. Install is <1s.

Usage: CLI (treesit.py)

Every call auto-scans, prints a tree overview, then runs any queries. No state to manage between calls.

TREESIT=/mnt/skills/user/tree-sitting/scripts/treesit.py

# Orient: root-level overview (default depth=1)
python3 $TREESIT /path/to/repo

# Featuring: complete tree, minimal detail
python3 $TREESIT /path/to/repo --depth=-1 --detail=sparse

# Explore a subdirectory in full detail
python3 $TREESIT /path/to/repo --path=src/core --detail=full

# Run queries (tree overview + query results)
python3 $TREESIT /path/to/repo 'find:Parser*' 'source:parse_input'

# Queries only, no tree
python3 $TREESIT /path/to/repo --no-tree 'refs:AuthToken'

Options

Detail Levels

All levels include line ranges (:start-end) so you can feed the window straight into Read --offset/--limit without another scan.

Queries

Append after the repo path. Multiple queries per call.

Workflow

1. treesit.py /repo                           → orient: what dirs, how big
2. treesit.py /repo --path=src/core           → drill into interesting directory
3. treesit.py /repo 'find:Parser*'            → find specific symbols
4. treesit.py /repo 'source:parse_input'      → read implementation
5. treesit.py /repo 'refs:ParseState'         → find usage across codebase

Each call is self-contained. No need to "scan first, query later" — scan happens automatically every time (~700ms).

Usage: Direct Python (single invocation)

For custom scripts that need the engine API directly:

import sys; sys.path.insert(0, '/mnt/skills/user/tree-sitting/scripts')
from engine import CodeCache

cache = CodeCache()
cache.scan('/path/to/repo')
# All queries in the SAME invocation:
print(cache.tree_overview())
print(cache.find_symbol('ClassName'))
print(cache.get_source_range('src/core/parser.c', 100, 150))

Important: The cache is in-memory only. All scan + query calls MUST happen in the same Python process. Splitting across separate python -c invocations loses the cache — use treesit.py instead.

Supported Languages

Bundled grammars (work out of the box): Python, JavaScript, TypeScript, TSX, Go, Rust, Ruby, Java, C, HTML, Markdown, Mojo.

Three-tier extraction for bundled languages:

1. Custom extractors (richest — signatures, hierarchy, docstrings): Python, C, Go, Rust, JavaScript, TypeScript, TSX, Ruby, Markdown (heading outline)
2. tags.scm queries (community-maintained — kinds, docs): Java, Mojo
3. Generic heuristic (names + kinds + locations): HTML and any future bundled grammars

Adding a grammar

Files with unsupported extensions are silently skipped (they show as SKIP (no parser) with --stats). To add a grammar, drop a compiled libtree_sitter_<lang>.so into parsers/ — the engine picks it up automatically on the next run. Build from the grammar's repo (each tree-sitter/tree-sitter-<lang> repo has a src/ directory you can compile with cc -shared -fPIC -I src src/parser.c src/scanner.c -o libtree_sitter_<lang>.so, or use tree-sitter build).

If you need a language urgently and can't build the .so, you can try installing tree-sitter-language-pack as a fallback (uv pip install --system --break-system-packages 'tree-sitter-language-pack<1.6.3') — but note 1.6.3 ships a broken wheel (only _native/, missing the tree_sitter_language_pack/ python module → ModuleNotFoundError despite pip showing it installed); 1.6.2 and earlier work, hence the <1.6.3 pin and try to download grammars at runtime from a domain that may not be in your network allowlist. Bundling the .so is the reliable path.

What It Extracts

• Symbols: functions, classes, structs, enums, methods, constants, defines, types
• Signatures: parameter lists and return types (Python, C; partial for others)
• Doc comments: first-line summaries from docstrings, JSDoc, Doxygen, ///, #
• Line ranges: start and end line for every symbol
• Imports: per-file dependency tracking
• Hierarchy: class→methods, struct→fields (Python, C)

Architecture

CodeCache (in-memory, per-invocation)
  ├── files: {relpath → FileEntry(source, tree, symbols, imports)}
  ├── _symbol_index: {name → [Symbol, ...]}  ← fast lookup
  └── methods: scan(), find_symbol(), file_symbols(), dir_overview(), ...
       │
       └── treesit.py CLI — auto-scan + progressive-disclosure tree + queries

Parse cost is paid once per invocation. The symbol index enables O(1) exact match and O(n) substring/glob search where n is the number of unique symbol names (not files).