sourcecrumb

Tree-sitter repository map in TOON format for LLM consumption.

What it does

sourcecrumb parses a codebase with tree-sitter, extracts symbols (classes, functions, methods, imports), builds a file-to-file dependency graph, and ranks files by PageRank. The output is a compact TOON-formatted map designed to fit in an LLM context window.

The goal: give an LLM agent a high-level map of a codebase so it can explore more effectively — knowing which files matter most, what symbols they define, and how they depend on each other.

Installation

Requires Python >= 3.13.

From PyPI

pip install sourcecrumb

Or with uv:

uv pip install sourcecrumb

Run without installing

uvx sourcecrumb .

From source

git clone https://github.com/phobologic/sourcecrumb.git
cd sourcecrumb
uv sync

Usage

sourcecrumb [ROOT] [OPTIONS]

Example

$ sourcecrumb . -n 3
repo: sourcecrumb
root: sourcecrumb
files[3]{path,language,rank}:
  sourcecrumb/models.py,python,0.2615
  sourcecrumb/languages.py,python,0.1155
  sourcecrumb/discovery.py,python,0.0590
symbols[17]{file,name,kind,line,signature}:
  sourcecrumb/models.py,TagKind,class,10,TagKind(enum.Enum)
  sourcecrumb/models.py,SymbolKind,class,17,SymbolKind(enum.Enum)
  sourcecrumb/models.py,Tag,class,27,Tag
  sourcecrumb/models.py,FileInfo,class,39,FileInfo
  ...
dependencies[1]{source,target,symbols}:
  sourcecrumb/discovery.py,sourcecrumb/languages.py,language_for_extension

Claude Code integration

The primary use case is running sourcecrumb as a Claude Code hook so every subagent automatically gets a repo map injected into its context.

Add this to .claude/settings.json:

{
  "hooks": {
    "SubagentStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "uvx sourcecrumb \"$CLAUDE_PROJECT_DIR\" --cache \"$CLAUDE_PROJECT_DIR/.cache/sourcecrumb.toon\""
          }
        ]
      }
    ]
  }
}

If you have sourcecrumb installed globally, you can use sourcecrumb directly instead of going through uvx.

The SubagentStart hook fires when any subagent launches. sourcecrumb's stdout is injected into the subagent's context, giving it an instant overview of the codebase.

--cache avoids re-parsing on every agent launch — the cache file is reused as long as no source files have changed. Add .cache/ to your .gitignore.

TOON format

The output uses TOON (Text Object Oriented Notation), a compact format designed for LLM consumption:

• Scalar fields — key: value
• Tabular arrays — name[count]{col1,col2,...}: followed by indented CSV rows
• Quoting — values containing special characters are double-quoted; numbers and plain strings are bare

How it works

1. Discover files — uses git ls-files when available, falls back to .gitignore-based filtering
2. Parse with tree-sitter — extracts classes, functions, methods, and imports from each file
3. Build dependency graph — creates file-to-file edges based on shared symbols (imports that resolve to definitions in other files)
4. Rank with PageRank — scores files by importance in the dependency graph
5. Select top N — when --max-files is set, keeps only the highest-ranked files
6. Encode to TOON — serializes the repo map into the compact output format

Supported languages

Python. Extensible by adding a .scm query file to sourcecrumb/queries/ and registering the language in sourcecrumb/languages.py.

Development

uv run pytest                            # run tests
uv run ruff check sourcecrumb/ tests/    # lint
uv run ruff format sourcecrumb/ tests/   # format