trailmap 0.10.0

Map a Python source tree into typed-edge markdown for fast AI navigation.

Trailmap

Map a source tree into typed-edge markdown so an AI can find any symbol and see how it connects without reading the source. Python and TypeScript. Design and rationale in DESIGN.md.

Python analysis uses the standard library only. Other languages are opt-in extras, so the base install stays dependency-free.

Install

pip install trailmap                # Python analysis, zero dependencies
pip install trailmap[typescript]    # adds TypeScript/TSX and JavaScript/JSX (tree-sitter)

Installs a trailmap console command. Or run it in place: python -m trailmap ....

Languages

Python (stdlib ast, always available), and TypeScript/TSX plus JavaScript/JSX (tree-sitter, the typescript extra — JS reuses the TS adapter and handles both ES modules and CommonJS require/module.exports). Trailmap maps whatever it recognizes in a tree and skips files whose language extra isn't installed, telling you what to install. Each new language is an adapter behind its own extra; the core stays pure Python.

Run

python -m trailmap <source-tree> -o <output-dir>

Map Trailmap's own source (the demo):

python -m trailmap trailmap -o trailmap-out

Output

• INDEX.md — front door and how-to-read.
• modules/<module>.md — one page per module; frontmatter has imports, body has a section per symbol with its edges.
• symbols.tsv — fqn kind file line signature doc. Grep it to locate a symbol, or to search intent: doc is a one-line summary harvested from the symbol's docstring (Python) or JSDoc/leading comment (TS/JS).
• edges.tsv — src type dst confidence resolved. Grep src for dependencies, dst for dependents.
• candidates.tsv — for a ?-unresolved call, the possible in-project targets (caller method candidate n_candidates). A lead, kept out of edges.tsv.
• entrypoints.tsv — where execution or the public API enters the tree (entry name fqn kind file line): Python __main__ guards and console scripts, JS/TS package.json bins, and shebang'd files. The "start here" list; also summarized at the top of INDEX.md.
• modules.tsv — the module-level dependency graph (src dst edges kinds): the resolved symbol edges rolled up to module granularity, so you can read the architecture — which module depends on which, and how much — without walking every symbol. Each module page also lists its depends: in frontmatter.
• trailmap.json — the full structured graph.

Edges are ast/parser (precise, trust) or heuristic (a lead). Unresolved targets keep a trailing ?.

Incremental runs

Re-running is incremental: each file's extraction is cached (by content hash) in .trailmap-cache.json in the output directory, so only changed files are re-parsed. Output is identical whether a module was parsed or reused. Add .trailmap-cache.json to your .gitignore.

trailmap <src> -o out --dry-run    # report added/modified/deleted files, no work
trailmap <src> -o out --no-cache   # re-parse everything

Discovery honors .gitignore and .hgignore (from the scanned dir up to the VCS root), on top of a built-in ignore list for .git, node_modules, .venv, and the like. Use --no-ignore to map every source file regardless.

Status

Early (0.1). Python (stdlib ast) and TypeScript/TSX (tree-sitter extra). Resolution is best-effort without a type checker: it follows imports, inferred receiver types (annotations, constructors, and declared return types), the inheritance chain, in-project star imports, and TypeScript workspace/tsconfig package aliases with barrel re-exports, and offers candidate leads for the rest. Limits are in DESIGN.md. The core is language-agnostic; new languages are adapters behind their own extra.

MCP server

pip install trailmap[mcp] adds a trailmap-mcp command: an MCP server that exposes the lookups as tools for any MCP host (Claude Desktop, Cursor, ...), so an agent can navigate a codebase without reading raw source. Point a host at it:

{
  "mcpServers": {
    "trailmap": { "command": "trailmap-mcp" }
  }
}

Tools: overview, find_symbol, callers, dependencies, candidates, module_dependencies (the architecture view for a module), and refresh (rebuild after code changes). Each takes a root (default .), so launch the server with the repo as its working directory or pass root per call. The graph is built in memory and cached per root.

Releasing

Maintainers: scripts/release.sh minor (or patch/major) bumps the version, builds a clean dist/, and publishes to PyPI; run with no argument to publish the current version as-is. Needs uv and a PyPI token in ~/.pypirc. Commit and push the version change separately.

License

MIT. See LICENSE.