anatomize 0.2.1

Deterministic, token-efficient codebase packs and skeleton maps for AI review (Python)

anatomize

Generate deterministic, token-efficient maps and review bundles for Python repositories.

anatomize has two complementary workflows:

1. Skeletons: structure-only “code maps” for navigation and architecture understanding.
2. Packs: single-file bundles (repomix-style) for external review, with filtering and slicing.

If you want the full guide (modes, slicing, config, determinism guarantees), see docs/GUIDE.md.

---

Installation

pip install anatomize

---

Quick Start (CLI)

Generate skeletons

# Scaffold config for the common workflow “src detailed, tests minimal”
anatomize init --preset standard

# Generate all configured outputs from .anatomize.yaml (writes into .anatomy/*)
anatomize generate

# Or run ad-hoc generation for a specific source path
anatomize generate ./src

# Choose resolution level
anatomize generate ./src --level hierarchy --output .anatomy
anatomize generate ./src --level modules --output .anatomy
anatomize generate ./src --level signatures --output .anatomy

# Write multiple formats
anatomize generate ./src --format yaml --format json --format markdown --output .anatomy

Estimate tokens

anatomize estimate ./src --level modules

Validate (and fix) skeleton output

# Validate all configured outputs (from .anatomize.yaml)
anatomize validate

# Rewrite configured outputs to match regenerated content (strict, atomic-ish replacement)
anatomize validate --fix

# Or validate a specific directory against explicit sources
anatomize validate .anatomy/src --source ./src

Pack a repository into an AI-friendly bundle

# If --format is omitted, it is inferred from --output when the extension is known
anatomize pack . --output codebase.jsonl
anatomize pack . --output codebase.md

# Full bundle
anatomize pack . --format markdown --output codebase.md

# Minimal prefix (lower token overhead)
anatomize pack . --prefix minimal --output codebase.md

# Explain selection (why files were included/excluded)
# (writes `codebase.md.selection.json` by default)
anatomize pack . --explain-selection --output codebase.md

# Filter by globs
anatomize pack . --include "src/**" --ignore "**/__pycache__/**" --output src-only.md

# Forward dependency closure (entrypoint + everything it imports)
anatomize pack . --entry src/anatomize/cli.py --deps --output slice.md

# Reverse dependency closure (module + everything that imports it)
anatomize pack . --target src/anatomize/cli.py --reverse-deps --output importers.md

# Reverse + forward (importers plus what they import)
anatomize pack . --target src/anatomize/cli.py --reverse-deps --deps --output importers-and-deps.md

# Token-efficient Python compression (signatures/imports/constants)
anatomize pack . --compress --output compressed.md

# Make markdown robust to embedded ``` fences (default)
anatomize pack . --content-encoding fence-safe --output safe.md

# Maximum robustness (content is base64-encoded UTF-8)
anatomize pack . --content-encoding base64 --output safe.base64.md

# Split output into multiple files (markdown/plain only)
anatomize pack . --split-output 500kb --output codebase.md

# Hard cap output (bytes or tokens)
anatomize pack . --max-output 20_000t --output codebase.md

# Print a per-file content token tree to stdout
anatomize pack . --token-count-tree --output codebase.md

# JSONL (stream-friendly)
anatomize pack . --format jsonl --output codebase.jsonl

# Hybrid mode (summaries + selective fill; token-efficient)
# - defaults to markdown when --format and the output extension are not specified
# - Python files default to summary; non-Python defaults to metadata-only
anatomize pack . --mode hybrid --output hybrid.md

# Hybrid: include full content for a slice and fit within a hard token budget (JSONL only)
anatomize pack . --mode hybrid --format jsonl --max-output 50_000t --fit-to-max-output \
  --content "src/pkg/**" --output hybrid.slice.jsonl

Reference-based usage slicing (requires Pyright language server):

anatomize pack . --target src/anatomize/cli.py --uses --slice-backend pyright --output uses.md

---

Python API

Generate skeletons in code

from anatomize import SkeletonGenerator
from anatomize.formats import OutputFormat, write_skeleton

gen = SkeletonGenerator(sources=["./src"])
skeleton = gen.generate(level="modules")

print("Modules:", skeleton.metadata.total_modules)
print("Classes:", skeleton.metadata.total_classes)
print("Functions:", skeleton.metadata.total_functions)
print("Estimated tokens:", skeleton.metadata.token_estimate)

write_skeleton(skeleton, ".anatomy", formats=[OutputFormat.YAML, OutputFormat.JSON])

Key exported objects

• anatomize.SkeletonGenerator: orchestrates discovery + extraction.
• anatomize.formats.write_skeleton: writes YAML/JSON/Markdown plus schemas and manifest.json.
• anatomize.validation.validate_skeleton_dir: strict validator with optional fix.

---

Configuration (.anatomize.yaml)

The CLI can auto-discover .anatomize.yaml. Generation commands use config from the current working directory (or explicit --config). pack discovers config relative to the chosen ROOT when --config is not provided.

Minimal config:

output: .anatomy

sources:
  - path: src
    output: src
    level: modules
  - path: tests
    output: tests
    level: hierarchy

# Defaults applied to sources that omit fields
level: modules
formats: [yaml, json, markdown]
exclude:
  - __pycache__/
  - "*.pyc"
symlinks: forbid # forbid|files|dirs|all
workers: 0 # 0 = auto

pack:
  format: markdown # markdown|plain|json|xml|jsonl (hybrid supports markdown|plain|jsonl)
  mode: bundle # bundle|hybrid (hybrid is token-efficient summaries + selective fill)
  prefix: standard # standard|minimal
  output: anatomize-pack.md # if the extension is known, it must match `format`
  include: []
  ignore: []
  ignore_files: []
  respect_standard_ignores: true
  symlinks: forbid # forbid|files|dirs|all
  max_file_bytes: 1000000
  workers: 0 # 0 = auto
  token_encoding: cl100k_base
  compress: false
  content_encoding: fence-safe # verbatim|fence-safe|base64 (markdown disallows verbatim)
  line_numbers: false
  no_structure: false
  no_files: false
  max_output: null # e.g. "500kb" or "20_000t"
  split_output: null # e.g. "500kb" or "20_000t"
  fit_to_max_output: false
  # Hybrid representation rules (repeatable patterns). Precedence: meta < summary < content.
  meta: []
  summary: []
  content: []
  summary_config:
    max_depth: 3
    max_keys: 200
    max_items: 200
    max_headings: 200
  python_roots: [] # defaults to ["src"] if present, else ["."]
  slice_backend: imports # imports|pyright
  uses_include_private: false
  pyright_langserver_cmd: "pyright-langserver --stdio"

Exclude patterns use gitignore-like semantics and are applied relative to each configured root.

Tip: anatomize init --preset standard scaffolds .anatomize.yaml with the common pattern “src detailed, tests minimal”.

---

Output artifacts

Skeleton output directory

write_skeleton(...) and anatomize generate ... --output DIR create:

• hierarchy.yaml|json|md / modules.* / signatures.* depending on selected formats and level
• schemas/*.json embedded with the package
• manifest.json (SHA-256 per output file and metadata for validation)

When anatomize generate runs from .anatomize.yaml, it writes one skeleton directory per configured source:

• .anatomy/src/...
• .anatomy/tests/...

Pack output file(s)

anatomize pack writes one or more files depending on splitting:

• anatomize-pack.md (or .txt|.json|.xml)
• if split: anatomize-pack.1.md, anatomize-pack.2.md, …

Each pack artifact starts with a lightweight, deterministic overview (and, if enabled, a structure tree) before file blocks/records.

Token reporting:

• Artifact tokens: exact tokens of the written output file(s) (returned by the Python API).
• Content tokens: tokens of file contents only (returned by the Python API; useful for budgeting).

Pack artifacts intentionally do not embed token counts (agents don’t need them; they waste tokens).

---

Determinism and strictness

• Deterministic ordering (paths and symbols sorted).
• No timestamps in outputs.
• Parse failures are hard failures (no partial output).
• Validation is strict; --fix replaces output with regenerated content.

---

Development

python -m venv .venv
. .venv/bin/activate
python -m pip install -U pip
python -m pip install -e ".[dev]"

python -m ruff check .
python -m mypy -p anatomize
python -m pytest

Optional local benchmark:

.venv/bin/python scripts/bench_pack.py . --compress --workers 0

---

Tests

Tests are indexed via pytest markers in pyproject.toml and documented in tests/README.md:

• unit: fast, isolated tests
• integration: filesystem-level tests
• e2e: CLI-level tests

---

License

MIT