docrot

Deterministic doc↔code drift checker for Python repositories.

docrot finds references in your prose documentation — READMEs, docs/, contributor guides, and agent context files like CLAUDE.md/AGENTS.md — that no longer resolve against your codebase: symbols that were removed or renamed, file paths that moved, make/tox/npm targets that are gone, and APIs that were documented but never shipped.

No LLM. No side-store. Every finding is backed by static analysis and git history, and comes with provenance: the commit where the claim was written, and the commit where it broke.

$ docrot
docs/advanced/transports.md:178:53  PY003 error  `httpx.Mounts` is documented but has never existed
    written at 3faa4a8 (2024-02-14)

docrot: 187 references | 82 resolved | 85 unknown (skipped) | 1 findings (1 error)

That output is real: on its first run against httpx (~14k stars), docrot found that the official transports documentation shows a complete usage example for httpx.Mounts — a class that has never existed in any commit of the codebase. On requests and flask, the same run produced zero findings — not because docrot checked less, but because their docs are actually in sync.

Why

A 2023 Empirical Software Engineering study (DOCER) measured that 28.9% of the top-1000 GitHub projects currently document a code element that no longer exists, and 82.3% have been in that state at some point. Since then the problem got teeth: AI coding agents read AGENTS.md/CLAUDE.md as ground truth on every session, so one stale path silently misdirects every run.

Adjacent tools deliberately stop short of this: doc example runners (sybil, xdoctest, pytest-examples) check fenced code blocks; docstring linters (docvet, pydoclint) check docstrings against signatures; link checkers (lychee) check URLs; AGENTS.md linters check file structure. None of them validate what your prose claims about your code.

How it stays quiet: the two-gate design

False positives are what killed every previous attempt at this, so docrot only reports when two independent gates agree:

Gate 1 — resolution (static analysis). References are resolved against a griffe-backed index of your package that understands star-import re-exports, alias chains (requests.Response → requests.models.Response), class inheritance (flask.Flask.route lives two base classes away), and instance attributes assigned in __init__. Every reference gets a tri-state verdict — RESOLVED, BROKEN, or UNKNOWN — and UNKNOWN never becomes a finding. External packages, dynamic attributes, tutorial placeholders: silence, not noise.

Gate 2 — history (git). A broken reference is only drift if it once resolved. docrot blames the doc line to the commit where the claim was made, re-resolves the reference at that commit (by parsing file blobs — no checkouts), and only then decides:

The result on these three: requests 0 findings, flask 0 findings, httpx 1 finding, a real bug (709 references checked in total, a few seconds per repository). The full corpus run is in BENCHMARKS.md.

Install & run

pip install docrot
docrot                  # zero config: checks tracked *.md, *.rst + agent files
docrot --format json    # machine-readable, stable schema
docrot --show-unknown   # audit what was skipped and why
docrot explain PY002    # rule documentation

Requires Python 3.10+ and git. Without git history docrot still runs, but it says so and drops every finding to medium confidence, because the drift-versus-fiction evidence is exactly what history provides.

Command reference

Rules

Suppress inline with <!-- docrot: ignore[PY002] -->; suppressions are counted in the report, never silent.

CI

GitHub Action (findings become PR annotations, optional SARIF upload):

- uses: actions/checkout@v4
  with:
    fetch-depth: 0        # full history = full provenance
- uses: dager23/docrot@v0

pre-commit:

repos:
  - repo: https://github.com/dager23/docrot
    rev: v0.1.0
    hooks:
      - id: docrot        # runs --changed-only for speed

Exit codes: 0 clean, 1 findings at/above fail_on severity, 2 usage error. Medium-confidence findings (no git history) fail only under --strict.

Configuration

Everything is optional. Use [tool.docrot] in pyproject.toml, or a standalone docrot.toml (same keys, no table header, takes precedence):

[tool.docrot]
packages = ["mypkg"]                 # default: auto-detected
exclude_docs = ["docs/archive/**"]   # changelogs are always excluded
severity = { PATH002 = "ignore" }
fail_on = "error"                    # error | warning | info
temporal = "auto"                    # auto | on | off
external_packages = ["werkzeug"]     # opt-in cross-package resolution
ignore_refs = ["flask.request.*"]     # glob-match references to skip
docs = ["**/*.md", "**/*.rst"]       # what counts as a doc
rules = { enable = ["core", "agents"], disable = ["LINK001"] }
strict = false

Unknown keys are reported and ignored rather than fatal. Changelogs, migration guides and upgrade notes are always skipped: naming removed APIs is their entire purpose.

Anchors

For prose whose meaning depends on specific code (not just references), bind a region to a symbol's implementation:

<!-- docrot:anchor symbol=mypkg.auth.DigestAuth hash=sha256:9f2ab8... -->
Digest auth performs two requests: the first receives the challenge...
<!-- /docrot:anchor -->

The hash covers the AST-normalized body, so formatting and comment churn never invalidate it — only structural change does. docrot anchors update recomputes hashes after you review the prose. Anchors are strictly opt-in: an absent anchor asserts nothing.

What docrot deliberately does not do

• No LLM calls, ever. Semantic truth of prose isn't deterministically checkable; references and anchored claims are. docrot never guesses.
• No docstring linting (see docvet, pydoclint), no example execution (see sybil, pytest-examples), no URL checking (see lychee). docrot is designed to sit alongside all of them.

Status

Alpha (0.x). Symbol resolution is Python-only; doc formats are Markdown, reStructuredText, and agent context files. Multi-language resolution (tree-sitter), CLI flag introspection, --fix rename suggestions, and an MCP server are planned.

Measured against 25 mature open-source projects: 6,073 references checked, 10 findings, every one confirmed by hand as a real documentation defect, and 139 of 139 seeded scenarios correct. The numbers and the method are in BENCHMARKS.md.

License

MIT