docvet

Better docstrings, better AI.

Why docvet?

docvet combines docstring presence metrics, structural checks, history-based staleness signals, and renderer diagnostics in one CLI, with editor and agent interfaces. These checks help reviewers find documentation gaps; they do not prove that prose describes runtime behavior correctly.

Comparison reviewed 2026-10-21 against primary documentation: Ruff's D103 presence rule and DOC501 exception completeness rule (requires --preview and rule selection), pydoclint's supported checks and styles, and interrogate's coverage documentation. Capabilities depend on enabled rules and configuration; -- means no corresponding check is described in these sources.

pydoclint checks arguments, returns, yields, raises, class attributes, and type hints across Google, NumPy, and Sphinx styles. docvet's enrichment alone has 21 rules, including Raises, Yields, Receives, Warns, Attributes, Examples, cross-references, parameter agreement, and more. Add presence (coverage metrics + threshold enforcement), freshness (git diff/blame staleness signals), griffe rendering compatibility, and mkdocs coverage: 32 rules across 5 checks. Rule totals measure catalog entries, not equivalent coverage between tools.

Structural agreement checks compare documented elements with code structure. Freshness uses change history to suggest review: code can change without invalidating prose, and simultaneous code/docstring edits can still leave incorrect prose. Neither establishes semantic correctness.

Quickstart | GitHub Action | Pre-commit | Configuration | AI Agent Integration | Docs

What It Checks

Presence (existence) -- 2 rules: missing-docstring overload-has-docstring

Enrichment (completeness) -- 21 rules: missing-raises missing-returns missing-yields missing-receives missing-warns missing-deprecation missing-param-in-docstring extra-param-in-docstring missing-other-parameters missing-attributes undocumented-init-params missing-typed-attributes missing-examples missing-cross-references extra-raises-in-docstring extra-yields-in-docstring extra-returns-in-docstring missing-return-type trivial-docstring prefer-fenced-code-blocks scaffold-incomplete

Freshness (staleness signals) -- 5 rules: stale-signature stale-body stale-import stale-drift stale-age

Griffe (rendering) -- 3 rules: griffe-unknown-param griffe-missing-type griffe-format-warning

Coverage (visibility) -- 1 rule: missing-init

Quickstart

pip install docvet && docvet check --all

For optional griffe rendering checks:

pip install docvet[griffe]

Example output:

src/mypackage/helpers.py:1: missing-docstring Module has no docstring [required]
src/mypackage/utils.py:42: missing-raises Function 'parse_config' raises ValueError but has no Raises section [required]
src/mypackage/models.py:15: stale-signature Function 'process' signature changed but docstring not updated [required]
src/mypackage/api.py:1: missing-init Package directory missing __init__.py (invisible to mkdocs) [required]

Configuration

Configure via [tool.docvet] in your pyproject.toml. All checks run and print findings. Checks listed in fail-on cause a non-zero exit code; unlisted checks are treated as warnings.

A check in fail-on that cannot run — most often griffe without the docvet[griffe] extra — never certified the gate you configured, so docvet reports it on stderr and exits 1. With --format json the run object reports status: "unavailable" and an unavailable_checks array, which is what distinguishes it from a gate that found problems. Set fail-on-unavailable = false (or pass --no-fail-on-unavailable) to warn and exit 0 instead. Unavailable checks that nothing gates on stay a quiet skip — though fail-on membership is not the only thing that gates: a min-coverage floor gates presence without naming it there.

[tool.docvet]
exclude = ["tests", "scripts"]
fail-on = ["griffe", "coverage"]

[tool.docvet.freshness]
drift-threshold = 30
age-threshold = 90

Pre-commit

Add to your .pre-commit-config.yaml:

repos:
  - repo: https://github.com/Alberto-Codes/docvet
    rev: v3.0.0
    hooks:
      - id: docvet

For griffe rendering checks, add the optional dependency:

repos:
  - repo: https://github.com/Alberto-Codes/docvet
    rev: v3.0.0
    hooks:
      - id: docvet
        additional_dependencies: [griffe]

GitHub Action

Add docvet to your GitHub Actions workflow — findings appear as inline annotations on your PR:

- uses: Alberto-Codes/docvet@v3.0.0

Select specific checks or pin a version:

- uses: Alberto-Codes/docvet@v3.0.0
  with:
    checks: 'enrichment,freshness'
    docvet-version: '3.0.0'
    python-version: '3.13'

The griffe rendering check needs no setup: the action installs docvet[griffe], pinned or not, so every check the checks input offers is available. Note that docvet releases before 1.7.0 declare that extra without an upper bound, so pinning one installs whatever griffe publishes at the time rather than a version docvet was released against.

Behavior change — this can turn a passing build red.

Earlier releases installed plain docvet, so the griffe check was skipped and contributed zero findings. It now runs. determine_run_outcome (src/docvet/reporting.py) returns exit code 1 as soon as any check listed in fail-on reports findings, so if your pyproject.toml has griffe in [tool.docvet] fail-on, your build goes from green to failing with no change on your side. This repository's own ci.yml docvet job is exactly such a consumer.

These are not new problems — it is the check finally running on docstrings that were always broken. To get back to green, fix the griffe findings or remove griffe from fail-on.

AI Agent Integration

For tool-specific integration snippets, see the full AI Agent Integration guide.

Add docvet to your AI coding workflow. Drop this into your CLAUDE.md, .cursorrules, or agent configuration:

## Docstring Quality

After modifying Python functions, classes, or modules, run `docvet check` and fix all findings before committing.

Recommended pyproject.toml configuration:

[tool.docvet]
fail-on = ["enrichment", "freshness", "coverage", "griffe"]

Subcommand Quick Reference

Better Docstrings, Better AI

AI coding agents rely on docstrings as context when generating and modifying code. Agents modify code but often leave docstrings stale, and research shows stale or incorrect documentation is actively harmful, worse than no docs at all:

• Incorrect docs degrade LLM task success by 22.6 percentage points
• Comment density improves code generation by 40-54%
• Misleading comments reduce LLM fault localization accuracy to 24.55%
• Performance drops substantially without docstrings

As the 2025 DORA report puts it: "AI doesn't fix a team; it amplifies what's already there." The only signal correlating with AI productivity is code quality.

docvet's freshness heuristics flag potential staleness for review, and its enrichment rules check supported docstring sections against code structure. Run docvet check in your CI, pre-commit hooks, or agent toolchain.

Badge

Add a badge to your project to show your docs are vetted:

[![docs vetted | docvet](https://img.shields.io/badge/docs%20vetted-docvet-purple)](https://github.com/Alberto-Codes/docvet)

Used By

Are you using docvet? Open a pull request to add your project here.

License

MIT -- see LICENSE for details.

mcp-name: io.github.Alberto-Codes/docvet