Review only the Python code you changed before pushing

avouch

Review the Python you changed, not the Python you inherited.

Avouch is a code reviewer for Python that only looks at the files your next commit touches. It asks git for that list, parses each changed .py with the standard ast module, and reports structural problems against limits you set in avouch.toml.

No daemon, no network, nothing to install alongside it. Run it seconds before git push:

pip install avouch
avouch

Why

Paid AI reviewers wasted half their report on legacy I never wrote and queued my diffs behind everyone else's. I wanted a local, diff-only check that runs before git push, costs nothing, and shows only your changes - so I built Avouch.

• Diff-only: Review set is git diff HEAD + untracked, not whole repo.
• Local: Three git calls + ast, no server.
• Yours: Baseline hides legacy; every finding is attributable to your diff.

Install

Requires Python 3.10+ (ast.Match, tomllib) and git on PATH.

pip install avouch

or from source:

git clone https://github.com/mukundzha/avouch.git
cd avouch
pip install -e .

Both register avouch (avouch.cli:main).

Quick start

Change a file, run it, read the compiler-style report:

$ avouch
AVOUCH · 2 FILES · 4 WARN
────────────────────────────────────────────────
bad.py:1: SCR002: Bare except detected.
  │
1 │ def connect(host, port, ...):
  │     ^^^^^^^ SCR002
2 │     try:
BY RULE  SCR002 Bare except 1
         SCR014 Too many parameters 1
PASSED  ✓ src/util.py

$ avouch
All clean.

file:line + rule + message, dimmed context, caret under name, BY RULE tally, capped PASSED grid. Duplicate (component, rule) collapses per file.

Exit codes: 0 clean, 1 findings, 2 error. Colors only on TTY; piped is plain.

Usage

Avouch takes no path argument - review set is Git-defined.

avouch --help        # all flags
avouch --docs        # full docs, no review (TTY browser or piped)
avouch --version     # avouch 0.3.4
avouch --watch       # watch and re-run on change

Review scopes (mutually exclusive)

Default - Changed vs HEAD + untracked. Best before git push.

Findings are limited to functions and classes whose source span overlaps an added or changed line. File-level findings still apply to the changed file. --all-files and --not-git review complete files.

--list-changed - Print each changed file path, one per line, and exit. Useful for scripts and tooling.

avouch --list-changed
# src/app.py
# tests/test_review.py

--display FILE - Display a file with syntax highlighting, line numbers, and an interactive pager. Press q to quit.

avouch --display src/avouch/cli.py

--changed - Diff view with +/- hunks and findings pinned inline. Local PR sketch.

avouch --changed
# [CHANGED FILES]
#   bad.py  +2 -1
#   ────────────────────
#     + def f(a,b,c,d,e,f):
#       ^ SCR014

--staged - Only staged (git add). Pre-commit uses this.

avouch --staged

--all-files - Every eligible .py via git ls-files. Use in CI where checkout is clean.

avouch --all-files
avouch --all-files --json

--not-git - Walk CWD for .py without Git. Skips .venv, __pycache__, dist, build, node_modules.

avouch --not-git
avouch --not-git --format github

Without a repo or on clean checkout:

error: no Git repository found
hint: use --not-git
error: nothing to review
hint: nothing changed vs HEAD; use --all-files

Output formats (mutually exclusive)

Default - Human report.

--json - Deterministic {"version":1,"tool":"avouch","violations":[...],"summary":{...}} with rule/severity/message/file/name/kind/line (null for file-level). Same exit codes, gate CI.

avouch --json | jq .summary.total
avouch --all-files --json > report.json

--format github - GitHub workflow commands for inline annotations.

avouch --all-files --format github
# ::warning file=bad.py,line=1,col=5,title=SCR014::Too many parameters...
# ::error for ERROR severity

Use in Actions to annotate PR diff.

--format sarif - SARIF 2.1.0 for code-scanning. Includes tool.driver.rules (20 rules) and physicalLocation spans.

avouch --all-files --format sarif > results.sarif
# upload with github/codeql-action/upload-sarif

--json/--format bypass --quiet and respect baseline filtering.

Diagnostics

--quiet - Only exit code (errors still emit, --json/--format still emit).

--verbose - Step-by-step to stderr: config: <resolved-path>, review set:, analyzing, suppressed, findings:.

--fix - Apply safe fixes before reviewing. Currently replaces bare except: clauses with except Exception: and converts mutable literal/constructor defaults to None sentinel initialization. Combine with any review scope, including --not-git.

--watch - Watch Python files and re-run on change (polling, Ctrl+C to quit). Polls 0.5s (AVOUCH_WATCH_INTERVAL=0.5), snapshots mtime+size of reviewable files + avouch.toml + .avouch/baseline.json; clears screen on TTY and prints ⟳ HH:MM:SS — change detected: path. Works with --all-files / --not-git / --staged / --changed; incompatible with --json / --format / --list-changed / --display. Premium TUI uses full-width block header (┌─ avouch 0.3.4 ─┐).

--ignore-path - Repeatable, component-wise (tests skips tests/ not tests.py). Combined with ignore_paths in avouch.toml.

--select RULES - Review only the comma-separated rule IDs. Repeatable; --ignore RULES is applied afterward.

--ignore RULES - Skip the comma-separated rule IDs for this run without changing avouch.toml.

--docs - Full docs (avouch --docs) - workflow, 20 rules, limits, examples. Works outside repo.

Init & Baseline

avouch init

Measure repo maxima and write avouch.toml with measured+1 headroom so first run is All clean. Rerun recomputes.

avouch init
avouch init --dry-run   # preview without writing

Example generated avouch.toml:

ignore_paths = ["tests"]
[limits]
max_parameters = 6
max_file_lines = 450

avouch baseline

Snapshot current findings to .avouch/baseline.json (rule+file+name+line fingerprint; moving a function re-flags). Commit it. Next runs show only new findings.

avouch baseline          # snapshot full review
avouch                   # only new
avouch --no-baseline     # show all
avouch baseline          # idempotent, recomputes

--verbose shows suppressed N finding(s) by baseline, BY RULE shows (+N suppressed). Malformed/wrong version → error: invalid baseline: exit 2. No file → no suppression. Composes with init.

avouch rule [ID]

Per-rule help from single registry (same source as --docs).

avouch rule              # list all: SCR001  async function without await ...
avouch rule SCR002       # show one: name, description, scope, config, Bad/Good, severity
avouch rule SCR014       # too many parameters

Bare avouch rule lists 20 IDs; unknown → error: unknown rule 'FOO' exit 2; works outside repo.

Configuration

avouch.toml is discovered by walking upward from CWD to filesystem root (so tests/ uses repo root), optional/partial, merged over defaults. Malformed/invalid → error: invalid avouch.toml configuration exit 2.

[limits]        # thresholds per rule
[rules]         # on/off per rule (bool, default true)
ignore_paths = ["tests", "migrations"]

Config file

• avouch.toml in upward-found directory, plain TOML; only AVOUCH_FONT env for terminal font.
• Missing/empty → silent defaults, no warning.

Changing a threshold

[limits]
max_parameters = 8    # allow 8 instead of 5
max_file_lines = 2500

Only named keys change.

Disabling a rule

[rules]
nested_function = false   # stop SCR015

One-line [rules] is valid.

Rule toggles (all true)

Limits (all tunable)

Missing limit falls back to hardcoded default - partial never disables.

Ignoring paths

Component-wise: tests skips tests/ but not tests.py; "." skips repo.

• avouch --ignore-path PATH (repeatable) or ignore_paths = ["tests"] in avouch.toml (must be list).

Combined, de-duplicated, string-based (is_ignored.py), no FS access.

Verifying & errors

avouch --verbose shows config: <resolved-path>, N ignore path(s) or defaults (no avouch.toml).

• Malformed/invalid: error: invalid avouch.toml configuration: limits.max_parameters must be a positive integer; got 'eight' exit 2.
• Unknown keys silently ignored.
• --ignore-path appends; no flag for [limits]/[rules]; applies to all modes; severity fixed WARNING/ERROR.

Example (this repo)

ignore_paths = ["tests"]
[limits]
max_parameters = 5
max_nesting = 5
max_function_lines = 300
max_class_lines = 200
max_file_lines = 1000
max_complexity = 40
max_boolean_conditions = 5
max_if_chain = 5
max_local_variables = 30
max_return_statements = 6
max_lambda_nodes = 10
max_large_comprehensions = 40
[rules]
max_parameters = true
# ... all true

GitHub Actions

Use --all-files in CI (fresh checkout has no diff):

name: Avouch
on: [pull_request, push]
jobs:
  avouch:
    runs-on: ubuntu-latest
    permissions: {contents: read}
    steps:
      - uses: actions/checkout@v6
      - uses: actions/setup-python@v5
        with: {python-version: "3.12"}
      - run: pip install avouch
      - run: avouch --all-files --format github  # annotations
      # or SARIF:
      # - run: avouch --all-files --format sarif > results.sarif
      # - uses: github/codeql-action/upload-sarif@v3
      #   with: {sarif_file: results.sarif}

checkout provides PR code; avouch --all-files --format github annotates diff; permissions: contents: read is enough. Pin avouch==0.3.4.

Why --all-files? Default is changed vs HEAD - fresh checkout is empty (nothing to review). --changed/--staged only make sense locally.

Exit codes 0 clean, 1 findings, 2 error - Actions fails non-zero, so --all-files --format github fails on findings. Fix or ignore_paths.

This repo ships .github/workflows/avouch.yml - enable in Actions, it does pip install -e . and runs --all-files --format github on PR.

Pre-commit

Add to pre-commit so every commit reviews staged changes:

repos: [{repo: https://github.com/mukundzha/avouch, rev: v0.3.2, hooks: [{id: avouch}]}]
# .pre-commit-hooks.yaml: entry avouch --staged, pass_filenames: false

Baselined findings don't fail the hook.

Baseline

avouch baseline        # snapshot to .avouch/baseline.json
avouch                 # only new findings
avouch --no-baseline   # show all

Fingerprint rule+file+name+line (moving re-flags). Commit .avouch/baseline.json. --verbose shows suppressed count, BY RULE shows (+N suppressed). Malformed → exit 2.

Other CI

Any CI: pip install avouch → avouch --all-files --json → check exit code.

Configuration

avouch.toml discovered by walking upward from CWD to FS root (so tests/ uses root config), optional/partial, merged over defaults. Invalid limits (must be positive int), rules (bool), ignore_paths (list[str]) → error: invalid avouch.toml configuration exit 2.

[limits]
max_parameters = 8
[rules]
nested_function = false
ignore_paths = ["tests"]

Limits and toggles are all true by default; see avouch --docs for full table. --ignore-path appends to ignore_paths. No env vars.

How it works

cli.py only orchestrates. Pipeline: load_config → git review set → analyzer.analyze_file (ast + walk cache + rules) → baseline.filter → report (or --json/--changed/--quiet). --docs/--version short-circuit.

See avouch --docs for mermaid diagrams and full pipeline.

Layout

avouch.toml  .avouch/baseline.json
src/avouch/
  cli.py              # entry, only orchestrates
  baseline.py         # snapshot + suppression
  git.py              # review set
  analyzer.py         # ast walk, dispatch
  report.py           # terminal / json / github / sarif / diff
  rules/              # one analyze(node,limits) per rule
    complexity.py  max_nesting.py  max_parameters.py ...
  utility/
    walk.py  docs.py  is_generated.py  is_ignored.py  measure.py
  config/
    default.py  loader.py  # DEFAULT_LIMITS, load_config
tests/
  test_git.py         # 74 tests + real git repo
  test_init.py        # init/baseline

Adding a rule

One file per rule with analyze(node, limits) -> list[issue]:

{"rule": "SCR017", "severity": "WARNING",
 "message": "Description (value/limit). Remediation."}

Add toggle in DEFAULT_RULES (+limit in DEFAULT_LIMITS if needed), wire in analyzer.py with guard, add violation + boundary tests. Renderer handles any (severity,message).

Example: mutable_default_args - flags def f(x=[]), suggests None.

Testing

pip install -e . && python -m pytest

86 tests, <1s, no network. Covers git helpers, config validation (upward, limits must be positive int, rules must be boolean, ignore_paths list), nesting/complexity/boolean metrics, rule boundaries, unreadable/syntax errors, report/diff, avouch rule, github/sarif golden, --docs outside repo, real temp git repo. Only subprocess.run mocked.

FAQ

Changed files only? Whole-repo buries your few findings under legacy noise; diff keeps it relevant to next push.

Why git diff HEAD not git diff? Plain diff is unstaged only; HEAD is staged+unstaged plus untracked - nothing missed.

Why AST not regex? Regex can't count cross-line parens, nesting, or def vs call; AST answers exactly.

Exit codes? 0 clean, 1 findings, 2 error. Reviews, doesn't gate - CI can still react.

Network/daemon? No - three git calls + stdlib; runtime ∝ diff, not repo.

Baseline vs ignore_paths? ignore_paths hides files forever; baseline hides findings at snapshot time but re-flags if moved - use baseline for legacy debt.

init vs baseline? init sets limits to measured+1 so first run is All clean; baseline snapshots findings. Compose both for greenfield.

rule vs --docs? Same registry; --docs shows all, avouch rule SCR002 shows one.