impact-gate

A merge gate that flags changes piling complexity onto code that is already complex, before a class quietly grows into a god class nobody can safely touch. Run it as a standalone CLI, a git pre-commit hook, or a plugin in GitHub, GitLab, and Jenkins CI.

Website: https://impactgate.officefloor.net

The pattern it catches is gradual. A class gains one more method, then another, then another. Each change looks reasonable on its own. But over dozens of them the class ends up doing five different jobs, and every edit gets riskier. That slow accretion is what we call structural decay (also considered cohesive erosion). The gate scores each change against a base (main by default), so it shows up while it is still cheap to fix:

impact = files_changed * Σ max(WMC_other, 1) * CC * Δlines      (over changed functions)

WMC_other is the complexity that was already in the file or class you are editing, measured before your change. Adding a brand new file is cheap. There was nothing there to make worse. Adding a complex method to an already heavy class is expensive. The gate measures that difference, not the raw size of the diff.

For the reasoning behind the formula, see Measuring the Blast Radius of Change on the OfficeFloor blog.

When impact is too high, the gate asks you to simplify the change or refactor the code it touches. It can warn (report only) or block (fail the build).

How this differs from SonarQube (and other static analysis)

SonarQube and linters score the state of the code. This file is too complex, this method too long, this block duplicated.

ImpactGate scores the marginal cost of a change, weighted by what it lands on. The WMC_other term in the formula is the complexity already in the class before you touched it. A complex method in a brand new file is cheap. The same method added to a class already carrying five responsibilities is expensive. It's the increment that turns a heavy class into a god class nobody can safely edit.

Languages

The gate is language agnostic. It scores changes in any of these languages:

Java, C#, C, C++, JavaScript, TypeScript, Python, Go, Kotlin, Swift, Ruby, PHP, Rust, Scala, Objective-C, Lua, and TTCN-3.

Complexity is parsed per function by lizard. A mixed language repo is scored the same way throughout.

You can add or remap extensions in the measure config (lang_by_ext).

The grading curve is calibrated per language for Python, Java, TypeScript, C#, JavaScript, C, Scala, and Go. The other languages are still measured and graded. They fall back to a pooled cross language distribution that holds until your project baseline builds up its own history.

Install

pip install impact-gate         # installs the `impact-gate` command

Or run it without installing anything, via the published image:

docker run --rm -v "$PWD:/repo" ghcr.io/officefloor/impact-gate \
  score --mode range --base origin/main

To hack on it locally, install from a checkout instead:

python -m venv .venv && . .venv/bin/activate
pip install -e '.[dev]'         # editable install plus the test deps

Use

# The commit you are about to make (pre-commit): staged vs HEAD. This is the default.
impact-gate score

# Uncommitted local edits: working tree vs HEAD.
impact-gate score --mode worktree

# CI or PR review: the committed branch vs main (merge-base..HEAD).
impact-gate score --mode range --base origin/main --format json

# Set thresholds and enforcement. You can also put these in .impact-gate.yml.
impact-gate score --warn-at 50000 --block-at 200000 --enforcement block

Exit codes. 0 means ok or warn (the change is allowed). 2 means blocked (impact too high under --enforcement block). 1 means a usage or environment error.

Every report also lists the files to consider for refactoring, ranked by their share of the impact. The change level number gates. The per file ranking points at where the decay is concentrating, so a file quietly growing into a god class surfaces as a candidate before it blocks anything.

A source file whose diff is larger than max_diff_lines (200,000 by default, in the measure config) is almost always a generated dump or a vendored blob. The gate skips it so it neither distorts the number nor slows scoring, and lists it under skipped so the result is never silently wrong.

Use as a git pre-commit hook

Gate every commit locally, before CI:

# Installs .git/hooks/pre-commit. It scores the staged change on each commit.
impact-gate install-hook

With enforcement: block in .impact-gate.yml, a commit whose impact is too high is blocked. On warn (or off) the report prints and the commit proceeds. Rerun with --force to overwrite an existing pre-commit hook.

Prefer the pre-commit framework? This repo ships a hook definition — add to your .pre-commit-config.yaml:

repos:
  - repo: https://github.com/officefloor/ImpactGate
    rev: v0.3.0
    hooks:
      - id: impact-gate

Grade against a distribution (the curve)

A raw threshold is hard to set. A typical change's impact varies by orders of magnitude across languages and projects. Instead of guessing a number, grade a change by its percentile against a distribution, and gate on the percentile.

# Build (or refresh) the project's own impact distribution from the merged history.
# Writes .impact-gate-baseline.json. Re-run it as the branch moves.
impact-gate baseline --base-ref main

# Gate on the grade instead of an absolute number.
impact-gate score --curve --warn-percentile 90 --block-percentile 98

The grade blends two distributions:

• a seed prior shipped with the tool. Per language percentile tables built from 20 open source repositories, with a pooled fallback for languages not in the table.
• the project baseline using the repo's own per change distribution, walked from the merged mainline.

The blend weights the project by w = n / (n + K), where n is the number of landed changes behind the baseline and K (curve_prior_weight, default 200). It is how much history it takes to trust the project over the seed. A fresh repo with no baseline file grades on the seed alone. A deep history leans on itself. The grade shows in every format next to the raw number.

Gate on cognitive complexity (deeply-nested methods)

Change-impact and the curve both measure how a change moves. It cannot see a method that is simply hard to read (working code with deep decision nesting). That is the shape AI code generators typically produce.

--cognitive-max N turns on an absolute per-method gate on Cognitive Complexity (Campbell 2018). A change blocks when any method in a changed file exceeds N. It is independent of the impact number. A small diff can pass on impact yet block here for leaving a tangled method behind.

# Block any change that leaves a method with cognitive complexity over 15 (SonarSource's line).
impact-gate score --enforcement block --cognitive-max 15

Configure with .impact-gate.yml (repo root)

warn_at: 50000          # impact above which to warn
block_at: 200000        # impact above which to block
enforcement: warn       # off, warn, or block. Start on warn. Flip to block when ready.
tolerance: 1.0          # CI-adjustable multiplier on both thresholds. Above 1 is more lenient.
# measure_config: .impact-measure.yml   # optional: ignore globs and language overrides
cognitive_max: 15       # block a method over this Cognitive Complexity (null/omitted = off)

# Grading curve (percentile gate). When enabled, warn_at/block_at are ignored and the
# gate uses the percentiles below instead.
curve_enabled: false           # gate on the percentile grade instead of absolute numbers
warn_percentile: 90            # grade at or above this warns
block_percentile: 98           # grade at or above this blocks
curve_prior_weight: 200        # K in w = n/(n+K): history needed to trust the project over the seed
baseline_file: .impact-gate-baseline.json   # where `impact-gate baseline` caches the distribution

The cognitive gate is independent of the curve: it can run alongside either the absolute or the percentile impact gate, or on its own.

CLI flags override the file. A CI job can pass --tolerance or --warn-at. So a team can dial tolerance without editing the repo. The curve dials have flags too: --curve, --warn-percentile, --block-percentile, --baseline-file.

Use in GitHub Actions

Add a workflow to your repo. The action scores the PR branch against its base and writes a summary. fetch-depth: 0 is required so the base branch and merge base are present.

name: Change impact
on: pull_request
permissions:
  contents: read
  pull-requests: write         # so the action can post the score as a PR comment
jobs:
  impact:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5
        with:
          fetch-depth: 0
      - uses: officefloor/ImpactGate@v0
        with:
          enforcement: warn        # switch to block when ready
          # warn-at: 50000
          # block-at: 200000
          # tolerance: 1.0

The score appears in the job summary and as a sticky comment on the PR (one comment, updated each run). In block mode the job fails when impact exceeds the block threshold. Make the check required in branch protection to gate merges. Adding the PR comment needs pull-requests: write. Without it the run still passes and just skips the comment.

Use in GitLab CI

A ready made job is in ci/gitlab-ci.yml. Copy it into your .gitlab-ci.yml, or include it remotely:

include:
  - remote: 'https://raw.githubusercontent.com/officefloor/ImpactGate/v0/ci/gitlab-ci.yml'

It runs on merge request pipelines, scores the MR against its base ($CI_MERGE_REQUEST_DIFF_BASE_SHA) with the published Docker image, and posts a sticky note to the MR (one note, updated each run). Without the required GITLAB_TOKEN token it still scores and gates. It just skips the note. In block enforcement the job fails when impact is too high. Make it required in the merge request settings to gate merges.

Use in Jenkins

A pipeline snippet is in ci/Jenkinsfile. It runs the Docker image on an agent with Docker, scoring the change against its target branch (origin/${CHANGE_TARGET:-main}) and archiving the report. In block enforcement the stage fails when impact is too high. Posting the score back to the PR/MR is left to your SCM integration. To post it with the tool itself, run impact-gate comment in the container with the provider's token and env set.