reprove

Verify a pull request with evidence, not opinions.

reprove runs on a repository and a pull request and answers one question: what evidence exists that this change is safe to release, and what evidence is missing? It is built for the case where the change was written quickly, often by an AI tool, and nobody has checked whether the tests actually bite.

Every finding carries a status:

reprove never claims more than it ran.

What it does

1. Diff. git diff base...head. Each changed file is classified as source, test or config. Changed source lines are the unit everything else measures against.
2. Tests. Detects and runs the project's test command (pytest if pyproject.toml/pytest.ini/tests/ exist; npm test, vitest or jest from package.json). Captures pass/fail counts, duration and the tail of the output, with a configurable timeout.
3. Diff coverage. Python: the suite runs once under coverage with per-test contexts, and executed lines are mapped onto the changed lines: changed lines executed by tests: X of Y, listing the uncovered ones. JS/TS: c8 or nyc if the project already has one, otherwise needs-evidence with the exact command that would produce it.
4. Mutation of changed lines only (Python). One AST mutator at a time on the changed lines: comparison swap (</<=, ==/!=, >/>=), arithmetic swap (+/-, *//), numeric boundary (n to n+1, n-1), boolean flip, negation removal, and/or. Each mutant is written into a temp copy of the repository and the tests that cover that line are run (the full suite when coverage is unavailable). Result: killed or survived, with the exact mutation and line. A surviving mutant on a changed line is the headline finding: the tests do not detect that change. Budgets: --max-mutants (default 50) and --time-budget seconds.
5. Hypotheses (optional, bring your own key). Only if ANTHROPIC_API_KEY is set and --no-llm is not passed: the diff, test summary and surviving mutants go to the model (default claude-sonnet-5, --model to change), which proposes a few specific risks, each with a pytest reproduction. reprove writes each reproduction into a temp copy and runs it. A hypothesis is shown only with its run result: reproduced (the test failed as predicted), not-reproduced, or could-not-run. Without a key this stage is skipped and the report says so in one line.
6. Report. Markdown sized for a PR comment, and JSON. A verdict line (release-safe evidence: strong | partial | weak with the specific gaps), findings ordered by severity (file:line, what was checked, the command run, an output excerpt, status), then the coverage table, the mutant table, the hypotheses table, and what reprove could not check here.

Honesty rules

• reprove proves what it ran. It does not prove absence of bugs.
• Coverage of a line is not correctness of a line; a killed mutant is one specific change the tests noticed.
• Surviving mutants are the point: they are places where a change could ship unnoticed.
• The hypothesis stage is the only part that uses a model, it is off by default without a key, and every hypothesis is executed before it is shown.

Install

Python 3.10 or newer. Install reprove into the same environment as your project's test dependencies (it runs python -m pytest with its own interpreter).

pip install reprove              # core: coverage only
pip install "reprove[llm]"       # + anthropic, for the optional hypothesis stage

From a checkout:

git clone https://github.com/ping-dev-ui/reprove
cd reprove
python -m venv .venv
# Windows: .venv\Scripts\activate    Linux/macOS: source .venv/bin/activate
pip install -e ".[llm,test]"

Usage

reprove verify [--base REF] [--head REF] [--format md|json|both] [--out PATH]
               [--max-mutants N] [--time-budget S] [--test-timeout S]
               [--no-llm] [--model NAME] [--max-hypotheses N]
               [--fail-on weak|partial|never] [--repo PATH]

Defaults: base is origin/main if it exists, else main; head is HEAD; --fail-on never; Markdown to stdout.

# the last commit, no model
reprove verify --base HEAD~1 --head HEAD --no-llm

# a branch against main, both formats, fail the job on a weak verdict
reprove verify --base origin/main --format both --out reprove-report --fail-on weak

Exit codes: 0 ok, 1 the verdict hit --fail-on, 2 reprove could not run (not a git repository, unknown ref).

Tests, coverage and mutation run against the working tree. If the working tree is not at --head, or has uncommitted changes, the report says so under what reprove could not check here.

GitHub Action

name: reprove
on:
  pull_request:

permissions:
  contents: read
  pull-requests: write   # to create/update the sticky comment

jobs:
  verify:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          ref: ${{ github.event.pull_request.head.sha }}   # run against the PR head, not the merge commit
          fetch-depth: 0
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - run: pip install -e ".[test]"        # your project's test dependencies
      - uses: ping-dev-ui/reprove@v0.1.1
        with:
          max-mutants: "50"
          time-budget: "300"
          fail-on: never                     # weak | partial | never
          no-llm: "false"
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}   # optional (BYOK)

The action installs reprove, runs verify with base ${{ github.event.pull_request.base.sha }} and head ${{ github.event.pull_request.head.sha }}, writes reprove-report.md/.json, uploads them as an artifact, and creates or updates one sticky PR comment (marker <!-- reprove -->) using the gh CLI with GITHUB_TOKEN. Inputs: max-mutants, time-budget, fail-on, no-llm.

Bring your own key. The hypothesis stage runs only if you add your own ANTHROPIC_API_KEY secret. reprove reads the key from the environment only; it never writes, logs or stores it, and never sends anything anywhere else. Without the secret the stage is skipped and the report says so.

What it proves, and what it cannot

Proves (status reproduced):

• the test suite passes or fails on this working tree, with counts and the command;
• which changed lines the tests execute, and which they never touch;
• for each mutant on a changed line, whether the covering tests noticed it;
• for each model hypothesis, whether its reproduction actually fails.

Cannot prove (and says so):

• absence of bugs, or correctness of any line: only that specific changes were or were not noticed;
• anything the tests do not exercise, including code paths reachable only in production;
• whether a test-only change made the tests weaker: a diff with no changed source lines gets at most a partial verdict;
• mutation and hypothesis execution for JS/TS in v1 (detected, reported as needs-evidence);
• a mutant survives trivially when no test executes its line: the report labels these no test executes this line;
• a file that the tests import from somewhere other than the working tree (for example an installed copy of a src/ package). reprove checks this with an import sentinel: it breaks the file in the temp copy first, and if the covering tests still pass, that file's mutants are reported as needs-evidence instead of survived.

Model-written reproductions are executed like any other test in your project, inside a temp copy: only enable the hypothesis stage on repositories whose tests you would run anyway.

Verdict

• strong: tests pass, every executable changed line is executed, every mutant run was killed, nothing skipped, no reproduced hypothesis.
• partial: tests pass but something is missing: uncovered changed lines, a few survivors, mutants skipped by a budget, coverage unavailable, or a test-only change.
• weak: no test command, failing or timed-out tests, fewer than half the changed lines executed, at least half the mutants surviving, or a reproduced hypothesis.

What it is for, and what it is not

reprove answers did the tests notice this diff. It does not answer is this diff safe. Those are different questions, and the second one still needs a person with judgement: whether a part should be thrown away, whether a finding matters to the business, whether to ship. reprove does the reviewer's first hours, the part where someone has to establish what the tests actually exercise and whether they would fail, and hands over evidence instead of opinions. If you are hiring an independent reviewer for AI-generated, vibe-coded or fast-built code before launch, this is what you would want them to run first.

It runs offline by default. Tests, coverage and mutation never leave your machine or your CI. The only stage that sends anything anywhere is the optional hypothesis step, and it sends the diff to a model under your own key, only when you set that key. If your code cannot go to a third-party model, do not set the key: everything else still runs.

Passing tests can lie. The case in docs/demo-weak-vs-strong.md has 100% line coverage of the change and a passing test, and the tests cannot tell > from >= at the boundary that the change is about. That is the regression reprove is built to surface before it reaches production.

Pricing

• Public repositories: free. MIT licence, no limits, no account.
• Private repositories: reprove Pro, $19 per month per repository. Same code, same licence. Pro is honour-based in this release: there is no key to paste and nothing phones home. It pays for the maintenance, gets you email support, and puts your repository first in line for the roadmap below. Subscribe at the link in the repository description once it is live, or open an issue and ask.

That number is set from what people actually pay for a single human review of one change on freelance marketplaces: a few hours at a reviewer's rate is more than a year of Pro.

Roadmap

In the order buyers of independent reviews ask for them:

1. Authorization and tenant isolation checks. Cross-tenant reads and writes attempted per route, and whether isolation is enforced in the database or only in each query.
2. Payment and webhook reproductions. Duplicate delivery, retry after timeout, and idempotency of anything that moves money, executed against the project's own handlers.
3. Dependency and licence audit on the changed manifest, with the exact advisory and version.
4. Mutation and hypothesis execution for JS/TS, matching the Python path.
5. A decline verdict: when the evidence says do not release, say so in one word at the top.

Findings with reproduced status will always be executed checks with logs. That rule does not change as the list grows.

The gaps are always listed next to the verdict.

Developing

pip install -e ".[llm,test]"
python -m pytest -q          # ~40 s; builds real git repos under tmp_path
ruff check .
reprove verify --base HEAD~1 --head HEAD --no-llm   # run reprove on itself

Tests never call the network; the model is mocked. See docs/demo-self.md for reprove run on its own repository.

License

MIT, copyright 2026 Miguel Jardim.