TestSeal

Deterministic test-integrity checks for Python and pytest diffs.

TestSeal compares tests before and after a change and reports concrete weakening signals: removed assertions, newly disabled tests, weaker comparisons, wider tolerances, swallowed exceptions, snapshot regeneration, and suspicious mocks. It runs locally, needs no model or API key, and never imports or executes the repository it scans.

- assert total == Decimal("19.99")
+ assert total

[HIGH] TS003 tests/test_checkout.py:42:5 - Assertion weakened
  A precise equality assertion became a truthiness assertion.
  Fingerprint: 6c48d147cbe59f553e224d7d

Why TestSeal

A green test suite is weak evidence when the same change made the suite easier to pass. Ordinary linters accept both sides of the example because both are valid Python. General code reviewers may notice the downgrade, but their output is probabilistic. TestSeal provides a narrow, reproducible signal dedicated to how the tests themselves changed.

• Deterministic: identical input and configuration produce identical output.
• Offline: zero runtime dependencies, model calls, telemetry, or accounts.
• Diff-aware: compares complete before/after syntax when Git blobs are available.
• Safe by design: reads source and Git data without importing the target project.
• Advisory by default: teams choose when findings should block a workflow.
• Portable: text, versioned JSON, SARIF 2.1.0, pre-commit, and GitHub Actions.

Install

TestSeal requires Python 3.11 or newer and Git 2.x for repository-backed scans:

python -m pip install testseal

For an isolated CLI installation:

uv tool install testseal
# or: pipx install testseal

Use the CLI

Scan tracked changes plus non-ignored untracked files in the working tree:

testseal scan

Scan a branch relative to its base:

testseal scan --base origin/main --head HEAD

Scan only staged changes:

testseal scan --staged

TestSeal remains advisory unless a threshold is configured:

testseal scan --fail-on high

Exit codes are 0 for a completed advisory scan, 1 when the selected finding threshold is met, and 2 for invalid configuration or an incomplete blocking scan.

Output

Text is the default. JSON and SARIF can be printed or written atomically:

testseal scan --base origin/main --format json --output testseal-report.json
testseal scan --base origin/main --format sarif --output testseal-report.sarif

Run testseal scan --help for the complete CLI contract.

Configure policy

TestSeal discovers testseal.toml first, then [tool.testseal] in pyproject.toml. An explicit --config PATH takes precedence.

[testseal]
fail_on = "high"
test_patterns = ["test_*.py", "*_test.py", "tests/**/*.py"]
source_roots = ["src"]
disabled_rules = ["TS008"]

# Copy a fingerprint from text, JSON, SARIF, or the Action output after review.
ignore_fingerprints = ["6c48d147cbe59f553e224d7d"]

[testseal.rules.TS006]
severity = "low"

Configuration is strict: unknown keys, rule IDs, severities, and malformed fingerprints fail with exit code 2 instead of silently weakening policy. See the repository's default, strict, and monorepo examples.

Pre-commit

The hook installs TestSeal in its own environment and scans the staged diff:

repos:
  - repo: https://github.com/satwiksps/testseal
    rev: v0.1.0
    hooks:
      - id: testseal
        args: ["--fail-on", "high"] # omit to remain advisory

GitHub Actions

The Action installs the Python core bundled in the same release, derives pull request refs from the event payload, annotates changed lines, and exposes a normalized JSON result.

name: Test integrity

on: [pull_request]

permissions:
  contents: read

jobs:
  testseal:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          fetch-depth: 0
      - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
        with:
          python-version: "3.12"
      - id: testseal
        uses: satwiksps/testseal@v0.1.0
        with:
          fail-on: high

Omit fail-on to honor repository configuration. The default install: true uses only the source bundled with the pinned Action release. Set install: false only when the same TestSeal version is already installed in the selected Python environment.

Action outputs include finding-count, severity counts, files-scanned, suppressed-count, outcome, and the complete normalized result JSON.

Rules

The table is a summary. The rule reference defines supported syntax, intentional non-findings, and precision limits. Context-heavy rules should be baselined before enabling a blocking threshold.

Trust boundary and limitations

TestSeal invokes Git to obtain refs, diffs, and blobs, then parses Python source with the running interpreter. It does not run tests, import changed modules, or execute hooks from the target repository. Use pull_request, read-only permissions, and no repository secrets when scanning contributions from forks.

The analyzer reports specific transformations; it does not decide whether an author is honest, prove that tests are complete, or replace code review, coverage, linters, type checkers, and security analysis. A finding can describe a legitimate refactor, so blocking is explicit and reviewed exceptions use stable fingerprints rather than hidden heuristics.

Read the full architecture and trust model and security policy.

Development

git clone https://github.com/satwiksps/testseal.git
cd testseal
python -m venv .venv
# POSIX: source .venv/bin/activate
# Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install -e ".[dev]"
python -m pytest

The TypeScript Action lives in packages/action and the Next.js/Tailwind site in site. See CONTRIBUTING.md for the complete verification commands and rule-change requirements.

License

TestSeal is licensed under the Apache License 2.0.