Repository navigation
Releases
v1.3.0
api-vv-toolkit v1.3.0
OpenAPI contract plus written requirements in; API tests, a requirement-to-verdict
traceability matrix and a CI gate out. This release adds OpenAPI 3.1, seeded negative and
property-based generation, contract drift checks, a reusable GitHub Action and a rebuilt HTML
report.
Highlights
- Use it from any workflow:
uses: muhib-karim/api-vv-toolkit@v1validates, generates, runs
and reports, and gates on requirement verdicts. Withbaseline-spec, it also fails on
breaking contract drift. Outputs:gate,failed-requirements,breaking-changes,
report-dirandjunit. api-vv diff old.yaml new.yaml: classifies every change as breaking or non-breaking from
an existing client's point of view, as text, JSON or Markdown. Exits 1 on any breaking change.
Renamed operationIds, renamed path placeholders and 3.0 → 3.1 migrations aren't reported as
changes. Changes it can't analyse fail closed.- OpenAPI 3.1 alongside 3.0:
typearrays,examplesmaps andwebhooks(listed, never
tested). Fixes a crash on 3.1 type-array parameters and lostnullfor 3.0nullable+
allOf. - Auth from environment variable names only: bearer, basic and apiKey in header, query or
cookie.credentials_envmust name variables; otherwiseAPI_VV_AUTH_<SCHEME>is read. Each
secured operation gets aninvalid credentialsnegative. - Seeded negatives:
patternsupport, and
generate --property-cases N --seed Sfor random valid (expect 2xx) and single-violation
(expect 4xx) requests. The same seed regenerates a byte-identical plan. - Reports: an accessible, mobile-friendly HTML matrix (requirement → tests → result with
"why it failed"), JUnit XML, CSV and Markdown. The Pages site is rebuilt by CI from a live run
against the bundled demo API. - Measured: a live LLM gate re-run on 2026-10-21 with claude-opus-5-5, claude-sonnet-5 and
gpt-6-sol (results/2026-10-21-llm-gate/). Acceptance was 95%, 76% and 95%.
Quality gates
These run in CI on every push: ruff, mypy --strict on the package, pytest with coverage ≥ 85%
on Python 3.11 and 3.12, CodeQL (security-and-quality), gitleaks, Dependabot, a wheel build and
smoke test, the seeded-defect demo, the Action dogfood workflow and a Docker build.
Compatibility
- Plans keep
version: "1.0". New fields (seed,property_cases,Spec.webhooks,
auth: invalid) are optional, so 1.2 plans still load in 1.3. Plans written by 1.3 aren't
readable by 1.2. - Generated plans gain one
invalid credentialscase per secured operation, which changes
untraced-test counts.
Install
pip install https://github.com/muhib-karim/api-vv-toolkit/releases/download/v1.3.0/api_vv_toolkit-1.3.0-py3-none-any.whlVerify the download against SHA256SUMS.txt attached to this release. Full details are in
CHANGELOG.md.
v1.2.0
api-vv-toolkit v1.2.0: LLM gate queries only opted-in requirements; live four-model study
v1.1.0
What's new in 1.1.0
- Fix: execution timeouts and per-requirement latency budgets must be finite.
NaNand infinity are rejected when the config or requirements file is loaded; an infinite value previously passed validation and effectively switched the limit off. New tests cover it. - Live demo report: CI runs the seeded-defect demo (a loopback Starlette API with two deliberate bugs) on every push to
mainand publishes the HTML traceability report, CSV matrix and JUnit XML to GitHub Pages: https://muhib-karim.github.io/api-vv-toolkit/ — the gate fails on exactly REQ-003 and REQ-008, as designed. - Packaging: author and project URLs in
pyproject.toml.
CI: ruff, mypy --strict, pytest on Python 3.11 and 3.12, the demo gate check, and a Docker smoke test.
Full changelog: CHANGELOG.md
v1.0.0
First tagged release.
api-vv-toolkit reads an OpenAPI 3.0/3.1 contract and a YAML file of written requirements, generates API tests that each trace to a requirement, runs them against a live service, and reports a requirements-to-test traceability matrix: which requirements are verified, failed, partially covered or not covered.
Included
- CLI
api-vvwithparse,generate,run,report,traceandvalidate-requirements. - Rule-based generators: happy path, schema and parameter boundaries (length, range, item counts, enums, wrong types, unexpected properties), required-field omission, auth-required checks, and status-code coverage.
- Optional LLM-assisted generator (Chat Completions or Messages API over httpx, keys from the environment). Every proposal must match a JSON schema and is validated against the spec before it enters the plan; the prompt, raw reply and accept/reject reason are stored in the plan. A recorded mock provider is used for tests, CI and the demo.
- A reviewable YAML test plan with stable ids and request-fingerprint de-duplication.
- An async executor with bounded concurrency and timeouts that checks status code, response body schema, documented headers and optional latency budgets.
- Reports in Markdown, self-contained HTML, CSV and JUnit XML, plus an exit-code gate (by default, any must-priority requirement that is failed, partially covered or not covered fails the run).
Demo
make demo starts a small demo API with two deliberately seeded defects, generates 31 cases (rule-based plus mock LLM: 1 proposal accepted, 1 rejected), runs them and writes the reports. Result: 29 passed, 2 failed, and the gate fails on exactly the two seeded defects:
- REQ-003: an off-by-one
maxLengthcheck (expected 422, got 201). - REQ-008: the wrong success status (expected 204, got 200).
Verified for this tag
- CI: https://github.com/ai-dev-2024/api-vv-toolkit/actions/runs/36503537583 (lint, format, mypy --strict and tests on Python 3.11 and 3.12; the demo job asserts the gate names exactly REQ-003 and REQ-008; Docker build and smoke test).
- Locally: 111 tests, 98.51% coverage.
Limitations
Response bodies are read fully into memory, and regex patterns from the spec run as written, so a hostile spec can be slow to validate. Remote $refs are rejected. See the README for the full list.