Repository navigation
Releases
OpenFix 0.4.0: on PyPI as openfix-cli
OpenFix is now on PyPI as openfix-cli:
pip install "openfix-cli[all]" # or: uv tool install "openfix-cli[all]"The command is still openfix, and the import is still import openfix.
New
- Several scanners in one run.
verify.scannercan map each tool to its own rescan command, so oneopenfix fixcan take a Snyk Code report and a Semgrep report and rescan each finding with the right scanner. See docs/policy.md. - Clearer terminal output. One line per finding (scanner, severity, what, where), then its checks and the outcome, plus the draft PR link. Colour is used in a terminal and turned off with
NO_COLOR. - Published from GitHub releases with PyPI trusted publishing. No API token is stored anywhere.
Fixed
--open-prsreruns now reuse the PR branch instead of failing with "branch already exists".
Docs
- New README diagrams: how a fix is proved, what a receipt holds, and the scheduled GitHub Action loop.
- A rebranded, dark-only README, with a terminal demo that fixes a real Snyk Code finding and a real Semgrep finding and opens a draft PR.
Full changelog: v0.3.0...v0.4.0
OpenFix 0.3.0: verified security fixes from any model
OpenFix 0.3.0 is a ground-up rewrite. It turns findings from the scanners you already run into fixes from the model you choose. A fix only counts after OpenFix checks it independently of the model.
Models propose. Code verifies.
Highlights
-
Any scanner in. SARIF 2.1.0, plus native JSON from Semgrep, Snyk, Trivy, OSV-Scanner and Gitleaks, plus GitHub code-scanning and Dependabot alerts. Formats are detected automatically, duplicates are merged and secret values are redacted.
-
Policy as code.
openfix.yamldecidesautofix,proposeorskipfor each finding. It also sets guards (protected paths, diff-size limits), budgets (turns, tokens, USD) and paths that may only be sent to local models. -
Any model. Anthropic, OpenAI, Gemini and xAI Grok; OpenRouter, Groq, DeepSeek, Mistral, Together and Fireworks; any OpenAI-compatible gateway; and local models through Ollama, vLLM or LM Studio. Choose one with
--model provider:model. -
A small blast radius. The agent works in a throwaway git worktree with a few policy-checked tools: read, search, edit, run tests and rescan. It has no shell, network or git access. Tests run in a Docker container with networking turned off.
-
Independent verification:
- Rescan: the finding is gone and nothing new appeared.
- Tests: no worse than on the original code.
- Regression test: fails on the original code and passes with the fix.
- Optional review by a second model.
Findings that a fresh scan can't reproduce are dropped before any model is called.
-
Deterministic dependency bumps. Vulnerable direct npm and pip dependencies are bumped to the lowest fixed version, with no model involved.
-
Evidence you can re-check. Each fix gets an in-toto receipt bound to the SHA-256 of its patch, with optional Sigstore signing.
openfix receipt verifyre-checks it, andopenfix.sarifcarriesfixes[]for your other tools. -
Fits your pipeline. Use the CLI, or the GitHub Action:
uses: markusweldon/openfix@v0.3.0. Verified autofixes become draft PRs with an evidence table. OpenFix never merges anything.
Install
pip install "openfix[all] @ git+https://github.com/markusweldon/openfix@v0.3.0"Try it without an API key using the bundled vulnerable app; see the README.
Status
0.3.0 is alpha. The full pipeline is tested end to end with real scanners and test suites, and CI covers Python 3.10, 3.12 and 3.13. Not yet tested in CI: calls to live model APIs, the Docker sandbox, and opening real PRs. The 0.2 prototype is kept at the v0.2.0 tag.