Codna GitHub Action

Run Codna in GitHub Actions. fix, review or secure a repository with the same codna command you run on your machine.

Codna docs · Contributing · Support · Security

---

Codna maps your repository before it spends a token, then sends an agent in with the exact context it needs. A fix arrives as a pull request with the root cause, the changed symbols and a confidence score. You review. You merge. Your key. Your infra. Your code stays yours.

This is the public wrapper for Codna's GitHub Action channel. The action installs the published codna package from PyPI and runs it. One install: the agent runtime ships inside the package, so the action adds no Node, Bun or second runtime to your job.

Fix mode

permissions:
  contents: write
  pull-requests: write

jobs:
  codna:
    runs-on: ubuntu-latest
    env:
      OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
    steps:
      - uses: thyn-ai/codna-action@v1
        with:
          mode: fix
          issue: "CI failed on checkout tests"
          model: openai/gpt-5
          github-token: ${{ github.token }}

Model calls run through the provider named in model (<provider>/<model-id>) and read that provider's key from the job environment — for example OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY or OPENROUTER_API_KEY; any provider in the bundled catalog works. When model is omitted (or carries no <provider>/ prefix) the default provider is Anthropic, so ANTHROPIC_API_KEY must be set in that case. api-key is the Codna engine API key (exported as CODNA_API_KEY). It is only needed when the run is pointed at a remote Codna engine via CODNA_ENGINE_URL; the default local run does not use it.

Review mode

permissions:
  contents: read
  pull-requests: write
  checks: write

on:
  pull_request:

jobs:
  codna_review:
    runs-on: ubuntu-latest
    env:
      OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: thyn-ai/codna-action@v1
        with:
          mode: review
          post: "true"
          model: openai/gpt-5

On pull_request events the action fetches the PR base branch and reviews origin/<base>...HEAD. For non-PR events, set diff explicitly, for example diff: origin/main...HEAD.

What a review posts:

• One inline comment per finding: severity (high, medium, low), category (correctness, security, performance), an explanation and a suggestion block when the fix is a plain replacement. Findings at confidence 0.75 or higher, ten at most; min-confidence and effort adjust this.
• One check run named codna review: success with no findings, neutral with findings, and failure only with blocking: "true" and a finding at a blocking severity.
• A verdict. Approve when the pass has no medium or high finding and no earlier Codna medium or high thread is unresolved; otherwise Comment, with the reason. Never Request changes. Set review.approve: false in the repository's codna.yaml to turn approvals off.
• A red-head note when a required check is failing at review time: the verdict is about the diff, not a merge go-ahead.
• Dependency claims checked against npm and PyPI before posting. A contradicted claim is dropped; an uncheckable one is posted as low and marked Unverified.
• Re-reviews cover the commits pushed since Codna's last review.

Secure mode

steps:
  - uses: actions/checkout@v4
  - uses: thyn-ai/codna-action@v1
    with:
      mode: secure
      from-sarif: results.sarif
      reachability-engine: local
      verification: codna-security.yaml

Secure mode is read-only in this wrapper: it classifies each SARIF 2.1.0 finding as reachable or not and opens no pull request. Opening security PRs uses Codna's privilege-separated evidence and writer workflow (Security Autofix).

What fix mode opens

mode: fix runs codna fix <repo> --ref <sha> --open-pr. When Codna finds a fix it pushes a codna/… branch and opens a pull request titled codna: fix … whose body states the issue, the root cause, the changed symbols and Codna's confidence, and ends with Review before merging. The step exposes the URL as pull-request-url. Codna opens pull requests. It never merges.

Limits

• The runner needs Python 3.10 or newer and a platform with a published wheel: Linux x86_64 or macOS on Apple silicon. The install is wheel-only, so other platforms fail at the install step.
• mode: fix opens the pull request once. It does not re-run your tests after the patch; your CI and your reviewers verify the PR. The test-and-re-fix loop is the CLI's codna fix --tests --apply.
• mode: review needs the PR base in the checkout (fetch-depth: 0) or an explicit diff.
• mode: secure is read-only and always leaves pull-request-url empty.

Pinning

By default the action installs the latest published codna package:

with:
  package-spec: codna

For deterministic CI, pin the package:

with:
  package-spec: codna==0.2.76

Either way the step installs codna as a wheel only (--only-binary=codna): pip will not build the package from a source distribution, so the runtime that executes is the one published to PyPI. The step does not install with --require-hashes. A hash lock names every distribution in the resolved set — codna and each of its dependencies, at one version — so a lock shipped inside this action would either fix the codna version, breaking the default that follows the latest release, or fall out of date with every codna release. Reproducibility is the caller's choice, made with two pins: package-spec: codna==X.Y.Z fixes the package, and the commit SHA below fixes the wrapper.

Pin the action

@v1 is a floating tag that maintainers move to each vetted change of the wrapper (see CONTRIBUTING.md). For a reproducible workflow, pin the full commit SHA and let Dependabot move it:

- uses: thyn-ai/codna-action@<full commit SHA> # v1

Contributing

Bug reports and pull requests for the wrapper are welcome — CONTRIBUTING.md explains how to check and smoke-test the action locally, and SUPPORT.md says where each kind of question goes. Bugs in what Codna itself did (the fix PR, the review findings, a secure classification) are triaged at thyn-ai/feedback. Security issues: SECURITY.md, never a public issue.

Related repositories

Open-source tooling around Algenta, from the Algenta team. The Algenta engine itself is proprietary; everything listed here is Apache-2.0. Issues and discussions are welcome in whichever repository owns the code.

• thyn-ai/algenta-sdk — Python and TypeScript SDKs for Algenta: self-hosted building blocks for AI applications, 6,000+ deterministic functions on custom Mojo kernels behind one API, SDK and MCP surface.
• thyn-ai/algenta-integrations — Framework integrations for Algenta: LangChain, LlamaIndex, pydantic-ai, MAF, Haystack, LiteLLM, Ray Serve, vLLM, Vercel AI SDK and n8n.
• thyn-ai/mojo-kernels — Clean-room Mojo kernels as drop-in accelerators for popular Python/TypeScript libraries, with bit-exact parity and pure-language fallbacks.
• thyn-ai/security-toolchain — The pinned, checksum-verified security toolchain (Gitleaks, Opengrep, OSV-Scanner, Trivy config, actionlint) that every thyn-ai repository runs locally and in CI.
• thyn-ai/feedback — Public issue intake for the open-source tooling around Algenta and for the Codna GitHub App.
• thyn-ai/codna-action (this repository) — GitHub Action for Codna: fix, review or secure a repository in CI with the same codna command you run on your machine.

License

This action wrapper is licensed under Apache-2.0. The codna package it installs is a separate product with its own license terms; see NOTICE.