Guide · Agent skill
How to use the Autofix Local Changes agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The Autofix skill is an agent skill for Qwen Code that reviews your current uncommitted git changes, checks each finding against the code, and applies minimal fixes. It is for developers who want a second review pass on work-in-progress code before committing. After this guide you will know what the skill does, how to install it, how to trigger it, and what safety conditions to check first.
Autofix Local ChangesQwenLM/qwen-code · Apache-2.0 Download the skill Quick start
- 1Check requirements. Confirm Git, the Qwen Code CLI, model credentials, a sandbox runtime, and a POSIX shell (Git Bash or MSYS on Windows) are available.
- 2Download and unzip the skill. Download the ZIP from this page and unzip it so the autofix folder is in ~/.claude/skills/ or a project's .claude/skills/.
- 3Open a git repository with changes. Move into the repository that has staged, unstaged, or untracked changes you want reviewed.
- 4Type /autofix. Run /autofix with no arguments. Arguments cause the skill to stop.
- 5Confirm trust. Read the warning about build or test commands and untracked files, then explicitly confirm that you trust the repository.
- 6Review the result. Read the final status (CONVERGED, BLOCKED, STALLED or NO_CHANGES), the findings' dispositions, and the list of changed files.
What is an agent skill, and what does Autofix make the agent do?
An agent skill is a folder with a SKILL.md file and optional scripts that an AI coding agent loads when a task matches it. Autofix is named autofix and is triggered by typing /autofix in a git repository.
When it runs in local mode, the agent starts a Qwen Code review of your staged, unstaged, and untracked changes. It reads the report, verifies each finding against the exact code, and applies minimal root-cause fixes. It reruns the review until the review approves or the work gets stuck.
- Fixes are left as working-tree changes.
- It does not commit, push, or stage files.
- It does not inspect remote CI, pull requests, or review comments in local mode.
How do you install the Autofix skill?
Download the ZIP from this page and unzip it. Then place the autofix folder in one of two locations: ~/.claude/skills/ for all your projects, or .claude/skills/ inside a single project.
If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.
- Personal install: ~/.claude/skills/autofix
- Project install: .claude/skills/autofix inside the repository
Before you run it: requirements
The skill needs several tools in place. The review will fail closed, meaning it reports an incomplete result rather than a clean one, if a requirement is missing.
- Git
- Qwen Code CLI, available as the qwen command or through the QWEN_CODE_CLI setting
- Model credentials configured for Qwen Code review
- A sandbox runtime for Qwen Code, since the review sets QWEN_SANDBOX=true
- A POSIX shell. On Windows, use Git Bash or MSYS; the bundled workflow requires it
How do you trigger Autofix?
Type /autofix with no arguments from inside the git repository. Arguments are not accepted in local mode; if you pass any, the agent explains this and stops without changing anything.
The skill is set so the model does not invoke it on its own (disable-model-invocation is true), so a bare /autofix is the trigger. A bare /autofix is not consent: the agent asks you to confirm that you trust the repository before the review starts.
- Type /autofix in a git repository with uncommitted changes.
- Read the warning about repository build or test commands and any untracked files it lists.
- Reply with your explicit confirmation to continue.
- If the working tree is empty, the agent finishes with NO_CHANGES and does not start a review.
What do the results mean?
Each run ends with exactly one status. CONVERGED means the review approved, or every suggestion was fixed or declined with evidence. BLOCKED means a decision needs a human or a required check could not be fixed safely. STALLED means changes oscillated or stopped making progress. NO_CHANGES means there was nothing to review.
The final report lists each finding's disposition (act, decline-with-evidence, or defer-to-human), the changed files, the checks that were actually run, and any remaining blocker.
- CONVERGED: review approved, nothing left to act on
- BLOCKED: a human decision or a safe fix is needed
- STALLED: no further progress is possible
- NO_CHANGES: working tree was clean
Safety notes: scripts and network
The skill runs scripts (scripts/run-agent.mjs is among its bundled files) and uses the network. The review runs repository-defined build or test commands inside the sandbox, and that process keeps your model credentials and network access. Only run Autofix on repositories you trust.
Untracked, non-ignored files are sent to the configured review models. If any contain sensitive data, move them out or add them to ignore rules before running the skill.
- Do not run it on repositories you do not trust.
- Check untracked files for secrets before confirming.
- The skill never runs git add, git commit, git push, or any GitHub write in local mode.
Tips for better results
Keep your change set focused. The skill is built to make minimal, scoped fixes and avoid refactors, so a smaller diff gives cleaner findings.
Run it before you commit, while your changes are still in the working tree. Review the resulting changes yourself before staging anything.
- Stage or leave changes as you prefer; the skill preserves your index.
- Expect findings marked defer-to-human to need your own decision.
What people use it for
- Catching a bug in a feature branch before you commit it
- Getting a second review pass on a refactor that touches several files
- Triaging a set of Qwen Code review findings and fixing the safe ones in one run
- Testing Qwen Code's own review and repair workflow as a maintainer
Frequently asked questions
How do I install the autofix agent skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder. In the Claude apps, you can upload it under Settings → Capabilities → Skills.
What do I need to run Autofix?
You need Git, the Qwen Code CLI (the qwen command or QWEN_CODE_CLI), model credentials for Qwen Code review, a sandbox runtime, and a POSIX shell. On Windows, that means Git Bash or MSYS.
Is Autofix free to use?
The skill is published under the Apache-2.0 licence. The source does not state pricing for the model services it calls, so check the terms of your Qwen Code model credentials.
Does Autofix commit or push my changes?
No. In local mode it leaves fixes as working-tree changes and never runs git add, git commit, git push, or other GitHub writes. It preserves your staging area.
Why does Autofix ask me to confirm before it starts?
The review runs repository-defined build or test commands inside the sandbox, and it sends the contents of untracked, non-ignored files to the review models. A bare /autofix is not treated as permission, so you must confirm you trust the repository.
What is the difference between Autofix and a normal code review?
A normal review only reports findings. Autofix runs a Qwen Code review, verifies each finding against the code, applies minimal fixes, and reruns the review until it converges or gets stuck.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Autofix suits developers and teams who already use Qwen Code review and want findings checked and fixed in one pass before committing, on repositories they trust. Skip it for untrusted repositories, for folders with sensitive untracked files, or for Windows setups without Git Bash or MSYS, and use a plain review tool or manual fixes instead.
Download the skill