apifix

Apifix migrates Python code across breaking library API changes. An LLM rewrites the affected code; apifix then runs the project's own tests before and after the change in an offline Docker sandbox, rejects removed or weakened tests, and measures how much of the changed code the tests actually ran.

What it does

When a library ships a breaking change (for example, the OpenAI Python SDK 1.0 removed openai.ChatCompletion.create), apifix:

1. finds the affected code (Python ast, no LLM);
2. rewrites the affected files with an LLM;
3. verifies the result: the project's tests run in a Docker sandbox with no network, before and after the change; tests are compared one by one, and removed or weakened tests are rejected;
4. reports a verdict, and can turn a verified fix into a pull request that states it was written by an AI model.

Installation

Requirements: Python 3.11+, Docker, git. The GitHub CLI (gh) is needed only to open pull requests.

pip install apifix

For development:

git clone https://github.com/theicelotus/apifix.git
cd apifix
pip install -e '.[dev]'

Usage

apifix scan path/to/repo --migration openai-v1            # what would change

export APIFIX_LLM_API_KEY=...                             # key for your endpoint
apifix fix path/to/repo --migration openai-v1 \
    --provider openai-compat --base-url <endpoint> --model <model> \
    --edit-mode whole-file --no-cost-limit                # fix and verify

apifix pr apifix-out/<run> --repo path/to/repo            # branch + PR text

--base-url takes any OpenAI-compatible endpoint. With a local model through Ollama (--base-url http://localhost:11434/v1 --model qwen2.5-coder:14b), your code never leaves your machine. Claude is the default provider (ANTHROPIC_API_KEY). Free hosted tiers may keep what you send, so use them only for public code.

Your repository is never modified: apifix works on a copy and writes fix.diff, report.md, verification.json and trajectory.json to apifix-out/<timestamp>/.

Tested on

7 public open-source repos, with the open-weight model gpt-oss-120b:

Verified: the tests pass before and after, and they run at least half of the changed lines. Weak evidence: the tests pass but run less than half of the changed lines, so those lines are listed for review by hand. Diffs, reports and how each run was made: docs/guide.md.

Documentation

docs/guide.md: all commands and options, the sandbox, model choices (local and hosted), pull requests, and detailed results.

License

MIT