auto-code-fixer 0.4.0

Automatically fix Python code using ChatGPT

Auto Code Fixer

Auto Code Fixer is a CLI tool that detects runtime failures and automatically fixes Python code using OpenAI.

It is designed for real projects where an entry script imports multiple local modules. The tool runs your code in an isolated sandbox + venv, installs missing external dependencies, and applies fixes only after the code executes successfully.

---

Installation

pip install auto-code-fixer

---

Quick start

auto-code-fixer path/to/main.py --project-root . --ask

If you want it to overwrite without asking:

auto-code-fixer path/to/main.py --project-root . --no-ask

---

What it fixes (core behavior)

✅ Local/internal imports are treated as project code

If your entry file imports something like:

from mylib import add

…and mylib.py exists inside the project, Auto Code Fixer will:

• copy main.py + mylib.py into a sandbox
• execute inside the sandbox
• if the traceback points to mylib.py, it will fix mylib.py
• then apply the fix back to your repo (with backups)

✅ External imports are auto-installed

If execution fails with:

ModuleNotFoundError: No module named 'requests'

…it will run:

pip install requests

…but only inside the sandbox venv (so your system env isn’t polluted).

---

Safety

Backups

Before overwriting any file, it creates a backup:

• file.py.bak (or .bak1, .bak2, ...)

Approval mode (diff review)

auto-code-fixer path/to/main.py --project-root . --approve

In patch-protocol mode, approvals are file-by-file (apply/skip).

Diff / size guards

To prevent huge edits from being applied accidentally:

• --max-diff-lines limits unified-diff size per file
• --max-file-bytes limits the proposed new content size per file
• --max-total-bytes limits total proposed new content across all files

Dry run

auto-code-fixer path/to/main.py --project-root . --dry-run

---

Advanced options

Fixpack mode (folder/package runner)

Fixpack runs a command (default: pytest -q) against a full-project sandbox, then iteratively fixes the failing file(s) until green or you hit --max-retries.

auto-code-fixer path/to/package --project-root . --fixpack --no-ask
# equivalent explicit command:
auto-code-fixer path/to/package --project-root . --fixpack --run "pytest -q" --no-ask

Run a custom command (pytest, etc.)

Instead of python main.py, run tests:

auto-code-fixer . --project-root . --run "pytest -q" --no-ask

When you use --run, the tool (by default) also performs a post-apply check: after copying fixes back to your project, it re-runs the same command against the real project files (using the sandbox venv for dependencies).

You can disable that extra check with:

auto-code-fixer . --project-root . --run "pytest -q" --no-post-apply-check

Model selection

export AUTO_CODE_FIXER_MODEL=gpt-4.1-mini
# or
auto-code-fixer main.py --model gpt-4.1-mini

Max retries / timeout

auto-code-fixer main.py --max-retries 8 --timeout 30

Optional AI planning (which file to edit)

auto-code-fixer main.py --ai-plan

This enables a helper that can suggest which local file to edit. It is best-effort.

Structured patch protocol (JSON + sha256) (default)

By default, Auto Code Fixer uses a structured patch protocol where the model returns strict JSON:

{ "files": [ {"path": "...", "new_content": "...", "sha256": "..."}, ... ] }

The tool verifies the SHA-256 hash of new_content before applying edits.

To disable this and use legacy full-text mode only:

auto-code-fixer main.py --legacy-mode

Ruff-first mode (best-effort)

Before calling the LLM, you can try Ruff auto-fixes (including import sorting) and Ruff formatting:

auto-code-fixer path/to/package --project-root . --fixpack --ruff-first

Optional formatting / linting (best-effort)

auto-code-fixer main.py --format black
auto-code-fixer main.py --lint ruff --fix

These run inside the sandbox venv and are skipped if the tools are not installed.

JSON report (explain)

Write a machine-readable report with attempts, files touched, and unified diffs:

auto-code-fixer main.py --explain
# or
auto-code-fixer main.py --report /tmp/report.json

---

Environment variables

• OPENAI_API_KEY (required unless you always pass --api-key)
• AUTO_CODE_FIXER_MODEL (default model)
• AUTO_CODE_FIXER_ASK=true|false
• AUTO_CODE_FIXER_VERBOSE=true|false

---

Notes

• This tool edits code. Use it on a git repo so you can review diffs.
• For maximum safety, run with --ask and/or --dry-run.