Writing Custom Autofix Modules

Relevant source files

The following files were used as context for generating this wiki page:

• scripts/autofix.py
• README.md
• .pr_agent.toml
• .github/workflows/pr_agent.yml
• package.json

The autofix system provides autonomous repair capabilities for pull requests by addressing findings from AI code reviews or repairing CI build and test failures. It operates as a standalone Python script that interfaces with the Google Gemini API to generate minimal, surgical code changes based on context extracted from the repository.

The system functions in two primary modes: review-fix, which targets specific inline comments and summaries, and ci-fix, which reacts to execution errors detected by local test runners. Every change undergoes local verification before the script commits and pushes the fix to the PR branch.

Sources: scripts/autofix.py:1-26, README.md:1-20

Core Architecture and Execution Flow

The autofix module follows a linear execution path that prioritizes explicit review findings before falling back to general CI failure repairs. It utilizes environment variables for configuration and relies on the GitHub CLI (gh) and Git for repository interactions.

Logic Flow Diagram

The following diagram illustrates the decision logic used to select the repair mode and the subsequent validation loop.

The workflow prioritizes review_fix_mode if review comments exist, then executes ci_fix_mode if validation commands fail. Sources: scripts/autofix.py:321-346

Operational Modes

Review-Fix Mode

The review_fix_mode function identifies issues flagged by PR-Agent or other reviewers. It collects inline comments and review summaries using the GitHub REST API and sends this text along with file contents to the LLM.

• Targeting: It focuses on files specifically mentioned in findings or modified in the current branch.
• Filtering: The system prompt instructs the AI to skip style nitpicks and focus on bugs or security issues.

Sources: scripts/autofix.py:157-183, scripts/autofix.py:228-264

CI-Fix Mode

The ci_fix_mode function triggers when a repository's standard check command (e.g., npm test, cargo check) fails. It extracts file paths from error logs to narrow the context provided to the LLM.

• Error Extraction: extract_relevant_files_from_error uses regular expressions to find paths and line numbers in compiler or runtime output.
• Context Generation: It includes modified files and files mentioned in the error logs.

Sources: scripts/autofix.py:277-318, scripts/autofix.py:267-275

LLM Integration and Prompting

The system communicates with Gemini using specific system prompts that enforce a strict JSON response schema. This structure ensures the script can programmatically apply file updates.

JSON Response Schema

All repairs must return the following structure:

{
  "summary": "Short explanation of the fix",
  "files": [
    {
      "path": "relative/path/to/file",
      "content": "Full, complete updated file contents"
    }
  ]
}

Sources: scripts/autofix.py:185-207

Configuration Parameters

Variable	Description	Default
GEMINI_API_KEY	Required Google AI Studio key	N/A
GEMINI_MODEL	The specific model to use for generation	gemini-3.8-flash
REPO	GitHub repository in owner/repo format	N/A
PR_NUMBER	The ID of the current Pull Request	N/A
BASE_REF	Base git reference for diffing	origin/main

Sources: scripts/autofix.py:28-36, scripts/autofix.py:322-331

Automated Test Detection

The script automatically determines how to validate changes based on the presence of specific configuration files in the repository.

The function detect_test_runner returns the working directory and the command string. Sources: scripts/autofix.py:65-92

Security and Validation Mechanisms

To prevent destructive changes or directory traversal attacks, the module implements several safeguards:

1. Path Normalization: safe_normpath ensures file paths do not escape the repository root or use absolute paths.
2. Size Limits: Files exceeding MAX_FILE_BYTES (80,000 bytes) are excluded from the LLM context to prevent token overflow.
3. Verification Loop: The verify_commit_push function runs the detected test command after applying changes. If the command fails, the script executes git checkout -- . to roll back changes.
4. Ownership Verification: Changes are only pushed to the HEAD_REF branch, and specific Git configurations are applied for the github-actions[bot].

Sources: scripts/autofix.py:38-39, scripts/autofix.py:150-155, scripts/autofix.py:228-235, scripts/autofix.py:241-260

Conclusion

The custom autofix module integrates LLM-driven code generation with standard Git workflows and language-specific test runners. By targeting both explicit review feedback and implicit CI failures, it provides a comprehensive autonomous repair loop that maintains repository integrity through strict post-repair validation.