ContribAI

Automatically discovers open-source repositories matching configurable criteria (language, topic, star count, activity level) by querying GitHub's API with intelligent filtering logic. The agent maintains state about previously analyzed repos to avoid redundant processing and applies heuristic scoring to prioritize high-impact contribution opportunities based on code quality signals and maintenance status.

Analyzes cloned repository code by feeding file contents and directory structure to an LLM (Gemini or compatible) with semantic understanding prompts. The agent extracts architectural patterns, identifies code quality issues, security vulnerabilities, and documentation gaps by leveraging the LLM's ability to reason about code intent and best practices without requiring static analysis tool chains.

Scans repository issue trackers and code analysis results to identify fixable problems that align with the agent's capabilities and contribution scope. Uses LLM reasoning to evaluate issue complexity, estimate effort, assess impact, and rank issues by likelihood of successful PR acceptance based on project activity patterns and maintainer responsiveness.

Generates code fixes by prompting an LLM with detailed context: the identified problem, relevant code snippets, project coding style, existing tests, and dependency constraints. The agent constructs context-aware prompts that include the full file being modified, related files, and project-specific patterns extracted from codebase analysis, enabling the LLM to generate fixes that align with project conventions and architecture.

Validates generated fixes by running the project's test suite, linters, and type checkers locally. If validation fails, the agent feeds error messages and test output back to the LLM with a refinement prompt, iteratively improving the fix until it passes all checks or reaches a maximum iteration limit. This closes the loop between generation and validation without human intervention.

Automatically creates pull requests on GitHub with semantically meaningful commit messages, detailed PR descriptions, and proper branch naming. The agent generates PR descriptions by summarizing the fix, explaining the rationale, linking to related issues, and highlighting any breaking changes or dependencies. Uses GitHub API to create branches, commit changes, and open PRs with proper metadata.

Abstracts LLM interactions behind a provider-agnostic interface that supports multiple LLM backends (Gemini, OpenAI, Anthropic, local Ollama) with automatic fallback. If one provider fails or hits rate limits, the agent transparently switches to an alternative provider without interrupting the workflow. Manages API keys, request formatting, and response parsing for each provider.

Automatically detects and infers project configuration by analyzing repository structure, manifest files (package.json, requirements.txt, Cargo.toml, etc.), CI/CD configuration (GitHub Actions, GitLab CI), and code patterns. Extracts coding style conventions, dependency constraints, test framework, build tools, and project-specific patterns without requiring explicit configuration files.

Orchestrates the entire contribution workflow as a state machine: repository discovery → codebase analysis → issue detection → fix generation → validation → PR creation. Maintains persistent state about processed repositories, attempted issues, and PR submissions to avoid redundant work. Implements retry logic, error recovery, and graceful degradation when individual steps fail.

Applies safety filters before submitting PRs to prevent contributing malicious or low-quality code. Checks for common security vulnerabilities (SQL injection patterns, hardcoded credentials, unsafe deserialization), code quality metrics (complexity, duplication, test coverage), and project-specific constraints. Blocks PR submission if generated code fails safety checks, with configurable thresholds.