PR-Agent

Analyzes pull request diffs using pluggable LLM providers (OpenAI, Anthropic, Ollama, Azure, etc.) to generate structured code review feedback. Routes requests to configured models via a provider abstraction layer that normalizes API calls, handles streaming responses, and manages token limits per model. Supports both synchronous review and asynchronous batch processing for large changesets.

Parses unified diff format to extract changed lines, identify affected functions/classes, and build a minimal code context window that includes only relevant surrounding code. Uses AST-aware language detection to understand code structure and avoid reviewing auto-generated or vendored code. Implements smart filtering to exclude low-risk changes (whitespace, comments, formatting) from detailed review.

Performs language-specific analysis using Abstract Syntax Tree (AST) parsing and semantic understanding for supported languages (Python, JavaScript, Java, Go, Rust, C++, etc.). Extracts code structure (functions, classes, imports, dependencies) to provide context-aware feedback that understands code semantics rather than just text patterns. Uses language-specific linters and type checkers (if available) to enhance analysis.

Caches analysis results for unchanged code sections to avoid redundant LLM calls and parsing. Uses content hashing to detect changes and invalidate cache entries only when necessary. Implements incremental analysis that focuses on changed sections while reusing cached results for unchanged code, reducing latency and token usage by 30-50% for typical PRs.

Analyzes code changes to detect new or modified functions, classes, and APIs, then generates documentation (docstrings, JSDoc, Javadoc, etc.) in the appropriate language format. Validates API contracts (function signatures, return types, exceptions) against documentation to detect inconsistencies. Suggests documentation updates when APIs change without corresponding documentation updates.

Analyzes PR title and description against the actual code changes to identify gaps, inconsistencies, or missing context. Uses LLM to generate improved descriptions that accurately reflect the changes, suggest better titles, and identify missing information (e.g., breaking changes, migration steps). Integrates with PR metadata to validate descriptions against commit messages and issue references.

Examines code changes to identify untested or under-tested logic, then suggests test cases or test file locations where coverage should be added. Parses existing test files to understand testing patterns and conventions, then generates test suggestions that match the project's style. Integrates with coverage reports (if available) to prioritize high-impact areas.

Scans PR diffs for common security vulnerabilities (SQL injection, XSS, hardcoded secrets, insecure cryptography, etc.) using pattern matching and LLM-based semantic analysis. Integrates with SAST tools (if available) and cross-references against known vulnerability databases. Provides severity ratings and remediation suggestions for each finding.

Analyzes code changes to identify potential performance regressions (algorithmic complexity increases, new database queries, memory leaks, etc.) and suggests optimizations. Uses heuristic analysis of code patterns (nested loops, database calls, memory allocations) combined with LLM reasoning to assess impact. Integrates with performance benchmarks (if available) to quantify expected impact.

Indexes the repository to build a codebase knowledge base (function signatures, class hierarchies, common patterns, architectural conventions) and injects relevant context into review prompts. Uses semantic search to find similar code patterns and architectural examples from the codebase, ensuring review feedback aligns with existing conventions. Supports custom context injection via configuration (e.g., architectural guidelines, coding standards).

Processes multiple PRs in batch mode to generate aggregated reports on code quality trends, review patterns, and team metrics. Tracks metrics over time (e.g., average review time, common issue types, code churn) and identifies trends (e.g., increasing complexity, declining test coverage). Generates visualizations and summaries for team dashboards or executive reporting.

Integrates with Git platform webhooks (GitHub, GitLab, Bitbucket) to automatically trigger PR analysis when new PRs are opened or updated. Posts review feedback directly as PR comments, suggestions, or reviews using platform-native APIs. Handles authentication, rate limiting, and idempotency to ensure reliable operation. Supports custom comment formatting and threading for readability.

Allows users to define custom review rules (e.g., 'flag PRs that increase file size by >20%', 'require tests for changes to auth module') and custom prompts for LLM analysis. Uses a rule engine to evaluate conditions against PR metadata and diffs, then applies custom prompts to focus LLM analysis on specific concerns. Supports rule composition and conditional logic for complex scenarios.