Adrenaline: Debugger that fixes errors and explains them with GPT-3

Parses runtime error stack traces and exception messages to identify root causes, then queries GPT-3 to generate contextual explanations of what went wrong. The system extracts file paths, line numbers, and error types from structured stack trace output, maps them to source code context, and uses that context window to prompt GPT-3 for diagnosis rather than sending raw traces.

Takes diagnosed errors and generates candidate code fixes by prompting GPT-3 with the error context, stack trace, and surrounding source code. The system constructs a multi-turn prompt that includes the error diagnosis, relevant code snippets (extracted via AST or line-range queries), and asks GPT-3 to propose specific code changes with explanations. Outputs are formatted as diffs or inline code suggestions.

Accepts free-form technical questions across programming concepts, GitHub repositories, documentation, and code snippets, then performs targeted internet searches to ground answers in authoritative sources. The system uses semantic understanding to decompose questions, search for relevant documentation/repositories, and synthesize GPT-3 responses that cite sources. Supports questions about algorithms, design patterns, API behavior, and implementation details.

Accepts user-provided code snippets (functions, classes, or full files) and generates detailed explanations of what the code does, how it works, and potential issues. The system parses the code to identify language, extracts key structures (functions, classes, control flow), and prompts GPT-3 with the code and metadata to generate line-by-line or block-level explanations. Can identify bugs, suggest optimizations, and explain algorithmic complexity.

Analyzes public GitHub repositories by fetching repository metadata, README files, and key source files, then generates explanations of repository architecture, function behavior, and implementation details. The system constructs a knowledge graph of the repository structure (identifying entry points, main modules, dependencies) and uses GPT-3 to synthesize explanations of how components interact and what the repository does.

Retrieves and parses technical documentation from websites (API references, language docs, framework guides) and generates clarifications or answers to specific questions about that documentation. The system fetches documentation pages, extracts relevant sections, and uses GPT-3 to explain concepts, provide examples, or answer questions grounded in the documentation text.

Decomposes complex technical questions into sub-questions, searches for information to answer each sub-question, and synthesizes a comprehensive answer by reasoning across multiple sources. The system uses chain-of-thought prompting with GPT-3 to break down questions like 'how do I implement X pattern in Y framework' into component questions about the pattern, the framework, and integration points, then retrieves information for each and synthesizes a complete answer.

Generates visual diagrams (ASCII art, structured descriptions, or references to diagram tools) to explain technical concepts, architectures, or workflows. The system uses GPT-3 to generate diagram descriptions or ASCII representations of system architectures, data flows, or algorithm visualizations based on technical questions or code analysis.

Tracks and attributes information in answers to specific sources (documentation pages, GitHub repositories, Stack Overflow posts, etc.), providing citations and source URLs alongside explanations. The system maintains a mapping of retrieved information to sources and includes source references in generated answers, enabling users to verify information and explore sources independently.