VSCode Extension

Accepts natural language bug descriptions and reproduction steps, then autonomously navigates the codebase using VSCode's Language Server Protocol (GoToDefinition, GetAllReferences) combined with custom backend-specific tools (GetFilesRelevantToEndpoint) to identify root causes. The agent performs code triangulation across multiple files, executes reproduction steps via terminal integration, and generates fixes using match-and-replace editing rather than line-number-based modifications, with user review gates before applying changes.

Leverages VSCode's Language Server Protocol to perform structural code navigation including GoToDefinition (jump to symbol definitions) and GetAllReferences (find all usages of a symbol across the codebase). Combined with custom backend-specific tooling (GetFilesRelevantToEndpoint), the agent can map dependencies and trace bug propagation across multiple files without regex-based heuristics, enabling structurally-aware debugging.

The extension acknowledges on its roadmap that 'proper unit and e2e testing' is needed, implying current test coverage is incomplete or absent. This is a documented maturity limitation affecting reliability and stability. Additionally, the extension is early-stage (576 installs, 1 review) with unimplemented roadmap items (Claude/LLaMA 3 support, history pruning, automated reproduction steps), indicating active development and potential for breaking changes.

Modifies backend code using a match-and-replace technique (matching code blocks by content rather than line numbers) to apply fixes identified by the debugging agent. All proposed edits are presented to the user in a VSCode sidebar UI for explicit accept/reject review before file modification, preventing unintended changes and enabling mid-execution feedback loops where users can reject changes and guide the agent toward alternative fixes.

Executes user-provided build and test commands via VSCode terminal integration to reproduce bugs in a live runtime environment. The agent captures terminal output (build logs, test failures, runtime errors) and uses this runtime context to perform dynamic debugging, identifying issues that static code analysis alone cannot detect. Requires user to manually specify reproduction steps and build/test commands, with unknown support for concurrent execution or port conflict management.

Integrates with OpenAI's API for LLM inference, requiring users to provide their own API key on first run. The API key is stored locally in VSCode configuration (not sent to external servers), and all agent reasoning is powered by OpenAI models. The extension currently supports OpenAI only, with planned (but unimplemented) support for LLaMA 3 and Claude via unknown API patterns.

Provides a chat-like UI embedded in the VSCode sidebar (accessed via ghost icon) where users can describe bugs in natural language and receive agent responses. The interface accepts bug descriptions, reproduction steps, and user feedback during agent execution, enabling conversational debugging workflows. The sidebar UI integrates with the agent loop to present change proposals and accept user accept/reject decisions.

As the agent executes debugging steps (navigating files, analyzing code, running tests), the prompt context grows unbounded by accumulating agent reasoning, file contents, and execution history. This causes documented performance degradation (slower LLM inference) and increased confusion (agent loses track of original bug context). The roadmap acknowledges this limitation but no mitigation (history pruning, summarization) is currently implemented, making long debugging sessions unreliable.

Enables users to provide feedback to the agent during execution, allowing them to reject proposed changes and guide the agent toward alternative fixes. The agent loop pauses at change proposal gates, waits for user accept/reject decisions, and incorporates user feedback into subsequent reasoning steps. This creates an interactive debugging workflow where users can steer the agent away from incorrect fixes without restarting the entire process.

Provides a custom tool (GetFilesRelevantToEndpoint) that maps backend endpoints to the files that implement them, enabling the agent to quickly identify which files are relevant to a specific bug without manual codebase exploration. This backend-specific tool is distinct from generic LSP navigation and allows the agent to focus debugging on endpoint-related code rather than searching the entire codebase.

Ports the core agent loop from SWE-Agent (an open-source autonomous software engineering agent) into VSCode as a sidebar extension. The agent loop follows a standard pattern: receive user input (bug description), execute debugging actions (LSP navigation, file analysis, test execution), propose changes, and iterate based on feedback. This porting maintains SWE-Agent's reasoning approach while adapting it for VSCode's UI and integration points.