AI Review Chat
Interact with the AI using a dedicated, persistent chat interface directly in VS Code.
Persistent Chat Sidebar
The AI Review Chat is always accessible from the VS Code Activity Bar ($(comment-discussion) icon).
- Conversation History: Your chats are saved and persist across VS Code sessions.
- Discuss Button: After any code review, click the 💬 Discuss button to send the full review into the sidebar for follow-up questions.
- Agentic Editing: When using models like Claude 3.7 or v0, the AI can autonomously edit files in your workspace (after your confirmation).
@-Context Mentions
Type @ in the chat input to instantly inject rich context into your conversation.
| Mention | Description |
|---|---|
@file | Include a specific file from your workspace. |
@diff | Include the current staged git changes. |
@selection | Include the text currently selected in the editor. |
@review | Include the most recent AI code review. |
@knowledge | Include entries from your Team Knowledge Base. |
Chat Commands
Use slash commands for quick actions:
/staged: Load the currently staged git diff as context./help: Show all available chat commands./gather <question>: Harvest relevant codebase context for a question and copy a paste-ready prompt to the clipboard.
/gather — Smart Codebase Context Harvester
/gather searches your workspace for files relevant to a plain-English question and assembles a self-contained prompt (question + file contents) that you can paste directly into Claude, Gemini, or any external LLM.
Usage:
/gather What is the billboard split screen feature?
/gather How does authentication work in this project?
/gather Where is the data-fetching logic for the dashboard?
How it works:
- Keyword extraction — meaningful terms are pulled from your question (stopwords removed, camelCase and PascalCase variants generated for class/file name matching).
- Semantic search (Strategy A) — if the workspace has been indexed for RAG, similar code chunks are retrieved by embedding similarity.
- Filename search (Strategy B) — files whose names match the extracted keywords are scored and collected.
- Content search (Strategy C) — when fewer than 5 candidates are found, all workspace source files are scanned for keyword matches.
- Budget & dedup — results are deduplicated (same file from multiple strategies keeps the highest-relevance entry), capped at 10 files / ~30 000 characters, and ordered: semantic results first, then filename matches, then content matches.
- Clipboard copy — the assembled prompt is written to your clipboard; a confirmation message shows the file count, approximate size, and which strategies fired.
Result in chat:
✅ Context copied to clipboard!
3 file(s) · ~12.4k chars via semantic (2 snippets) + filename (1 file)
Paste directly into Claude, Gemini, or any LLM — the question and all
relevant file contents are included.
:::tip Enable RAG for best results
Run Ollama: Index Codebase for RAG once to build a semantic index. /gather will then use embedding-based search (Strategy A) in addition to filename and content matching, giving significantly more accurate results for abstract questions.
:::
:::note No question provided
Typing /gather without a question displays usage instructions instead of searching.
:::