PromptFlow AI — VS Code Extension

Analyzes, validates, and optimizes prompts in the editor before they reach Copilot or any other model. The rule engine (dedupe, filler-strip, token counting, validation, scoring) runs in-process in the extension itself — no external server, no Python, nothing to install separately. It's a TypeScript port of apps/ide-mcp-server's original Python implementation; see src/engine/.

This is a separate product from this repo's apps/api gateway. See docs/03-ide-extension-product-brief.md and docs/04-ide-extension-architecture.md.

Features (V1)

• Sidebar (activity bar → PromptFlow icon): paste a prompt and click Optimize to run the full optimization pipeline — validate, analyze, detect redundancy, optimize structure, reduce tokens, re-validate — with a live animated progress view of every stage. Then see:
  • a token comparison (original vs. optimized, exact tokenizer for GPT, clearly-labeled estimates for other providers),
  • an estimated LLM cost impact (before/after/savings, always labeled as an estimate — there is no live billing integration),
  • session savings accumulated across every optimize run this window,
  • a polished summary card with a Send to AI Chat button that opens the optimized prompt directly in VS Code's Chat view (falling back to the clipboard on hosts without that command).
• Status bar: shows the last optimize result; click it to re-run Optimize.
• Commands (Command Palette):
  • PromptFlow: Optimize Prompt — runs the same pipeline against the active selection (or whole document), reporting token/cost savings and a 0-100 quality score.
  • PromptFlow: Validate Prompt — flags missing objective/output format, ambiguity, duplicate requirements, unsafe instructions (PromptFlow output channel).
  • PromptFlow: Compare Original vs Optimized — diff view of the last optimize result.
  • PromptFlow: Copy Optimized Prompt — sends the optimized (or current) prompt to VS Code Chat, or copies it to the clipboard as a fallback.
• On activation, writes a .vscode/mcp.json entry (if the workspace doesn't already have one) so GitHub Copilot Agent mode can call the same tool set directly, over a bundled standalone MCP server (node dist/mcp-server.js — see src/mcpServerEntry.ts). This is the one part of the extension that still needs an external process, since Copilot Agent mode connects independently and can't share the extension's JS runtime; it requires node on PATH. The sidebar and PromptFlow: commands don't use this path at all — they call the engine in-process.

Contributing

See CONTRIBUTING.md for local setup, build, and test scripts.