AI Preflight

See what your AI coding assistant sees — before you hit send.

AI Preflight is a VS Code extension that gives you visibility into the context being sent to your AI tool. It estimates token usage, detects waste, identifies risks, and suggests fixes — so you get better AI responses with fewer wasted tokens.

Works with GitHub Copilot, Cursor, Claude Code, Windsurf, Amazon Q, Gemini, and ChatGPT.

Why?

When you use an AI coding assistant, it silently gathers context — your open files, selections, tabs, terminal output — and sends it as part of your prompt. You have no visibility into:

• How much context is being sent (and what it costs in tokens)
• Whether that context is noisy, wasteful, or irrelevant
• What's missing that would help the AI give a better answer

AI Preflight makes this invisible layer visible — and actionable.

Features

Always-On Status Bar

A persistent indicator in the bottom-right corner shows your current context risk level (LOW / MEDIUM / HIGH) with the top reason — like "lock file open" or "large data file". Click to open the full panel.

Sidebar Panel

A dedicated panel showing:

• Risk level with color-coded badge
• Token estimate with per-file breakdown showing what's consuming the budget
• Context window usage — percentage of your AI tool's context window in use (e.g., "~68k of 75k tokens")
• Waste patterns detected with severity
• Positive signals — things you're doing right
• Actionable suggestions with 1-click fixes — close tabs, create instruction files, select a function, and more

@preflight Chat Participant

Type @preflight in GitHub Copilot Chat followed by your prompt to get prompt-aware analysis:

@preflight refactor the auth service

Returns:

• Task type classification (debugging, refactoring, coding, etc.)
• Relevant files — which open files match your intent
• Missing context — files mentioned in your prompt that aren't open
• Low-relevance files — open tabs unlikely related to your task, with wasted token estimates
• Scope hints — warnings when prompts are too broad or too vague
• Workspace search — finds related files you haven't opened yet:
  • Imported by active file — follows import paths from your current file
  • Filename matches — searches the workspace for files matching your prompt keywords
  • Test companions — finds matching test or source files
  • Nearby files — sibling files in the same directory as matched files
  • Content matches — files containing your prompt keywords
  • Results are ranked by relevance and grouped into strongly related and possibly related with actionable tips

Security: Sensitive File Detection

Detects 30+ file patterns that may contain secrets or private keys — and warns before they're sent to AI:

Security: Data Flow Transparency

When sensitive files are detected, AI Preflight tells you which company will receive your data and how much:

⚠ Sensitive files in context (~15k tokens) will be sent to Anysphere (Cursor)

Security: Integrity Scanner

Scans AI instruction files (.cursorrules, CLAUDE.md, etc.) for supply-chain attacks:

• Hidden unicode — zero-width characters used to hide malicious instructions
• Bidirectional overrides — text reordering attacks (CVE-2021-42574)
• Prompt injection — "ignore previous instructions", role hijacking, data exfiltration patterns
• Compound attack detection — multiple techniques on same line auto-escalated to error severity

Tool-Aware Analysis

Auto-detects which AI tool you're using and adapts analysis:

• Context window limits per tool and model
• Instruction file checks — missing .cursorrules, CLAUDE.md, .github/copilot-instructions.md, etc.
• Instruction file quality — warns on empty, too-short, or too-long instruction files
• Ignore file checks — missing .cursorignore, .copilotignore, etc.
• Tab suppression — skips tab-related warnings for tools that don't use tabs as context
• Truncation risk — warns when context is likely to be cut off
• Injection surface warning — alerts when large context + instruction files increase attack surface

Waste Detection

20+ rules that run automatically on every context change:

1-Click Fixes

Many suggestions include an executable action — close sensitive files, close unrelated tabs, save the file, create a missing .cursorrules, select the current function instead of the whole file, and more. One click, problem solved.

Outcome Intelligence

After 5+ analysis sessions, AI Preflight starts showing insights about your patterns — like how often high-risk sessions lead to re-prompts. Helps you build better habits over time.

Quick Start

1. Install from the VS Code Marketplace
2. Open any project — analysis starts automatically
3. Check the status bar for your risk level
4. Open the sidebar (Cmd+Shift+I / Ctrl+Shift+I) for full details
5. Try @preflight fix the auth bug in Copilot Chat for prompt-aware analysis

Commands

Settings

Privacy

All analysis is local. No code leaves your machine. Ever.

• No network requests
• No telemetry
• No API keys required
• No data sharing

See PRIVACY.md for the full privacy contract.

Contributing

git clone https://github.com/deepwissen/ai-preflight.git
cd ai-preflight
npm install

Press F5 in VS Code to launch the Extension Development Host for testing.

See DEV_SETUP.md for architecture details and coding conventions.

Feedback

Found a bug? Have a feature request? Open an issue on GitHub.

License

MIT — Copyright (c) 2026 Deep Wissen