Turn TaskTape into an agent-operated debugging instrument

[codeswithroh]

Summary

• embed a loopback-only MCP server in TaskTape with copyable Claude Code and Codex setup
• let external agents reproduce local browser bugs through an isolated headed Chrome session
• capture ordered actions, screenshots, DOM snapshots, console events, network failures, and a Playwright trace
• compile completed evidence sessions into visible review-required Replay checks
• preserve the existing GPT-5.6 computer replay, visual outcome evaluation, scheduling, and history loop
• reframe the Build Week submission and demo around agent-operated bug reproduction

Verification

• pnpm check: 61 tests across 15 files, formatting, lint, TypeScript, and production build passed
• pnpm test:e2e: all 10 Electron journeys passed
• real SDK MCP client completed the broken fixture and created a saved check
• packaged arm64 app exposed 12 tools, observed Uncategorized, persisted evidence, and returned the saved check
• pnpm package:mac produced the DMG and ZIP
• earlier paid gates remain valid: 10 consecutive visual evaluations and paired live broken/fixed computer replay

Known limits

• agent capture accepts only loopback development URLs
• one agent browser session can be active at a time
• Google Chrome is required and is not bundled
• the macOS package is unsigned and still uses Electron's default icon
• no additional paid model calls were used for this milestone