Current — August 2026
Taskloopr records real browser actions and replays them reliably — including while the window is minimized, the tab is unfocused, and across MV3 service-worker restarts.
This is the changelog tracking what's been fixed, improved, and added over time.
Recent Features
AI Decision steps
Multimodal proceed/skip checks using page text and live screenshots. Configure pass/fail actions and human-in-the-loop approval.
AI-generated values
Let AI decide what to type into fields — fresh values on every run for search terms, test data, and more.
AI-driven clicks
Vision model chooses what to click on dynamic pages — essential for clicks that follow varied search results.
Conversational editing
Update workflows by describing changes in plain language — "remove the like-button step" or "add an AI check after step 5."
Reliability Fixes
8-tier fallback chain
id → data-testid → saved position → aria-label → exact text → similar text → toggle-label → matching class → only visible match.
Trusted input via CDP
Real OS-level events via Chrome DevTools Protocol for more reliable typing and clicking. Falls back gracefully if unavailable.
Contenteditable support
Fixed recording and replay in rich-text editors (Notion, Docs, comment boxes) that use contenteditable instead of inputs.
SPA navigation capture
Polls location.href to catch client-side route changes that don't trigger history APIs — fixes YouTube search and similar SPAs.
Notable Bug Fixes
Navigate step undoing AI
Fixed recorded navigate steps that were force-navigating back to the original URL, undoing AI-generated searches and clicks.
AI clicks on detached elements
Added document.contains() check before clicking — prevents silent failures when pages re-render during AI decision time.
Enter-to-search replay
Enter key now properly replays for search submission, including form.requestSubmit() and page transition timing.
React-controlled inputs
Fixed typing by using native prototype setter instead of direct value assignment — bypasses framework state overrides.
Build & Architecture
The extension now has a build step using TypeScript and Vite. The repo root is no longer loadable as an unpacked extension — load dist/ instead.
Run npm run dev for watch+rebuild while working, or npm run build for a clean production bundle.
Unit tests use Vitest for core logic; E2E tests use Playwright with the extension actually loaded in Chrome.
Documentation
The MVP scope, technical audit, and target architecture are documented in docs/GOAL.md, docs/AUDIT.md, and docs/DESIGN.md.
These documents define the scope, UX requirements, reliability taxonomy, performance budgets, and migration path for future features like composable sub-workflows and live tutorials.
