TASKS.md

Tasks

Completed

• Project scaffolding (SPEC, CLAUDE, ARCH, context, claude config)
• Reference brainstorm archived to docs/reference/
• Full framework implementation (Phases 1–9) — see .context/progress.md
• CDP raw access: debugPort config, --remote-debugging-port launch arg, getCdpUrl() in SessionManager, surfaced in get_page_state
• README, ARCH.md updated with raw Playwright access via CDP use-case

In Progress

Backlog

Foundation

• Decide stack: TypeScript + Playwright vs Python + Playwright
• Init package.json / pyproject.toml with Playwright dependency
• Implement SessionManager — browser launch, storage state persist/restore
• Implement SiteAdapter base interface / abstract class
• Implement MCPServer entry point (stdio transport first)
• Implement ToolRegistry — auto-discover and register adapters

First Site Adapter: LinkedIn

• LinkedIn isLoggedIn detection
• LinkedIn login flow (manual-first via HumanHandoff)
• LinkedIn tool: get_feed — fetch top N feed items
• LinkedIn tool: get_messages — list unread messages
• LinkedIn tool: search_people — search by name/company

First Site Adapter: Shufersal

• Shufersal login flow
• Shufersal tool: search_products(query) — return name, price, unit
• Shufersal tool: get_cart — return current cart contents

Human Handoff

• Implement HumanHandoff — foreground browser, wait for signal
• Integrate handoff into adapters for 2FA and order confirmation

Testing

• Unit tests for SessionManager (mock browser)
• Unit tests for each site adapter tool
• Integration test: full login + tool call flow

Developer Experience

• README with quickstart
• Example: multi-site grocery price comparison in 10 lines
• CLI: browserkit start [adapter...]

Follow-up Ideas

• Plugin registry for community-contributed site adapters
• Cross-site composition tools (search N sites, merge results)
• Replay / test harness from recorded sessions
• Rate-limit and jitter config per adapter