epic: one canonical AutoBot browser — unify stacks + full agent-browser feature set #12756

Goal

Converge AutoBot's browser automation into one canonical browser control plane that carries the full agent-browser feature set — and fully implement it. Today "browser control for AI agents" is spread across multiple, non-shared Playwright subsystems with real feature gaps. This umbrella tracks the convergence and the missing features, so there is exactly one Browser in AutoBot with all capabilities behind a single interface.

Sourced from a /research comparison against PinchTab (a Go browser-automation control plane for AI agents). PinchTab is not adopted as a product (Go vs Python; it would be a duplicate control plane) — this umbrella mines its proven patterns onto AutoBot's existing Playwright stack.

Current state (audit — file:line evidence)

There are two/three distinct, non-shared browser subsystems:

• Live interactive browser — Node/Playwright autobot-browser-worker/ (one Chromium, one BrowserContext per chat session_id), driven by chat_workflow/tool_handler.py → api/browser_mcp.py.
• Research browser — in-process Python Playwright autobot-backend/research_browser_manager.py (page snapshot / API intercept / scrape-template tools).
• Docker/remote render — autobot-backend/services/playwright_service.py (web_fetch JS-render fallback).

element-index.js:16-23 explicitly notes the live and research stacks do not share code.

What already works well (keep):

• OpenManus-style stable indexed element refs with server-side resolution + stale-index guard (autobot-browser-worker/element-index.js:37-117).
• Hybrid HTTP→browser render cascade for single-URL fetch (Jina→BS4→Playwright, web_fetch/fetcher.py:345-420).
• Strong DNS-resolving SSRF guard (autobot_shared/url_safety.py), enforced at every fetch primitive.
• NPU-accelerated local captcha OCR + human loop (agents/browser/captcha_solver.py).

Workstreams

A. Canonical interface / consolidation

• tech-debt(backend): unify two Playwright browser stacks behind one interface #12651 — unify the two Playwright stacks behind one canonical browser interface (in-proc vs Docker as a strategy). This is the backbone of "one Browser".

B. Navigate the real internet (existing epic, folds in here)

• epic: enable AutoBot to reliably navigate the internet (KB web ingestion beyond plain-HTTP — browser render, captcha/anti-bot, auth, non-HTML, coverage) #12492 — epic: reliably navigate the internet, with children kb/web-connector: web_crawler yields 0 docs + opaque 'empty/failed' on Cloudflare/JS-challenged sites — FAST fetch only, no browser-render fallback, error drops status_code #12486 (browser-render fallback), kb/web: handle CAPTCHA/interactive anti-bot (detect + session-cookie/solver + honest 'captcha-blocked') #12493 (captcha/anti-bot incl. Cloudflare/Turnstile/DataDome), kb/web: authenticated & gated-site support (cookies/headers/browser-login) for the web connector #12494 (authenticated/gated sites), kb/web: extract non-HTML content (PDF/Office/feeds) instead of returning empty/failed #12495 (non-HTML), kb/web: crawl coverage + transparency (sitemap.xml, depth/page-cap visibility, robots opt-out/reporting) #12496 (crawl coverage), kb/web: fetch reliability (proxy/egress, SSL surfacing, per-domain politeness/rate-limit) #12497 (fetch reliability).

C. Perception + ergonomics (existing)

• Umbrella: OpenManus research adoption — browser perception loop, budgets, external MCP #11536 — OpenManus research adoption: browser perception loop, budgets, external MCP.
• feat(browser): ergonomic actions — scroll_to_text, get_dropdown_options, multi-tab support #11544 — ergonomic actions: scroll_to_text, get_dropdown_options, multi-tab.

D. Wire the GUI (existing)

• product decision (#12326): aspirational UI panels have no backend (browser-automation, dev-speedup, system restart/shutdown/backup) — build backend or remove the dead controls #12378 — browser-automation UI panels have no backend; wire to the canonical browser or remove dead controls.

E. New gaps surfaced by the PinchTab research (children filed under this umbrella):

• security(browser): sanitize + trust-boundary-wrap untrusted web content on the agent scrape/browser path (IDPI) #12757 — SECURITY: sanitize + trust-boundary-wrap untrusted web content on the agent scrape/browser tool path (IDPI)
• feat(browser): progressive disclosure (preview-before-expand) for scraped page bodies #12758 — Progressive disclosure (preview-before-expand) for scraped page bodies
• feat(browser): persistent browser profiles (storage_state) surviving restart / idle-GC #12759 — Persistent browser profiles (storage_state) surviving restart / idle-GC
• feat(browser): site audit + visual-diff surface (screenshot-diff, Core Web Vitals, console errors, a11y) #12760 — Site audit + visual-diff surface (screenshot-diff, Core Web Vitals, console errors, a11y)
• arch(browser): multi-instance orchestration / pooling behind the canonical interface #12761 — Multi-instance orchestration / pooling behind the canonical interface

Non-goals / constraints

• No new parallel browser framework — reuse autobot-browser-worker / research_browser_manager / web_fetch / Playwright, converged behind the tech-debt(backend): unify two Playwright browser stacks behind one interface #12651 interface.
• No stealth / fingerprint-injection bypass (e.g. CloakBrowser-style) — ToS/abuse liability and an unwinnable arms race. Cloudflare/Turnstile handling stays detection-first (wait-out + human-loop), per kb/web: handle CAPTCHA/interactive anti-bot (detect + session-cookie/solver + honest 'captcha-blocked') #12493.
• Canonical-coding discipline: collapse to ONE implementation, no _v2/variant browsers.

Done when

One Browser abstraction with pluggable in-proc/Docker execution; all agent + research + KB callers use it; the workstream/child boxes above are delivered with tests green; no second browser stack remains.