Research stealth mode

Settings → RESEARCH controls the hidden pipeline used by search_web, deep_crawl, and research_topic. It is a reliability layer, not a CAPTCHA bypass.

What it changes

Control	Effect
Enable stealth	Applies the research defaults below; when off, the existing tool arguments are left alone.
Engine chain	Default provider order. The shipped stealth order is DuckDuckGo → Brave → Google. A custom template is tried first when configured.
Custom template	A SearXNG/self-hosted URL containing `{query}`.
Research proxy	Applied only to the isolated `persist:kuang-research` session; normal browsing tabs keep the default session and direct route.
Domain cooldown	After a WAF/challenge page or HTTP 403/429, the host is skipped for this many seconds.
Politeness delay	Minimum gap between crawled pages and research sources.
Crawl concurrency	Parallel page fetches inside `deep_crawl` (1–4). `1` is sequential and safest; higher values raise WAF risk.
Prefer Markdown/LLM	Try content-negotiated Markdown/Text and advertised alternates before rendering.
Default crawl pages / research sources / search results	Applied when the tool call omits its own limit.
Activate hidden view	Uses CDP to keep the off-screen view in an `active` lifecycle, emulates focus, and gives it a 1280×900 layout viewport.

Isolation

Research renderers use persist:kuang-research, not the normal tab session. That means research cookies, cache, and aws-waf-token values do not mix with the visible browser tabs. The session is persistent so consent/challenge state can survive between calls; Reset research cookies clears it.

search_web and crawl rendering use the isolated session. The Markdown/Text content-negotiation fast path uses a second LlmSurface bound to the same research session, so prefer_llm does not silently bypass the proxy or isolation.

Challenge handling

Search hub already reports when an engine returns no results plus a bot-check/consent wall. Stealth mode records a cooldown for that engine.
Rendered pages are checked for AWS WAF, Cloudflare, DataDome, CAPTCHA, “unusual traffic”, 403, and 429 signatures. A detected challenge marks the host as cooling and skips the page.
The crawler calls the cooldown gate before the cheap Markdown path too, so a cooled host is never contacted by prefer_llm.

Limits

A hidden Chromium view is still detectable. TLS/JA4, hidden visibility, lack of human input, IP reputation, and missing cookies can all matter.
The search path still loads SERP URLs directly; humanized typing into a search box is not implemented yet.
CAPTCHAs are surfaced as failures/notes, never auto-solved.
A proxy only helps if it is reliable and not itself flagged. Prefer a sticky residential/mobile exit or a separate network for agent traffic.
For sites that aggressively challenge, use the visible tab and request_human_intervention instead.

Tests

npm run test:research:settings   # schema, persistence, challenge detection
npm run test:crawl:pool          # crawl concurrency / budget / prefer-llm
npm run test:research            # bridge-level research_topic regression

Files

src/main/research/settings.ts — settings schema, persistence, challenge detector.
src/main/browser-chassis.ts — isolated session, proxy, cooldown gate, hidden-view activation.
src/main/research/researcher.ts — per-source delay.
Settings → RESEARCH — controls.