chrome-agent-bridgelisted

# chrome-agent-bridge (cab) You control a real Chrome through shell verbs. Every verb prints compact text made for you — refs stay stable per page, every command is hard-bounded by `CAB_TIMEOUT` (default 45s) so nothing hangs your turn. ## Startup reflex 1. `cab tabs` → if ERROR no-session: `cab launch` (add `--headless` for CI/quiet work, `--profile <name>` to keep logins). 2. `cab goto <url>` then `cab tree`. ## Verb playbook | Goal | Do | |---|---| | See what's clickable | `cab tree` (returns `ref_N: tag "text" [x,y w×h]`) | | Find something specific | `cab find "sign in"` (text/aria/placeholder search) | | Read page content | `cab read` | | Click | `cab click ref_7` (real mouse events; output = post-click tree) | | Fill a field | `cab type ref_12 "text"` (clicks ref first) | | Submit / keys | `cab press Enter`, `Escape`, `Tab`, `ArrowDown`… | | Wait for UI | `cab waitfor "Dashboard" --timeout 20` | | Visual check | `cab see` → returns PNG path; READ the file to actually look | | Debug network | `cab net [filter]` · console: `cab console` | | Arbitrary JS | `cab js "document.title"` | ## Discipline (non-negotiable) 1. **Refs go stale after navigation** — re-run `tree`/`find` after any `goto` or a click that navigates; never reuse old refs blindly. 2. **Two strikes rule**: if a verb errors twice, change approach (different selector text, `js` fallback, or report the blocker) — never retry-loop. 3. **Verify before claiming success**: after flows (login, submi