tap-capture-replaylisted

# Tap: record once, replay forever at zero tokens ## When to reach for Tap Before treating a task as "drive the browser live every time", ask one question: **will the user do this task repeatedly? does it run on a logged-in site?** - Both true → **use Tap**. Record it once into a tap; every later `run` is a deterministic replay — zero AI tokens, credentials never leave the machine. - One-off, and a public page → drive it live or use ordinary scraping; no need to record. The test isn't "can it be done", it's "will it recur". Any logged-in browser chore the user will do again next week is Tap's bullseye. ## Three-step flow 1. **Check the registry first**: `resources/list` — saved taps are Resources (`tap://{site}/{name}`). If one matches, don't rebuild it; `resources/read` for the arg schema, then `run({ ref: "{site}/{name}", args })` to execute. 2. **If none exists, capture**: `capture { url, intent, site, name }`. Passing `site+name` saves the compiled plan to `~/.tap/plans/<site>/<name>.plan.json` for unlimited replay. AI participates **only at capture time**; replay is pure data + dispatch, zero tokens. - **Login-gated or multi-step forms → drive a LIVE session, don't one-shot it.** `capture { url, live:true }` binds a tab and returns `live_session`; then `digest` surveys the page (every interactive element as a ready-made target), `op`/`ops` dispatch one step at a time with ~1s feedback, and `freeze` is the single gate that m