pi-ego

A persistent browser-session layer for the ego-lite ego-browser skill. Where the skill is just a prompt that drives the browser through one opaque bash call, pi-ego turns every browse into indexed, browsable, replayable session state.

It works whether the agent drives the browser with raw ego-browser nodejs heredocs or the ego_* tools this extension registers.

Install

Requires the ego-lite app + ego-browser skill, and the Pi coding agent.

As a Pi package via settings.json:

{
  "packages": ["git:github.com/hjanuschka/pi-ego@main"]
}

Or drop it in as a local extension:

git clone https://github.com/hjanuschka/pi-ego ~/.pi/agent/extensions/pi-ego

(the entry point is src/index.ts).

What it does

A tool_result hook watches bash/ego_* output and the script source, and indexes:

• Screenshots -- copied into <cwd>/.pi/ego/<sessionId>/shots/ with URL, title, timestamp, and task space.
• Navigations -- from logged pageInfo() and from openOrReuseTab(...) etc.
• Task spaces -- useOrCreateTaskSpace / completeTaskSpace / handOff... lifecycle. The built-in ego_* tools bind to a unique pi-agent:<sessionId> space, so concurrent Pi sessions do not share page or tab state.

Everything persists to index.json, so the views survive /reload.

Commands

| Command | What | |---|---| | /ego-gallery (ctrl+g) | Filmstrip overlay of every screenshot. Inline image + URL/title/time. ←/→ browse, o open file, e export, Esc close. | | /ego-spaces | Dashboard of task spaces (open/closed, event counts, last URL) + recent nav trail. | | /ego-export | Self-contained HTML report (base64-embedded shots + space timeline), opened in the browser. | | /ego-learn | Distill a visited domain into a reusable ego-lite learnings pack under ~/agent-config/skills/ego-browser-learnings/. | | /ego-inline | Toggle inline screenshot thumbnails in the transcript. |

Tools (callable by the model)

• ego_open(url) -- open/reuse a URL in this Pi session's isolated agent task space, return pageInfo + snapshot.
• ego_snapshot() -- fresh semantic snapshot of the current page.
• ego_screenshot(full?) -- capture a PNG (auto-added to the gallery).
• ego_click(target, label?) -- click a selector/ref/locator, re-snapshot.

Flags

• --ego-inline -- start with inline thumbnails on.
• --ego-autocapture -- after an ego navigation that produced no screenshot, auto-grab one by reusing the agent task space.

Notes

• Inline images and the gallery render in image-capable terminals (Kitty, iTerm2, Ghostty, WezTerm, Warp); elsewhere they degrade to text.
• Storage lives under .pi/ which is gitignored.