v0.1.0 — the half that remembers

Browser tasks in dsh cost the same every time. The agent opens the page, takes a
snapshot, hunts for the element, clicks, scrolls, extracts — and next week it
does all of it again, at the same price, because nothing it learned survived the
session.

ego lite's own README lists the fix as still to come:

"Experience accumulation that makes your agent faster the more you use it (coming soon)"

The reader already ships — site.learnContext, site.runTool and
site.runBrowserTool load a learnings/ directory. This is the half that
fills it.

ego_recall  ──▶  ego_run  ──▶  ego_learn  ──▶  ego_site_run
                                    │
                                    └──▶ ego_handoff

Seven tools. The second time you scrape that dashboard, the agent calls a tool
it wrote last week.

Try it in 60 seconds

Needs ego lite installed and onboarded, so ego-browser is on the PATH.

dsh plugin --profile web add -w @dshworks/dsh-ego-browser
# restart dsh
curl -s localhost:8090/dsh-ego-browser/memory | jq .tools   # seven names, or [] if it did not take

Then ask the agent to run ego_doctor once. On a fresh install it comes up
already holding three sites and eight tools, inherited from the ego skill on
your machine.

What is proven

Verified on a real dsh 0.1.1-rc.2 boot of the published package — installed
with dsh plugin add -w @dshworks/dsh-ego-browser, so the artifact that was
tested is the one you install: all seven tools register, the store self-seeds,
our validator finds zero problems in ego's own shipped learnings, and the
memory route returns 403 for a rebound Host. Verified
against the built ego-browser bundle: both capability probes, the output
classification, and the promotion gate. 74 tests, no browser needed.

Not verified: anything past the wire — a real page load, a real task space,
a real user takeover. ego lite is macOS-only and this was built in a Linux
container. If you have a Mac,
a wire report
is the most useful thing you can send, including "matches exactly".

Three things about the ego wire worth knowing

1. The argv shape is not stable. The shipped skill documents
   ego-browser nodejs; the CLI at citrolabs/ego-lite@main takes no argv and
   exits 2 on a stray one. This plugin probes rather than assumes.
2. The helper surface is not stable. Flat globals (cliLog,
   snapshotText) versus facades (console.log, page, taskSpaces). A
   script for one throws ReferenceError on the other. Probed, and the answer
   is handed to the model.
3. A hard stop discards the run's whole output. Take a task space back and
   ego prints only its own guidance — so a result sentinel never arrives. Here
   its absence is the signal, not a parse failure.

On npm

@dshworks/dsh-ego-browser,
public, MIT. Installing straight from the repo works too
(github:dshworks/dsh-ego-browser) — the built files are committed, so there is
nothing to build either way.

Credit

ego lite by CitroLabs is the good idea
here. The store format, the temp-ref rule and the handoff protocol are all
theirs. And if what you want is to watch the agent browse,
Fisfzy/dsh-ego-browser is the one
to install — a live screencast panel and a Linux port of the ego runtime.

MIT. Not affiliated with DeepSeek or CitroLabs.