REAL SOURCES · CONTEXT THAT STAYS

Do the research.
Keep what matters.

Find open-source tools, remember a shortlist, and come back to see what changed.

FIRST REAL TASK
PUBLIC PROJECT RESEARCH

Research open-source tools READ ONLY

What are you looking for?

LanguageMinimum stars

Loading remembered preferences…

Checks up to five projects against GitHub’s public API. No account changes or model calls. This first task uses structured search, not a general chat planner.

Your workspace THIS BROWSER

Tasks and memory are saved on this server for this browser.

Protect this workspace or open an existing one

Workspace name Passphrase

Use the name and passphrase on another device to continue. Keep the passphrase yourself; password recovery is not available yet. Opening another workspace keeps this browser’s previous data separate.

Task activity 0 tasks

Start a live search. Every source check and failure will appear here, even if you leave this page.

YOU DECIDE WHAT STAYS

A memory you
can inspect.

Preferences guide the next search. Notes keep context. Source snapshots let you compare a later result.

Research preferences USER PROVIDED

Preferred languageDefault minimum stars

Remember something EXPLICIT SAVE

Context you want to keep

Matching notes appear as task context. Only the structured preferences above change filters. Page descriptions never become instructions or notes automatically.

Saved context

Find in memory

Forgetting a note also removes its text from recalled task context. Public source results and task inputs remain in task history until you forget those receipts. Export includes memory and task evidence, never account credentials.

OWNED FIXTURE · NO API KEY NEEDED

Learn the page.
Keep the understanding.

Observe once. Replay a typed operation. Check what actually changed.

EXPERIMENT 001
SEMANTIC REPLAY

This lab operates the sandbox below with real Chromium and Juno’s semantic reader/actor. It does not connect to your accounts. The general compiler and real Rolo benchmark are still ahead.

1 The websiteInteractive sandbox

Page variation

2 Compile a bindingNot observed

Read live accessibility semantics and save a binding for this fixture’s known operation. Recompile explicitly after label changes.

Inspect saved operation

No model saved yet.

3 Execute & verifyWRITE · SANDBOX

capture_interaction(...)

PersonOccurred on

Interaction noteNext action

Five semantic actions. Zero model calls. Outcome checked against fixture state.

Latest resultReady when you are

Start with Observe & compile, then run the operation. Change the page variation to test replay and failure.

EVERY ATTEMPT COUNTS

Evidence, not impressions.

Receipts include failures, inputs, observations and measured runtime. Stored for this browser on the server.

Run history

Selected receipt

Select a run to inspect its evidence.

What should improve?

Feedback is attached to the selected run, or to the lab if none is selected.

THE PATH TO A USEFUL JUNO

Small experiments.
Compounding evidence.

A testable lab is a starting point. These gates keep the larger claims honest.

Your lab checklist

What this experiment proves

A stable capture_interaction contract can drive a changing, controlled UI through observed semantic references, then verify the resulting state.

Compilation here is a small deterministic template for this fixture. This is not general website discovery, learning from a demonstration, an LLM benchmark, or real Rolo integration.

Zero model cost describes this execution path. It is not evidence of a cost advantage over another agent.

Next experiments · GSD is the source of truth

Browser foundation

Persistent profile, accessibility perception, strict semantic actions.

BUILT

LAB

Hosted controlled experiment

Saved bindings, replay, UI mutations, verified receipts and feedback. Validate with the checklist above.

TRY IT

1.1 / 2

Windows runtime + full execution guard

Complete the planned CDP runtime; add confirmations, audit, injection boundaries and kill-switch guarantees.

PENDING

Real Rolo → read-only Canvas

Run capture_interaction in real Rolo, survive a UI mutation, then get_upcoming_coursework across real courses. Compare repeated runs to a baseline.

PENDING

3–7

General compiler, replay & teaching

Per-site knowledge, explicit uncertainty, local repair, teach-by-showing, portal reuse, then voice.

PENDING

Direction: Learn once, replay fast ↗ · Rolo experiment ↗ · GSD plans ↗

Juno / a semantic website compiler Observe → Bind → Execute → Verify → Remember

Do the research.Keep what matters.

A memory youcan inspect.

Saved context

Learn the page.Keep the understanding.

Evidence, not impressions.

Small experiments.Compounding evidence.

Browser foundation

Hosted controlled experiment

Windows runtime + full execution guard

Real Rolo → read-only Canvas

General compiler, replay & teaching

Do the research.
Keep what matters.

A memory you
can inspect.

Learn the page.
Keep the understanding.

Small experiments.
Compounding evidence.