Article 11 Workspace
Technical preview · not a hosted serviceA task should survive the gap between asking for work and having a runner available to do it.
Workspace explores a shared, authorized work record: save the request first, run only the permitted check, then return to the same result. It builds on Article 11’s existing local task engine.
This is a technical demonstration.No hosted workspace is available here. The demonstrated executor is a mechanical plan checker—not Ember or Lumen.
A synthetic authorized client submits a non-sensitive plan. Storage commits the task before acceptance is reported.
DURABLE TASK · QUEUEDThe browser can still read the queued task. Reading the record does not execute it.
SAME TASK · NO DISPATCHAn explicit runner invocation performs the mechanical check. This demonstration calls no model.
EXPLICIT CHECKER EXECUTIONReload or return from the task list to read the committed result. A completed check still needs human judgment.
PERSISTED RESULT · REOPENEDBased on the retained synthetic demonstration and scoped independent Chrome review. No new job was run to make this page.
Why we are building it
Today, a useful request can get stranded in a chat tab, a local process or a handoff someone has to carry. The Workspace prototype separates the saved work from the process that executes it.
The same authorized task and artifact were exercised through browser, MCP and A2A entrances in synthetic tests. These are prototype interfaces, not public endpoints to connect to. Clear permission, an inspectable result and an honest record of uncertainty matter more than a claim that everything is autonomous.
Two different things
Available path today
The existing Windows preview provides a local window for saved tasks and deliberate stage-by-stage work. It is a separate product path from this hosted-style Workspace prototype.
Demonstrated in isolation
A durable synthetic task, readable while its runner is stopped, followed by an explicit mechanical check and a saved result you can reopen.
Intended next operating release
A proposed operator-only release, with one account and non-sensitive synthetic tasks. No model or private-memory integration.
The operating plan is a next-step requirement, not a delivery-date promise. This page has no submission form, home-runner connection or workspace sign-up.
Evidence, not a readiness badge
Evidence snapshot · 3 October 2026
The frozen SLICE003 prototype has author evidence and a scoped independent review. Those are different kinds of evidence. Neither constitutes a production security audit or permission to run a pilot.
Client permissions & browser flow · scoped acceptanceTally independently accepted Fix A, measured submit-only behavior, replay confidentiality and the tested Chrome workflow in a synthetic environment.
Runner credential separation · review incompleteFix B has no completed independent acceptance. Author tests and a repeated author suite do not replace that review.
The author recorded 72 of 72 author tests passing. Tally reproduced 71 of 72: the remaining encoded-request test ended in a client-side transport error on two attempts. It is not counted as a pass. Tally’s separate final targeted checks met 19 of 19 expected outcomes; those counts are not added together.
Tally measured the ordinary sign-in, consented submission, queued read with the runner stopped, explicit checker execution and result reopening in Chrome 154 headless. Foreign and opaque/null Origin submissions were rejected with HTTP 403 and no session cookie. Other browsers and the in-app browser are not qualified by that result.
The author’s earlier browser-control attempt was blocked before a verifiable response. That original record remains unchanged; the later independent Chrome evidence is separately attributed. This page is a sanitized summary of retained evidence, not a replay of the test.
A submit-only client can invoke the known submission tool directly, but retained MCP discovery and tool listing require read permission. Operation permission and fresh-client discovery are separate. The prototype has not solved that discovery limitation.
Replaying a completed or closed-unknown task with submit-only permission reveals the acknowledgement and task state, but withholds the artifact, artifact hash, attempt identifier and recovery note. The encoded-request size mismatch remains open. These observations do not grant public access to the prototype.
The tested paths preserve uncertainty and prevent automatic repetition when the outcome is unknown. A missing acknowledgement does not prove that no work happened. Some missing-evidence cases keep the shared execution slot on hold.
Closing a page is not cancellation. Stopping a runner does not prove that an already-dispatched effect was undone. Unsafe restore, lost execution history, broader database credential authority and recovery authorization remain operating-release concerns. No universal exactly-once guarantee is claimed.
This candidate adds static explanatory text, an illustrative diagram and a machine-readable status document. It adds no task-submission endpoint, model call, private-memory access, account system or analytics script. Standard site hosting and existing site policies still apply.
No private screenshots, source package, evidence archive, credentials or session records are offered for download. Opening this page does not dispatch a Workspace job.
This is documentation, not an invitation to execute. Read the machine-readable preview status for the same boundary. A saved record is evidence, not fresh authority. Keep a known refusal as a refusal, and keep an unknown result unknown.
Leave a clear place for the next mind to pick up the work—and a clear choice not to.