Daily Edition Sources +4

A Child Agent’s Approval Request Shouldn’t Hide in Another Tab

Openwork now brings a child session’s blocked permission into the parent task; fresh Goose state-machine work shows why seeing the ask is only half the control.

A rough cream-paper diagram shows a child agent's permission request moving into a parent task, then one answer passing through a Goose confirmation checkpoint.
Diagram Punkthe approval request is task state, not a missing tab.
repos different-ai/openwork + aaif-goose/goose evidence
4 source signals 2 repos 2 linked commits
Evidence: 2 linked commits / September 2, 2026 / Daily Edition
Open Edition Evidence below

On September 2, a commit to different-ai/openwork added parent-level visibility for a child session’s blocked permission request. Its task row can now say Needs permission, the approval panel names the requesting child session, and the selected reply returns to the session that owns the pending request.

A delegated agent can otherwise look busy while work has stopped at a permission prompt in another session. That is more than awkward interface design: the person supervising the parent task cannot decide what to do until they know which child is blocked and what it is asking for.

The useful change is not the amber shield by itself. Openwork's interaction path collects permissions from the selected session and its child sessions, then replies with the pending request's own session ID.

Its public opt-in end-to-end test spells out the intended scene: a parent sees a blocked child called “Investigate the deployment failure,” gets Deny, Allow once, and Allow for session controls, and watches that child row return to its running treatment after an Allow once answer. The parent has become a place to find and route the interruption, not merely a progress display.

“Working” is no longer allowed to hide “waiting”

That ordering matters. The task-row logic gives a pending permission precedence over both its normal running shimmer and a degraded-sync treatment. The parent should not reassure an operator that a child is working when the next useful human action is an approval.

The test also checks that the parent names the source of the request rather than making a reader open a separate child session and guess which task caused it.

This is a modest claim with a real operational consequence. It does not make the command safe, decide whether a broad session grant is wise, or prove that Openwork has changed its policy engine. It makes the decision point legible to the person who owns the parent task. For a system built around delegation, that is the difference between an invisible stall and a request an operator can actually assess.

Visibility still needs a durable answer

The first reading of this patch could have been a sweeping claim that agent tools are moving permission upward. The source does not support that. On the same day, Goose changed a different layer: its state machine records tool-confirmation requests and responses, then waits for answers before resuming a halted turn.

Its focused test accepts the same outstanding answer idempotently, rejects a conflicting answer, runs the tool once after resumption, and rejects a later response. That is not evidence that the projects share a design or policy. It is a useful contrast: the parent must be able to see the ask, and the runtime must know exactly what one answer means.

Builders of delegated workflows can turn that contrast into a small audit. Can the parent identify the exact child that is waiting? Does the reply return to that child rather than the visible parent? Can a double click, reconnect, or stale modal change the decision or run the tool twice? These public sources show intended main-branch behavior and tests, not a released feature or safer approval policy. The next signal worth watching is regression coverage for several children, reconnects, and retries after a parent has answered.

Evidence Trail

Receipts below the story

The article above is the public narrative. This section keeps the source trail and limits on the same page.

Edition
DateSeptember 2, 2026
LaneDaily Edition
Confidence78%
Sources4
Reposdifferent-ai/openwork, aaif-goose/goose

Primary Evidence

Evidence Limits

  • These public receipts establish intended main-branch behavior and focused test coverage, not release availability, adoption, production reliability, or the safety of a project's underlying permission policy.
  • The projects work at different layers. Their concurrent changes do not prove a shared standard, product coordination, or that parent-level approval should be broader than the child request.
  • Watch for release notes and regression coverage around multiple children, reconnects, stale requests, policy changes, and retries after a parent has answered.
Letters & Corrections

Send a note to the desk

Corrections, missing context, or a follow-up lead.