Skip to content

[User Feedback] actions: automated tasks lack an explicit handoff for guardian-only validation #2137

Open

Description

@yunfanye

Situation

This came from user feedback. An app author cannot finish validating a guardian-only action from an automated private task. They want an explicitly authorized, auditable validation path without weakening the app’s identity checks.

Pain

The reporter observed that execute_action discovers and invokes a custom read-only action, but the action returns its identity rejection. Its guard requires actorKind === "guardian", a private channel context, and a guardian sender bond. The same app’s guardian-authenticated HTTP UI works. No literal error string or custom guard source was supplied.

The platform explains this distinction:

  • Discovery does not grant guardian identity. romeToolPorts() resolves actions through the agent’s allow-list, clears unrelated ambient session identity, and passes the current turn’s actor and thread context. An automated turn without an actor remains actorless. See action discovery and execution.
  • The SDK accessor works, but does not authorize delegation. The host projects store.actor?.kind into getCurrentActionContext(). The supported context patch permits only sharedContext, not identity changes. See host projection and SDK contract.
  • Private conversation metadata is not authenticated identity. Webchat supplies private-thread and guardian-bond metadata, but separately omits the actor for messages said as another account. This permits a private automated task to fail an actor check. See thread metadata and turnActor(). The report does not identify its task’s admission path, so this is a code-derived explanation, not proof of that particular entry point.
  • Approval is not an identity upgrade. ApprovalHandler runs outside the approver’s session and replays with the original requester’s actor and channel context. Adding requiresApproval alone does not solve this request. See approval replay.
  • Ordinary guardian delegation already exists. Subagents inherit their parent turn’s actor, with an explicit guardian-delegation test. Guardian-authenticated HTTP requests separately receive a host-resolved app caller. This is not evidence of general worker identity loss.

Validation recipe, not executed during triage:

  1. Reporter-provided: Install a read-only custom action with the three checks above and invoke it from an automated private task.
  2. Reporter-provided: Compare its rejection with the app’s working guardian-authenticated HTTP UI.
  3. Derived from code: Compare an actorless turn with a new message actually submitted by the signed-in guardian in private webchat. The latter supplies the guardian actor and thread metadata. Use that as the manual-validation control, not an app-supplied identity override.

The remaining gap is a documented handoff from the failed automated validation to a guardian-run execution, or a scoped delegation contract. The general actor documentation defines attribution, but not that workflow.

Related: #841 concerns a supported guardian-only action boundary and missing SDK identity exposure. This report already uses the exposed identity and asks how to authorize validation from an automated task. PR #1620 added actorKind before the reported build, not the requested handoff.

Version and environment: Agent-filed report from core:main, build 1.1.193 at 62d788beb5d369c4498f6ebc08315ee71fbfd028, 24 commits behind inspected main 47a0cfa0e1fad38a34432c33c57fd28e7192fa1f. Relevant identity and approval behavior is unchanged. Database healthy, no broken or failed apps, uptime 84,544 seconds, upgraded from 1.1.192 since boot. Relay configured. No browser, viewport, route, theme, or language was supplied. Nothing points to a browser or startup failure.

Possible directions

  • Document guardian-submitted private-chat validation, expected context fields, and returning the result to the automated task. This reuses existing identity propagation but requires manual participation.
  • Add a guardian-authenticated validation handoff for one named action and argument set, linked to the requesting task and execution log. This improves usability but needs result-delivery and authorization design.
  • Define scoped, expiring delegation if manual execution is insufficient. Preserve requester and authorizer attribution separately, and define how existing guardian-only guards recognize an authorized execution. This costs more than documentation and must not turn approval into automatic guardian inheritance.

Triage confidence: high — Source and tests establish deliberate actorless automated execution, SDK-visible actor identity, inherited guardian identity for subagents, and approval replay without identity elevation. The missing explicit authorization/handoff workflow remains on main; the exact custom guard rejection is not independently verified. Estimated size: large.
Source: Rome Cloud user feedback, submitted 2026-11-13T16:53:11.354069Z, build 1.1.193. Analyzed against 47a0cfa0e1 (main); reporter build commit 62d788b, 24 commits behind main. Filed automatically by Feedback Triage.

Activity

  1. added
    feature-requestThe change is undecided — a pain, gap, or feature idea filed to start a design conversation
    needs-triageAwaiting triage / needs a maintainer to categorize
    size:largeRough size: crosses packages or needs design work, several days or more
    on Nov 13, 2026
  2. yunfanye commented on Nov 13, 2026

    @yunfanye
    ContributorAuthor

    Another user report (feedback 6a522d67-48f8-47e2-af3f-1cafde4a276f)

    Supplement to #2137: the reporter now identifies Concierge task delivery as the admission path missing from the original report. Static inspection confirms a narrower actor-propagation defect, not merely a missing authorization handoff.

    Triage confidence: high — The source explicitly connects Concierge task sends to nonempty saidBy, which unconditionally removes the receiving turn actor. The action port then clears ambient identity and forwards only that missing actor, deterministically explaining guardian-only rejection. The relevant behavior remains on main.
    Source: Rome Cloud user feedback, submitted 2026-11-13T17:11:38.003251Z, build 1.1.193. Analyzed against 47a0cfa0e1 (main); reporter build commit 62d788b, 24 commits behind main. Filed automatically by Feedback Triage.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    feature-requestThe change is undecided — a pain, gap, or feature idea filed to start a design conversationneeds-triageAwaiting triage / needs a maintainer to categorizesize:largeRough size: crosses packages or needs design work, several days or more

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions