Skip to content

bug(webchat): Concierge task delivery drops the authenticated chain actor #2199

Closed

Description

@Asuka109

Problem

A signed-in guardian can ask Concierge to delegate a task, but the task's turn loses the authenticated actor. A guardian-only app action therefore refuses the delegated call, although a directly submitted guardian turn works.

This is a concrete admission-path bug related to #2137, not a request to grant guardian identity to autonomous work. Ordinary subagents already inherit their parent's actor.

Reproduction

  1. In private guardian webchat, ask Concierge to start or continue one of its task chats.
  2. Have that task invoke a synthetic read-only action requiring the SDK action context's actorKind === "guardian".
  3. Compare the rejection with the same action invoked after a message submitted directly by the signed-in guardian.

No provider key or paid request is needed. An in-process runtime-route harness reproduces the missing actor with a fake repository and fake model-session submission.

Cause on inspected main

Inspected commit: 78beb244.

  • Concierge's sendInput() calls dashboard.send() with CONCIERGE_ACCOUNT as the message source.
  • Dashboard post() passes that source as saidBy to handleChatSend().
  • turnActor() returns undefined whenever saidBy exists, even when the chain has a verified guardian or visitor actor.
  • The action port correctly clears unrelated ambient request identity and uses the task turn's actor, so it cannot repair the missing admission metadata.

Relevant source: concierge/tools.ts, api/routes/webchat.ts, core/rome-tool-ports.ts.

Expected boundary

  • Keep the authenticated accountable actor separate from the automated message source. This is identity propagation, not a model-extracted quotation or a spending approval.
  • Only a live, host-verified private source may pass its actor to its own same-owner task. Visitor identity must remain visitor identity.
  • Reject forged model/HTTP fields, unrelated tasks, changed ownership, expired/replayed handoffs and cancellation before admission. Unknown, group and autonomous sources must not gain an actor.
  • Keep action catalogs, app guards and approval binding unchanged. Do not infer authority from a private-thread label or an unrelated loopback/HTTP actor.

Candidate and limits

A draft fix will use a host-only, short-lived, single-use handoff bound to the source turn, target chat and exact text/skill input. It checks ownership again immediately before synchronous submission. An admitted task keeps its own cancellation lifetime.

This needs security review of the delegation contract. Full container startup and deployed Concierge-to-app worker validation remain separate acceptance gates. The candidate does not change a running instance, app credentials, Core deployment or app publication.

Activity

  1. yunfanye commented on Nov 14, 2026

    @yunfanye
    Contributor

    Another user report (feedback aeb5288f-01ed-400d-8454-6bdba1abc68e)

    Additional reproduction for #2199: the reporter delegated a public social-post read into a private task chat. The task discovered TikHub actions, called tikhub_search_endpoints, and stopped because no verified guardian actor was supplied. The reporter describes this as free public-catalog discovery, so the failure blocks endpoint selection before the requested post read. No literal error response was supplied.

    Version check: Rome 1.1.194, commit 8a71a43a9f27ab14f86de5cbe852d3190f06a31e, 40 commits behind inspected main b5aa4ec906b8577ff598c1d7f9ac27feff74989e. The relevant actor behavior remains unchanged:

    • Concierge sends task input as its own account in sendInput(). Webchat post() passes that account as saidBy. turnActor() unconditionally removes the receiving turn's actor when saidBy exists.
    • Discovery checks the agent's action allow-list, not the app's runtime identity guard. Execution clears unrelated ambient identity and forwards only the receiving turn's actor. This explains discoverability followed by identity rejection, rather than an action-registration mi

    Triage confidence: high — Both the reporter's build and main explicitly strip the actor from agent-sourced task input. Action discovery ignores runtime actor requirements, while execution forwards only the turn actor, explaining the reported discovery-then-denial sequence and matching open #2199.
    Source: Rome Cloud user feedback, submitted 2026-11-14T14:27:50.799433Z, build 1.1.194. Analyzed against b5aa4ec906 (main); reporter build commit 8a71a43, 40 commits behind main. Filed automatically by Feedback Triage.

  2. yunfanye commented on Nov 14, 2026

    @yunfanye
    Contributor

    Another user report (feedback 0b1dced9-d209-4914-a2ea-5623104ecbec)

    Additional evidence for #2199:

    • The reporter identifies two immediately rejected calls to tikhub:tikhub_search_endpoints: {"platform":"twitter","query":"search","limit":4} and {"platform":"xiaohongshu","query":"search notes","limit":3}. Both occurred in a private task automatically handed off by Concierge on the guardian's behalf.
    • The reported rejection requires a host-verified guardian in a private context. The record describes that rejection but does not include the literal action response. The reporter says these were free public-directory lookups and no paid data request occurred. This adds a cross-platform example that fails before endpoint selection, rather than during provider access.
    • Static inspection confirms the host-side mechanism. Concierge's sendInput() sends as its own account. Webchat post() passes that account as saidBy. turnActor() then removes the receiving turn's actor unconditionally. The TikHub guard implementation is absent from the supplied snapshots, so its exact predicate was not independently verified.

    Version: Rome 1.1.194, commit `8a71a4

    Triage confidence: high — The complete Concierge-to-webchat path deterministically removes the receiving actor for agent-sourced task input on both the reporter build and main. This exactly matches open #2199 and explains the guardian-only rejection without requiring a provider request.
    Source: Rome Cloud user feedback, submitted 2026-11-14T16:24:00.669438Z, build 1.1.194. Analyzed against b5aa4ec906 (main); reporter build commit 8a71a43, 40 commits behind main. Filed automatically by Feedback Triage.

  3. added a commit that references this issue on Nov 16, 2026
    915cb0f
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

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions