Skip to content

[User Feedback] people: person lookup rejects reportedly guardian-authorized agent work #2270

Open

Description

@yunfanye

Situation

This came from user feedback. The guardian’s agent could not look up someone during an authorized task. The lookup returned a permission error instead of the person’s information.

Symptom

The reporter observed that core:main discovered core:read_person, then invoked it through execute_action with the documented name argument. It returned status:error and “Only the guardian’s agents can read a person’s notes.” No person data was returned. Other connected-provider reads succeeded.

The source emits the same message with straight apostrophes: Only the guardian's agents can read a person's notes.

How to reproduce

  1. Derived from code: Use an instance with core:main and core:read_person registered. For a guardian-authorized control, start a fresh message as the signed-in guardian in an existing guardian-owned dashboard chat.
  2. Reporter-provided: Call read_action with {"action_name":"core:read_person"}.
  3. Reporter-provided: Call execute_action with {"action_name":"core:read_person","json_args":{"name":"Example Person"}}. The reporter received the permission error. The report does not specify how the failing task started.
  4. Derived from code: To reproduce the rejection branch independently, use the existing action-engine test fixture and invoke core:read_person without an actor or channel context. It rejects before consulting the person database. A person record is unnecessary.
  5. Derived from code: Compare with an explicit guardian actor, or an actorless call carrying an existing guardian-owned webchat context. Both should succeed. An unknown name should return an empty people array, not a permission error.

These are static reproduction recipes. No tests or live action probes were run during triage.

Initial triage

  • The action handler rejects the call, not discovery or argument parsing. execute_action forwards the object arguments through the action port. The port checks the agent’s catalog and passes its turn actor and thread context to the engine. See tool handler and action port.
  • Three conditions produce this exact error: an actor exists but is not a guardian, neither actor nor thread exists, or a webchat thread lacks a resolvable guardian owner. Even a guardian actor does not override a missing or non-guardian webchat owner. The guard does not authorize by agent name. See complete guard.
  • The ownership lookup uses stored chat ownership. Registration delegates to ConciergeService.ownerOfChat(), which reads the dashboard chat and returns null when it cannot find it. See registration and owner resolution.
  • A name mismatch, missing notes, or provider failure cannot produce this rejection. Name validation and database reads follow authorization. Unknown names return an empty array. Missing or unreadable note files become null. See lookup and note reader. Successful provider reads do not establish this action’s caller identity.
  • The engine preserves supplied context. It resolves the actor and installs actor and channel context around the handler. Ordinary subagents also inherit their parent’s actor and thread. There is no evidence here of general worker identity loss. See actor resolution, handler context, and subagent inheritance.
  • Existing tests establish an important counterexample. Guardian actors and actorless guardian chats already pass. Actorless, chatless calls deliberately fail. These tests call the engine directly rather than reproducing the reported task admission path. See authorization tests.
  • Version and history: PR #2158, commit f7098818, introduced this action before the reported build. Its handler, tests, and engine are unchanged between that build and inspected main. PR #2200, commit 915cb0f4, subsequently fixed Concierge task actor propagation and closed #2199. That is not a demonstrated fix here: read_person already permits actorless guardian-owned chats. Searches of open and closed issues and PRs found no matching person-lookup report. #2137 concerns authorization handoffs for custom-action validation, not this first-party lookup.

Suspected root cause

Confidence: medium — The call reaches the person-reader authorization guard and fails before accessing person data. Its effective context must satisfy one of the three rejection conditions above. Missing guardian provenance or an unresolved webchat owner is consistent with the report, but neither is established. The missing link is the failing turn’s host-side actor kind, channel-context presence, and resolved chat-owner classification. The agent name alone cannot distinguish an authorized guardian turn from an actorless invocation. The code does not support claiming that ordinary guardian calls always fail, or that the post-build Concierge fix resolves this report.

Possible fixes

  • If task admission loses verified guardian provenance, preserve it at that specific admission boundary and add an integration regression through execute_action and the real person-reader guard.
  • If a valid guardian chat is passed under the wrong identifier, correct thread-context construction or owner resolution. Cover missing, deleted, visitor-owned, and guardian-owned chats.
  • If the invocation is intentionally actorless and chatless, retain the refusal. Document an authorized retry path and distinguish the rejection reasons through non-sensitive diagnostics rather than granting authority from core:main alone.

Environment

  • Agent-filed report from core:main, category bug.
  • Build 1.1.198, commit 801166febe32da75cc7fb54d6aa57c4cab4a13f1, 35 commits behind inspected main 4acc2001d3dbfa660203aecb13a1e6c7b992492e.
  • Build timestamp: 2026-11-15T08:23:30Z. Feedback submitted: 2026-11-16T02:03:20Z.
  • Uptime: 53,468 seconds. Upgraded since boot from 1.1.194.
  • Database healthy. No failed or broken apps among 22 installed apps.
  • Relay and deposit URL configured. Email and webchat connected.
  • No route, browser user agent, viewport, theme, mode, language, or subject supplied. No evidence identifies a browser-specific or startup failure.

Triage confidence: medium — The exact error uniquely identifies the person-reader authorization guard, and source plus tests establish every rejection condition before database access. The failing turn’s effective actor and chat ownership are absent, so the specific provenance failure remains unverified. Suggested priority: P2. Estimated size: medium.
Source: Rome Cloud user feedback, submitted 2026-11-16T02:03:20.143857Z, build 1.1.198. Analyzed against 4acc2001d3 (main); reporter build commit 801166f, 35 commits behind main. Filed automatically by Feedback Triage.

Activity

  1. added
    bugSomething isn't working
    P2Fix in the normal course of work
    needs-triageAwaiting triage / needs a maintainer to categorize
    size:mediumRough size: several files in one area, about one to three days
    on Nov 16, 2026
  2. yunfanye commented on Nov 19, 2026

    @yunfanye
    ContributorAuthor

    Another user report (feedback bc19d087-2918-443f-be5f-0d6810f11df2)

    Additional occurrence on 1.1.207, commit ee078961d436689a8b3bf7e9d3e607f93e5ad9b4, reported 2026-11-18. This build is 29 commits behind inspected main 0b9935d1db2b53e666483b5688769bede049321f.

    The agent-filed diagnostics identify core:main. The reporter explicitly places the failure in a private guardian webchat, while following the required person-lookup instruction through execute_action. The reported error is “Only the guardian’s agents can read a person’s notes.” No lookup value was included.

    This build already contains PR #2200, commit 915cb0f4, which preserves the actor in Concierge task handoffs. That change therefore does not resolve this occurrence. The person-reader handler and tests, action-tool adapter, action port, session implementation, action engine, and task-input actor implementation have no changes between this build and inspected main.

    The authorization guard still rejects before name validation or database access. The report does not expose the effective actor, channel context, or resolved chat owner, so it cannot distinguish which rejection branch fired. Describing the chat as private does not establish those host-side values.

    Diagnostics: database healthy, no failed or broken apps among 76 installed apps, uptime 38,915 seconds, upgraded since boot from 1.1.202.

    Triage confidence: medium — The exact error identifies the person-reader authorization guard, and the occurrence matches open issue #2270 on a newer build that already includes the task-handoff fix. The failing call's effective actor and resolved chat ownership are absent, so the specific rejection branch remains unverified.
    Source: Rome Cloud user feedback, submitted 2026-11-18T22:02:10.246602Z, build 1.1.207. Analyzed against 0b9935d1db (main); reporter build commit ee07896, 29 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

    P2Fix in the normal course of workbugSomething isn't workingneeds-triageAwaiting triage / needs a maintainer to categorizesize:mediumRough size: several files in one area, about one to three days

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions