Task action attribution cannot distinguish an agent actor from a human, so an agent approval is recorded as KindHumanApproved #1303

Closed
opened 2026-10-27 13:39:51 +00:00 by sam · 0 comments
Owner

Sam has approved exposing approve/reject to agents, with one requirement: it must be auditable that the ACTION was performed by an agent while the AUTHORITY came from him ('approved by came from me'). Today the model cannot express that. internal/domain/taskactions records only a source string - 'the dashboard' or 'chat' - and stamps events.KindHumanApproved / KindHumanRejected with detail 'approved via '. No actor identity is stored. So an agent approving through the dashboard's WebMCP tools would be recorded as a human approval with source 'the dashboard': the audit would assert something false, which is worse than recording nothing. Required: two attributions, not one - the identity that acted, and the principal that authorised it. identities already exist with a bot kind and an identity.Audit{ActorID, Source, RequestID} pattern, so the vocabulary is there; what is missing is the second field and its use on task actions. HARD LIMIT to design around: a WebMCP tool call is NOT attributable to an agent, because the page can invoke its own tools via document.modelContext.executeTool, so an agent's call and a page script's call are indistinguishable from archie's side. Real agent attribution therefore needs the agent to authenticate with its own credential - which is the A2A work (an A2A agent is an Identity) and the identity requirement that keeps recurring. Until then the honest record is: actor = the signed-in identity (verifiable), a source that says webmcp (a hint, not proof), and NO claim that a human approved.

Close reason

Closed on GitHub; syncing bead state


Imported from beads archie-core-ejpi · GitHub #1038 · type bug · priority P1 · created 2026-09-22 · closed 2026-09-26

Sam has approved exposing approve/reject to agents, with one requirement: it must be auditable that the ACTION was performed by an agent while the AUTHORITY came from him ('approved by came from me'). Today the model cannot express that. internal/domain/taskactions records only a source string - 'the dashboard' or 'chat' - and stamps events.KindHumanApproved / KindHumanRejected with detail 'approved via <source>'. No actor identity is stored. So an agent approving through the dashboard's WebMCP tools would be recorded as a human approval with source 'the dashboard': the audit would assert something false, which is worse than recording nothing. Required: two attributions, not one - the identity that acted, and the principal that authorised it. identities already exist with a bot kind and an identity.Audit{ActorID, Source, RequestID} pattern, so the vocabulary is there; what is missing is the second field and its use on task actions. HARD LIMIT to design around: a WebMCP tool call is NOT attributable to an agent, because the page can invoke its own tools via document.modelContext.executeTool, so an agent's call and a page script's call are indistinguishable from archie's side. Real agent attribution therefore needs the agent to authenticate with its own credential - which is the A2A work (an A2A agent is an Identity) and the identity requirement that keeps recurring. Until then the honest record is: actor = the signed-in identity (verifiable), a source that says webmcp (a hint, not proof), and NO claim that a human approved. ### Close reason Closed on GitHub; syncing bead state --- _Imported from beads `archie-core-ejpi` · GitHub #1038 · type bug · priority P1 · created 2026-09-22 · closed 2026-09-26_
sam added the highbug labels 2026-10-27 13:39:51 +00:00
sam self-assigned this 2026-10-27 13:39:51 +00:00
sam closed this issue 2026-10-27 13:39:51 +00:00
sam added a new dependency 2026-10-27 13:42:22 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Reference: sam/archie#1303