Chat authorisation, identity and memory security #2066
Notifications
Due Date
No due date set.
Depends on
Reference: sam/archie#2066
Reference in new issue
Block a user
Chat turns are enforced deterministically outside the model: a turn can only do what its person, the agent and the session all allow, and what may be disclosed to each audience and destination is decided separately from which tools may run.
Normative source: the revised RFC exported verbatim at transcripts/2026-11-20-chat-authorization-rfc-3fa43269.md in sam/archie-product (see the 2026-11-20 comment on this issue). It supersedes the original spec paste. Its amendments are normative for the existing platform, not just the chat frontend: roles grant business-action eligibility, not blanket tool or credential access; forbidden is distinguished from permitted-but-awaiting-confirmation, and a confirmation never converts the former into the latter; authorization is enforced on every entry point, not only the chat tool router.
Children track each gap:
Done when every child is closed.
Normative source update (2026-11-20): the maintainer approved the architectural direction but not the original spec as written. The revised RFC with the normative amendments is exported verbatim at transcripts/2026-11-20-chat-authorization-rfc-3fa43269.md in sam/archie-product.
Key amendments: disclosures must be authorized per audience and destination, not just tool calls, and a confirmation click never turns a forbidden operation into a permitted one; roles grant eligibility for business actions, not blanket access to tool implementations or credentials; authorization is enforced on every entry point, not only the chat tool router.
Children #2067-#2075 are being re-based onto the revised RFC, and new children are filed for the newly normative gaps (disclosure authorization, entry-point enforcement parity, chat-triggered workflow execution contracts).