Skip to main content

ADR-037 β€” Shared-asset promotion is distinct from ordinary AgentFS work

Status: Corrected by ADR-046 Originally accepted: 2026-07-07

Decision that remains​

An automated correction pipeline must not silently publish model-generated content into the shared Specialist catalog or a client's curated KB. A proposal to change shared published assets is a separate product operation with explicit authority, versioning, provenance, and audit. Human approval may be required for that publication operation.

This rule protects shared configuration. It is not a managed-turn approval gate.

Conversation-scoped AgentFS views are mutable​

The organization has one cloud AgentFS authority. Each Hermes conversation owns its private config, state, work, artifacts, and logs subtree and may use the deliberately exposed shared memory/skills. It may create and update authorized files there through normal tools. Those writes do not require Humanwork to stage a host home, copy files per turn, classify the task, or ask an Expert to release the reply.

Published baseline persona/skill assets and conversation overlays are written through the API's direct cloud AgentFS boundary, not hermes-run. A later proposal to promote conversation-derived material back into a shared catalog/curated asset follows the separate publication workflow above; ordinary AgentFS mutation does not.

The removed correction-to-KB auto-ingest and its feature flag remain historical implementation detail. They must not be reintroduced as a prompt few-shot, pre-turn KB write, or delivery condition.