Restricted runtime building and skill snapshots¶
Build operation |
Input and effect |
|---|---|
context |
No input. Resolves creation/revision facts from trusted conversation and target. |
resources |
No input. Reads selectable resources for the trusted editing target. |
propose |
input contains candidate operation content without its own operation. The backend adds propose, submits the candidate, and publishes a visible server event. |
workspace_mcp |
input contains this owner’s operation. Calls admitted tools and may import or prepare capabilities. |
Outer runtime proof must allow Agent Builder. The gateway still validates original task/manifest/tool sets per call; operation cannot expand admission scope.
propose derives call_id from normalized arguments; workspace_mcp import also uses a digest call identifier, while other suboperations may use new UUIDs. These are not one global idempotency scheme. Repairable feedback retains owner data; conflicts/invalid candidates may have different HTTP statuses. Candidate events come from server results/private receipt verification, not terminal stdout.
Skill ZIP endpoints take only skills bound in verified claims with one positive version. Sources are limited to the current workspace or system and RawQuery is rejected. ArchiveSkillFiles still crosses the permission boundary before reading snapshots/files at the immutable version. Caches retain verified content only and do not bypass authorization. Workspace instruction versions may inherit package files; system packages require exact file-version matches. Do not claim both always read current versions or both permit old files.
ZIP generation also validates file count, duplicate paths, digests/sizes, and archive limits. This does not establish acceptance of all package metadata, version recovery, or object-storage lifecycle behavior.