Programs access selector
Reusable configuration, optional long-term memory, consistent execution semantics.
中文Select callable Programs without duplicating their catalog
Agents configures access; Programs manages the actual callable entities. The selected Agent’s main panel displays policy and counts, with an optional effective-entry view. A collapsed list is cheap; opening Browse Programs loads the detailed catalog. Preserve the existing target of one-level Workflow composition preview and explicit access editing as separate actions.
Selector and structure preview
| Surface | Content | Interaction |
|---|---|---|
| Agent panel | Functions / Workflows / Applications counts and collapsible effective entries | Expansion reads structure; it does not enable/disable anything |
| Selector left | Type and category navigation | Search, Selected only, Issues only |
| Selector center | Names, descriptions, selected and availability states | Checkboxes edit the Agent draft, not global enablement |
| Inspector | Entry signature, input/output, direct dependencies and validation evidence | Open Programs for source/full graph; expand at most two levels |
| Workflow tree | Stable circular expansion controls and continuous hierarchy lines | Keyboard-expandable; honor reduced motion |
| MCP entries | Not duplicated in Program authorization choices | Managed through the MCP section; runtime tools still use shared invocation |
Selection semantics and failures
- All, Selected, None and Inherit follow the shared capability contract. A preset is a reference; preview expanded names and version. Switching to explicit selection copies only names into the draft, not Program definitions.
- Selected=[] remains None in effect. Missing entries appear in Issues with their stored names, remain removable, and cannot be invoked. Catalog failure is displayed separately from an empty search result.
- Closing Done retains local draft changes; Cancel restores the selector-open draft snapshot. Neither saves the Agent. Global Save commits every section under the same expected_revision.
- Do not mark a workflow validated from a title, source status or stale test count. Show available evidence separately from capability readiness. Production catalog APIs must provide the required metadata before the inspector is marked complete.
- The first-turn model context contains resident schemas and scoped search guidance, not the full deferred inventory. Nested workflow execution must enforce the existing program/trust boundary; selecting a workflow does not authorize arbitrary hidden tool calls.
Implementation status and acceptance
Current AgentsPage has a mode summary and lazy catalog dialog with search and checkboxes. The two-level Workflow structure inspector and richer metadata are target UI, not verified production behavior. Acceptance uses actual expand/search/select/close/reopen actions, checks preserved scroll and drafts, rejects denied calls through the public executor, and checks narrow-screen and keyboard behavior. No additional workflow execution engine is introduced.