Add production contract refresh sections and normative contracts: 017/050 capture-reuse trust, 036 evidence promotion, 037 catalog lifecycle/revision, 038 browser actions/decision policy, 044-047 production chain/evidence/atomic triage, and 050 MCP interface artifacts. Requirements marked OPEN pending executable evidence.
3.0 KiB
#region AgentTestStabilization.UxReference [C:3] [TYPE Tombstone] [SEMANTICS ux,reference,agent,dashboard-testing] @DEPRECATED Former agent-workspace UX reference retired by the 2026-09-08 external-MCP-only decision; no product frontend agent interaction is allowed. @REPLACED_BY ScenarioEditor.Modules (manual editor/review) and ScenarioRunMonitor.Modules (read-only results, human checkpoints); see specs/036-agent-test-stabilization/contracts/ux/screen-models.md. @RATIONALE Explicit 2026-09-08 user decision prohibits agent chat, prompt/request input, assistant editing, proposal generation, agent workspace/start and handoff controls in product UI, not merely a Gradio transport change.
Feature Branch: 036-agent-test-stabilization
Created: 2026-07-07 | Superseded: 2026-09-08 | Status: Historical record; not an active UX contract
1. Current boundary (active)
- AgentRun recovery, draft-artifact ownership, tool-run records and durable ActionApprovalGate contracts continue server-side and are consumed by external MCP clients (see
specs/050-mcp-interface/spec.md). - Product frontend keeps only: manual CRUD/editor (042/043), human approval/review cards bound to the same gate DTOs, monitoring (045/047) and read-only evidence/evaluation rendering (045). No control may start or direct an agent.
- Existing agent workspace/proposal panels in runtime code are drift; removal plus negative DOM/route/network acceptance remains OPEN in tasks.md. This spec-only refresh changes no runtime code.
2. Retired interaction (historical narrative)
The 2026-07-07 design placed a persistent Svelte /agent workspace on the dashboard page: agent context header with run/intent binding, a stage progress strip, draft-artifact preview with request-save, and inline approval cards. Scenario creation, inspection and draft generation were agent-driven inside the product UI. Every element of that flow is superseded: authoring moved to external MCP clients, and the product UI now loads server-stored proposals as read-only diffs for human review with manual editing (043) and human-in-the-loop approval (036 gates).
The retired mockups (agent context header, progress strip, draft preview, inline gate) are preserved in Git history only; they must not be reimplemented or cited as current test evidence.
3. Still-valid server-side UX obligations
- High-risk decisions (baseline candidate publish, repository save, promotion to 037 baseline) render as human approval cards bound to a durable gate: target paths, evidence refs, risk, reconciliation policy, approve/deny/open-diff. Denial records a cancellation with zero side effects.
- Permission-denied presents the required role and no confirm control; viewer roles never see confirm buttons.
- Recovery semantics survive restarts: run/gate snapshots reload deterministically; no write occurs while a decision is unauthenticated or disconnected.
- Terminology stays explicit and side-effect aware: "run", "draft artifact", "gate", "confirmation", "approval reason", "dashboard context".
#endregion AgentTestStabilization.UxReference