Skill: Agent Run Verdict Workflow
Facilitates agent-run features that deliver precise verdicts and traceable evidence through the app UI.
Goal
Ship agent-run features that can produce exact verdicts, open-source citations, trace-backed evidence, and explicit next actions through the existing app UI.
When to Use
- Any task-session or execution-trace workflow change
- Any feature that claims verified results, citations, or proof packs
- Any UI slice that should let an operator inspect and act on an agent run
- Any work involving contract wiring, backend issue enrichment, workflow panels, verdict surfacing, or verification
Inputs
- Existing NodeBench task/session substrate
- Convex session and trace contracts
- UI surface where the run will be inspected
- User request or product requirement for verified, cited, traceable output
Outputs
- Updated contract and data-model wiring
- Backend derivation or enrichment for verdict-ready context
- UI surface exposing verdict, citations, evidence, and next actions
- Deterministic tests and verification artifacts
Protocol
- Start with the contract. Define the session output shape before coding.
- Reuse the existing session and trace substrate before adding persistence.
- Derive verdict state from evidence, verification, approvals, and drift.
- Put the operator summary in the UI above the trace drill-down.
- Surface exact verdicts, not vague status theater.
- Add deterministic tests for derivation and UI rendering.
- Run the closed loop: codegen, typecheck, targeted tests, build, dogfood if the UI changed.
Required workflow sections
- Contract + data model wiring
- Backend issue context enrichment
- Frontend live workflow panels
- Verdict exactness UI surfacing
- Tests + verification
Progressive disclosure rule
If the workflow expects the agent to resolve tool usage itself, prefer surfacing or recording:
discover_toolssmart_select_toolsget_tool_quick_refget_workflow_chainfindTools
Source of truth
See docs/agents/AGENT_RUN_VERDICT_WORKFLOW.md.