菜单
Diagnose an exact OPL EvidenceBundle without taking evaluator or runtime authority.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Assess an existing Agent version and design a frozen, independent evaluation specification.
Use only when Codex is explicitly asked to create, take over, assess for evolution, or improve an OPL-compatible Agent, including its architecture, prompts, skills, contracts, evaluations, or authority boundaries. Do not use because OMA was mentioned or @-mentioned, another Agent is producing a deliverable, or a deliverable validator, render, or QA check failed.
Admit design evidence and turn an Agent mission into a coherent Stage and authority architecture.
Produce an evidence-bound complete next AgentBlueprint and semantic change set.
| name | oma-evidence-diagnosis |
| description | Diagnose an exact OPL EvidenceBundle without taking evaluator or runtime authority. |
Bind the diagnosis to the exact blueprint, candidate, baseline, frozen test plan, and evidence digests. Separate design defects from evaluator blocks, platform failures, target-owner rejection, safety regression, cost regression, and latency regression.
Each root cause must cite direct evidence refs and identify the earliest owning semantic surface. Do not infer hidden-test bodies from aggregates and do not reinterpret an OPL verdict as an OMA verdict.
Platform failure without a completed evaluation must remain in OPL retry/failure handling and must not enter OMA diagnosis. If completed evidence establishes no admissible Agent-semantic change, say so explicitly and do not invent one.
