Critical failures across all 10 tools. All input schemas are empty objects with NO properties defined (e.g., SessionCreateArgs, ObserveArgs all resolve to {}). No parameter descriptions visible, tools declare inputSchema but the actual Zod schemas are not shown in source, making validation impossible. Tool names follow verb_noun pattern (good), but lack clarity on required vs optional parameters. No output schemas documented. Error handling guidance absent.
按 Agent 声明的视觉类别做结构化检测(JSON bbox)。仅当 Provider 已验证 structured_detection 能力时方可执行,否则返回可行性评估事实。
通用视觉观察:返回可观察的视觉事实(默认聚焦图片主体,不转录水印等细小文字)。默认不主动输出主观评价;用户明确要求时,基于可观察特征参考普遍审美标准给出,并说明依据。profile=deep 时预置深入指令(纳入水印/细小文字/细粒度特征;仅当用户明确要求更深入时使用)。observation_schema 提供声明式结构化观察:声明维度(如 color/shape),逐字段返回 value 或 unknown+reason(须模型验证通过)。需要综合理解多张图片时,请优先使用 vision.summarize。文字转录请用 vision.ocr。只陈述事实,不推断诊断。
提取图中文字,返回 text_block Observation。
取消 Operation。已 committed 的 Observation/Artifact 保留(取消不是错误)。
查询 Operation 生命周期状态(completed / cancelled / failed / running)。
Session 审计汇总(专业模块,按需调用):仅当用户明确要求审计/记录/汇总/导出时使用,绝不主动输出。默认返回操作级汇总(时间/工具/状态/执行者/失败原因);include_observations=true 附已提交观察的完整元数据(label/location/confidence/limitations/source,location 即区域对应)。
All 10 tools have empty input schemas ({}). No properties, types, or constraints defined.
Descriptions are verbose (200-400+ chars) and mix implementation concerns (e.g., 'profile=deep', 'observation_schema', 'structured_detection capability') with user-facing guidance. Descriptions should be 10-1024 chars and focus on WHAT/WHEN/WHY, not internal mechanics.
No parameter descriptions visible in source. Tools reference Zod schemas (SessionCreateArgs, ObserveArgs, etc.) but actual schema definitions are not shown. Cannot verify parameter types, constraints, or descriptions. Violates: 'Every parameter needs a description.'
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 36 | 2026-07-28+ | v2 |
创建新的 Vision Session(绑定当前 Principal/Tenant)。返回 vision_session_id。
关闭 Vision Session(已提交证据按 retention 保留)。
查询 Vision Session 状态(须归属当前 Principal/Tenant)。
批量图片综合概述:对 1~16 张图输出一段平实事实化的综合描述(共同主题、显著差异、整体印象),不逐张罗列细节,默认忽略水印等细小文字;只陈述可观察事实。**若未能覆盖全部输入图片,输出必须如实声明实际覆盖范围**(如「仅覆盖前 4 张」),不得假装已分析全部。profile=deep 时纳入水印/细小文字异同的深入比较。任一图片不可读则整体失败并指明第几张。单张细节请用 vision.observe,文字转录请用 vision.ocr。
No output schemas documented. Tools return observations, artifacts, and operation states but no structured response format is specified. LLMs cannot plan downstream calls or extract required fields.
No error handling guidance. Tools declare risk levels (WRITE, DESTRUCTIVE, READ_ONLY) but provide no recovery instructions. E.g., vision.session.delete is DESTRUCTIVE but description does not explain what happens on failure or how to recover.