任务域决策包
引擎特定优化是 Sonnet 与 Opus 的能力分界
c
我的 JOIN 查询返回重复行，但我找不到原因
r
locate@structured_data/sql_repair
C'
根因通常是多对多 JOIN 且缺少 DISTINCT。常见错误用 Sonnet 4.6，窗口函数和引擎特定行为用 Opus 4.8。
OCG Lab任务域决策系统
D7
结构化数据 / SQL 修复
@structured_data/sql_repair
返回任务域
决策栈
使用
standard SQL errors, high volume, cost-sensitive
claude/sonnet-4.6使用
complex window functions, performance optimization, engine-specific
claude/opus-4.8避免
engine-specific constraint satisfaction required
openai/gpt-5.5认知转移
模型层| ID | 关系 | 执行体 | 结果 | 输出 |
|---|---|---|---|---|
| CT-D7-001 | locate@query_layer | claude/opus-4.8 | 已接受 | distinguishes WHERE vs JOIN vs subquery vs window function; explains engine constraints |
| CT-D7-002 | locate@query_layer | claude/sonnet-4.6 | 有边界 | strong on common JOIN and WHERE issues; weaker on complex window functions |
| CT-D7-003 | rewrite@equivalent | claude/opus-4.8 | 已接受 | explains performance tradeoffs between equivalent rewrites and optimizer behavior |
E-D7-001有边界
Opus 4.8 leads on GDPval-AA economically valuable knowledge work including structured data tasks
Anthropic 发布, May 2026E-D7-002不确定
Sonnet 4.6 sufficient for standard SQL errors in developer preference tests
根据相邻证据推断： Sonnet coding preference data