answer_commit drops requested elements that appear only in `basis`, and loses trailing zeros in numeric cells
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 3/5
- 预计耗时
- 1-2 天
- 新手友好度
- 38/100
- Issue 类型
- 缺陷
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 技术栈
- sql, typescript
- 领域
- tooling
调研方向
Start in packages/opencode/src/altimate/tools/answer-commit.ts on the feat/analyst-v10-commit branch, where the numeric 15-significant-digit check (about line 31), the rows union schema (about lines 56-61), and the optional precision field (about lines 37-46) live, and where answer_only rendering skips basis. The three defects need separate changes: scale-preserving number rendering, rejection that names missing targets or falls back to basis, and rounding instead of rejecting long floats. Done when the three acceptance unit tests in the issue pass, but the maintainers still need to decide how compound requests are detected.
由索引模型根据 Issue 内容生成。
描述
Problem
Two independent defects in answer_commit (packages/opencode/src/altimate/tools/answer-commit.ts, present on the analyst-commit branches, not yet on main) cause a correct computation to be reported as an incomplete or differently formatted final answer.
- Compound requests lose elements. With
answer_only: truethebasisis never rendered; onlyfields/rowsreach the final output. When a question asks for two things ("the repository and the most frequently copied file"), models commit one field and place the second item only inbasis. The second item is silently dropped. - Trailing zeros are lost. Cell values are
z.union([z.string(), number, z.null()])and rendered withString(value). A JSON number5.00reaches the tool as JS5and renders as5, next to sibling cells such as4.18.precisionis opt-in, so nothing restores the scale. - Valid numbers are rejected. Numbers with more than 15 significant digits (for example a float32 artefact
18.440000534057617printed bysql_execute) are rejected until the model re-sends them as strings.
Mechanism
- Numeric validation: the 15-significant-digit check at about
answer-commit.ts:31; rows schema union at about:56-61;precisionoptional at:37-46(line numbers from branchfeat/analyst-v10-commit). answer_onlyrendering skipsbasis, so a requested element that exists only there never reaches the output.
Evidence
Recorded analyst sessions:
- Compound request: 5 of 5 sessions that used
answer_commitcommitted only the repository field and returned an 18-byte answer, while the same model without the tool answered with both items in all 5 sessions. 4 of the 5 committed the same repository as the correct answers. - Trailing zero: in 5 sessions on a question whose answer has averages like
4.18and5.00, 3 sessions passed the number5and returned an incorrect answer; the 2 that passed the string"5.00"were correct. The source SQL had printed5.000000and5.0. - 9 further calls were rejected for passing 15+ significant digit doubles or float32 artefacts.
Impact
Answer-losing: about 8 incorrect answers in the recorded sessions that were otherwise computed correctly. No time or cost effect.
Proposed fix
- Before rendering with
answer_only, compare the number of requested outputs (parse "X and Y", "for each ...") with the committed fields; on mismatch reject with the list of missing targets, or fall back to rendering the basis. - Render each numeric column with the maximum decimal scale present in that column (or the source result's scale).
- Accept numbers with more than 15 significant digits and round to
precisioninstead of rejecting.
Acceptance check
- Unit test:
rowswith cells5and4.18renders5.00and4.18when the column scale is 2. - Unit test: a two-target question committed with one field is rejected naming the missing target (or renders the basis).
- Unit test: a float32 artefact value is accepted and rounded.
Applies to the answer_commit tool introduced in #1449.
- 主要语言
- TypeScript
- 星标
- 810
- 派生
- 121
- 平均合并
- 1 天 12 小时
- 30 天内合并 PR
- 94
环境准备
在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
AltimateAI/altimate-code 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 68/100
AltimateAI/altimate-code#1454 ·
维护者通常 1 天内回复
-
test: MCP tests fail when the developer's ~/.claude.json has MCP servers (HOME is not sandboxed)未关闭
难度 2/5 1-3 小时 新手友好度 78/100
AltimateAI/altimate-code#1386 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 84/100
AltimateAI/altimate-code#1384 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 84/100
AltimateAI/altimate-code#1323 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 86/100
AltimateAI/altimate-code#1288 ·
维护者通常 1 天内回复
查看 AltimateAI/altimate-code 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 77/100
melgarafael/DeskcommCRM#3104 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 72/100
维护者通常 1 天内回复
-
bug service: webapp
难度 2/5 1-3 小时 新手友好度 82/100
Ocelot-Social-Community/Ocelot-Social#10684 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
nicobailon/pi-subagents#2978 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 66/100
getsentry/sentry-conventions#732 · 1 条评论 ·
维护者通常 1 天内回复