codegraph_explore: a character or symbol-only budget beside maxFiles
维护者通常 1 天内回复
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 48/100
- Issue 类型
- 功能
- 描述清晰度
- 基本清楚
- 活跃度
- 冷清
- 技术栈
- typescript
- 领域
- tooling
调研方向
从 src/mcp/tools.ts 开始，#1699 的 doc tier 已经在按排名顺序消耗每个文件的上限，然后追踪 codegraph_explore 的输入 schema (maxFiles) 如何流入 section 渲染。决定要构建哪种变体 (maxChars 预算 vs. 仅符号邻居，形式为 name (file:line))，再对照 issue 正文中实测的回答大小进行验证。完成的标准是：一个新的预算旋钮按照表格预测的方式改变回答的字符数，同时现有的 maxFiles 行为保持不变。
由索引模型根据 Issue 内容生成。
描述
Summary
maxFiles is the only size knob on codegraph_explore, and it counts files, not characters. Each file that makes the cut contributes whole symbol bodies plus its share of the blast-radius and relationships sections, so a small maxFiles does not give a small answer. Ask: a maxChars budget, or a mode that returns only the named symbols' source with caller names and no bodies for the neighbours.
Measured (1.6.0, Windows, ~500-file TypeScript/Vue repo)
One symbol query, the same every time, varying only maxFiles:
maxFiles |
answer size |
|---|---|
| 1 | 5.4k |
| 3 | 8.1k |
| 6 | 20.5k |
| 12 | 20.7k |
| unset | 7.7k |
Natural-language queries are worse: at maxFiles: 3 the answer was still 13–24k characters in 7 of 8 calls, because a file that matches on a prose term renders its matched sections whole.
Why it matters on the client side
Every later turn re-reads the whole context, so answer size is the cost that compounds. On a paired A/B (headless Claude Code, Opus, n = 12) the tool cut a discovery task from 5.67 to 3.33 calls and still raised cost per task by a median $0.13, and the ~21.7k-character mean answer is the reason (a bounded Read averages ~4.6k on the same repo). Telling the model about maxFiles in the repo rules did not help: a rule naming the cap changed calls +0.67 and cost −$0.14 per task, neither under p .3 at n = 9.
What did shrink answers was a shape change, not a count: the section-first doc tier in #1699 renders a markdown hit's top sections under a per-file source cap and drops the graph sections when no code file rendered, and the doc-question chain went from a median 4 calls to 1 with cost per cell $0.30 against $0.54. A code-side equivalent would be the ask here: a character budget the tool spends in rank order, or a symbols-only mode where neighbours appear as name (file:line) rather than as source.
Happy to send a PR for either if the maintainer has a preference; #1699's doc tier in src/mcp/tools.ts already spends a per-file cap in rank order, so maxChars is mostly plumbing a value from the input schema into the same place.
- 主要语言
- C
- 星标
- 73.2k
- 派生
- 4.7k
- 平均合并
- 4 小时 53 分钟
- 30 天内合并 PR
- 626
环境准备
这个项目没有提供开发容器、Dockerfile 或贡献指南，环境需要你自己搭建：先看它的 README，通用步骤见我们的新手贡献指南。
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。
colbymchenry/codegraph 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 68/100
colbymchenry/codegraph#1900 · 2 条评论 · 5 个 reaction ·
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
colbymchenry/codegraph#1700 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 70/100
colbymchenry/codegraph#1535 · 2 条评论 · 2 个 reaction ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 63/100
colbymchenry/codegraph#1403 ·
维护者通常 1 天内回复
-
agent-target
难度 2/5 1-3 小时 新手友好度 68/100
colbymchenry/codegraph#649 · 5 条评论 · 50 个 reaction ·
维护者通常 1 天内回复
查看 colbymchenry/codegraph 的全部 Issue
相似的 Issue
-
difficulty: easy
难度 2/5 1-3 小时 新手友好度 64/100
-
难度 2/5 1-3 小时 新手友好度 65/100
-
performance solver
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 76/100
ChefKissInc/Inferno#319 ·
-
llvm-tools
难度 1/5 1 小时以内 新手友好度 72/100
llvm/llvm-project#235522 ·
维护者通常 1 天内回复