Generated client files (AGENTS.md block, SKILL.md) teach a call syntax no harness can execute

未关闭 适合新手
#3,050 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
75/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
markdown

调研方向

首先定位安装程序代码,找到生成 codebase-memory SKILL.md 和 agent 指令文件中 codebase-memory-mcp 区块的部分。阅读现有的工具调用示例,并根据 CLI 或 MCP 工具定义确认支持的调用形式。更新生成的指南,为原生 JSON 参数和代码执行调用添加可执行示例;当新生成的文件解释简写形式并展示这两种形式时,工作即告完成。

由索引模型根据 Issue 内容生成。

描述

bug editor/integration parsing/quality windows
Version

0.12.1

Platform

Linux (x64)

Install channel

GitHub release archive / install.sh / install.ps1

Binary variant

standard

What happened, and what did you expect?

The client instruction files generated by the installer (the codebase-memory-mcp block written into agent instruction files such as AGENTS.md/CLAUDE.md, and the generated codebase-memory SKILL.md) teach tool calls with a syntax that no harness can execute:

  • search_graph(name_pattern=".*OrderHandler.*")
  • trace_path(function_name="OrderHandler", direction="inbound")
  • get_code_snippet(qualified_name="pkg/orders.OrderHandler")

The tool(param="...") keyword-args-in-parens form matches no actual invocation path. MCP tools are invoked either as native tool calls (a JSON object of arguments), or — in harnesses that route MCP tools through a code-execution tool — as code like tools["codebase-memory-mcp"].search_graph({ name_pattern: "..." }). An agent that tries to follow the example literally cannot produce a valid call and tends to fall back to grep, which quietly defeats the "ALWAYS prefer MCP graph tools over grep/glob" instruction in the same block.

Expected: the generated guidance should be executable. Concretely:

  • keep the tool(param=...) notation as shorthand in tables/workflows, but state explicitly that it is tool + params shorthand, not literal syntax;
  • add one concrete example per real invocation form (native tool call with JSON args, and the code-execution form);
  • optionally note that only a couple of the 17 tools appear in a code-execution tool catalog by default, so agents should list the server's tools before assuming one is missing.

Related but distinct: #3029 fixes the installed skill's query guidance content (its tool inventory also omits compare_graphs and get_file_outline — the generated SKILL.md still says "15 MCP Tools" on 0.12.1). This report is about the invocation syntax, which #3029 does not cover.

Reproduction
  1. Install 0.12.1 via the release archive / install.sh path.
  2. Open the generated codebase-memory skill's SKILL.md, and the codebase-memory-mcp block in the generated agent instructions file (AGENTS.md/CLAUDE.md).
  3. Observe the examples use tool(param="...") syntax, and the section heading "15 MCP Tools".
  4. Compare with any real invocation: e.g. codebase-memory-mcp cli search_graph '{"project":"...","name_pattern":"..."}' takes a JSON object of arguments; an MCP-native tool call likewise takes a JSON object. Neither matches the documented parenthesised keyword-arg form.
  5. In a harness that routes MCP tools through a code-execution tool, ask an agent to "find who calls X using the graph tools": it cannot map the examples to a real call and typically falls back to grep.

No indexed code is involved — the defect is in the generated files themselves.

Logs

Diagnostics trajectory (memory / performance / leak issues)

Project scale (if relevant)

No response

Confirmations
  • I searched existing issues and this is not a duplicate.
  • My reproduction uses shareable code (a dummy snippet or a public OSS repository), not proprietary code.
主要语言
C
星标
45.8k
派生
3.8k
平均合并
1 天 23 小时
30 天内合并 PR
83

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

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

DeusData/codebase-memory-mcp 的其他 Issue

查看 DeusData/codebase-memory-mcp 的全部 Issue

相似的 Issue

更多 C Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。