MCP server: silent wrong matches on ambiguous/near-miss names, noisy query_graph, misleading list_prs error

已关闭
  • 哪里不对？
#3,917 1 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

@nothariharan 已经在做这个了。

开始于 2026年10月21日。

  • #4618 来自 @nothariharan —— 未关闭

评估

难度
5/5
预计耗时
一周以上
新手友好度
35/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
php, python, symfony
领域
backend-api-design, devtools

调研方向

Start by reproducing the listed calls against the explicit graph.json, then trace the MCP tool entry points get_node, query_graph, get_community, list_prs, get_pr_impact, and get_neighbors. Done means ambiguous and non-exact labels are explicit, graph results are relevance-aware, communities are labelled, repository errors explain repo=, and neighbor truncation preserves incoming edges.

由索引模型根据 Issue 内容生成。

描述

Version: graphifyy 0.9.71 (installed via uv tool), graphify-mcp over stdio, started with an explicit graph.json path.
Codebase: PHP / Symfony bundle, ~11k nodes / ~33k edges, 614 communities.

1. get_node silently picks one match for an ambiguous name

  • Steps: get_node(label="edit") on a graph where several classes define edit().
  • Actual: returns a single node (Controller/VisitController.php L427) with no hint that other candidates exist.
  • Expected: a list of the candidates (path + line), or an explicit "ambiguous, N matches" error. The caller can't tell whether the answer is the right edit().

2. get_node silently resolves a non-existent name to a different node

  • Steps: get_node(label="MailController") when no MailController exists in the codebase.
  • Actual: returns EmailController as if it were an exact match.
  • Expected: "no exact match", optionally followed by fuzzy suggestions clearly labelled as such.

Side note: the path::symbol form works (Controller/VisitController.php::edit → found). get_node(node_id=...) does not find a generated-looking id for a non-existent symbol, which is correct.

3. query_graph is dominated by god nodes

  • Steps: query_graph(question="who sends the convocation emails?"), default depth 3.
  • Actual: BFS starts from 5 seeds and reaches 2,708 nodes. The output is TRUNCATED and mostly made of the highest-degree nodes (framework interfaces, core entities, enums). With token_budget=8000 it is still truncated (240 of 2,708). With context_filter=['call'] only 20 nodes remain, all unrelated test helpers.
  • Expected: rank results by relevance or distance to the seeds rather than degree, or damp hubs during traversal (similar to exclude_hubs_percentile in god_nodes).

4. get_community returns no community label

get_community(3) returns "Community 3 (205 nodes)" with no name or summary. A generated label (e.g. dominant class or directory) would make the output usable.

5. list_prs reports "gh CLI not found" when the real issue is repo detection

  • Steps: list_prs() with no repo argument, the server started on a graph.json outside any git repo.
  • Actual: Error: gh CLI not found or not authenticated. Run: gh auth login. gh is installed and authenticated. Passing repo=owner/name makes it work. get_pr_impact with the same repo works too.
  • Expected: an error such as "could not detect repository; pass repo=".

6. Minor: get_neighbors truncation hides incoming edges

With the default budget, outgoing edges are listed first, so all incoming edges (callers) are cut off on high-degree nodes. Listing incoming edges first, or balancing both directions, would help.

主要语言
Python
星标
124k
派生
11.9k
平均合并
2 小时 35 分钟
30 天内合并 PR
6

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。

Graphify-Labs/graphify 的其他 Issue

查看 Graphify-Labs/graphify 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。