Skip to content

Frequent codegraph_explore calls in v1.16 make Codex slow for small tasks #1080

Description

@qg-hs

Hi, thanks for building CodeGraph. It is a very useful tool for understanding call chains, impact scope, and cross-module relationships.

However, after upgrading to the latest version v1.16, I noticed that Codex now frequently calls codegraph_explore even for very small tasks.

For example, for simple UI changes such as:

  • removing one line of text from a card
  • updating loading skeletons on a few pages
  • changing labels or copy
  • small style/layout changes

Codex still tends to call codegraph_explore first instead of using a simple rg / read / edit flow.

This causes a noticeable slowdown because codegraph_explore often returns a lot of source code, relationship information, and blast radius data. That returned content enters the model context, which increases token usage and makes Codex much slower.

In my use case, codegraph_explore is great for structural tasks, such as:

  • call chain analysis
  • data flow / state flow analysis
  • permission boundary review
  • impact scope before refactor
  • security review
  • architecture exploration

But for small known-file edits, UI text changes, skeleton loading updates, simple config changes, or literal string lookups, it feels too heavy as the default tool.

Would it be possible to improve the MCP guidance or tool selection behavior so that codegraph_explore is not recommended as the default tool for almost every task?

Ideally, the behavior could be something like:

Known file / small edit      → read/edit directly
Literal lookup               → rg or lightweight search
Local module issue           → lightweight context/search first
Cross-module structural task → codegraph_context / trace / impact
Full source + relationship map needed → codegraph_explore

A configuration option to make codegraph_explore opt-in or lower priority would also be helpful. For example, allowing users to prefer lightweight tools by default and only use explore when explicitly needed.

The main issue is not that codegraph_explore is wrong, but that it is too heavy for small Codex tasks and causes unnecessary latency and token consumption.

Thanks again for the project. I think CodeGraph is very valuable, but this behavior makes the Codex experience much slower in day-to-day development.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions