[BUG] Cursor aggressively injects local other agents skills into context window, consuming >50% tokens unexpectedly

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

Cursor appears to automatically inject local agent/tool “skills” into the active context window, even when they are unrelated to the current task.

This causes the context window usage to grow extremely quickly with Claude Opus 4.7, despite: only a few conversation turns, no heavy code analysis, and no large file reads.

In my case:

  • Context usage quickly reached ~77%, ~230K / 300K tokens used
  • while the actual conversation content was relatively small.

After inspecting the context breakdown, I found that “Skills” alone occupied ~127K tokens, which is abnormally large. It appears Cursor may be:

  • recursively loading local skills from other agents/tools, including skills from Claude/Codex/Cursor ecosystems.
  • and not deduplicating or filtering them aggressively enough.

Observed Context Breakdown
Approximate values from Cursor context inspector:

System prompt:            4.4K
Tool definitions:        23.4K
Rules:                    3.1K
Skills:                 127.6K
MCP:                      5.3K
Subagent definitions:     1.5K
Summarized conversation: 18.8K
Conversation:            45.9K

Steps to Reproduce

  1. install both cursor / claude code / codex, and install the same skills in each tool independenly.
  2. using cursor to start a session, any session will be fine, and then check the cursor context windows usage statistic.

Expected Behavior

  • avoid automatically injecting unrelated local skills from other agents directory without user explictly say so.
  • deduplicate overlapping skills, and lazily load skills only when actually needed.

Operating System

MacOS

Version Information

3.5.33 (Universal)

For AI issues: which model did you use?

Claude Opus 4.7

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

Hey there.

Thanks for the report. I think this is the same issue as reported here:

And, there’s a workaround!

Hi, Colin.

Thanks! I see. I think I might have accidentally turned on this config a long time ago >.<

That probably explains the unexpected context usage behavior I’m seeing now.