[BUG] /context command shows full skill tokens instead of loaded metadata tokens
What went wrong?
The /context command displays the absolute token count for skills (the full SKILL.md content), not the minimal amount of tokens they actually consume at startup (only name and description from the YAML frontmatter).
This confuses users because it appears as though all skill tokens are loaded into context, when in fact skills use progressive disclosure—only their metadata (~100 tokens per skill) is loaded initially, with full content loaded on-demand when triggered.
What should happen?
The /context command should either:
- Display the actual tokens consumed (just name + description metadata), or
- Clearly indicate that the displayed token count is the "potential maximum" vs "currently loaded" amount, or
- Show both values (e.g., "50 / 2,500 tokens" meaning 50 loaded, 2,500 if fully activated)
<img width="300" src="https://github.com/user-attachments/assets/1694c99a-0b62-4950-aef3-22f7d269601a" />
<img width="300" src="https://github.com/user-attachments/assets/29e7c515-fdb0-4783-b608-168e9cb60dbf" />
Steps to reproduce
- Install several skills (user or project level)
- Run
/contextcommand - Observe that skills show their full token counts
- Compare to actual context usage—skills aren't fully loaded until triggered
Environment
- Version: 2.0.74 (Claude Code)
- OS: macOS 26.2 (Tahoe)
- Terminal: Warp v0.2025.12.17.17.17.stable_02
Is this a regression?
Don't know - this may have always been the behavior, but it's misleading given the progressive disclosure architecture.
15 Comments
Found 3 possible duplicate issues:
This issue will be automatically closed as a duplicate in 3 days.
🤖 Generated with Claude Code
And I think this bug applies to slash commands and plugins, too.
Hope they get to this soon. It's been several days of increased token usage and degraded experience as a result. This should be close to top priority / urgent. Each passing day is probably causing millions of wasted tokens for both users and Claude. I've also felt a decrease in quality due to the large amount of mixed context loaded into conversations.
can confirm this.
can reproduce, MacOS, Claude Code 2.0.76
Your increased token usage is due to a different reason.
As far as I can see, the bug I reported here is merely a visual/informational issue (albeit an important one). The skills are loaded correctly (i.e. only their metadata).
You're right, I think I'm confusing your issue with a separate problem. Claude Code now treats slash commands in a similar way as skills, and "eager loads" their metadata into context regardless of them using frontmatter or not. The issue with the
/contextcommand is separate, they just need to clarify it's showing the "total token count that would load if invoking a given skill or command" and not the "currently loaded tokens for a given skill or command".yep.
at any rate, it's a mess
This completely undermines the integrity of the
/contextcommand and user's confidence in the "_progressive disclosure_" aspect of skills - a key part of why they're useful, otherwise we might as well include them in ourCLAUDE.md/AGENTS.mdin full.Definitely feels like a something that should be patched asap.
Yep. It also seems like a trivial fix
Ok,I was working for days figuring out the cause and found this issue. Rocky Linux9.7 cc2.0.76 reproduced.
@m-marone please thumb up the original issue so that Anthropic prioritizes this
@PaulRBerg Done！Hope this fixes on next version.
They fixed it in v2.1.1 🎉
<img width="300" src="https://github.com/user-attachments/assets/ae76a7f9-a23e-425d-983f-d4aee1881f10" />
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.