Claude TokenUsagePercent is out of sync

[MelGrubb]

Code of Conduct

• I agree to follow this project's Code of Conduct

What happened?

I've been working with the Claude Code support for a day or two, and I'm noticing that after issuing a /clear or restarting Claude Code, the context percentage doesn't seem to update. Calling /context shows that I'm on a fresh, empty context, but the gauge and text percentage are still showing that I'm at like 80%. I've run several commands in the last half hour, and it's just not catching up.

Theme

Stock theme with only minor adjustments to the segment layout.

    {
      "type": "prompt",
      "alignment": "right",
      "segments": [
        {
          "type": "claude",
          "style": "diamond",
          "leading_diamond": "\ue0b6",
          "trailing_diamond": "\ue0b4",
          "foreground": "p:black",
          "background": "p:green",
          "background_templates": [
            "{{ if ge (atoi .TokenUsagePercent.String) 75 }}p:red{{ end }}",
            "{{ if and (ge (atoi .TokenUsagePercent.String) 51) (lt (atoi .TokenUsagePercent.String) 75) }}p:yellow{{ end }}"
          ],
          "template": "\udb82\udfc9 {{ .Model.DisplayName }} {{ .TokenUsagePercent.Gauge }} {{ .TokenUsagePercent.String }}%"
        }
      ]
    }

What OS are you seeing the problem on?

Windows

Which shell are you using?

powershell

Log output

No errors, just enexpected behavior.

[JanDeDobbeleer]

@MelGrubb unfortunately we do not control when claude refreshes the statusline and injects new data. You could add --trace to see the logs inside oh-my-posh cache path (logs folder) to see if there are any errors.

[MelGrubb]

It's not that simple. I had Claude observe raw numbers and its best guess is that there are two problems. It appears that the gauge is adding InputTokens to OutputTokens and dividing by context size to come up with its value, but Claude's own calculation doesn't count output tokens. It does, however, include things like system prompt size, tools, agents, memory files, and skills, subtracting them as overhead from your overall context size. I don't know if those are visible to you or not, but they play into Claude's internal context usage calculations.

So, the gauge and string properties would need to be something more like (systemPrompt + systemTools + customAgents + memoryFiles + skills + messages) / ContextSize to come up with the same numbers as Claude does.

[JanDeDobbeleer]

@MelGrubb these are the docs, the answer is no.

[MelGrubb]

The docs show "used_percentage" as an option. I think that would put all the math back on Claude's shoulders. It had better be accurate at that point, I would think.

[MelGrubb]

I really think this needs to go back to being a bug. The "out of sync" issue is worse than simply lagging or leading. This is my status line this morning.

   Mel   PathToMyProject   CurrentBranch ≡  󰯉 Opus 4.5 ▱▱▱▱▱ 324% 

Obviously the fonts aren't going to work here, but you get the point. Despite multiple compactions and /clears, the token usage percent just keeps climbing forever. Whatever math this segment is doing, it's not working. I think using claude code's own "used_percentage" might be more correct. I can't see what you're seeing under the covers though.

[ahmetb]

It's fixed now in master.