Context tokens and context percent are underreported

[polczak-itt]

I frequently see that the status line shows lower token count / usage percent in the status line compared to what Claude reports.

This can be tested by running some task which will occupy eg. 50% of the context window and then using the /context slash command.

Output of /context:

/context 
  ⎿  
      Context Usage
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   claude-sonnet-4-5-20250929 · 146k/200k tokens (73%)
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛀ ⛀ 
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ System prompt: 2.7k tokens (1.4%)
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ System tools: 14.2k tokens (7.1%)
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ MCP tools: 21.2k tokens (10.6%)
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ Custom agents: 400 tokens (0.2%)
     ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁ ⛁   ⛁ Memory files: 5.6k tokens (2.8%)
     ⛁ ⛁ ⛁ ⛁ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛁ Messages: 101.7k tokens (50.8%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶   ⛶ Free space: 54k (27.1%)
     ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ 

Status line:

Ctx: 131.9k | Ctx: 66.0%

This is discrepancy is visible even though Claude paused, was waiting for a new prompt and the /context was processed. This means that even if the status line does not update IRL there were already at least two points in time for it to update / correct itself.

It's also interesting that the difference between the reported values is really similar to: System tools: 14.2k tokens (7.1%). My hypothesis is that maybe it's not being counted. Will report back if further testing confirms it.

[sirmalloc]

Sadly there's no easy way to fix this. Claude Code does not provide the context usage to us in the JSON passed to the statusline, we have to parse it out of the JSONL for the session transcript and calculate it ourselves. If there are discrepancies between our calculation and the value shown in /context, there's not much we can do about it. Our calculations are simple: input tokens + cache read tokens + cache create tokens from the most recent entry in the JSONL that has a usage property on it and that has isSidechain set to false. We don't get a granular breakdown like /context gives. When the user executes /context an API call is made to Claude's backend that returns the breakdown for the current session.

I will add that as context usage approaches the context window limit, the numbers seem to converge and the difference trends towards zero. For an actual proper fix to this, I'd suggest filing a bug with Anthropic to provide that data to statuslines via the JSON they get via stdin so they can be as accurate as possible.