openclaw - 💡(How to fix) Fix [Bug]: Web UI msg-meta contextPercent shows 100% after long sessions due to _j() accumulating cacheRead across messages [1 comments, 2 participants]

Bug Description

After a long session (821 assistant messages, ~20 hours of conversation), the Web UI message metadata bar shows 100% ctx with a red/danger indicator. This makes it look like the 1M context window is full, but the actual context usage is only ~12.8%.

Root Cause

In control-ui/assets/index-ckUmEo1l.js, the _j() function accumulates usage from all assistant messages:

function _j(e,t){
  let n=0,r=0,i=0,a=0,o=0,s=null,c=!1;
  for(let{message:t}of e.messages){
    let e=t;
    if(e.role!==`assistant`)continue;
    let l=e.usage;
    l&&(c=!0,
      n+=l.input??...,           // ← accumulates across ALL messages
      i+=l.cacheRead??...,       // ← BUG: cacheRead is per-API-call cache hit size
      a+=l.cacheWrite??...);
  }
  let l=n+i+a,
    u=t&&l>0?Math.min(Math.round(l/t*100),100):null;
  return {..., contextPercent:u};
}

The issue is:

1. cacheRead is accumulated across all API calls: Each API response reports cacheRead as how many tokens were read from the cache for that specific call. After 821 calls, this sum reaches ~69M tokens.
2. The formula l = input + cacheWrite + cacheRead / contextTokens produces 7120% which gets capped to 100% by Math.min(..., 100).
3. kw() function (context bar at the top of the chat pane) correctly uses the session-level totalTokens and shows ~128K/1M = 13%.

Impact

Users see a misleading red "100% ctx" badge on every assistant message in a long session, suspecting the context window is full when it isnt.

Reproduce

1. Use a model with reasoning (e.g. deepseek-v4-pro) in a long conversation
2. Accumulate 200+ assistant messages with API usage data
3. Observe _j() returning contextPercent: 100 from summary stats
4. Check kw() context bar — shows correct lower value

Expected Behavior

_j() should either:

• Use the latest (current) API response usage instead of accumulating across all messages, OR
• Use the same session-level totalTokens approach as kw(), OR
• Remove cacheRead from the contextPercent denominator since it is already included in the per-call context sizing

Environment

• OpenClaw version: 2026.4.24
• Model: deepseek-v4-pro
• Session: 821 assistant messages, 67M cumulative API tokens