Summary

On current origin/master (v0.50.36, 37850a4), the WebUI context indicator can prefer stale per-session metadata over fresher usage data, and older session/index entries may also lack the usage fields needed to render the indicator correctly.

In practice this shows up after provider/model changes or after reopening older sessions:

• the active provider/model is correct
• the backend is healthy
• but the UI can still show an old context window (for example 128k) or stale usage values from the stored session object

This is related to stale-session UI state, but distinct from the already-open stale-model issues.

Why this still appears on origin/master

In static/sessions.js on origin/master, loadSession(...) still merges context usage like this:

const _s=S.session;
if(_s&&typeof _syncCtxIndicator==='function'){
  const u=S.lastUsage||{};
  _syncCtxIndicator({
    input_tokens:_s.input_tokens||u.input_tokens||0,
    output_tokens:_s.output_tokens||u.output_tokens||0,
    estimated_cost:_s.estimated_cost||u.estimated_cost,
    context_length:u.context_length||0,
    last_prompt_tokens:u.last_prompt_tokens||0,
    threshold_tokens:u.threshold_tokens||0
  });
}

Problems with this shape:

1. It mixes session-level and latest-usage fields inconsistently.
2. It uses ||, so valid falsy values are not preserved correctly.
3. Older session/index entries may not contain context_length, threshold_tokens, or last_prompt_tokens at all.
4. There is no single latest-first merge helper for usage payloads.

Reproduction pattern

1. Use WebUI with provider/model A and create a session.
2. Switch provider/model (for example to openai-codex / gpt-5.4) or reopen an older session created under a different provider/model path.
3. Confirm backend state is correct via /api/models.
4. Load the session in the UI.
5. Observe that the context indicator can still reflect stale values from the session snapshot rather than the latest usage/model path.

A real observed symptom was an old 128k indicator surviving even after provider/model had been corrected, because the UI merge favored stored session state over fresher usage information.

Expected behavior

When loading a session, the UI should prefer the freshest usage payload over stale session metadata for:

• input_tokens
• output_tokens
• estimated_cost
• context_length
• last_prompt_tokens
• threshold_tokens

Also, session/index compaction should preserve these fields so older session loads do not silently drop them.

Suggested fix

A local patch that resolved this used two pieces:

1. Latest-first usage merge in static/sessions.js

function _pickUsageValue(primary, fallback, defaultValue=0){
  if(primary!==undefined&&primary!==null) return primary;
  if(fallback!==undefined&&fallback!==null) return fallback;
  return defaultValue;
}

function _mergedCtxUsage(sessionUsage, latestUsage){
  const s=sessionUsage||{};
  const u=latestUsage||{};
  return {
    input_tokens:_pickUsageValue(u.input_tokens,s.input_tokens,0),
    output_tokens:_pickUsageValue(u.output_tokens,s.output_tokens,0),
    estimated_cost:_pickUsageValue(u.estimated_cost,s.estimated_cost,null),
    context_length:_pickUsageValue(u.context_length,s.context_length,0),
    last_prompt_tokens:_pickUsageValue(u.last_prompt_tokens,s.last_prompt_tokens,0),
    threshold_tokens:_pickUsageValue(u.threshold_tokens,s.threshold_tokens,0),
  };
}

Then in loadSession(...):

_syncCtxIndicator(_mergedCtxUsage(_s,S.lastUsage));

2. Preserve usage fields in session/index models

The same local fix also added these fields to the session model / compacted session payloads in api/models.py and populated them from the context compressor in api/streaming.py:

• context_length
• threshold_tokens
• last_prompt_tokens

Related issues

Potentially related but not the same root bug:

• When a user switches the model via the WebUI during an active chat session, the UI updates to show the new model, but the agent continues processing messages using the original model from session start. #419 active chat continues using the original model after UI model switch
• Stale active session model can break chats after switching to a custom endpoint #433 stale active session model can break chats after switching to a custom endpoint

This issue is specifically about the context/usage indicator and stale session usage metadata on load.