anthropics / anthropics/claude-code

Desktop app: compact context indicator silently changed meaning — now tracks plan usage, not context window

Offen Anfängerfreundlich
#91,075 5 Kommentare 4 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen
area:desktop area:ui bug platform:windows
Vorherrschende Sprache
Python
Sterne
145k
Forks
23.1k
PR-Merge-Kennzahlen
PR-Kennzahlen ausstehend

Beschreibung

### Documentation Type

Missing documentation (feature not documented)

### Documentation Location

_No response_

### Section/Topic

compact context indicator

### Current Documentation

_No response_

### What's Wrong or Missing?

## Environment
- **Platform:** Windows 11 Pro (desktop app)
- **CLI core:** 2.1.72 (`claude --version`); desktop app build:
- **Plan:** Max (20x)

## Summary
The small circular gauge at the bottom-right of the chat input (next to the model/plan
chip) previously visualized **context-window fill %**. It now appears to visualize a
**plan usage-limit %** instead (matching "Weekly · " in the expanded panel).
The change is not mentioned in the changelog, and no setting toggles which metric the
compact indicator shows.

## Observed
Same session, two data points:

1. Expanded panel read **Context window: 806.8k / 1M (81%)** — the circular icon showed
~44–50% filled, matching **Weekly · Fable: 44%**, not the 81% context figure.
2. After a manual transcript compaction, the panel read **160.6k / 1M (16%)** — the
icon was **unchanged** (still ~44%). Weekly usage of course doesn't decrease, which
is consistent with the icon now tracking usage rather than context.

In earlier builds, the same icon visibly shrank when context was reduced by manual
compaction — that at-a-glance signal is what regressed/changed.

## Expected
Either of:
- the compact indicator reflects context-window fill (previous behavior), or
- a documented setting chooses which metric it shows, or
- at minimum, the behavior change is documented in the changelog.

## Impact
Workflows that manage context **manually** (compacting deliberately rather than relying
on auto-compact) lose the only always-visible context signal. In this instance the icon
reading "~under 50%" while actual context was at 81% nearly caused a missed manual
compaction — the mismatch was only caught by opening the expanded panel. A gauge whose
meaning changes silently while its appearance stays identical is worse than no gauge:
operators calibrated to the old meaning read it with confidence.

I understand usage limits may be the more relevant metric for users who rely on
auto-compact — but a toggle (or a documented statusline-style option for the desktop
app) would serve both audiences. Related: #28962 (context-window visibility).

## Notes
- The circular indicator itself appears to be undocumented; the statusline docs expose
both `context_window.used_percentage` and `rate_limits.*.used_percentage`
programmatically, but the desktop app's built-in indicator has no documented contract.

### Suggested Improvement

Either of:
- the compact indicator reflects context-window fill (previous behavior), or
- a documented setting chooses which metric it shows, or
- at minimum, the behavior change is documented in the changelog.

### Impact

Medium - Makes feature difficult to understand

### Additional Context

_No response_

Beitragsleitfaden

Für dieses Repository ist kein Beitragsleitfaden indexiert

Rechercherichtung

Start with the statusline documentation and changelog, the documentation surfaces identified in the issue, and compare their descriptions with the desktop app's compact indicator and expanded panel values. Done means clearly documenting which metric the indicator represents and the behavior change, or documenting an agreed setting if maintainers choose to provide one.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Bewertung

Bereich
documentation
Issue-Typ
Dokumentation
Schwierigkeit
2/5
Geschätzter Aufwand
1-3 Stunden
Aktivitätsstatus
Aktiv
Klarheit
Größtenteils klar
Anfängerfreundlichkeit
62/100

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.