Greptile Summary

This PR fixes two related math bugs: the ctx% context-window badge was dividing only input tokens by contextWindow (leaving it near 0% when prompt caching is warm), and the cache-hit-rate denominator was omitting cacheWrite tokens (inflating the ratio). Both expressions now use the full prompt-token sum (input + cacheRead + cacheWrite), which is the correct base in both cases. The English tooltip is updated to match.

• The one outstanding gap is that pnpm ui:i18n:sync was not run — 12 non-English locale files retain the old formula in their cacheHint tooltip string. Per ui/CLAUDE.md, regenerating and committing the locale bundles is a required pipeline step when en.ts is edited.

Confidence Score: 4/5

Safe to merge after running pnpm ui:i18n:sync and committing the regenerated locale files.

All logic changes are correct. The only gap is the missing i18n sync step mandated by ui/CLAUDE.md — non-English locale tooltips will show the old formula until regenerated. This is a process step that should be completed before landing per project rules.

ui/src/i18n/locales/en.ts — missing pnpm ui:i18n:sync to regenerate the 12 non-English locale bundles.

This is a comment left during a code review.
Path: ui/src/i18n/locales/en.ts
Line: 514

Comment:
**Missing `pnpm ui:i18n:sync` run**

`ui/CLAUDE.md` specifies that after editing `en.ts`, the pipeline step `pnpm ui:i18n:sync` must be run and the regenerated locale bundles committed. The 12 non-English locale files (`de.ts`, `es.ts`, `fr.ts`, etc.) still contain the old `cacheHint` formula. The tooltip mismatch is cosmetic, but the process rule is explicit — regenerate and commit before landing.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "fix(ui): include cache tokens in ctx% ba..." | Re-trigger Greptile}

Closing this as duplicate or superseded after Codex automated review.

PR #68557 is superseded by merged PR #71462. Current main contains the requested Control UI runtime fixes and regression coverage for both cache-token-aware ctx% and cache-hit-rate math. The latest release tag predates the Usage overview denominator fix, so there is no shipped release to cite yet.

Best possible solution:

Close #68557 as superseded by merged PR #71462 and keep the current main implementation/tests. Track the remaining non-English cacheHitRate tooltip drift separately if maintainers want localized tooltip parity.

What I checked:

• Superseding PR: Provided GitHub context shows PR fix(ui): use current context usage in Control UI #71462 was merged on 2026-04-25, explicitly related fix(ui): include cache tokens in ctx% badge and cache-hit-rate #68557, and included current context usage plus cache-write tokens in the Usage overview cache-hit-rate denominator. (4282c591775b)
• Ctx badge implementation: Current main accumulates input, cacheRead, and cacheWrite per assistant call and uses their sum as maxPromptTokens for contextPercent, so warm prompt-cache turns are counted in the ctx badge. (ui/src/ui/chat/grouped-render.ts:442, d54d2d6b9b8a)
• Cache-hit-rate implementation: Current main computes cacheBase as totals.input + totals.cacheRead + totals.cacheWrite before deriving cacheHitRate and the displayed prompt-token base. (ui/src/ui/views/usage-render-overview.ts:500, d54d2d6b9b8a)
• English tooltip updated: The English cacheHitRate tooltip now documents the denominator as input + cache read + cache write. (ui/src/i18n/locales/en.ts:538, d54d2d6b9b8a)
• Regression tests: Usage overview test expects 300 cache-read tokens over 100 input + 300 read + 600 write tokens to render 30.0%, and grouped-render tests expect cached prompt tokens to produce a 44% ctx badge. (ui/src/ui/views/usage-render-overview.test.ts:44, d54d2d6b9b8a)
• Release boundary: Latest release tag v2026.4.24 still has the old Usage overview cacheBase formula without cacheWrite, so the fix is current-main evidence rather than a released-version close. (ui/src/ui/views/usage-render-overview.ts:500, cbcfdf62c729)

So I’m closing this here and keeping the remaining discussion on the canonical linked item.

Codex Review notes: model gpt-5.5, reasoning high; reviewed against d54d2d6b9b8a; fix evidence: commit 4282c591775b.