Greptile Summary

This PR fixes a display bug where the context-window percentage indicator showed ~0% during heavily-cached sessions. The Anthropic API only puts uncached tokens in input; cached tokens come back in cacheRead/cacheWrite. By computing totalInput = input + cacheRead + cacheWrite and using that for the percentage, the indicator now reflects real context utilisation.

• Root cause fixed: contextPercent was computed from input alone, which can be a small fraction of the total when prompt caching is active.
• Minimal scope: only the contextPercent calculation changes; the raw input, cacheRead, and cacheWrite values returned in GroupMeta are unchanged, preserving the existing token-breakdown display.
• Edge cases handled: token counters are already initialised to 0 and guarded with ?? 0 on accumulation, so totalInput is always a valid number; the totalInput > 0 guard prevents a division-by-zero when no usage is available.
• No concerns found: the fix is correct, well-commented, and non-breaking.

Confidence Score: 5/5

• This PR is safe to merge — it is a one-expression bug fix with no side-effects on data flow, rendering, or API contracts.
• The change is small, well-reasoned, and correct. Token counters are already zero-initialised and null-guarded, so no new NaN/undefined risk is introduced. The fix only affects contextPercent and does not mutate any other part of the returned GroupMeta, keeping the rest of the UI unchanged.
• No files require special attention.

_{Last reviewed commit: 8541ed1}

Closing this as implemented after Codex review.

Current main already includes the requested behavior. ui/src/ui/chat/grouped-render.ts:427 computes prompt/context usage as input + cacheRead + cacheWrite before deriving contextPercent, matching this PR’s change, and ui/src/ui/chat/grouped-render.test.ts:263 covers the cached-token case with an assertion for 44% ctx.

What I checked:

• Main code already sums cached tokens: extractGroupMeta uses const promptTokens = input + cacheRead + cacheWrite; and computes contextPercent from promptTokens, which is the same semantic fix proposed by PR fix: include cached tokens in context window % calculation #48827. (ui/src/ui/chat/grouped-render.ts:427, 01bc49c88c0c)
• Regression test exists on main: There is a targeted UI test named includes cache tokens when rendering assistant context usage asserting cached token counts are included and that the rendered context display is 44% ctx. (ui/src/ui/chat/grouped-render.test.ts:263, 01bc49c88c0c)
• Blame points to existing main commit: git blame attributes the implemented promptTokens/contextPercent lines to commit 8fade9df27a9a25cd017db1154409b3ad9bd3741, showing the behavior is already present on current main independently of this open PR. (ui/src/ui/chat/grouped-render.ts:427, 8fade9df27a9)
• Checkout remained clean: Repository status was clean after inspection. (01bc49c88c0c)

So I’m closing this as already implemented rather than keeping a duplicate issue open.

Review notes: reviewed against 01bc49c88c0c; fix evidence: commit 8fade9df27a9.