feat(ui): show last-turn input + cache hit % in context panel (instead of cumulative)

[lxistired]

Problem

The context panel currently displays input/output token counts as cumulative sums across the entire session. After ~10 turns of normal usage, the input number reads "hundreds of thousands of tokens" — even though the cached prefix is being reused at deep discount on each turn. Users panic and assume they're being billed full rate every turn.

The context-usage bar has the inverse problem: it uses only tokenUsage.input (uncached portion), so the bar visually shrinks when a cache hit lands. But the prompt is still that big — the model still has to attend to all of it. Users see a session that feels "lighter" right when it gets fuller.

Fix

Switch the display to last-turn input + last-turn output + overall cache-hit percentage. That honestly reflects what the user is paying for.

Three changes:

1. types/index.ts — TokenUsage gains optional cacheRead / cacheWrite fields. The provider/SDK already produces these; the UI was discarding them.

2. agent-runner.ts normalizeTokenUsage — extract cacheRead / cacheWrite from common provider field-name variants:

   • Anthropic: cache_read_input_tokens / cache_creation_input_tokens
   • OpenAI Responses: prompt_tokens_details.cached_tokens
   • pi-ai-core normalized: cacheRead / cacheWrite

3. ContextPanel.tsx —

   • tokenUsage now exposes lastInput + lastOutput as input/output, plus aggregates (totalCacheRead, cacheHitRate) for the header.
   • contextUsage bar uses lastInput + lastCacheRead (true prompt size), so the bar doesn't shrink when a cache hit lands.
   • Header shows cache N% inline next to input when totalCacheRead > 0; absent for non-cache providers (existing behavior preserved).

Files

• src/renderer/types/index.ts (+4)
• src/renderer/components/ContextPanel.tsx (+47/-18)
• src/main/claude/agent-runner.ts (normalizeTokenUsage only, +24/-3)

Test plan

• tsc --noEmit passes
• First message: input/output reflect that single turn (no aggregation)
• After 10 turns with high cache hit rate: input shows current-turn prompt + 'cache 90%' style label, NOT 1.5M
• Context bar % grows with chat length (no shrink on cache hit)
• Provider with no cache info (e.g. Ollama, OpenAI Chat-Completions w/o caching): no 'cache N%' label, behavior identical to before

[github-actions]

Findings

• [Major] TokenUsage.input is being treated as "uncached-only" in the UI, but normalizeTokenUsage() preserves provider-native semantics. That makes the new panel wrong in both directions: OpenAI-style usage already includes cached prompt tokens in the prompt total, while Anthropic-style usage splits cache writes into cache_creation_input_tokens, which the new input + cacheRead calculation never adds back. Evidence src/main/claude/agent-runner.ts:395-417, src/renderer/components/ContextPanel.tsx:117-119, src/renderer/components/ContextPanel.tsx:142-149, src/renderer/types/index.ts:95-98.
  Suggested fix:

  const cacheRead = typeof cacheReadCandidate === 'number' ? cacheReadCandidate : 0;
  const cacheWrite = typeof cacheWriteCandidate === 'number' ? cacheWriteCandidate : 0;
  const baseInput = raw.input ?? raw.input_tokens ?? raw.inputTokens;
  
  const normalizedInput =
    raw.prompt_tokens_details !== undefined
      ? (baseInput as number)
      : (baseInput as number) + cacheRead + cacheWrite;
  
  return { input: normalizedInput, output, cacheRead, cacheWrite };

  Then use u.input directly in ContextPanel for both the header input value and contextUsage.used.

Summary

• Review mode: initial. 1 issue found. CLAUDE.md was not found in repo/docs. No PR-side regression tests were added for provider-specific token accounting in the context panel.

Testing

• Not run (automation). Suggested tests: Anthropic cache-write first turn, Anthropic cache-read follow-up turn, and OpenAI cached prompt totals.

Open Cowork Bot

[github-actions]

[MAJOR] ContextPanel now assumes tokenUsage.input is the uncached tail and rebuilds prompt size as input + cacheRead, but normalizeTokenUsage() is still passing provider-native values through unchanged. That breaks in both directions: OpenAI-style usage already includes cached prompt tokens in the prompt total, while Anthropic-style usage splits cache writes into cache_creation_input_tokens, which this calculation never adds back.

Suggested fix:

const cacheRead = typeof cacheReadCandidate === 'number' ? cacheReadCandidate : 0;
const cacheWrite = typeof cacheWriteCandidate === 'number' ? cacheWriteCandidate : 0;
const baseInput = raw.input ?? raw.input_tokens ?? raw.inputTokens;

const normalizedInput =
  raw.prompt_tokens_details !== undefined
    ? (baseInput as number)
    : (baseInput as number) + cacheRead + cacheWrite;

return { input: normalizedInput, output, cacheRead, cacheWrite };

Then use u.input directly in ContextPanel for both the header input value and contextUsage.used.