feat: decompose the token "in" headline into fresh vs cached

CHANGELOG.md

@@ -5,6 +5,16 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+### Added
+
+- **Token breakdown: decompose the "in" headline into fresh vs cached.** The headline
+  "NN in" is dominated by cheap `cache_read` tokens; a new `Input  X new · Y cached`
+  line separates freshly-billed input (`input + cache_creation`) from cache reads, so the
+  cost line is interpretable (a big "in" at 97% cache is mostly $1.50/M reads, not $15/M
+  fresh input). Applied to the single-session and live views.
+
 ## [1.0.0] - 2026-02-18
 
 ### Added

src/formatter.test.ts

@@ -413,3 +413,19 @@ describe('formatter: insights', () => {
     expect(output).toContain('Warmup overhead');
   });
 });
+
+describe('formatSession: input decomposed into fresh vs cached', () => {
+  it('shows freshly-billed input (input + cache writes) and cache reads separately', () => {
+    // input 5K + cacheCreation 10K = 15K fresh; cacheRead 50K cached.
+    const output = strip(formatSession(makeAnalysis()));
+    expect(output).toContain('Input');
+    expect(output).toContain('15.0K new');
+    expect(output).toContain('50.0K cached');
+  });
+
+  it('the headline "in" remains the full context total (fresh + cached)', () => {
+    // input+cacheRead+cacheCreation = 65K — the line a user sees as scale.
+    const output = strip(formatSession(makeAnalysis()));
+    expect(output).toContain('65.0K in');
+  });
+});

src/formatter.ts

@@ -162,6 +162,12 @@ export function formatSession(analysis: SessionAnalysis): string {
     const inBar = Math.round(inFrac * BAR_WIDTH);
     const outBar = BAR_WIDTH - inBar;
     lines.push(` Tokens  ${chalk.cyan('\u2588'.repeat(inBar))}${chalk.green('\u2588'.repeat(outBar))}  ${chalk.cyan(formatTokens(tokens.input + tokens.cacheRead + tokens.cacheCreation) + ' in')} ${chalk.green(formatTokens(tokens.output) + ' out')}`);
+    // Decompose "in": the headline is dominated by cheap cache reads. Show the
+    // freshly-billed input (input + cache writes) separately so the cost line
+    // is interpretable (a huge "in" at 97% cache is mostly $1.50/M reads, not
+    // $15/M fresh input).
+    const freshIn = tokens.input + tokens.cacheCreation;
+    lines.push(` Input   ${chalk.gray(formatTokens(freshIn) + ' new')} ${chalk.gray('\u00b7')} ${chalk.gray(formatTokens(tokens.cacheRead) + ' cached')}`);
   }
   const cachePct = Math.round(analysis.cacheHitRate * 100);
   lines.push(` Cache   ${renderCacheBar(analysis.cacheHitRate)}  ${cachePct}% hit`);
@@ -263,6 +269,8 @@ export function formatSessionLive(analysis: SessionAnalysis): string {
     const inBar = Math.round(inFrac * BAR_WIDTH);
     const outBar = BAR_WIDTH - inBar;
     lines.push(` Tokens  ${chalk.cyan('\u2588'.repeat(inBar))}${chalk.green('\u2588'.repeat(outBar))}  ${chalk.cyan(formatTokens(liveTok.input + liveTok.cacheRead + liveTok.cacheCreation) + ' in')} ${chalk.green(formatTokens(liveTok.output) + ' out')}`);
+    const liveFreshIn = liveTok.input + liveTok.cacheCreation;
+    lines.push(` Input   ${chalk.gray(formatTokens(liveFreshIn) + ' new')} ${chalk.gray('\u00b7')} ${chalk.gray(formatTokens(liveTok.cacheRead) + ' cached')}`);
   }
   const liveCachePct = Math.round(analysis.cacheHitRate * 100);
   lines.push(` Cache   ${renderCacheBar(analysis.cacheHitRate)}  ${liveCachePct}% hit`);