fix(hooks): prefer fresh harness cost cache (#2054)

Uses a fresh harness cost cache when available and keeps transcript pricing as the fallback. Focused cost-tracker tests passed locally before merge.

scripts/hooks/cost-tracker.js

@@ -20,15 +20,54 @@
  * Each row therefore represents the cumulative session total up to that point.
  * To get per-session cost, take the last row per session_id. To get per-day
  * spend, aggregate.
+ *
+ * Harness-cost contract (optional, opt-in by the statusline):
+ *   If the user's statusline (which receives `cost.total_cost_usd` directly
+ *   from Claude Code) writes `{ts, cost_usd}` to
+ *   `<os.tmpdir()>/harness-cost-<session_id>.json` on each render, this hook
+ *   prefers that authoritative value over the transcript-sum estimate when
+ *   the cache is fresh (≤ 300s). The transcript-sum is kept as a safe
+ *   fallback because:
+ *     - the hard-coded rate table cannot represent Opus 4.7's >200K-token
+ *       2x tier or the 1h-cache 2x tier (under-counts on long sessions);
+ *     - summing the full transcript double-counts work done across
+ *       `--resume` boundaries while `cost.total_cost_usd` is per-process.
+ *   Absent a writer, behavior is unchanged.
  */
 
 'use strict';
 
 const fs = require('fs');
+const os = require('os');
 const path = require('path');
 const { ensureDir, appendFile, getClaudeDir } = require('../lib/utils');
 const { sanitizeSessionId } = require('../lib/session-bridge');
 
+const HARNESS_COST_MAX_AGE_SECONDS = 300;
+
+/**
+ * Read authoritative harness cost from the per-session cache file.
+ * @param {string} sessionId
+ * @param {number} maxAgeSeconds
+ * @returns {number|null} cost in USD, or null on miss / stale / parse error
+ */
+function readHarnessCost(sessionId, maxAgeSeconds) {
+  if (!sessionId) return null;
+  try {
+    const fp = path.join(os.tmpdir(), `harness-cost-${sessionId}.json`);
+    if (!fs.existsSync(fp)) return null;
+    const obj = JSON.parse(fs.readFileSync(fp, 'utf8'));
+    const ts = Number(obj && obj.ts);
+    const cost = Number(obj && obj.cost_usd);
+    if (!Number.isFinite(ts) || !Number.isFinite(cost) || cost < 0) return null;
+    const age = Math.floor(Date.now() / 1000) - ts;
+    if (age < 0 || age > maxAgeSeconds) return null;
+    return cost;
+  } catch {
+    return null;
+  }
+}
+
 // Approximate per-1M-token billing rates (USD).
 // Cache creation: 1.25x input rate. Cache read: 0.1x input rate.
 const RATE_TABLE = {
@@ -125,13 +164,23 @@ process.stdin.on('end', () => {
     } = usageTotals || {};
 
     const rates = getRates(model);
-    const estimatedCostUsd = Math.round((
+    const transcriptCostUsd = Math.round((
       (inputTokens      / 1e6) * rates.in +
       (outputTokens     / 1e6) * rates.out +
       (cacheWriteTokens / 1e6) * rates.cacheWrite +
       (cacheReadTokens  / 1e6) * rates.cacheRead
     ) * 1e6) / 1e6;
 
+    // Prefer the harness's authoritative `cost.total_cost_usd` when the
+    // statusline has written it to the per-session cache (see contract in
+    // the file header). The harness number reflects API-billed truth
+    // (correct rates, 1h-cache 2x, >200K tier 2x) and is per-process so it
+    // does not drift across `--resume`. Cache miss → transcript-sum.
+    const harnessCost = readHarnessCost(sessionId, HARNESS_COST_MAX_AGE_SECONDS);
+    const estimatedCostUsd = harnessCost !== null
+      ? Math.round(harnessCost * 1e6) / 1e6
+      : transcriptCostUsd;
+
     const metricsDir = path.join(getClaudeDir(), 'metrics');
     ensureDir(metricsDir);