apps/codex/src/data-loader.ts (1)

295-306: ⚠️ Potential issue | 🟠 Major

Advance previousTotals when falling back to last_token_usage.

After Line 302 uses lastUsage, previousTotals is not advanced. If a later event resumes total_token_usage, Line 300 diffs against stale/null totals and re-emits already-counted usage. Example: a last=100 event followed by a total=150 event currently emits 100 and 150, not 100 and 50. Please also add a regression for a last-only -> total transition.

Suggested fix

+function addRawUsage(base: RawUsage | null, delta: RawUsage): RawUsage {
+	return {
+		input_tokens: (base?.input_tokens ?? 0) + delta.input_tokens,
+		cached_input_tokens: (base?.cached_input_tokens ?? 0) + delta.cached_input_tokens,
+		output_tokens: (base?.output_tokens ?? 0) + delta.output_tokens,
+		reasoning_output_tokens:
+			(base?.reasoning_output_tokens ?? 0) + delta.reasoning_output_tokens,
+		total_tokens: (base?.total_tokens ?? 0) + delta.total_tokens,
+	};
+}
...
 let raw: RawUsage | null = null;
 if (totalUsage != null) {
 	// Prefer cumulative totals when available. Codex can emit duplicate token_count
 	// snapshots (e.g. with different rate-limit buckets) where last_token_usage repeats.
 	// Diffing totals collapses those duplicates into zero-delta events.
 	raw = subtractRawUsage(totalUsage, previousTotals);
-} else {
-	raw = lastUsage;
-}
-
-if (totalUsage != null) {
 	previousTotals = totalUsage;
+} else {
+	raw = lastUsage;
+	if (lastUsage != null) {
+		previousTotals = addRawUsage(previousTotals, lastUsage);
+	}
 }

Based on learnings, "Use payload.info.total_token_usage as cumulative totals and payload.info.last_token_usage as per-turn delta; when only cumulative totals exist, compute delta by subtracting previous totals".

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@apps/codex/src/data-loader.ts` around lines 295 - 306, When falling back to
last_token_usage (the lastUsage branch) we must advance previousTotals so future
total_token_usage diffs against an updated cumulative baseline; update the
branch that sets raw = lastUsage to also set previousTotals = previousTotals ?
addRawUsage(previousTotals, lastUsage) : lastUsage (or equivalent) so the
per-turn delta is incorporated into previousTotals, and add a regression test
covering a last-only -> total transition (e.g., last=100 then total=150 should
emit 100 then 50). Reference variables/functions: lastUsage, totalUsage,
previousTotals, subtractRawUsage (and use or add an addRawUsage helper if
needed).