From 8b403c54df729e255b574b6b0b7af691176d4cda Mon Sep 17 00:00:00 2001
From: seonNoh <dark.pearl.nst@gmail.com>
Date: Thu, 8 Oct 2026 10:53:08 +0900
Subject: [PATCH] fix(agent): record token usage summed over all turns
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

previous_response_id로 잇는 Responses API는 턴마다 전체 컨텍스트를 입력으로 과금하는데
세션에는 마지막 턴의 usage만 남아 비용이 과소 기록되었다(실측 2026-10-08: 12턴, 마지막 턴 입력
10,997). 턴별 input/output/total을 합산하고 턴 수도 함께 기록한다.
---
 ai/internal/agent/agent.go | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/ai/internal/agent/agent.go b/ai/internal/agent/agent.go
index 0715b99..149b58f 100644
--- a/ai/internal/agent/agent.go
+++ b/ai/internal/agent/agent.go
@@ -158,7 +158,9 @@ func (a *Agent) Diagnose(ctx context.Context, req Request) (*Result, error) {
 	var prevID string
 	var diagnosis *diag.Diagnosis
 	pf := false
-	var usage any
+	// usage is summed over every turn: with previous_response_id each turn re-bills the whole context, so the
+	// last response alone understates the cost (seen 2026-10-08: 12 turns, last turn 11K input tokens).
+	usage := map[string]int{"input_tokens": 0, "output_tokens": 0, "total_tokens": 0, "turns": 0}
 	for turn := 1; turn <= a.cfg.MaxTurns && diagnosis == nil; turn++ {
 		res.Turns = turn
 		r := openai.Request{Model: a.cfg.Model, Input: input, Tools: defs, ToolChoice: "auto", MaxOutputTokens: a.cfg.MaxOutputTokens, ParallelToolCalls: &pf}
@@ -173,7 +175,10 @@ func (a *Agent) Diagnose(ctx context.Context, req Request) (*Result, error) {
 			finish(err)
 			return nil, err
 		}
-		usage = resp.Usage
+		usage["input_tokens"] += resp.Usage.InputTokens
+		usage["output_tokens"] += resp.Usage.OutputTokens
+		usage["total_tokens"] += resp.Usage.TotalTokens
+		usage["turns"]++
 		prevID = resp.ID
 		fcs := resp.FunctionCalls()
 		if len(fcs) == 0 {