Local usage overcounts resumed sessions by treating inherited cumulative totals as new usage
評価
調査の方向性
Start in src-tauri/src/shared/local_usage_core.rs, where previous_totals is initialized per file and the parser prefers total_token_usage over last_token_usage. Reproduce the double count with the two-file example in the issue, then check how token_usage_record entries and their response_id could be read and deduplicated. Done means the continuation file yields 10 input tokens for the second day, and regression cases cover resumed files, duplicate records, and cached input.
索引モデルが issue の本文から書いたものです。
説明
Summary
The local usage scanner can count a resumed session's inherited cumulative token usage as new usage. This substantially inflates daily totals when one logical session spans multiple rollout files.
Environment and scope
- macOS, installed Codex Monitor version: 0.7.67.
- The same scanner logic is present in the latest fetched
main, commitdd61b9abd37de5ded86e82b9fe8a83fd49d46fa5(package version 0.7.68). - This report is based on source inspection and replaying the scanner's accounting algorithm against local rollout usage records. The numbers below are an algorithm reproduction, not a live Codex Monitor UI measurement or an invoice reconciliation.
Root cause
In src-tauri/src/shared/local_usage_core.rs, previous_totals is initialized to None for every file.
The parser prefers total_token_usage over last_token_usage, and subtracts a zero default for the first cumulative record.
A continuation file may start with the logical session's historical cumulative total. Subtracting zero attributes that entire history to the new record's date, even though most of it was already accounted for in earlier files. The scanner also does not currently consume token_usage_record or deduplicate those request records by response_id.
Minimal reproduction
For a legacy stream, place these usage events in two rollout files for the same logical session. Other metadata can follow the usual rollout format.
File A (previous day):
{"timestamp":"2026-10-08T10:00:00Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":1000,"cached_input_tokens":0,"output_tokens":0},"last_token_usage":{"input_tokens":1000,"cached_input_tokens":0,"output_tokens":0}}}}
File B (continuation on the next day):
{"timestamp":"2026-10-09T10:00:00Z","type":"event_msg","payload":{"type":"token_count","info":{"total_token_usage":{"input_tokens":1010,"cached_input_tokens":0,"output_tokens":0},"last_token_usage":{"input_tokens":10,"cached_input_tokens":0,"output_tokens":0}}}}
Expected usage for October 9: 10 input tokens. Current scanner result: 1,010 input tokens.
Impact on actual local data
Using the same cutoff, 2026-10-09 20:19:54 UTC+08:00, for both calculations:
| Accounting method | Input | Output | Total |
|---|---|---|---|
| Current scanner algorithm | 198,138,953 | 1,014,542 | 199,153,495 |
Unique token_usage_record.payload.usage records, deduplicated by response_id |
28,898,641 | 146,226 | 29,044,867 |
The request-based calculation contains 342 unique requests and includes 26,920,192 cached input tokens within the input total. Cached input is not added again. The current algorithm produces approximately 6.86 times the request-based total.
Suggested fix and regression coverage
- Prefer per-request
token_usage_record.payload.usagewhere available, with deduplication byresponse_idacross files. Count input plus output; cached input remains a subset of input, and reasoning output must not be added twice. - Avoid counting both a request usage record and its accompanying
token_countdisplay event. Matching solely by timestamp proximity can be unreliable because those events may be delayed. - Retain a conservative fallback for older logs: use valid per-request
last_token_usagewhen available, suppress repeated cumulative snapshots, and treat an initial cumulative-only record as a baseline rather than new usage. - Add regression cases for resumed files, duplicate request records across files, mixed request/display records, legacy cumulative resets, and cached input accounting.
The goal is accurate local request usage, including cached input, rather than interpreting session history as new consumption. It should still be distinguished from provider-side billing or account quota data.
- 主要言語
- TypeScript
- スター
- 4.3k
- フォーク
- 415
- PR マージ指標
- 30日以内にマージされた PR はありません
環境構築
このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
Dimillian/CodexMonitor のほかの issue
-
難易度 1/5 1時間未満 初心者へのやさしさ 78/100
Dimillian/CodexMonitor#590 ·
-
難易度 3/5 1〜2日 初心者へのやさしさ 63/100
Dimillian/CodexMonitor#615 ·
-
Codex Monitor does not surface MCP browser tool approval prompts, making browser tools appear hungオープン
難易度 4/5 3〜5日 初心者へのやさしさ 52/100
Dimillian/CodexMonitor#612 ·
-
難易度 4/5 3〜5日 初心者へのやさしさ 55/100
Dimillian/CodexMonitor#602 · コメント 1 件 · リアクション 1 件 ·
-
難易度 4/5 3〜5日 初心者へのやさしさ 38/100
Dimillian/CodexMonitor#599 ·
Dimillian/CodexMonitor の issue をすべて見る
似ている issue
-
bug
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
vercel/streamdown#673 · リアクション 1 件 ·
メンテナーはふだん 1 日以内に返信
-
docs(stream-protocol): document messageMetatadata on start/finish and the message-metadata chunkオープンfactory-active task-identify-harness-labels-done task-identify-issue-type-done
難易度 2/5 1〜3時間 初心者へのやさしさ 82/100
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
agilepathway/label-checker#721 ·
メンテナーはふだん 1 日以内に返信
-
CLI Empathy needs triage
難易度 2/5 1〜3時間 初心者へのやさしさ 74/100
chromaui/chromatic-cli#1524 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
メンテナーはふだん 1 日以内に返信