Coding agents have complementary strengths. In our own day-to-day work, Codex is often the stronger code generator, while Claude Code is remarkably good at reconstructing intent from terse instructions and repository context. Fusuma's /sub-agent skill lets you use both in a single flow: Claude Code sits in front as the interpreter and orchestrator, and delegates the heavy lifting to a second agent — Codex or another Claude — running right below the chat.
From four words to a complete brief
Here is the entire instruction we gave Claude Code:
Plz review diff with codex /sub-agent
Claude Code checks whether a sub-agent pane already exists, launches a fresh Codex panel below the chat, and composes the actual task prompt from the conversation and the state of the repository:
This is what Codex actually received:
Code-review the diff of commit 6cf72f9 in this repo (git show 6cf72f9). It's a fix commit in the Fusuma Tauri app (Rust backend in desktop/, React+TS frontend in src/): off-screen vt100 parser replacing a raw PTY byte ring for widget read, rate-limit snapshot invalidation on account/plan changes, xterm terminal focus/scroll fixes, hidden-widget PTY sizing, and a taller resize floor for chat/terminal widgets. Look for correctness bugs, race conditions, resource leaks (especially vt100 parser / PTY lifecycle), and edge cases in the focus/scroll handling. Output concrete findings with file:line references ranked by severity, or an explicit all-clear.
The target commit, a summary of the changes, the review focus areas, and the expected output format were all filled in from context — none of it was in the original four-word input.
The run is synchronous, the follow-up is automatic
After dispatching the brief, Claude Code waits for the sub-agent to finish, reads the result, and continues on its own — for a code review, that means triaging the findings and applying fixes without another round trip through you. The sub-agent pane stays open in the workspace, so you can also pick up the Codex session directly and keep working in it.
Both budgets stay on screen
Delegation consumes tokens on both sides, so Fusuma keeps the 5-hour and weekly rate limits for Claude and Codex visible in the top bar at all times. When one provider is running low, hand the next task to the other — the decision takes a glance, not a settings page.
Try it
Open any AI widget and end an instruction with /sub-agent — "review this diff with codex", "have claude write the tests for this". The orchestrating agent picks the target, writes the brief, runs it, and reports back. Short prompts stay short; the specificity is generated where it belongs.
Coding Agent には得意分野の違いがあります。私たち自身の開発では、コード生成は Codex が強い一方、短い指示とリポジトリの文脈から意図を復元する力は Claude Code が際立っています。Fusuma の /sub-agent skill は、この両方をひとつの流れで使うための機能です。Claude Code が前段で解釈と指揮を担い、実作業をチャットの下で動く 2 つ目のエージェント——Codex またはもうひとつの Claude——へ委譲します。
4 語の入力から、完全な指示書へ
Claude Code に与えた指示の全文がこちらです。
Plz review diff with codex /sub-agent
Claude Code は既存のサブエージェントの有無を確認し、チャットの下に新しい Codex パネルを起動して、会話とリポジトリの状態から実際のタスクプロンプトを組み立てます。
Codex が実際に受け取った内容がこちらです。
Code-review the diff of commit 6cf72f9 in this repo (git show 6cf72f9). It's a fix commit in the Fusuma Tauri app (Rust backend in desktop/, React+TS frontend in src/): off-screen vt100 parser replacing a raw PTY byte ring for widget read, rate-limit snapshot invalidation on account/plan changes, xterm terminal focus/scroll fixes, hidden-widget PTY sizing, and a taller resize floor for chat/terminal widgets. Look for correctness bugs, race conditions, resource leaks (especially vt100 parser / PTY lifecycle), and edge cases in the focus/scroll handling. Output concrete findings with file:line references ranked by severity, or an explicit all-clear.
対象コミットの特定、変更内容の要約、レビューの観点、期待する出力形式——すべて文脈から補完されたもので、元の 4 語の入力には含まれていません。
実行は同期的に、後続は自動で
指示を送った後、Claude Code はサブエージェントの完了を待ち、結果を読んで自律的に作業を続けます。コードレビューであれば、指摘の選別から修正の適用まで、あなたを経由せずに進みます。サブエージェントのパネルはワークスペースに残るので、Codex のセッションを直接引き継いで作業することもできます。
両方の残量が常に見える
委譲は両側のトークンを消費します。そのため Fusuma では、Claude と Codex それぞれの 5 時間／週間の rate limit をトップバーに常時表示しています。一方の残量が少なくなってきたら、次のタスクをもう一方へ——設定画面を開かず、一瞥で判断できます。
使ってみる
任意の AI ウィジェットで、指示の末尾に /sub-agent を付けてください。「review this diff with codex」「have claude write the tests for this」——指揮側のエージェントが委譲先を選び、指示書を書き、実行し、結果を報告します。プロンプトは短いまま。具体性は、それを書くべき側が生成します。