Orchestrator should own decisions/log.md writes; subagents return drafted entries, never write directly

オープン
#21 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

まだ誰も着手していません。

評価

難易度
3/5
見積もり時間
1〜2日
初心者へのやさしさ
68/100
issue の種類
リファクタリング
明瞭さ
おおむね明確
活発さ
活発
領域
tooling

調査の方向性

Start by reading the ace-protocol skill, the ace agent definition, and the ACE synthesis protocol section in user-testing SKILL.md. Trace every instruction that lets subagents read or write decisions/log.md, then audit for other shared files with concurrent writers. Done means the orchestrator is the sole writer, subagents return drafted entries, and serialized appends assign numbers without collisions.

索引モデルが issue の本文から書いたものです。

説明

Problem

During the Pass 3 user-testing battery (2026-09-28), 10 chapters' worth of ACE synthesis agents ran concurrently (up to ~20 at once), each independently tasked with checking decisions/log.md's current tail, picking the next free DL number, and appending its own entry.

This raced badly. Over the course of one battery we hit at least four real collisions: two, then three, then four separate entries independently landed on the same DL number (DL-054, then DL-056, then DL-058), because each agent's "read the tail, then write" sequence wasn't atomic against the others doing the same thing concurrently. The orchestrator had to do several rounds of manual detection (grep -o "^## DL-[0-9]*" decisions/log.md | sort | uniq -c) and renumbering, and even a "renumber it now" fix sometimes landed after a new concurrent collision had already appeared at the newly-assigned number.

The eventual workaround (switching later ACE agents to return their fully-drafted entry text with a placeholder number, rather than writing to the file at all, so the orchestrator does one single serialized append) worked cleanly and produced zero further collisions. But it was discovered mid-battery, not planned for.

Proposed fix

Make this the standing rule, not a one-off workaround: the orchestrator (or a single designated writer) is the only actor that ever writes to decisions/log.md. Every subagent — ACE synthesis, builder, reviewer, whatever role — composes its DL entry text as part of its report and returns it; it never reads the file's tail to pick a number, and never edits the file itself. The orchestrator serializes all appends, assigning numbers in the order entries actually arrive.

This likely means:

  • Updating the ace-protocol skill and the ace agent definition to state this as a hard rule (some agent definitions in this session already say "the orchestrator numbers and commits the log entry" — that text exists but wasn't consistently followed, or wasn't present in every agent's launch context).
  • Possibly updating user-testing SKILL.md's "ACE synthesis protocol" section, which currently has the ACE both drafting and logging.
  • Considering whether this generalizes to any other file multiple concurrent subagents might write to (none identified yet beyond decisions/log.md, but worth a quick audit).

Not urgent

This didn't cause any lost work or wrong content — every collision was caught and cleanly resolved before commit — but it cost real orchestrator turns doing manual dedup/renumbering that a structural fix would avoid entirely. Deferred for now per Z's own note; revisit before the next large parallel-agent battery (e.g. a future Pass 3 run, or a similarly large fan-out).

主要言語
Jupyter Notebook
スター
0
フォーク
0
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

  • Dockerfile・Docker Compose ファイルなし
  • プルリクエストのテンプレートあり
  • コントリビューションガイドなし

はじめの一歩

  1. 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。
  2. 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。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

Open-MBEE/toaster のほかの issue

Open-MBEE/toaster の issue をすべて見る

似ている issue

DevTools の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。