skills/task-subagent-delegation: prohibit recursive subagent spawn#1529
Merged
lipluscodex merged 1 commit intoJul 10, 2026
Hidden character warning
The head ref may contain hidden characters: "1528-task-subagent-delegation-\u518d\u5e30\u7684\u30b5\u30d6\u30a8\u30fc\u30b8\u30a7\u30f3\u30c8\u5897\u6b96\u306e\u9632\u6b62\u30eb\u30fc\u30eb\u660e\u6587\u5316"
Merged
Conversation
Adds a "Bounded delegation: prohibit recursive subagent spawn" section to skills/task-subagent-delegation/SKILL.md. On 2026-07-06, 3 parallel general-purpose subagents delegated a bounded read-only audit task each spawned their own nested Agent-tool children unprompted, cascading into API rate-limit exhaustion with almost no usable output. The root cause was the delegation prompt never stating "do this directly, do not spawn further subagents." 2026-07-06 に実際に観測された再帰的サブエージェント増殖インシデントを踏まえ、 委譲 prompt 作成時に孫サブエージェント生成を明示的に禁止するルールを追加した。 brake 1 (N=3 evaluator) の指摘を反映し、skill description のスコープ拡張と procedure/tool-authority-bound の切り分け一文も併せて追加。 Closes #1528
lipluscodex
commented
Jul 10, 2026
lipluscodex
left a comment
Member
Author
There was a problem hiding this comment.
Self-review (AI)
- 変更範囲確認:
skills/task-subagent-delegation/SKILL.mdのみ、L3-task layer。L1 Model Layer 非該当のため brake 2 対象外。 - brake 1 (
skills/evolution-parallel-agent-eval, N=3, opus, M=all axes) 実施済み。結果は PR body に記載の通り、全評価者 mergeable 判定。Axis A/B の軽微指摘 2点 (description スコープ拡張 / procedure との切り分け一文) は反映済み。Axis C (impression-literal detection) は 3/3 none flagged。 - CI (governance) 2件 pass 確認済み。
- patch 判定の exception 根拠: 委譲 prompt hygiene の内部ルール追加であり、user/system から観測可能な大規模構造変更ではない (
rules/operations/release-version-rule.mdpatch 基準に該当)。semi_auto の per-PR exception を適用し、human check を待たず AI 直接 merge とする。
Approve for merge.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #1528
変更内容
skills/task-subagent-delegation/SKILL.mdに「Bounded delegation: prohibit recursive subagent spawn」セクションを新規追加。委譲 prompt 作成時、Agent tool を持つ subagent (general-purpose等) へ bounded read-only investigation (audit / consistency check 等) を委譲する場合、「これ以上 subagent を生成するな」という明示を求めるルール。併せて:
背景
2026-07-06、scheduled evolution タスクの full-refactor audit ステップで、read-only な監査作業を 3 並列の
general-purposesubagent に委譲したところ、委譲 prompt に Agent tool 禁止の明示が無かったため 3/3 全てが自発的に孫 Agent を生成し、一部はさらに孫の孫まで生成。15+ のバックグラウンドタスク通知が発生し API rate limit に到達、成果物を得られたのは当初 3 体中 1 体のみだった (ワークスペースメモリfeedback_subagent_delegation_recursive_cascade.mdに記録済み)。brake 1 結果 (N=3, opus, M=all axes)
全評価者が "mergeable" と判定。2点の軽微な指摘は反映済みのため再evaluationは実施せず、本コメントに記録する。
スコープ
L3-task layer の spec 修正。L1 Model Layer には触れないため brake 2 (L1 root-criteria evaluator) は対象外。release-version-rule.md 上は patch 相当 (delegation prompt hygiene の内部ルール追加、user/system observable な大規模構造変更ではない)。