Hacktoberfest 2026：メンテナが10月に向けて印を付けた、オープンで初心者向けの issue。 Hacktoberfest の issue を見る

perf: cut what reaches the orchestrator's context across plugin skills, hooks and agents

オープン
  • 何か問題がありますか？
#6,322 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
22/100
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
技術スタック
shell
領域
ai-infra-agents, performance, tooling

調査の方向性

This is a scan-and-cut program, not one patch: walk plugin orchestration paths for skills, hooks, agents, handoff/compaction/resume. Compare against the listed patterns (long skill bodies, verbose subagent returns like discovery:research’s file-pointer-plus-summary, per-turn hook text from context-guard and rate-limit-guard, large command output). Measure with /context-budget:audit and /harness-ops:observability plus session transcripts; file each concrete cut separately. Related: issue #6316 and the forthcoming context-guard 128K/250K band PR.

索引モデルが issue の本文から書いたものです。

説明

needs-human needs-triage

Problem

Current long-context data says an orchestrating session degrades well before its window fills. Claude Opus 5 holds 91.3% on Context Arena's 8-needle MRCR at 128K but averages 45.7% across the full 1M window (https://contextarena.ai/?needles=8, fetched 2026-10-04). Practitioners put the start of degradation at about 125-150K tokens whatever the window size (Matt Pocock's AI Hero smart-zone entry, Geoffrey Huntley). context-guard now ships 128K/250K token bands on 1M windows to match.

An orchestrator that reads verbose output, re-reads large skill bodies or files, or relays subagent transcripts reaches those edges early and has to hand off or compact. Subagents and workflow agents are fine to use: each has its own window. The orchestrator's window is the scarce one.

Ask

Scan the plugins' orchestration paths and cut what reaches the orchestrator's context:

  • Skills that load long bodies or reference files the orchestrating session does not act on (progressive disclosure, smaller hubs).
  • Skills and agents whose subagent returns are verbose; they should return a file pointer plus a short summary, as discovery:research does.
  • Hook and mod text sent to the model on every turn or tool call (context-guard, rate-limit-guard, others): only what Claude acts on, in the fewest tokens.
  • Commands in skill bodies whose output is large (logs, diffs, listings) where a filter, a count or a subagent would do.
  • Handoff, compaction and resume flows that re-load more than the next step needs.

Measure before and after (/context-budget:audit, /harness-ops:observability, session transcripts), and file each concrete cut as its own change.

Related

  • #6316 (rate-limit and context-guard defaults)
  • The context-guard 128K/250K band change (PR to follow)

🤖 Generated with Claude Code

主要言語
Shell
スター
22
フォーク
2
平均マージ
8時間 27分
マージ済み PR（30日）
957

環境構築

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。

melodic-software/claude-code-plugins のほかの issue

melodic-software/claude-code-plugins の issue をすべて見る

似ている issue

Shell/Bash 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。