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(discovery): slice-only docs reads lose recall on scattered questions

クローズ
  • 何か問題がありますか？
#6,487 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
4/5
見積もり時間
3〜5日
初心者へのやさしさ
42/100
issue の種類
機能追加
明瞭さ
おおむね明確
活発さ
活発
技術スタック
markdown
領域
performance, search

調査の方向性

Start from the artifact introduced in #6348: the shared docs lookup skill /discovery:read-docs and its escalate_bytes (default 60 KB) and section-count (25%) escalation rules, plus the Phase 6 acceptance kit, note, cache and grader used for the confirmation run. Identify where the thresholds and the slice-reading procedure live, then run the measured comparison of default vs. lower threshold vs. procedure nudge vs. combination, reporting n per arm and the same interval rule as the confirmation run. Done means the slice-only versus escalated recall gap closes within noise at acceptable bytes/tokens cost, or the null result is recorded and defaults kept.

索引モデルが issue の本文から書いたものです。

説明

needs-triage priority: medium

No related issue: measured follow-up to #6348 (part of #6020).

Summary

The shared docs lookup (/discovery:read-docs, #6348) returns verbatim sections and escalates to the whole page past a size or section-count threshold (escalate_bytes, default 60 KB; 25% of sections). The Phase 6 acceptance run showed that runs reading only slices lose recall on questions whose answer is scattered across a page:

  • Slice-only runs: new-core recall .838 (n=4).
  • Runs that escalated to the whole page: .927 (n=6).
  • The worst slice-only run scored .74; it never read 3 scattered sections.
  • The same direction appeared in the original batch (.93 vs .86).

The pre-registered confirmation passed (10 fresh runs, mean .891 vs the no-cache arm's .936, lower bound -.072 against a -.10 margin) and thresholds were kept at defaults, so this is an improvement, not a regression fix.

Fix

Candidates, each to be measured before adopting:

  1. Lower the escalate_bytes threshold so more reads return the whole page.
  2. Add a procedure nudge: when a question spans many sections, read the whole page instead of slices.

Verification

  • A measured comparison on the same acceptance kit (same questions, same note and cache, same grader) of the default, each candidate, and the combination.
  • Acceptance: the slice-only versus escalated recall gap closes (target: slice-only runs within the noise of escalated runs) without losing the cost benefit, measured as bytes or tokens read per run against the current default. Report n per arm and use the same interval rule as the confirmation run.
  • If neither candidate closes the gap at acceptable cost, record the result and keep the defaults.

Related

  • #6020 (umbrella)
  • #6348 (shared docs cache, fetcher, /discovery:read-docs)
主要言語
Shell
スター
22
フォーク
2
平均マージ
7時間 42分
マージ済み PR（30日）
923

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。