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

Keeper triages results with a decision model

オープン
  • 何か問題がありますか？
#64 コメント 0 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
25/100
issue の種類
機能追加
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
csharp
領域
ai, backend

調査の方向性

Start with parent issue #21 and the prerequisite #63, then trace the existing triage flow, Scout registration, run recording, model gate, settings path, and research note handling. The issue specifies unit, handler, client, and TestWiki integration tests; done means the Keeper replaces Scout triage, its calls and verdicts appear in Runs, kept results produce notes with quotes, and the listed documentation and settings are updated.

索引モデルが issue の本文から書いたものです。

説明

enhancement ready-for-agent

Parent

#21

What to build

The Keeper, a decision model served by llama-server's /v1/systemone, takes over judging fetched results from the Scout. It keeps or discards each result from probabilities, flags injection attempts and picks exact quotes. Kept results still become one research note per topic. Until the Clerk lands (next ticket), notes carry title, URL, kind and quotes without a summary, and the Scout is no longer run.

Acceptance criteria

  • A Keeper port (e.g. IKeeper) with a /v1/systemone client, configured with the model name on the shared llama-server (router mode) or its own endpoint, and generous timeouts; it is not an AIAgent
  • One call per fetched result: state = topic brief plus delimited fetched content; questions useful (noul), kind (choice with described options: release, announcement, discussion, tutorial, opinion, beginner question, spam, unrelated), injection (noul) and quotes (choice over the content's paragraphs, top three, length-capped in code)
  • Kept when useful ≥ Research:KeepThreshold and injection < Research:InjectionThreshold; the verdict stores the probabilities, kind and quotes; a discard's reason is derived in code
  • A failed or timed-out call leaves the result waiting for the next cycle; a page that could not be fetched is discarded without a call, as today
  • Each call is recorded as a run (Triage trigger, no tool calls) and shows on the Runs page with its probabilities
  • Calls go through the model gate ahead of Pleb runs and count against a per-cycle cap; the cycle is skipped when the model server is unhealthy
  • Research notes hold title, URL, kind and quotes per kept result, never page text; existing header, file naming and awaiting/completed moves unchanged
  • New settings and AppHost parameters (model, thresholds, cap) follow the usual path: AppHost, CD workflow env: with || 'default' fallbacks, scripts/setup-github.ps1
  • The Scout is no longer run by triage (its registration is removed in the Clerk ticket)
  • docs/ubiquitous-language.md gains Keeper; Verdict, Triage and Research note updated; README research section updated
  • Unit tests: threshold and injection rules, quote capping and derived reasons as pure policies; handler tests with a substituted Keeper cover keep, discard, injection, failure/resume and aggregation; the client tested with a substituted HttpMessageHandler
  • Integration test over TestWiki with a scripted Keeper drops the expected note in the inbox

Blocked by

  • #63
主要言語
C#
スター
0
フォーク
0
平均マージ
40分
マージ済み PR（30日）
47

環境構築

このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。

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。

DigiBanks99/knowledge-base のほかの issue

DigiBanks99/knowledge-base の issue をすべて見る

似ている issue

C# 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。