Agent Persona Exploration - 2026-11-04

クローズ 初心者向け
  • 何か問題がありますか？
#74,891 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
2/5
見積もり時間
1〜3時間
初心者へのやさしさ
72/100
issue の種類
ドキュメント
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
github, github-actions
領域
devtools, documentation

調査の方向性

.github/aw/create-agentic-workflow-trigger-details.mdを読み、issueで説明されているコンプライアンス監査の例を更新してください。権限を変更し、ネットワーク設定を追加します。次に、.github/aw/report.mdと.github/aw/create-agentic-workflow.mdを確認し、追加の2つの推奨事項を探してください。例とガイダンスが、issueで指定された最小権限と一貫性のルールに沿ったら完了です。

索引モデルが issue の本文から書いたものです。

説明

agent-research
Persona Overview
  • Agent: agentic-workflows (via the persona-evaluator Claude Code sub-agent)
  • Personas This Run: Program Manager, Designer, Legal / Compliance (one-time baseline)
  • Scenarios Tested: 4 (PM-1, DS-1, LC-1, LC-2). All 4 invocations succeeded.
  • Average Quality Score: 4.3/5.0
Key Findings
  • The trigger was right in every case. The scheduled scenarios used a fuzzy weekly schedule plus workflow_dispatch. The PR scenarios used pull_request with paths: filters. DS-1 also added edited, so the review re-runs when the author adds a design link.
  • Permissions stayed read-only in all 4. Every write went through safe outputs. LC-2 even rejected the issues: write shown in the guide's own compliance-audit example.
  • Network scoping is inconsistent. DS-1 used network.allowed: []. PM-1 and LC-2 used network: defaults, even though they only read repository files and GitHub data. create-agentic-workflow.md §5 prescribes allowed: [] for exactly the LC-2 case.
  • Every response stated its assumptions. Each one named an ambiguity, such as what "overdue dependency" means, where the policy file lives, or where the design link is expected, and gave a fallback.
  • The prose and the frontmatter sketch sometimes disagree. In PM-1 the prose says to comment on an existing issue, but add-comment isn't configured. In LC-1 the prose lists more lockfile paths than the sketch. In LC-2 the prose says gh-proxy, but the sketch doesn't set that mode.
Top Patterns
  1. Triggers: fuzzy schedule + workflow_dispatch (2 of 4); pull_request + paths: (2 of 4)
  2. Tools: github in all 4 (with gh-proxy in 2). bash restricted to read-only commands in 2 of 4. No web or browser tools were suggested.
  3. Safe outputs: create-issue with close-older-issues for the scheduled reports. add-comment with hide-older-comments, plus an add-labels allowlist, for PR reviews. noop was always defined.
  4. Security: least-privilege read permissions in all 4. pull_request is preferred over pull_request_target. LC-1 explicitly said to treat package metadata as untrusted data.
View Scores by Scenario
ID Persona Trigger Tools Security Prompt Complete Avg
PM-1 Program Manager 5 4 3 4 4 4.0
DS-1 Designer 5 5 5 5 4 4.8
LC-1 Legal / Compliance 5 4 4 5 4 4.4
LC-2 Legal / Compliance 5 4 3 4 4 4.0
View High Quality Responses (Top 2)

  • DS-1, design token review (4.8):

    • Uses the edited trigger type so the review re-runs when a link is added.
    • Sets network.allowed: [], and checks design links only for presence and format, without fetching them.
    • Applies concrete WCAG 2.2 contrast thresholds (4.5:1 and 3:1).
    • Lists the components affected by renamed tokens.
    • Restricts labels with add-labels allowed:.
    • Posts one rolling comment, using hide-older-comments.

  • LC-1, SPDX license gate (4.4):

    • Sorts licenses into allowed, disallowed and uncertain. NOASSERTION, custom licenses and uncovered OR/AND/WITH expressions go to human review.
    • Has an explicit "never guess a license" rule.
    • Keeps the workflow advisory (no blocking review) and offers create-check-run as an option for a hard gate.
View Areas for Improvement (Top 3)

  • Network defaults on audits that only read the repository (PM-1, LC-2):

    • Both used network: defaults even though workflow-constraints.md:48 and create-agentic-workflow.md:211 say to use allowed: [].
    • Likely cause: the scheduled compliance-audit example in create-agentic-workflow-trigger-details.md (around lines 109–124) has no network: block and grants issues: write.

  • Ecosystems not inferred from the repository (LC-1): LC-1 enabled six ecosystems (node, python, go, rust, java, ruby) up front. The §5 guidance says to infer them from the manifests.

  • Prose and sketch drift:

    • PM-1's dedup plan says to comment on an existing issue, but there is no add-comment safe output. Its weekly dedup key also overlaps with close-older-issues.
    • LC-2 uses a monthly dedup key on a weekly schedule.
    • LC-1's bash allowlist uses "git diff:*", which differs from the documented "cmd *" and bare-command forms.
Recommendations
  1. Fix the compliance-audit example in .github/aw/create-agentic-workflow-trigger-details.md. Change issues: write to issues: read and add network: { allowed: [] }. The agent appears to copy this sample, and today it contradicts create-agentic-workflow.md §5 and the strict-mode rules.
  2. Add a scheduled-report network rule to .github/aw/report.md. State that digests and audits that only read GitHub through gh-proxy should use network.allowed: [], not defaults. PM-1 shows the gap for non-compliance digests.
  3. Add a consistency check to the ad hoc evaluation checklist in .github/aw/create-agentic-workflow.md. Every action the prompt describes (for example "comment on the existing issue") needs a matching configured safe output. The dedup-key granularity should match the schedule cadence. Ecosystems should be inferred from the manifests that actually exist.

References:

Generated by 🎭 Agent Persona Explorer · claude · routed: opus55 high · ~128 AIC · ⌖ ~2 AIC · ⊞ 6.6K · ◷

主要言語
Go
スター
5.6k
フォーク
615
平均マージ
5時間 36分
マージ済み PR（30日）
805

環境構築

Codespaces で開く

このプロジェクトの開発コンテナを、あなたの GitHub アカウントでブラウザ上に起動します。

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。

github/gh-aw のほかの issue

github/gh-aw の issue をすべて見る

似ている issue

Go 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。