anthropics / anthropics/claude-code

First-class "adversarial-verify + live-source check" review mode for multi-agent workflows

オープン
#91,045 コメント 1 件 リアクション 0 件 担当者 0 名 GitHub で見る
主要言語
Python
スター
145k
フォーク
23.1k
PR マージ指標
PR 指標を取得中

説明

**Type:** enhancement · **Area:** multi-agent workflows / verification patterns

### Motivation
A repeatedly high-value workflow shape is: **produce, then adversarially verify against the source of truth.** Concretely, in a codebase audit:
1. A "mapper" agent extracts claims (invariants, risks, behaviors) from code + history + docs.
2. A separate "verifier" agent is instructed to *refute* each claim by opening the actual files.
3. Claims that survive are confirmed; claims marked `needs-live-db` are settled by read-only queries against the running system.

In real runs this caught defects the mapper asserted confidently, **and** corrected the verifier itself — a claim derived from static source (a migration file) was wrong about the live system, and only the read-only production check surfaced the truth. Neither single agent, and no static-only pass, would have caught it.

### Proposed feature
A documented, first-class pattern / helper for "generate → adversarial-verify → live-source settle," including:
- a verifier role prompt template that defaults to skepticism (refute-by-default, cite `file:line`),
- a structured verdict schema (`confirmed` / `refuted` / `partly-true` / `needs-live-db`),
- guidance/guardrails for a **read-only** live-source settle step (e.g. enforced read-only transactions) so verification against a running system is safe by construction.

### Why it helps everyone
"One agent confidently produces a plausible-but-wrong artifact" is the central failure mode of LLM automation. Making adversarial verification a supported, low-friction mode — rather than something each author hand-rolls — raises the floor on trustworthiness for audits, reviews, migrations, and research.

*Filed from real multi-agent workflow usage.*

コントリビューションガイド

このリポジトリのコントリビューションガイドは索引されていません

調査の方向性

The issue names no files or tests; start by getting maintainer agreement on whether this should be a documented pattern, a helper, or both. Use the proposed pieces as acceptance criteria: skeptical verifier prompt citing file:line, structured verdict schema, and read-only live-source guardrails. Done means the workflow is first-class and documented with safe settle behavior.

索引モデルが issue の本文から書いたものです。

評価

技術スタック
ai-infra-agents, python
領域
ai-infra-agents, developer-experience, documentation
issue の種類
機能追加
難易度
5/5
見積もり時間
1週間以上
活発さ
活発
明瞭さ
おおむね明確
初心者へのやさしさ
28/100

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。