Safety prompts recur across run_skill and transfer_task sub-sessions

未关闭
#4,882 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@onatozmenn 已经在做这个了。

开始于 2026年11月12日。

  • #4915 来自 @onatozmenn —— 未关闭

评估

难度
4/5
预计耗时
3-5 天
新手友好度
45/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
docker, go
领域
security, testing

调研方向

Start by tracing how run_skill classifies read-only operations and how safety approvals are scoped when transfer_task creates a child session; the issue does not name source files or tests. Reproduce the reported flow, confirm both behaviors from the source, then add a focused regression test for a read-only skill crossing the sub-session boundary while preserving write confirmations and safety hooks.

由索引模型根据 Issue 内容生成。

描述

area/runtime area/security area/sessions area/skills

Problem

A prior session using Docker Agent dev build bbe70f53 observed repeated safety prompts across delegated work: 10/10 runs in the surviving log window prompted again. This report is about runtime safety-state/tool classification, not agent prompt design.

Working hypothesis from the observed behavior: launching a skill with run_skill does not mark its execution read-only, so the default safety mode prompts; an approve-autonomous decision appears scoped to the current sub-session, while transfer_task creates a new sub-session that prompts again. These details need source-level confirmation and a minimal reproduction by maintainers; the surviving observations alone do not establish the mechanism.

Reproduction to investigate

  1. In default safety mode, run a skill via run_skill that performs only read operations.
  2. Observe whether the tool launch prompts for approval despite being read-only.
  3. Grant the applicable autonomous/safety approval in that session, then delegate via transfer_task to a child session that performs the same read-only operation.
  4. Observe whether the new sub-session prompts again.

Reported observation: repeated prompts occurred in 10/10 runs visible in the surviving log window. This count is limited to that window and is not a broader rate estimate. No raw logs, private content, local paths, or credentials are attached.

Expected

Provide safe, explicit, least-privilege handling for read-only operations and/or scoped reuse of a user-granted safety policy across related sub-sessions, with clear visibility of scope. Preserve safety hooks and their enforcement. Reuse must not automatically authorize writes or bypass explicit confirmation for external, irreversible, or production-reaching writes.

Actual observation / uncertainty

Prompts recurred across skill/delegation runs. The proposed run_skill classification and per-sub-session approval scope are hypotheses, not independently verified source conclusions.

Environment / limitations

  • Docker Agent dev build: bbe70f53
  • macOS arm64
  • Prior-session observation; exact source revision and a standalone reproduction were not retained.

Please confirm the classification and approval-scope behavior, then add a focused regression test for a read-only skill crossing a transfer_task sub-session boundary.

主要语言
Go
星标
4.7k
派生
565
平均合并
18 小时 52 分钟
30 天内合并 PR
221

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

  • 提供 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 没有贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

docker/docker-agent 的其他 Issue

查看 docker/docker-agent 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。