Five Workspace Anti-Patterns That Fake Agent Skill

说明与用途

文章描述了五种工作区反模式，导致代理看似成功。建议使用隔离工作树、路径门控、原子补丁应用、运行前快照和清理，以确保可靠安全的代理演示。

问题

Coding agents often appear successful because the workspace is not isolated, leading to hidden failures and security leaks.

结果

Implementing isolated worktrees, path gates, atomic patch application, and pre-run snapshots eliminates hidden failures, secures secrets, and ensures reproducible agent demos.

工作流

• Create a unique worktree or clone for the run
• Configure path gates to allow only safe prefixes and deny secrets
• Generate a unified diff patch from the agent
• Check the patch with git apply --check
• Apply the patch atomically if check passes
• Record HEAD and git status before the run
• Run gates.sh to validate changes
• If gates fail, perform git reset --hard to the recorded HEAD and git clean -fd
• Delete the temporary worktree or clone after the run

证据

Give every run a throwaway worktree. Apply patches only after a gate passes.

来源

Deny by default. Allow one path prefix. Reject reads that match a deny list.

来源

Stage one unified diff. Apply it atomically. Reject the whole patch when any gate fails.

来源

Snapshot before tools. Revert on red. Record HEAD and git status --porcelain as first-class outputs.

来源

One run, one directory, one wipe. Cold clone or a fresh worktree every time.

来源