[建议 / Feature] Allow editing proposed actions and typing custom answers in permission prompts

问题类别 · Category: 对话 / Agent 交互 · Agent chat / interaction
涉及框架 · Framework: 多个 / 通用 · Multiple / General

使用场景 · Use case

When ZCode requests permission (e.g. to run a command, apply a file edit, or answer a yes/no question during a task), the only available actions are accept or reject. There is no way to:

1. Modify the proposed action before approving — e.g. tweak the command line, adjust a file edit, or change parameters — and then approve the modified version.
2. Reply with a free-text alternative to a yes/no prompt — e.g. "yes, but do X instead", "no, use path Y", "run this other command".

Current workaround: reject, then re-explain the desired change in the chat. This costs an extra round-trip and is easy to mis-communicate, often leading to several iterations before the agent does the right thing.

建议方案 · Proposal

On permission prompts:

• Add an "Edit" option on proposed tool calls (commands, file edits, etc.) that opens the proposed action for inline editing before approval, similar to how some other coding agents let you revise a proposed command.
• Allow free-text input alongside the accept/deny buttons for yes/no prompts, so the user can redirect the agent inline (e.g. "yes, but exclude the node_modules folder").

预期价值 · Expected value

• Fewer reject → re-explain → retry cycles: lower token usage and faster task completion.
• Safer approvals: users can correct a slightly-wrong command/edit instead of choosing between "accept the wrong thing" or "reject and start over".
• More natural, expressive interaction — especially for users who already know what change they want but currently have to phrase it indirectly.

我认为的优先级 · Priority

Medium

补充材料 · Additional context

Applies to permission prompts across all integrated agent frameworks, since the prompt UI is owned by ZCode itself.

No source files, tests, or entry points are named. Start by locating ZCode’s permission-prompt UI and tracing how proposed tool calls and yes/no prompts are rendered across integrated agent frameworks. Done means users can edit a proposed action before approval and submit free-text alternatives without rejecting and re-explaining.