hear-me

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Hear Me

Speech-to-text errors cluster in brittle terms: names, commands, flags, paths, identifiers, acronyms, versions, numbers, negation, and action words. Focus on a brittle term only when a wrong interpretation would change the action, target, scope, result, or safety. Accept ordinary wording as given.

语音转文本的错误集中出现在易出错术语上：名称、命令、标志、路径、标识符、首字母缩写、版本号、数字、否定词以及动作词汇。仅当错误解读会改变操作、目标、范围、结果或安全性时，才需要关注这类易出错术语。常规表述则直接按原文接受。

Invocation

调用方式

Match

on

and

off

without case sensitivity after trimming surrounding whitespace. Treat them as controls only when they are the entire argument of an explicit skill invocation. Reply to a control with only a confirmation of the new state.

No argument or
```
on
```
: Enable dictation mode for the current conversation. Reply
```
Hear Me is on for this conversation.
```
```
off
```
: Disable dictation mode. Reply
```
Hear Me is off.
```
Any other non-empty argument: Enable dictation mode and apply the protocol to that request. When this changes the state from off to on, state
```
Hear Me is on for this conversation.
```
after the
```
Heard:
```
readback and before other skill announcements or task work. Do not repeat the state for a later request while the mode is already on.

While the mode is on, apply the protocol to every later user message, including typed text and explicit skill invocations. Controls are the one exception: do not apply the protocol to an

on

off

control, and reply with only the confirmation. Only an explicit

off

control disables it. Start each new conversation with the mode off, and keep its state only in that conversation's history.

匹配

on

和

off

时忽略大小写，且会先去除周围空白字符。只有当它们作为显式技能调用的完整参数时，才将其视为控制指令。回复控制指令时仅需确认新状态即可。

无参数或参数为
```
on
```
：启用当前对话的听写模式。回复
```
Hear Me is on for this conversation.
```
```
off
```
：禁用听写模式。回复
```
Hear Me is off.
```
其他非空参数：启用听写模式并对该请求应用协议。当状态从关闭切换为开启时，在
```
Heard:
```
回读之后、其他技能通知或任务执行之前，告知
```
Hear Me is on for this conversation.
```
。若模式已处于开启状态，后续请求无需重复告知该状态。

当模式处于开启状态时，对后续所有用户消息应用协议，包括输入文本和显式技能调用。控制指令是唯一例外：不对

on

或

off

控制指令应用协议，仅回复确认信息。只有显式的

off

控制指令可禁用该模式。每次新对话开始时，模式默认处于关闭状态，且仅在该对话的历史记录中保留其状态。

Protocol

执行协议

Make a concise
```
Heard:
```
readback the first text of the first assistant message for each dictated request, before any skill announcement, tool call, or task work. Restate the action, target, scope, and all consequential constraints. Use one line when it preserves them; otherwise use a short block.
- Format each correction or assumption as code. Include the original transcript when the change is not obvious:
```
ruff check
```
  (from "rough check").
- Mark an unresolved phrase and its credible candidates instead of choosing one:
```
[unclear: feature/cash-sync or feature/cache-sync?]
```
  .
- Replace a secret or sensitive value with a role label such as
```
[API token]
```
  .
Resolve every consequential brittle term before work that depends on it. Use the conversation, repository state, and cheap read-only checks to test candidates. Treat a term as resolved only when one candidate has direct support and no credible alternative remains.
The readback is an error check, not a confirmation gate: proceed in the same turn when all dependent terms are resolved. When two or more credible interpretations remain, ask one focused question, complete every independent part of the request while that part is blocked, and mutate state only when the action and its target are both resolved.

针对每一条听写请求，在助手的第一条消息开头先添加简洁的
```
Heard:
```
回读内容，置于任何技能通知、工具调用或任务执行之前。重述操作、目标、范围以及所有相关约束条件。若一行文本可完整保留这些信息则使用单行；否则使用简短区块。
- 将每一处修正或假设格式化为代码块。若修改内容不明显，需附上原文转录内容：
```
ruff check
```
  (from "rough check")。
- 若短语存在歧义且无法确定，需标记该未解析短语及其可信候选选项，而非自行选择：
```
[unclear: feature/cash-sync or feature/cache-sync?]
```
  .
- 用角色标签替换机密或敏感值，例如
```
[API token]
```
  。
在执行依赖易出错术语的任务之前，必须先解析所有相关的易出错术语。利用对话内容、仓库状态以及低成本的只读检查来验证候选选项。只有当某一候选选项有直接依据且无其他可信替代选项时，才视为该术语已解析完成。
回读是错误检查机制，而非确认闸门：当所有依赖术语均已解析完成时，可在同一轮对话中继续执行任务。若存在两种或更多可信解读，需提出一个针对性问题，在该部分被阻塞时完成请求中所有独立的部分，且仅当操作及其目标均已解析完成时，才可修改状态。