Back to Details

workbench-token-context-discipline

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Workbench Token Context Discipline

Workbench 令牌上下文规范

Use this skill when a task may pull large docs, long issue histories, many skills, big logs, or expensive model context.

当任务可能涉及调取大型文档、冗长的问题历史、大量技能、庞大日志或高成本模型上下文时，使用本技能。

Context Budget Rules

上下文预算规则

Start with indexes, file lists, IDs, summaries, and targeted searches.
Read the smallest slice that can answer the current question.
Prefer
```
rg
```
,
```
jq
```
, concise table output, and explicit IDs over dumping entire histories.
Summarize large sources before passing them to another agent.
Attach only role-relevant skills; more skills are not automatically better.
Stop and report if two consecutive reads are too large or mostly irrelevant.

从索引、文件列表、ID、摘要和定向搜索开始。
读取能够回答当前问题的最小内容片段。
优先使用
```
rg
```
、
```
jq
```
、简洁表格输出和明确ID，而非导出完整历史记录。
在将大型数据源传递给其他Agent之前先进行摘要处理。
仅关联与角色相关的技能；并非技能越多越好。
如果连续两次读取的内容过大或大多无关，应停止操作并上报。

SDD Handoff Compact Mode

SDD 交接精简模式

When entering an SDD-gated issue:

Check whether the prior stage comment includes
```
HANDOFF_SUMMARY
```
.
If it exists, read the summary header and the listed
```
SCOPED_EVIDENCE
```
before reading any wider history.
If no handoff exists, read at most the latest three relevant comments, request or create a compact handoff, then stop broad scanning.
Never read full issue lists, full agent rosters, or unrelated repo docs for a single-stage task unless the stage explains why.
Output must fit in one reviewable comment. If it does not, compress rather than split.

进入SDD管控的问题时：

检查前一阶段的评论是否包含
```
HANDOFF_SUMMARY
```
。
如果存在，先读取摘要标题和列出的
```
SCOPED_EVIDENCE
```
，再读取更广泛的历史记录。
如果没有交接内容，最多读取最新的三条相关评论，请求或创建精简交接内容，然后停止大范围扫描。
除非阶段说明原因，否则不要为单阶段任务读取完整问题列表、完整Agent名单或无关的仓库文档。
输出内容必须能放入一条可审核的评论中。如果无法容纳，应进行压缩而非拆分。

Cache And Quota Awareness

缓存与配额感知

Distinguish fresh input, output, cache read, and cache write when usage is under investigation.
Do not assume dashboard usage is wrong without local/request-level evidence.
Avoid repeated full-context reruns when a comment ID, file path, or artifact can be cited instead.
Keep Workbench Max and existing private agents out of broad binding changes unless explicitly assigned.

当调查使用情况时，区分新鲜输入、输出、缓存读取和缓存写入。
不要在没有本地/请求级证据的情况下假设仪表板使用数据有误。
当可以引用评论ID、文件路径或工件时，避免重复进行全上下文重新运行。
除非明确分配，否则不要将Workbench Max和现有私有Agent纳入广泛的绑定变更中。

Output Contract

输出约定

Return:

```
CONTEXT_USED
```
: exact files, comments, commands, or IDs read.
```
SKIPPED
```
: intentionally avoided broad reads and why.
```
SUMMARY
```
: compact facts needed for the next step.
```
NEXT_SMALL_READ
```
: the next targeted evidence query, if any.
```
TOKEN_RISK
```
: normal / elevated / blocked.

返回：

```
CONTEXT_USED
```
：读取的确切文件、评论、命令或ID。
```
SKIPPED
```
：有意避免的大范围读取及其原因。
```
SUMMARY
```
：下一步所需的精简事实。
```
NEXT_SMALL_READ
```
：（如有）下一个定向证据查询。
```
TOKEN_RISK
```
：正常 / 升高 / 受阻。