imperatives

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

/imperatives

Extract atomic imperatives (MUST/SHOULD/MAY) from markdown instruction files into JSONL.

从Markdown指令文件中提取原子命令式语句（MUST/SHOULD/MAY）并转换为JSONL格式。

Usage

使用方法

/imperatives                                # default: ai-workspace/rules/*.md + AGENTS.md
/imperatives .claude/rules/*.md             # specific globs
/imperatives --output imperatives.jsonl     # write to file

/imperatives                                # 默认路径：ai-workspace/rules/*.md + AGENTS.md
/imperatives .claude/rules/*.md             # 指定通配符路径
/imperatives --output imperatives.jsonl     # 写入文件

Steps

步骤

Resolve files. Default:
```
ai-workspace/rules/*.md
```
and
```
AGENTS.md
```
. Expand globs to absolute paths. Zero matches → error, stop.
Pass 1 — Script extraction (fast, deterministic).
bash
```
node --import tsx scripts/extract-imperatives.ts <files...> --output /tmp/imperatives-pass1.jsonl
```
Keyword-based: catches RFC 2119 terms + imperative verbs. ~80% coverage, <1s.
Pass 2 — Subagent reasoning (catches implicit imperatives). For each file, dispatch a subagent to identify imperatives the regex missed — contextual rules, implicit constraints, prose-embedded obligations.

Subagent prompt:
Read
```
<file>
```
. Here are the imperatives already extracted by the script for this file:
```
<pass 1 JSONL lines where source.file matches>
```
Identify additional imperative statements the regex missed. Look for:
- Implicit obligations hidden in prose (e.g., "The workflow is a menu, not a pipeline" implies MUST NOT enforce order)
- Contextual constraints (e.g., section context that makes a statement imperative)
- Conditional rules not triggered by keyword patterns
For each new imperative found, output one JSON line (same schema — id, level, polarity, subject, predicate, when, source, tool_scope, tags, raw). Do NOT duplicate entries already in pass 1. If none found, output nothing.
Use
```
subagent_type: "explorer"
```
(Haiku) for most files. If a file had zero pass-1 imperatives OR contains dense prose (>50 lines without a keyword match), escalate that file to Sonnet for deeper reasoning.
Merge passes. Concatenate pass 1 + pass 2 outputs. Deduplicate by
```
raw
```
text. Write final output to
```
--output <path>
```
or present inline.
Present summary. Report in a table:
- Total count (pass 1 baseline + pass 2 additions)
- Breakdown by
```
level
```
  and
```
tool_scope
```
- Files with zero imperatives after both passes
- Pass 2 additions highlighted separately
Downstream callers. If invoked by
```
/policy-algebra
```
or
```
/distill
```
, return the JSONL path directly — skip the summary.

解析文件。默认路径：
```
ai-workspace/rules/*.md
```
和
```
AGENTS.md
```
。将通配符路径扩展为绝对路径。如果匹配到零个文件→报错并终止。
第一轮——脚本提取（快速、确定性）
bash
```
node --import tsx scripts/extract-imperatives.ts <files...> --output /tmp/imperatives-pass1.jsonl
```
基于关键词：捕获RFC 2119术语和命令式动词。覆盖率约80%，耗时<1秒。
第二轮——子代理推理（捕获隐含命令式语句） 针对每个文件，调度一个子代理识别正则表达式遗漏的命令式语句——上下文规则、隐含约束、嵌入在散文式文本中的义务。

子代理提示语：
读取
```
<file>
```
。以下是脚本已从该文件中提取的命令式语句：
```
<pass 1中source.file匹配的JSONL行>
```
识别正则表达式遗漏的其他命令式语句。请查找：
- 隐藏在散文式文本中的隐含义务（例如：“工作流是菜单而非流水线”意味着MUST NOT强制顺序）
- 上下文约束（例如：使语句成为命令式的章节上下文）
- 未被关键词模式触发的条件规则
对于每个找到的新命令式语句，输出一行JSON（遵循相同的 schema — id、level、polarity、subject、predicate、when、source、tool_scope、tags、raw）。请勿重复第一轮中已有的条目。如果未找到任何新条目，请输出空内容。
大多数文件使用
```
subagent_type: "explorer"
```
（Haiku）。如果某个文件在第一轮中未提取到任何命令式语句，或包含密集的散文式文本（超过50行无关键词匹配），则将该文件升级到Sonnet进行深度推理。
合并结果。将第一轮和第二轮的输出拼接起来。根据
```
raw
```
文本去重。将最终输出写入
```
--output <path>
```
或直接展示在界面中。
展示摘要。以表格形式报告：
- 总数（第一轮基准数 + 第二轮新增数）
- 按
```
level
```
  和
```
tool_scope
```
  分类的明细
- 两轮后仍未提取到命令式语句的文件
- 单独高亮显示第二轮新增的内容
下游调用方。如果被/policy-algebra或/distill调用，则直接返回JSONL路径——跳过摘要展示。

Failure modes

失败场景

Condition	Behavior
No files matched	Error message. Stop.
File not found	Script warns on stderr, continues with remaining files.
Zero imperatives	Report zero. Not an error.
Script exits non-zero	Surface stderr to user.

条件	行为
未匹配到任何文件	输出错误信息并终止。
文件未找到	脚本在stderr中发出警告，继续处理剩余文件。
未提取到任何命令式语句	报告数量为零，不视为错误。
脚本非零退出	将stderr内容展示给用户。

Cross-tool notes

跨工具说明

Codex / Cursor: run the script directly — it's tool-agnostic.

Codex / Cursor：直接运行脚本——该工具与编辑器无关。