dsh-plain-talk 0.1.0 → 0.1.2

package/dsh/check.js

@@ -1,250 +1,250 @@
-// Reply checker for dsh-plain-talk.
-//
-// This is deliberately a checker, not a gate: it reports what a reader would
-// stumble on, with the exact replacement form. The model decides what to keep.
-//
-// Three principles learned from real sessions in this workspace:
-//  - Common terms and project names are not errors. Only opaque codes the
-//    reader has never seen (M11, dev-1200, c590) need a first-use gloss.
-//  - A pointer is not a citation: `file.py:42` alone sends the reader away.
-//  - Cross-turn wording ("上述", "上一轮") is the main reason a report cannot
-//    be read on its own.
-
-const ALLOW = new Set([
-  'api', 'json', 'jsonl', 'yaml', 'yml', 'html', 'css', 'http', 'https', 'url', 'uri',
-  'cli', 'ui', 'ux', 'gpu', 'cpu', 'ram', 'ssd', 'hdd', 'ai', 'llm', 'sdk', 'ide',
-  'csv', 'tsv', 'ts', 'js', 'py', 'sh', 'md', 'txt', 'png', 'jpg', 'jpeg', 'svg',
-  'gif', 'webp', 'sql', 'xml', 'toml', 'npm', 'pnpm', 'git', 'ssh', 'dns', 'tcp',
-  'udp', 'tls', 'ssl', 'jwt', 'oauth', 'cors', 'orm', 'crud', 'rest', 'grpc',
-  'wasm', 'dom', 'smtp', 'imap', 'vpn', 'utf8', 'utf-8', 'utf16', 'utf-16',
-  'sha1', 'sha256', 'sha-256', 'sha512', 'sha-512', 'md5', 'x86', 'x64', 'arm64',
-  'mp3', 'mp4', 'pdf', 'h264', 'hevc', 'ipv4', 'ipv6', 'k8s', 'i18n', 'l10n',
-  '2d', '3d', '4k', 'v1', 'v2', 'v3', 'v4', 'ci', 'pr', 'id', 'ok', 'vs', 'iso',
-  'doi', 'arxiv', 'p50', 'p95', 'p99', '2x', '4x', '8x', '16x', 'gpt', 'llm',
-  'linux', 'windows', 'macos', 'node', 'react', 'github', 'python', 'javascript',
-  'typescript', 'docker', 'sqlite', 'postgres', 'redis',
-])
-
-const MODEL_FAMILY = /^(?:gpt|claude|gemini|llama|qwen|deepseek|mistral|phi|o1|o3|o4|glm|kimi|mimo|hy)\d*[-.]?\d*/i
-
-const BACKREF_ZH = [
-  '上述', '上面', '前述', '前文', '前面提到', '前面说', '之前提到', '之前说',
-  '此前', '上一步', '上一轮', '前一轮', '刚才', '该方案', '该项', '该任务',
-  '前者', '后者', '如上', '见上文', '先前', '那次', '前面', '之前',
-]
-const BACKREF_EN = [
-  'as mentioned', 'aforementioned', 'as discussed', 'the above', 'previous step',
-  'earlier', 'previously', 'the former', 'the latter',
-]
-
-const PLUMBING = [
-  /\bsession[-_][0-9a-z-]{6,}/i,
-  /子代理|子会话|父代理|父会话|\bsubagent\b|\bparent agent\b|report sent to/i,
-  /\b(?:plain_talk_check|run_code|ops_bash|ops_grep|todo_write)\b/,
-]
-
-const PREAMBLE = /^(好的|好|当然|没问题|收到|明白|我来|让我|让我先|我们先|首先，我|先看一[下看])/
-const CLOSER = /(需要我|要不要|如果你愿意|希望(这|以上)|随时告诉我|还需要我|要我继续)/
-
-const FILE_EXT = 'py|js|mjs|cjs|ts|tsx|json|jsonl|md|ya?ml|toml|sh|ps1|bat|csv|tsv|txt|ipynb|log|html|css|tex|bib|pdf|zip|sql|db|pt|pth|ckpt|npy|npz|parquet|docx|xlsx|pptx'
-
-const IDENT_PATTERNS = [
-  ['file', new RegExp(`(?:[A-Za-z]:[\\\\/]|\\/)?[A-Za-z0-9_.-]+(?:[\\\\/][A-Za-z0-9_.-]+)*\\.(?:${FILE_EXT})\\b`, 'gi')],
-  ['hash', /\b[0-9a-f]{7,40}\b/gi],
-  ['snake', /\b[A-Za-z][A-Za-z0-9]*(?:_[A-Za-z0-9]+)+\b/g],
-  ['camel', /\b[a-z]+[A-Z][A-Za-z0-9]*\b|\b(?:[A-Z][a-z0-9]+){2,}[A-Za-z0-9]*\b|\b[A-Z][a-z]+[A-Z]{2,}[A-Za-z0-9]*\b/g],
-  ['acronym', /\b[A-Z]{2,8}\b/g],
-  ['code', /\b[A-Za-z]{1,10}[-_.]?\d{1,6}[A-Za-z]?\b|\b\d{1,4}[A-Za-z]{1,6}\b/g],
-]
-
-export const LIMITS = {
-  boxChars: 300,
-  boxLines: 12,
-  listRun: 5,
-  identifierWarns: 8,
-}
-
-const cjkCount = (s) => (s.match(/[\u4e00-\u9fff]/g) ?? []).length
-const latinCount = (s) => (s.match(/[A-Za-z]/g) ?? []).length
-const len = (s) => [...s].length
-
-function stripCode(text) {
-  return text
-    .replace(/```[\s\S]*?```/g, ' ')
-    .replace(/`[^`\n]*`/g, ' ')
-}
-
-export function isCommonTerm(term) {
-  const text = String(term ?? '').trim()
-  if (!text) return false
-  return ALLOW.has(text.toLowerCase()) || MODEL_FAMILY.test(text)
-}
-
-export function scanIdentifiers(input) {  const text = stripCode(String(input ?? ''))
-  const found = []
-  for (const [kind, re] of IDENT_PATTERNS) {
-    re.lastIndex = 0
-    let m
-    while ((m = re.exec(text))) {
-      const raw = m[0].replace(/[.,;:]+$/, '')
-      if (!raw) continue
-      found.push({ kind, token: raw, start: m.index, end: m.index + raw.length })
-    }
-  }
-  found.sort((a, b) => a.start - b.start || (b.end - b.start) - (a.end - a.start))
-  const accepted = []
-  for (const item of found) {
-    if (accepted.some((a) => item.start < a.end && item.end > a.start)) continue
-    accepted.push(item)
-  }
-  return accepted
-}
-
-function isAllowed(token, known) {
-  const lower = token.toLowerCase()
-  if (ALLOW.has(lower)) return true
-  if (MODEL_FAMILY.test(token)) return true
-  // Durations and multipliers ("10m", "3s", "4x") are measurements, not names.
-  if (/^\d+(?:ms|s|m|h|d|w|k|m|b|x)$/i.test(token)) return true
-  // Product-style names (MechDW_V1, GeoEvidence_V2) are project vocabulary the
-  // reader already holds; a lower-case internal key (tier_reaudit) is not.
-  if (/^[A-Za-z0-9]+(?:_[A-Za-z0-9]+)+$/.test(token) && (token.match(/[A-Z]/g) ?? []).length >= 2) return true
-  if (known && (known.has(token) || known.has(lower))) return true
-  return false
-}
-
-function explainedAfter(text, end) {
-  const after = text.slice(end, end + 60)
-  const m = /^\s*[（(]([^）)]{4,})[）)]/.exec(after)
-  if (!m) return false
-  return cjkCount(m[1]) >= 3 || latinCount(m[1]) >= 8
-}
-
-function knownFrom(text, known) {
-  return { text, known }
-}
-
-/**
- * Analyze one reply draft.
- *
- * @param draft - the reply text the model is about to send.
- * @param options - `knownTokens` (project/user vocabulary) and `language`
- *                  (`zh` when the user writes Chinese).
- * @returns issues plus the measurements the README benchmark reports.
- */
-export function analyze(draft, options = {}) {
-  const text = String(draft ?? '')
-  const known = new Set((options.knownTokens ?? []).map((t) => String(t)))
-  const knownLower = new Set([...known].map((t) => t.toLowerCase()))
-  const language = options.language ?? 'unknown'
-  const issues = []
-
-  const add = (severity, rule, message, fix) => issues.push({ severity, rule, message, fix })
-
-  const narrative = stripCode(text)
-
-  // 1. Cross-turn wording.
-  const backrefs = [
-    ...BACKREF_ZH.filter((w) => narrative.includes(w)),
-    ...BACKREF_EN.filter((w) => narrative.toLowerCase().includes(w)),
-  ]
-  if (backrefs.length) {
-    add('error', 'backref', `出现回指词：${backrefs.slice(0, 5).join('、')}。`, '把指代的东西重新写一遍，读者没有前文可翻。')
-  }
-
-  // 2. Internal plumbing.
-  const plumbing = PLUMBING.find((re) => re.test(narrative))
-  if (plumbing) {
-    add('error', 'internal', '出现会话号、子代理或工具名等内部信息。', '删掉；只写用户能看见的对象。')
-  }
-
-  // 3. Pointer-only citations.
-  const pointers = []
-  const pointerRe = new RegExp(`(?:[A-Za-z]:[\\\\/]|/)?[\\w.-]+(?:[\\\\/][\\w.-]+)*\\.(?:${FILE_EXT}):\\d+`, 'gi')
-  pointerRe.lastIndex = 0
-  let pm
-  while ((pm = pointerRe.exec(text))) pointers.push(pm[0])
-  if (pointers.length) {
-    const unquoted = pointers.filter((pointer) => {
-      const at = text.indexOf(pointer)
-      const window = text.slice(Math.max(0, at - 220), at + pointer.length + 220)
-      return !/「[^」]{4,}」|"[^"]{4,}"|『[^』]{4,}』|`[^`]{4,}`|```/.test(window)
-    })
-    if (unquoted.length) {
-      add('error', 'citation', `只有位置没有原文：${unquoted.slice(0, 3).join('、')}。`, '把那几行原文贴出来，再标位置。')
-    }
-  }
-
-  // 4. Unexplained opaque identifiers.
-  const seen = new Set()
-  const unknown = []
-  for (const token of scanIdentifiers(narrative)) {
-    if (seen.has(token.token)) continue
-    seen.add(token.token)
-    if (isAllowed(token.token, known) || isAllowed(token.token, knownLower)) continue
-    if (explainedAfter(narrative, token.end)) continue
-    unknown.push(token.token)
-  }
-  if (unknown.length) {
-    add(
-      'warn',
-      'identifier',
-      `有 ${unknown.length} 个代号没有说明：${unknown.slice(0, 6).join('、')}${unknown.length > 6 ? ' 等' : ''}。`,
-      '第一次出现写成「M11（一句话说明它是什么）」；用户在本对话用过的词和项目名不用解释。',
-    )
-  }
-
-  // 5. Shape.
-  const lines = text.split('\n').filter((l) => l.trim())
-  const cjk = cjkCount(narrative)
-  if (cjk > LIMITS.boxChars || lines.length > LIMITS.boxLines) {
-    add('warn', 'length', `回复有 ${cjk} 个汉字、${lines.length} 行，超过默认的 ${LIMITS.boxChars} 字 / ${LIMITS.boxLines} 行。`, '删掉读者不决策所需的内容；用户要细节时再展开。')
-  }
-
-  let run = 0
-  let maxRun = 0
-  for (const line of text.split('\n')) {
-    if (/^\s*(?:[-*+]|\d+[.)])\s+/.test(line)) { run++; maxRun = Math.max(maxRun, run) } else if (!line.trim()) { run = 0 }
-  }
-  if (maxRun > LIMITS.listRun) {
-    add('warn', 'list', `有一段列表 ${maxRun} 条，超过 5 条。`, '合并或只留最要紧的五条。')
-  }
-
-  const first = lines[0] ?? ''
-  if (PREAMBLE.test(first)) {
-    add('warn', 'preamble', `开头是「${first.slice(0, 20)}」，属于寒暄或过程宣告。`, '第一行直接给结论。')
-  }
-  const tail = text.slice(-180)
-  const closer = CLOSER.exec(tail)
-  if (closer) {
-    add('warn', 'closer', `结尾出现「${closer[0]}」，在征求额外任务。`, '删掉；需要下一步就只给一个具体动作。')
-  }
-
-  // 6. Language mismatch.
-  const total = cjk + latinCount(narrative)
-  const cjkRatio = total === 0 ? 0 : cjk / total
-  if (language === 'zh' && cjkRatio < 0.2 && latinCount(narrative) > 40) {
-    add('warn', 'language', '用户用中文提问，但这条回复几乎全是英文。', '用中文写；命令、路径、报错原文保持原样。')
-  }
-  const hasNext = /下一步|接下来要|接下来请|你可以先|建议你先|Next step/i.test(text)
-  if (!hasNext && (cjk > 200 || lines.length > 10)) {
-    add('warn', 'next', '回复较长，但没有给出一个明确的下一步。', '结尾补一个用户现在就能做的事；没有就写"没有待办"。')
-  }
-
-  return {
-    ok: !issues.some((i) => i.severity === 'error'),
-    issues,
-    stats: {
-      chars: text.length,
-      cjk,
-      lines: lines.length,
-      cjkRatio: Number(cjkRatio.toFixed(3)),
-      identifiers: unknown.length,
-      backrefs: backrefs.length,
-      maxListRun: maxRun,
-      hasNext,
-    },
-  }
-}
-
-export { cjkCount, latinCount, len }
+// Reply checker for dsh-plain-talk.
+//
+// This is deliberately a checker, not a gate: it reports what a reader would
+// stumble on, with the exact replacement form. The model decides what to keep.
+//
+// Three principles learned from real sessions in this workspace:
+//  - Common terms and project names are not errors. Only opaque codes the
+//    reader has never seen (EXP-7, DEV-1K, run-3) need a first-use gloss.
+//  - A pointer is not a citation: `file.py:42` alone sends the reader away.
+//  - Cross-turn wording ("上述", "上一轮") is the main reason a report cannot
+//    be read on its own.
+
+const ALLOW = new Set([
+  'api', 'json', 'jsonl', 'yaml', 'yml', 'html', 'css', 'http', 'https', 'url', 'uri',
+  'cli', 'ui', 'ux', 'gpu', 'cpu', 'ram', 'ssd', 'hdd', 'ai', 'llm', 'sdk', 'ide',
+  'csv', 'tsv', 'ts', 'js', 'py', 'sh', 'md', 'txt', 'png', 'jpg', 'jpeg', 'svg',
+  'gif', 'webp', 'sql', 'xml', 'toml', 'npm', 'pnpm', 'git', 'ssh', 'dns', 'tcp',
+  'udp', 'tls', 'ssl', 'jwt', 'oauth', 'cors', 'orm', 'crud', 'rest', 'grpc',
+  'wasm', 'dom', 'smtp', 'imap', 'vpn', 'utf8', 'utf-8', 'utf16', 'utf-16',
+  'sha1', 'sha256', 'sha-256', 'sha512', 'sha-512', 'md5', 'x86', 'x64', 'arm64',
+  'mp3', 'mp4', 'pdf', 'h264', 'hevc', 'ipv4', 'ipv6', 'k8s', 'i18n', 'l10n',
+  '2d', '3d', '4k', 'v1', 'v2', 'v3', 'v4', 'ci', 'pr', 'id', 'ok', 'vs', 'iso',
+  'doi', 'arxiv', 'p50', 'p95', 'p99', '2x', '4x', '8x', '16x', 'gpt', 'llm',
+  'linux', 'windows', 'macos', 'node', 'react', 'github', 'python', 'javascript',
+  'typescript', 'docker', 'sqlite', 'postgres', 'redis',
+])
+
+const MODEL_FAMILY = /^(?:gpt|claude|gemini|llama|qwen|deepseek|mistral|phi|o1|o3|o4|glm|kimi|mimo|hy)\d*[-.]?\d*/i
+
+const BACKREF_ZH = [
+  '上述', '上面', '前述', '前文', '前面提到', '前面说', '之前提到', '之前说',
+  '此前', '上一步', '上一轮', '前一轮', '刚才', '该方案', '该项', '该任务',
+  '前者', '后者', '如上', '见上文', '先前', '那次', '前面', '之前',
+]
+const BACKREF_EN = [
+  'as mentioned', 'aforementioned', 'as discussed', 'the above', 'previous step',
+  'earlier', 'previously', 'the former', 'the latter',
+]
+
+const PLUMBING = [
+  /\bsession[-_][0-9a-z-]{6,}/i,
+  /子代理|子会话|父代理|父会话|\bsubagent\b|\bparent agent\b|report sent to/i,
+  /\b(?:plain_talk_check|run_code|ops_bash|ops_grep|todo_write)\b/,
+]
+
+const PREAMBLE = /^(好的|好|当然|没问题|收到|明白|我来|让我|让我先|我们先|首先，我|先看一[下看])/
+const CLOSER = /(需要我|要不要|如果你愿意|希望(这|以上)|随时告诉我|还需要我|要我继续)/
+
+const FILE_EXT = 'py|js|mjs|cjs|ts|tsx|json|jsonl|md|ya?ml|toml|sh|ps1|bat|csv|tsv|txt|ipynb|log|html|css|tex|bib|pdf|zip|sql|db|pt|pth|ckpt|npy|npz|parquet|docx|xlsx|pptx'
+
+const IDENT_PATTERNS = [
+  ['file', new RegExp(`(?:[A-Za-z]:[\\\\/]|\\/)?[A-Za-z0-9_.-]+(?:[\\\\/][A-Za-z0-9_.-]+)*\\.(?:${FILE_EXT})\\b`, 'gi')],
+  ['hash', /\b[0-9a-f]{7,40}\b/gi],
+  ['snake', /\b[A-Za-z][A-Za-z0-9]*(?:_[A-Za-z0-9]+)+\b/g],
+  ['camel', /\b[a-z]+[A-Z][A-Za-z0-9]*\b|\b(?:[A-Z][a-z0-9]+){2,}[A-Za-z0-9]*\b|\b[A-Z][a-z]+[A-Z]{2,}[A-Za-z0-9]*\b/g],
+  ['acronym', /\b[A-Z]{2,8}\b/g],
+  ['code', /\b[A-Za-z]{1,10}[-_.]?\d{1,6}[A-Za-z]?\b|\b\d{1,4}[A-Za-z]{1,6}\b/g],
+]
+
+export const LIMITS = {
+  boxChars: 300,
+  boxLines: 12,
+  listRun: 5,
+  identifierWarns: 8,
+}
+
+const cjkCount = (s) => (s.match(/[\u4e00-\u9fff]/g) ?? []).length
+const latinCount = (s) => (s.match(/[A-Za-z]/g) ?? []).length
+const len = (s) => [...s].length
+
+function stripCode(text) {
+  return text
+    .replace(/```[\s\S]*?```/g, ' ')
+    .replace(/`[^`\n]*`/g, ' ')
+}
+
+export function isCommonTerm(term) {
+  const text = String(term ?? '').trim()
+  if (!text) return false
+  return ALLOW.has(text.toLowerCase()) || MODEL_FAMILY.test(text)
+}
+
+export function scanIdentifiers(input) {  const text = stripCode(String(input ?? ''))
+  const found = []
+  for (const [kind, re] of IDENT_PATTERNS) {
+    re.lastIndex = 0
+    let m
+    while ((m = re.exec(text))) {
+      const raw = m[0].replace(/[.,;:]+$/, '')
+      if (!raw) continue
+      found.push({ kind, token: raw, start: m.index, end: m.index + raw.length })
+    }
+  }
+  found.sort((a, b) => a.start - b.start || (b.end - b.start) - (a.end - a.start))
+  const accepted = []
+  for (const item of found) {
+    if (accepted.some((a) => item.start < a.end && item.end > a.start)) continue
+    accepted.push(item)
+  }
+  return accepted
+}
+
+function isAllowed(token, known) {
+  const lower = token.toLowerCase()
+  if (ALLOW.has(lower)) return true
+  if (MODEL_FAMILY.test(token)) return true
+  // Durations and multipliers ("10m", "3s", "4x") are measurements, not names.
+  if (/^\d+(?:ms|s|m|h|d|w|k|m|b|x)$/i.test(token)) return true
+  // Product-style names (AcmeService_V2, DataHub_V3) are project vocabulary the
+  // reader already holds; a lower-case internal key (tier_reaudit) is not.
+  if (/^[A-Za-z0-9]+(?:_[A-Za-z0-9]+)+$/.test(token) && (token.match(/[A-Z]/g) ?? []).length >= 2) return true
+  if (known && (known.has(token) || known.has(lower))) return true
+  return false
+}
+
+function explainedAfter(text, end) {
+  const after = text.slice(end, end + 60)
+  const m = /^\s*[（(]([^）)]{4,})[）)]/.exec(after)
+  if (!m) return false
+  return cjkCount(m[1]) >= 3 || latinCount(m[1]) >= 8
+}
+
+function knownFrom(text, known) {
+  return { text, known }
+}
+
+/**
+ * Analyze one reply draft.
+ *
+ * @param draft - the reply text the model is about to send.
+ * @param options - `knownTokens` (project/user vocabulary) and `language`
+ *                  (`zh` when the user writes Chinese).
+ * @returns issues plus the measurements the README benchmark reports.
+ */
+export function analyze(draft, options = {}) {
+  const text = String(draft ?? '')
+  const known = new Set((options.knownTokens ?? []).map((t) => String(t)))
+  const knownLower = new Set([...known].map((t) => t.toLowerCase()))
+  const language = options.language ?? 'unknown'
+  const issues = []
+
+  const add = (severity, rule, message, fix) => issues.push({ severity, rule, message, fix })
+
+  const narrative = stripCode(text)
+
+  // 1. Cross-turn wording.
+  const backrefs = [
+    ...BACKREF_ZH.filter((w) => narrative.includes(w)),
+    ...BACKREF_EN.filter((w) => narrative.toLowerCase().includes(w)),
+  ]
+  if (backrefs.length) {
+    add('error', 'backref', `出现回指词：${backrefs.slice(0, 5).join('、')}。`, '把指代的东西重新写一遍，读者没有前文可翻。')
+  }
+
+  // 2. Internal plumbing.
+  const plumbing = PLUMBING.find((re) => re.test(narrative))
+  if (plumbing) {
+    add('error', 'internal', '出现会话号、子代理或工具名等内部信息。', '删掉；只写用户能看见的对象。')
+  }
+
+  // 3. Pointer-only citations.
+  const pointers = []
+  const pointerRe = new RegExp(`(?:[A-Za-z]:[\\\\/]|/)?[\\w.-]+(?:[\\\\/][\\w.-]+)*\\.(?:${FILE_EXT}):\\d+`, 'gi')
+  pointerRe.lastIndex = 0
+  let pm
+  while ((pm = pointerRe.exec(text))) pointers.push(pm[0])
+  if (pointers.length) {
+    const unquoted = pointers.filter((pointer) => {
+      const at = text.indexOf(pointer)
+      const window = text.slice(Math.max(0, at - 220), at + pointer.length + 220)
+      return !/「[^」]{4,}」|"[^"]{4,}"|『[^』]{4,}』|`[^`]{4,}`|```/.test(window)
+    })
+    if (unquoted.length) {
+      add('error', 'citation', `只有位置没有原文：${unquoted.slice(0, 3).join('、')}。`, '把那几行原文贴出来，再标位置。')
+    }
+  }
+
+  // 4. Unexplained opaque identifiers.
+  const seen = new Set()
+  const unknown = []
+  for (const token of scanIdentifiers(narrative)) {
+    if (seen.has(token.token)) continue
+    seen.add(token.token)
+    if (isAllowed(token.token, known) || isAllowed(token.token, knownLower)) continue
+    if (explainedAfter(narrative, token.end)) continue
+    unknown.push(token.token)
+  }
+  if (unknown.length) {
+    add(
+      'warn',
+      'identifier',
+      `有 ${unknown.length} 个代号没有说明：${unknown.slice(0, 6).join('、')}${unknown.length > 6 ? ' 等' : ''}。`,
+      '第一次出现写成「EXP-7（一句话说明它是什么）」；用户在本对话用过的词和项目名不用解释。',
+    )
+  }
+
+  // 5. Shape.
+  const lines = text.split('\n').filter((l) => l.trim())
+  const cjk = cjkCount(narrative)
+  if (cjk > LIMITS.boxChars || lines.length > LIMITS.boxLines) {
+    add('warn', 'length', `回复有 ${cjk} 个汉字、${lines.length} 行，超过默认的 ${LIMITS.boxChars} 字 / ${LIMITS.boxLines} 行。`, '删掉读者不决策所需的内容；用户要细节时再展开。')
+  }
+
+  let run = 0
+  let maxRun = 0
+  for (const line of text.split('\n')) {
+    if (/^\s*(?:[-*+]|\d+[.)])\s+/.test(line)) { run++; maxRun = Math.max(maxRun, run) } else if (!line.trim()) { run = 0 }
+  }
+  if (maxRun > LIMITS.listRun) {
+    add('warn', 'list', `有一段列表 ${maxRun} 条，超过 5 条。`, '合并或只留最要紧的五条。')
+  }
+
+  const first = lines[0] ?? ''
+  if (PREAMBLE.test(first)) {
+    add('warn', 'preamble', `开头是「${first.slice(0, 20)}」，属于寒暄或过程宣告。`, '第一行直接给结论。')
+  }
+  const tail = text.slice(-180)
+  const closer = CLOSER.exec(tail)
+  if (closer) {
+    add('warn', 'closer', `结尾出现「${closer[0]}」，在征求额外任务。`, '删掉；需要下一步就只给一个具体动作。')
+  }
+
+  // 6. Language mismatch.
+  const total = cjk + latinCount(narrative)
+  const cjkRatio = total === 0 ? 0 : cjk / total
+  if (language === 'zh' && cjkRatio < 0.2 && latinCount(narrative) > 40) {
+    add('warn', 'language', '用户用中文提问，但这条回复几乎全是英文。', '用中文写；命令、路径、报错原文保持原样。')
+  }
+  const hasNext = /下一步|接下来要|接下来请|你可以先|建议你先|Next step/i.test(text)
+  if (!hasNext && (cjk > 200 || lines.length > 10)) {
+    add('warn', 'next', '回复较长，但没有给出一个明确的下一步。', '结尾补一个用户现在就能做的事；没有就写"没有待办"。')
+  }
+
+  return {
+    ok: !issues.some((i) => i.severity === 'error'),
+    issues,
+    stats: {
+      chars: text.length,
+      cjk,
+      lines: lines.length,
+      cjkRatio: Number(cjkRatio.toFixed(3)),
+      identifiers: unknown.length,
+      backrefs: backrefs.length,
+      maxListRun: maxRun,
+      hasNext,
+    },
+  }
+}
+
+export { cjkCount, latinCount, len }

package/dsh/rules.js

@@ -1,44 +1,44 @@
-// The system-prompt section injected while "说人话" is on for the session.
-//
-// Design notes:
-//  - The rules target the reader's cost, not the model's output length alone:
-//    a short reply that still needs the previous turns is still unreadable.
-//  - Numbers in the comment block below come from a scan of 197 real sessions
-//    in this workspace (median 2,076 characters, 96% with unexplained codes,
-//    22% with back-references, 43% almost entirely English).
-//  - Keep the text shorter than the replies it is meant to shorten.
-
-export const RULES = `# 说人话（回复优化）
-
-读者只看这一条回复就要能做决定。先给结论，再给必要细节。
-
-## 长度
-- 能一句说清就一句；默认不超过 12 行、300 字。用户明确要细节时放开深度，但结论仍在第一行。
-- 不写"好的/我来看看/让我先…/希望这有帮助"；不复述问题；不汇报你准备怎么做。
-- 不写过程日记（"我先…然后…"）。只写结果、判断和可验证的事实。
-- 列表最多 5 条；只给一个下一步。结尾不要问"还需要我做什么"。
-
-## 自包含
-- 不用"上述/前面/之前提到的/上一轮/前者/后者/刚才"；要指代就重说一遍。
-- 代号、任务号、版本号、提交号、模块名（如 M11、dev-1200、c590）第一次出现时写清它是什么：M11（把实例验证加进推理流程的实验编号）。之后可以直接用。
-- 但用户在本对话里用过的词、项目名、常见术语（arXiv、DOI、API、GPU 等）默认读者已经懂，不要解释，也不要改写成别的说法。
-- 用户让你"调研 / 研究 / 查一下 X"时，第一遍回复默认读者还不了解 X，先把它讲清楚；这条回复交付之后，X 才进入词汇清单。
-- 不出现会话号、子代理、父会话、工具名等内部信息。
-
-## 引用与数字
-- 引用文件、日志、网页时直接贴出那几行原文；只给"文件名:行号"不算证据。
-- 数字给单位和对照："61.2% → 63.8%"，不写"明显提升"。
-
-## 语言
-- 用用户最近一条消息的语言回复。思考过程可以用英文。
-- 代码、命令、路径、报错原文保持原样，不翻译、不缩写。
-
-## 执行类回复
-用户给过计划、任务清单或验收要求时，第一行给结论：达标 / 部分达标 / 未达标，并点明哪几条没达标。
-验收标准只用用户原话，不自己发明。然后按顺序：做了什么 → 没做到什么 → 用户怎么自己确认（贴出应看到的内容）→ 一个下一步。
-
-## 收尾自检
-完成类回复或超过 12 行的回复，发送前调用 plain_talk_check；把 error 全部改掉，warn 逐条判断。
-
-## 覆盖规则
-用户说"展开讲/详细说明"时，保留结论先行，正文可以长。破坏性操作仍然先确认。harness 规则优先于本节。`;
+// The system-prompt section injected while "说人话" is on for the session.
+//
+// Design notes:
+//  - The rules target the reader's cost, not the model's output length alone:
+//    a short reply that still needs the previous turns is still unreadable.
+//  - Numbers in the comment block below come from a scan of 197 real sessions
+//    in this workspace (median 2,076 characters, 96% with unexplained codes,
+//    22% with back-references, 43% almost entirely English).
+//  - Keep the text shorter than the replies it is meant to shorten.
+
+export const RULES = `# 说人话（回复优化）
+
+读者只看这一条回复就要能做决定。先给结论，再给必要细节。
+
+## 长度
+- 能一句说清就一句；默认不超过 12 行、300 字。用户明确要细节时放开深度，但结论仍在第一行。
+- 不写"好的/我来看看/让我先…/希望这有帮助"；不复述问题；不汇报你准备怎么做。
+- 不写过程日记（"我先…然后…"）。只写结果、判断和可验证的事实。
+- 列表最多 5 条；只给一个下一步。结尾不要问"还需要我做什么"。
+
+## 自包含
+- 不用"上述/前面/之前提到的/上一轮/前者/后者/刚才"；要指代就重说一遍。
+- 代号、任务号、版本号、提交号、模块名（如 EXP-7、DEV-1K、run-3）第一次出现时写清它是什么：EXP-7（把实例校验加进推理流程的实验编号）。之后可以直接用。
+- 但用户在本对话里用过的词、项目名、常见术语（arXiv、DOI、API、GPU 等）默认读者已经懂，不要解释，也不要改写成别的说法。
+- 用户让你"调研 / 研究 / 查一下 X"时，第一遍回复默认读者还不了解 X，先把它讲清楚；这条回复交付之后，X 才进入词汇清单。
+- 不出现会话号、子代理、父会话、工具名等内部信息。
+
+## 引用与数字
+- 引用文件、日志、网页时直接贴出那几行原文；只给"文件名:行号"不算证据。
+- 数字给单位和对照："61.2% → 63.8%"，不写"明显提升"。
+
+## 语言
+- 用用户最近一条消息的语言回复。思考过程可以用英文。
+- 代码、命令、路径、报错原文保持原样，不翻译、不缩写。
+
+## 执行类回复
+用户给过计划、任务清单或验收要求时，第一行给结论：达标 / 部分达标 / 未达标，并点明哪几条没达标。
+验收标准只用用户原话，不自己发明。然后按顺序：做了什么 → 没做到什么 → 用户怎么自己确认（贴出应看到的内容）→ 一个下一步。
+
+## 收尾自检
+完成类回复或超过 12 行的回复，发送前调用 plain_talk_check；把 error 全部改掉，warn 逐条判断。
+
+## 覆盖规则
+用户说"展开讲/详细说明"时，保留结论先行，正文可以长。破坏性操作仍然先确认。harness 规则优先于本节。`;