dsh-plain-talk 0.1.3 → 0.1.4

package/dsh/rules.js

@@ -23,6 +23,7 @@ export const RULES = `# 说人话（回复优化）
 - 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。
+- 内部代号（EXP-7、DEV-1K、run-3 这类）不进入词汇清单：每次出现都按未知处理，第一次出现时解释。
 - 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

package/dsh/vocab.js

@@ -1,215 +1,229 @@
-// The user vocabulary file.
-//
-// One file for the whole profile, not per Session: "the user already knows this
-// word" is a property of the reader, not of one conversation. Sessions only add
-// observations; the checker reads the result.
-//
-// Categories, from broadest to most specific:
-//   common        常识词汇 — built in, never needs an explanation
-//   domain        专业词汇 — field terms the user has already met
-//   project       项目 / 产品名 — repositories, products, workspace names
-//   internal      内部代号 — experiment ids, task numbers, build codes
-//   unclassified  待归类 — recorded but not yet classified
-//
-// Two-phase learning for research requests: when the user says "调研 X", X is
-// marked pending for that Session, so the first reply must still explain it;
-// the next turn end promotes it into the vocabulary.
-
-import { mkdirSync, readFileSync, renameSync, writeFileSync } from 'node:fs'
-import { dirname } from 'node:path'
-
-export const CATEGORIES = {
-  common: '常识词汇',
-  domain: '专业词汇',
-  project: '项目 / 产品名',
-  internal: '内部代号',
-  unclassified: '待归类',
-}
-
-const MAX_TERMS = 2000
-const MAX_PENDING_PER_SESSION = 20
-
-const cjkCount = (s) => (s.match(/[\u4e00-\u9fff]/g) ?? []).length
-
-/** Classify one learned term from its shape alone. */
-export function classify(term, cwdName) {
-  const text = String(term ?? '').trim()
-  if (!text) return 'unclassified'
-  if (cwdName && text === cwdName) return 'project'
-  if (/^[A-Za-z]{1,10}[-_.]?\d{1,6}$/.test(text)) return 'internal'
-  if (/^[A-Za-z0-9]+(?:_[A-Za-z0-9]+)+$/.test(text) && (text.match(/[A-Z]/g) ?? []).length >= 2) return 'project'
-  if (/^[A-Z][A-Za-z0-9]+$/.test(text) && /[A-Z]/.test(text.slice(1))) return 'project'
-  if (cjkCount(text) >= 2) return 'domain'
-  if (/^[A-Z]{2,8}$/.test(text)) return 'domain'
-  return 'unclassified'
-}
-
-export function emptyVocab() {
-  return { version: 1, updatedAt: new Date(0).toISOString(), terms: {}, pending: {} }
-}
-
-export function loadVocab(file) {
-  try {
-    const parsed = JSON.parse(readFileSync(file, 'utf8'))
-    const terms = {}
-    const raw = parsed && typeof parsed === 'object' && parsed.terms && typeof parsed.terms === 'object' ? parsed.terms : {}
-    for (const [key, value] of Object.entries(raw)) {
-      if (!key || key.length > 80) continue
-      if (!value || typeof value !== 'object') continue
-      const category = Object.prototype.hasOwnProperty.call(CATEGORIES, value.category) ? value.category : 'unclassified'
-      terms[key] = {
-        category,
-        firstSeen: typeof value.firstSeen === 'string' ? value.firstSeen : new Date(0).toISOString(),
-        lastSeen: typeof value.lastSeen === 'string' ? value.lastSeen : new Date(0).toISOString(),
-        count: Number.isFinite(value.count) ? value.count : 1,
-        sessions: Array.isArray(value.sessions) ? value.sessions.filter((s) => typeof s === 'string').slice(-5) : [],
-        source: typeof value.source === 'string' ? value.source : 'user',
-      }
-    }
-    const pending = {}
-    const rawPending = parsed && typeof parsed === 'object' && parsed.pending && typeof parsed.pending === 'object' ? parsed.pending : {}
-    for (const [sessionId, list] of Object.entries(rawPending)) {
-      if (!Array.isArray(list)) continue
-      pending[sessionId] = list
-        .filter((item) => item && typeof item.term === 'string' && item.term.length > 0 && item.term.length <= 80)
-        .slice(-MAX_PENDING_PER_SESSION)
-        .map((item) => ({ term: item.term, at: typeof item.at === 'string' ? item.at : new Date(0).toISOString() }))
-    }
-    return { version: 1, updatedAt: typeof parsed.updatedAt === 'string' ? parsed.updatedAt : new Date(0).toISOString(), terms, pending }
-  } catch {
-    return emptyVocab()
-  }
-}
-
-export function saveVocab(file, vocab) {
-  const next = { ...vocab, version: 1, updatedAt: new Date().toISOString() }
-  mkdirSync(dirname(file), { recursive: true })
-  const tmp = `${file}.${process.pid}.${Date.now()}.tmp`
-  writeFileSync(tmp, `${JSON.stringify(next, null, 2)}\n`, 'utf8')
-  renameSync(tmp, file)
-  return next
-}
-
-function evict(vocab) {
-  const keys = Object.keys(vocab.terms)
-  if (keys.length <= MAX_TERMS) return
-  keys
-    .sort((a, b) => String(vocab.terms[a].lastSeen).localeCompare(String(vocab.terms[b].lastSeen)))
-    .slice(0, keys.length - MAX_TERMS)
-    .forEach((key) => delete vocab.terms[key])
-}
-
-/** Record one term as known. Existing terms keep their category. */
-export function addTerm(vocab, term, options = {}) {
-  const key = String(term ?? '').trim()
-  if (!key || key.length > 80) return vocab
-  const now = options.now ?? new Date().toISOString()
-  const session = typeof options.session === 'string' ? options.session : undefined
-  const existing = vocab.terms[key]
-  if (existing) {
-    existing.lastSeen = now
-    existing.count += 1
-    if (session && !existing.sessions.includes(session)) existing.sessions = [...existing.sessions.slice(-4), session]
-    if (options.category && CATEGORIES[options.category]) existing.category = options.category
-  } else {
-    vocab.terms[key] = {
-      category: options.category && CATEGORIES[options.category] ? options.category : classify(key, options.cwdName),
-      firstSeen: now,
-      lastSeen: now,
-      count: 1,
-      sessions: session ? [session] : [],
-      source: options.source ?? 'user',
-    }
-  }
-  evict(vocab)
-  return vocab
-}
-
-export function removeTerm(vocab, term) {
-  delete vocab.terms[String(term ?? '').trim()]
-  return vocab
-}
-
-export function setCategory(vocab, term, category) {
-  const key = String(term ?? '').trim()
-  if (!vocab.terms[key] || !CATEGORIES[category]) return vocab
-  vocab.terms[key].category = category
-  return vocab
-}
-
-/** Mark a research subject as not-yet-known for this Session. */
-export function markPending(vocab, sessionId, term, now = new Date().toISOString()) {
-  const key = String(term ?? '').trim()
-  if (!sessionId || !key || key.length > 80) return vocab
-  const list = vocab.pending[sessionId] ?? []
-  if (!list.some((item) => item.term === key)) list.push({ term: key, at: now })
-  vocab.pending[sessionId] = list.slice(-MAX_PENDING_PER_SESSION)
-  return vocab
-}
-
-export function pendingFor(vocab, sessionId) {
-  return (vocab.pending?.[sessionId] ?? []).map((item) => item.term)
-}
-
-/** Turn end: the reply has been delivered, so the research subject is now known. */
-export function promotePending(vocab, sessionId, options = {}) {
-  const list = vocab.pending?.[sessionId] ?? []
-  if (list.length === 0) return []
-  for (const item of list) addTerm(vocab, item.term, { ...options, session: sessionId, category: options.category ?? 'domain', source: 'research' })
-  delete vocab.pending[sessionId]
-  return list.map((item) => item.term)
-}
-
-/**
- * Terms the checker may treat as known for one Session.
- * Pending research subjects are excluded even when the profile learned them
- * earlier: the user asked to research them, so the first reply must not assume.
- */
-export function knownSet(vocab, sessionId) {
-  const suspended = new Set(pendingFor(vocab, sessionId))
-  const known = new Set()
-  for (const key of Object.keys(vocab.terms)) {
-    if (!suspended.has(key)) known.add(key)
-  }
-  return known
-}
-
-export function listByCategory(vocab) {
-  const groups = {}
-  for (const key of Object.keys(CATEGORIES)) groups[key] = []
-  for (const [term, value] of Object.entries(vocab.terms)) {
-    const category = CATEGORIES[value.category] ? value.category : 'unclassified'
-    groups[category].push({ term, count: value.count, lastSeen: value.lastSeen })
-  }
-  for (const key of Object.keys(groups)) groups[key].sort((a, b) => b.count - a.count || a.term.localeCompare(b.term))
-  return groups
-}
-
-/** Extract "research X" subjects from one user message. */
-export function researchTargets(text) {
-  const found = []
-  const input = String(text ?? '')
-  // A captured phrase may carry a code inside it ("调研 XR-77 这个实验"):
-  // the code is the subject, not the surrounding words.
-  const codePattern = /\b[A-Za-z]{1,10}[-_.]?\d{1,6}[A-Za-z]?\b/g
-  const push = (phrase) => {
-    const codes = phrase.match(codePattern) ?? []
-    if (codes.length > 0) {
-      for (const code of codes) if (code.length >= 2) found.push(code)
-      return
-    }
-    if (phrase.length >= 2) found.push(phrase)
-  }
-  const zh = /(?:调研|研究|调查|了解一下|了解下|查一下|查查|梳理一下|分析一下)\s*([^\n。！？；;]{1,40})/g
-  let m
-  while ((m = zh.exec(input))) {
-    const term = m[1].trim().replace(/^(?:一下|下)\s*/, '').replace(/[，,、\s]+$/, '')
-    push(term)
-  }
-  const en = /\b(?:research|investigate|look into|study)\s+([A-Za-z0-9][A-Za-z0-9 ._/-]{1,40})/gi
-  while ((m = en.exec(input))) {
-    const term = m[1].trim().replace(/[.,;:\s]+$/, '')
-    push(term)
-  }
-  return [...new Set(found)]
-}
+// The user vocabulary file.
+//
+// One file for the whole profile, not per Session: "the user already knows this
+// word" is a property of the reader, not of one conversation. Sessions only add
+// observations; the checker reads the result.
+//
+// Categories, from broadest to most specific:
+//   common        常识词汇 — built in, never needs an explanation
+//   domain        专业词汇 — field terms the user has already met
+//   project       项目 / 产品名 — repositories, products, workspace names
+//   unclassified  待归类 — recorded but not yet classified
+//
+// Codes (EXP-7, DEV-1K, run-3, commit hashes) are never stored: the plugin's
+// job is to make the model explain them, so remembering one would hide the
+// next unexplainable use.
+// Two-phase learning for research requests: when the user says "调研 X", X is
+// marked pending for that Session, so the first reply must still explain it;
+// the next turn end promotes it into the vocabulary.
+
+import { mkdirSync, readFileSync, renameSync, writeFileSync } from 'node:fs'
+import { dirname } from 'node:path'
+
+export const CATEGORIES = {
+  common: '常识词汇',
+  domain: '专业词汇',
+  project: '项目 / 产品名',
+  unclassified: '待归类',
+}
+
+const MAX_TERMS = 2000
+const MAX_PENDING_PER_SESSION = 20
+
+const cjkCount = (s) => (s.match(/[\u4e00-\u9fff]/g) ?? []).length
+
+// Codes never enter the vocabulary: explaining them is the plugin's job, so
+// remembering one would silence the next warning.
+const CODE_LIKE = /^[A-Za-z]{1,10}[-_.]?\d{1,6}[A-Za-z]?$|^\d{1,4}[A-Za-z]{1,6}$|^[0-9a-f]{7,40}$/i
+
+export function isCodeLike(term) {
+  return CODE_LIKE.test(String(term ?? '').trim())
+}
+
+/** Classify one learned term from its shape alone. */
+export function classify(term, cwdName) {
+  const text = String(term ?? '').trim()
+  if (!text) return 'unclassified'
+  if (isCodeLike(text)) return null
+  if (cwdName && text === cwdName) return 'project'
+  if (/^[A-Za-z0-9]+(?:_[A-Za-z0-9]+)+$/.test(text) && (text.match(/[A-Z]/g) ?? []).length >= 2) return 'project'
+  if (/^[A-Z][A-Za-z0-9]+$/.test(text) && /[A-Z]/.test(text.slice(1))) return 'project'
+  if (cjkCount(text) >= 2) return 'domain'
+  if (/^[A-Z]{2,8}$/.test(text)) return 'domain'
+  return 'unclassified'
+}
+
+export function emptyVocab() {
+  return { version: 1, updatedAt: new Date(0).toISOString(), terms: {}, pending: {} }
+}
+
+export function loadVocab(file) {
+  try {
+    const parsed = JSON.parse(readFileSync(file, 'utf8'))
+    const terms = {}
+    const raw = parsed && typeof parsed === 'object' && parsed.terms && typeof parsed.terms === 'object' ? parsed.terms : {}
+    for (const [key, value] of Object.entries(raw)) {
+      if (!key || key.length > 80) continue
+      if (isCodeLike(key)) continue
+      if (!value || typeof value !== 'object') continue
+      if (value.category === 'internal') continue
+      const category = Object.prototype.hasOwnProperty.call(CATEGORIES, value.category) ? value.category : 'unclassified'
+      terms[key] = {
+        category,
+        firstSeen: typeof value.firstSeen === 'string' ? value.firstSeen : new Date(0).toISOString(),
+        lastSeen: typeof value.lastSeen === 'string' ? value.lastSeen : new Date(0).toISOString(),
+        count: Number.isFinite(value.count) ? value.count : 1,
+        sessions: Array.isArray(value.sessions) ? value.sessions.filter((s) => typeof s === 'string').slice(-5) : [],
+        source: typeof value.source === 'string' ? value.source : 'user',
+      }
+    }
+    const pending = {}
+    const rawPending = parsed && typeof parsed === 'object' && parsed.pending && typeof parsed.pending === 'object' ? parsed.pending : {}
+    for (const [sessionId, list] of Object.entries(rawPending)) {
+      if (!Array.isArray(list)) continue
+      pending[sessionId] = list
+        .filter((item) => item && typeof item.term === 'string' && item.term.length > 0 && item.term.length <= 80)
+        .slice(-MAX_PENDING_PER_SESSION)
+        .map((item) => ({ term: item.term, at: typeof item.at === 'string' ? item.at : new Date(0).toISOString() }))
+    }
+    return { version: 1, updatedAt: typeof parsed.updatedAt === 'string' ? parsed.updatedAt : new Date(0).toISOString(), terms, pending }
+  } catch {
+    return emptyVocab()
+  }
+}
+
+export function saveVocab(file, vocab) {
+  const next = { ...vocab, version: 1, updatedAt: new Date().toISOString() }
+  mkdirSync(dirname(file), { recursive: true })
+  const tmp = `${file}.${process.pid}.${Date.now()}.tmp`
+  writeFileSync(tmp, `${JSON.stringify(next, null, 2)}\n`, 'utf8')
+  renameSync(tmp, file)
+  return next
+}
+
+function evict(vocab) {
+  const keys = Object.keys(vocab.terms)
+  if (keys.length <= MAX_TERMS) return
+  keys
+    .sort((a, b) => String(vocab.terms[a].lastSeen).localeCompare(String(vocab.terms[b].lastSeen)))
+    .slice(0, keys.length - MAX_TERMS)
+    .forEach((key) => delete vocab.terms[key])
+}
+
+/** Record one term as known. Existing terms keep their category. */
+export function addTerm(vocab, term, options = {}) {
+  const key = String(term ?? '').trim()
+  if (!key || key.length > 80) return vocab
+  if (isCodeLike(key)) return vocab
+  const category = options.category && CATEGORIES[options.category] ? options.category : classify(key, options.cwdName)
+  if (category === null) return vocab
+  const now = options.now ?? new Date().toISOString()
+  const session = typeof options.session === 'string' ? options.session : undefined
+  const existing = vocab.terms[key]
+  if (existing) {
+    existing.lastSeen = now
+    existing.count += 1
+    if (session && !existing.sessions.includes(session)) existing.sessions = [...existing.sessions.slice(-4), session]
+    if (options.category && CATEGORIES[options.category]) existing.category = options.category
+  } else {
+    vocab.terms[key] = {
+      category,
+      firstSeen: now,
+      lastSeen: now,
+      count: 1,
+      sessions: session ? [session] : [],
+      source: options.source ?? 'user',
+    }
+  }
+  evict(vocab)
+  return vocab
+}
+
+export function removeTerm(vocab, term) {
+  delete vocab.terms[String(term ?? '').trim()]
+  return vocab
+}
+
+export function setCategory(vocab, term, category) {
+  const key = String(term ?? '').trim()
+  if (!vocab.terms[key] || !CATEGORIES[category] || isCodeLike(key)) return vocab
+  vocab.terms[key].category = category
+  return vocab
+}
+
+/** Mark a research subject as not-yet-known for this Session. */
+export function markPending(vocab, sessionId, term, now = new Date().toISOString()) {
+  const key = String(term ?? '').trim()
+  if (!sessionId || !key || key.length > 80) return vocab
+  const list = vocab.pending[sessionId] ?? []
+  if (!list.some((item) => item.term === key)) list.push({ term: key, at: now })
+  vocab.pending[sessionId] = list.slice(-MAX_PENDING_PER_SESSION)
+  return vocab
+}
+
+export function pendingFor(vocab, sessionId) {
+  return (vocab.pending?.[sessionId] ?? []).map((item) => item.term)
+}
+
+/** Turn end: the reply has been delivered, so the research subject is now known. */
+export function promotePending(vocab, sessionId, options = {}) {
+  const list = vocab.pending?.[sessionId] ?? []
+  if (list.length === 0) return []
+  for (const item of list) addTerm(vocab, item.term, { ...options, session: sessionId, category: options.category ?? 'domain', source: 'research' })
+  delete vocab.pending[sessionId]
+  return list.map((item) => item.term)
+}
+
+/**
+ * Terms the checker may treat as known for one Session.
+ * Pending research subjects are excluded even when the profile learned them
+ * earlier: the user asked to research them, so the first reply must not assume.
+ */
+export function knownSet(vocab, sessionId) {
+  const suspended = new Set(pendingFor(vocab, sessionId))
+  const known = new Set()
+  for (const key of Object.keys(vocab.terms)) {
+    if (!suspended.has(key)) known.add(key)
+  }
+  return known
+}
+
+export function listByCategory(vocab) {
+  const groups = {}
+  for (const key of Object.keys(CATEGORIES)) groups[key] = []
+  for (const [term, value] of Object.entries(vocab.terms)) {
+    const category = CATEGORIES[value.category] ? value.category : 'unclassified'
+    groups[category].push({ term, count: value.count, lastSeen: value.lastSeen })
+  }
+  for (const key of Object.keys(groups)) groups[key].sort((a, b) => b.count - a.count || a.term.localeCompare(b.term))
+  return groups
+}
+
+/** Extract "research X" subjects from one user message. */
+export function researchTargets(text) {
+  const found = []
+  const input = String(text ?? '')
+  // A captured phrase may carry a code inside it ("调研 XR-77 这个实验"):
+  // the code is the subject, not the surrounding words.
+  const codePattern = /\b[A-Za-z]{1,10}[-_.]?\d{1,6}[A-Za-z]?\b/g
+  const push = (phrase) => {
+    const codes = phrase.match(codePattern) ?? []
+    if (codes.length > 0) {
+      for (const code of codes) if (code.length >= 2) found.push(code)
+      return
+    }
+    if (phrase.length >= 2) found.push(phrase)
+  }
+  const zh = /(?:调研|研究|调查|了解一下|了解下|查一下|查查|梳理一下|分析一下)\s*([^\n。！？；;]{1,40})/g
+  let m
+  while ((m = zh.exec(input))) {
+    const term = m[1].trim().replace(/^(?:一下|下)\s*/, '').replace(/[，,、\s]+$/, '')
+    push(term)
+  }
+  const en = /\b(?:research|investigate|look into|study)\s+([A-Za-z0-9][A-Za-z0-9 ._/-]{1,40})/gi
+  while ((m = en.exec(input))) {
+    const term = m[1].trim().replace(/[.,;:\s]+$/, '')
+    push(term)
+  }
+  return [...new Set(found)]
+}