rcs-context 0.0.0-stage → 0.1.39

package/web/memory-durability.mjs

@@ -0,0 +1,431 @@
+// Deterministic durability policy for Cycle 2 proposals.
+//
+// The model is good at copying text and bad at deciding whether a user message is
+// a standing project rule or a one-off task/question. That decision is therefore
+// made here, from the archived text alone, before any model call and again inside
+// the evidence validator. The model still chooses the exact quote; this module
+// only bounds where a quote may come from and how an overlapping predicate is
+// named.
+//
+// Nothing here reads a clock, a model, or mutable state. Every result is a pure
+// function of the input text, so a replay produces the same spans and predicates.
+
+export const DURABILITY_POLICY_VERSION = 'durable-v2'
+
+const WHITESPACE_CHAR = /\s/
+
+const RE = {
+  // Prompt-injection / role confusion. Explicit instructions to rewrite memory are
+  // never evidence, even when the text uses a durable-sounding label.
+  injection: [
+    /^\s*(SYSTEM|SYSTEM\s+OVERRIDE|系统指令|系统提示|SYSTEM\s+PROMPT)\s*[:：]/i,
+    /(忽略|无视|忘记).{0,8}(之前|以上|前面|原有).{0,6}(规则|指令|约束|设定)/,
+    /(把|将)(这条|这句话|上面这句|此条).{0,12}(写成|记为|记录为|设为|当作).{0,8}(约束|规则|偏好|决策|事实)/,
+  ],
+  assistantProse: [
+    /^\s*(我会记住|我已记住|已经记住|已为你|已完成|已处理|总结如下|根据以上|以下是)/,
+    /(我(会|将)(记住|记录)|已(经)?(记录|记住)|根据(你|您)的(要求|偏好))/,
+  ],
+  codeDump: [
+    /^\s*```/,
+    /^\s*(const|let|var|function|class|import|export)\s/m,
+  ],
+  question: [
+    /[？?]\s*$/,
+    /(吗|呢)\s*[。.]?\s*$/,
+    // A request for the current turn ("简单些好吗", "这样可以吗") is not a standing
+    // fact even when it carries a modal like 不要. The marker can sit mid-sentence.
+    /(好吗|行吗|可以吗|好不好|成吗|行不行|可不可以)/,
+    // A choice or a direct question is not a rule even without a question mark.
+    /^\s*(?:要不要|是否要|是否应该|为什么|为何|哪种|哪个|哪一个)/,
+    /(?:用|选|采用|使用|走)[^。！？；]{0,12}还是[^。！？；]{0,16}/,
+    /(?:有何|有哪些)[^。！？；]{0,12}[？?]/,
+    /(?:是什么|为什么)[^。！？；]{0,12}[？?]/,
+    /(?:是否(?:还)?(?:需要|要|应该|可以|能))/,
+    /(?:不是|难道)[^。！？；]{0,20}吗/,
+  ],
+  uncertain: [
+    /(我(还)?没(想好|定|确定)|还没想好|还没定|未定|待定|等有结论|再说吧|回头再说|具体方案还没定|拿不准|不确定(要|是否|用|选|改|做)|你(觉得|看)呢)/,
+  ],
+  transient: [
+    /(这次|本次|临时|暂时|先试|试试看|试一下|跑一下|抓一次|测试完|先这样)/,
+    // "这条回复/这个任务" scopes an otherwise durable-looking preference or rule
+    // to the current request.
+    /(?:这|本)(?:条|轮|次|个)(?:回复|回答|任务|请求|消息)/,
+  ],
+  plan: [
+    /(打算|计划|准备(?!好)|回头|下次再|以后再说|等有结论)/,
+    /(如果|假如).{0,24}(发布|上线|发版|合并|关掉|恢复原设置)/,
+  ],
+  transientError: [
+    /(刚才|刚刚|这次|本次).{0,24}(报错|报了|错误|失败|超时|异常).{0,12}(重试|再试|试试|一下|再跑|看下)/,
+    /(构建|测试|运行|安装).{0,8}(失败|报错).{0,12}(重试|再跑|再试)/,
+  ],
+  fileObservation: [
+    /(里写着|上写着|里已经写着|历史记录|只是历史|已经有.{0,16}(脚本|配置|文件|依赖))/,
+    /(README|package\.json|源码|文档)(里|中)\s*(写|说|提到|记录|配置)/i,
+  ],
+  // A tool or runtime that is unavailable, paired in the same sentence with the
+  // substitute to use. The substitute is required: "bash 不可用" alone is a
+  // status report, and "bash 不可用，试试 pwsh" is a one-off suggestion.
+  environment: [
+    /(?:shell|终端|命令行|解释器|运行时|运行环境|工具链|pwsh|powershell|bash|cmd(?:\.exe)?|python\d*(?:\.\d+)?|node(?:\.js)?)[^。！？；;]{0,40}?(?:不可用|用不了|不能用|没装|无法使用|不支持|不适配)[^。！？；;]{0,40}?(?:得|必须|只能|应当|应该|改用|统一|一律|用)\s*(?:用|使用|走)?\s*[A-Za-z]/i,
+  ],
+  assignmentOpen: [
+    /^\s*(你负责|你来做|你来|请你|请|帮我|麻烦|先|现在|这次|本次|临时|实现|修复|补足|运行|读取|写入|检查|启动|关闭|把|将|给我|重试|再试)/,
+  ],
+  // A command verb at the start of the message.
+  startImperative: [
+    /^\s*(?:改成|改为|调成|换成|部署|安装|卸载|创建|新建|添加|删除|移动|重命名|读取|写入|运行|执行|检查|启动|关闭|拉取|提交(?!前|后|信息|说明)|推送(?!前|后)|合并(?!前|后)|升级|降级|提升|达到|优化|降低|减少|增加|实现|修复)/,
+  ],
+  // A work chain ("…，然后重新部署") is request-scoped even when its opening clause
+  // looks like a decision ("端口改成 8080").
+  actionChain: [
+    /(然后|再|完成后|做完|之后|最后|接着)/,
+    /并(?:把|将|告诉|输出|贴|汇报|确认|运行|执行|检查|部署|安装|创建|添加|删除|修改|修复|重命名|读取|写入|提交|推送|合并|升级)/,
+  ],
+  directive: [
+    /(改成|改为|调成|换成|转为|部署|安装|卸载|创建|新建|添加|删除|移动|重命名|读取|写入|运行|执行|检查|启动|关闭|拉取|提交|推送|合并|升级|降级|提升|达到|优化|降低|减少|增加|实现|修复|清理|去除|制造|写|建|搭|明确|设置|配置|监测|打包|审验|传递|解决|调集|整理|修正|跑|做|看|改|加)/,
+  ],
+  // A first/second-person request frame anywhere in the message. It is only a task
+  // when paired with a directive and no standing-rule signal.
+  requestFrame: [
+    /(?:我(?:需要|要|想|希望)|你(?:需要|要|得|应该|可以)|你的任务|你要做的|我说(?:把|将|要)?|我们|请你|帮我|麻烦|给我|替我|可以做|再审|重新)/,
+  ],
+  // A status report, observation, pasted log, or acknowledgement. These carry
+  // project nouns and modals but describe what happened, not what must hold.
+  statusFrame: [
+    /^\s*(?:我注意到|我发现|我的理解|从全局来看|我不得不|已经|刚刚|目前|我在|我向|啊[，,]|都没有用|没有用|还是不行|已验证|已修复|已解决|已完成|成功了|搞定了|NO[，,])/i,
+    /^\s*(?:🧠|🟢|📋|✅|\[当前任务\]|Context Inject)/i,
+    /^(?:A|B|C)[，,]\s*我们(?:已经|摸清|知道)/,
+  ],
+  strongDurable: [
+    /(规则|约定|规范|标准|原则|约束)/,
+    /(记下|记住|记录一下|请记录|写进记忆|加入记忆)/,
+    /(以后|今后|往后|未来|长期|永久|始终|永远|每次|下一次)/,
+    /\b(?:add|put|keep)\s+(?:this|it|that)\s+(?:in|into|to)\s+(?:your\s+)?memory\b/i,
+    /\bremember\s+(?:that|this)\b/i,
+    /\bkeep\s+in\s+mind\b/i,
+  ],
+  explicitRule: [
+    /(规则|约定|规范|标准|原则|约束)\s*[:：]/,
+    /(项目|业务|接口|部署|测试|命名|结算|开发|团队|代码|提交|评测|安全|数据|协作|发布|架构|设计)[^，。；：]{0,6}(规则|约定|规范|标准|原则|约束)/,
+    /(这是|那是|作为|按照|根据|遵循)[^。；]{0,6}(规则|约定|规范|标准|原则|约束)/,
+  ],
+  designConstraint: [
+    /(?:不能|不得|不可|必须|应当|应该)(?:丢失|丢|破坏|省略|跳过|绕过|混用|忽略|覆盖)/,
+    /(?:是必要的|是危险的|必不可少|必须保留|不可替代)/,
+  ],
+  memo: [
+    /(记下|记住|记录一下|请记录|写进记忆|加入记忆)/,
+    /\b(?:add|put|keep)\s+(?:this|it|that)\s+(?:in|into|to)\s+(?:your\s+)?memory\b/i,
+    /\bremember\s+(?:that|this)\b/i,
+    /\bkeep\s+in\s+mind\b/i,
+  ],
+  future: [
+    /(以后|今后|往后|未来|长期|永久|始终|永远|每次|下一次)/,
+  ],
+  universal: [
+    /(一律|统一|总是|所有|任何|全部|默认|优先|从不|绝不)/,
+  ],
+  modal: [
+    /(必须|不得|不允许|禁止|不能|(?<!要)不要|别(?!人)|只允许|只能|最多|至少|应当|应该|可回滚|全绿|不得超过|保持|视为|固定为|需要固定)/,
+  ],
+  decision: [
+    /(决定|确定|定版|拍板|采用|改用|选用|迁移到|改为|改成|为准|最终方案|而不是|不引入|不再接受|不再适用|作废|撤回|取消|废弃|停用|撤销|回退|改回|不再使用)/,
+  ],
+  preference: [
+    /(偏好|更喜欢|喜欢|习惯|倾向|希望|不喜欢|prefer|would rather)/i,
+  ],
+  interaction: [
+    /(回答|回复|答复|沟通|先给结论|先给细节|先说明|先提问|提问|不要猜|命名|格式|风格|行文|报告|绝对路径|相对路径)/,
+  ],
+  retraction: [
+    /(作废|撤回|取消|废弃|停用|撤销|不再适用|不再使用|回退|改回|不用了|不要了)/,
+  ],
+  english: [
+    /\b(decided|decision|prefer|preference|must|never|always|should|requires?|required|avoid|convention|rule|constraint|from now on)\b/i,
+  ],
+  domain: [
+    /(金额|折扣|优惠券|结算|测试|依赖|接口|部署|生产|字段|单位|四舍五入|端口|目录|路径|文件|版本|编码|日志|数据库|缓存|并发|安全|权限|响应|请求|API|JSON|错误|时间|时区|精度|上限|下限|类型|编号|序号|schema)/i,
+  ],
+}
+
+function any(patterns, text) {
+  for (const pattern of patterns) {
+    if (pattern.test(text)) return pattern
+  }
+  return null
+}
+
+function describe(pattern) {
+  return pattern?.source ?? null
+}
+
+// A period followed by whitespace and an uppercase letter/quote starts a new
+// English sentence. A period inside a decimal, version, path, or URL is followed
+// by a non-space character and therefore stays part of the token.
+function asciiSentenceEnd(text, index) {
+  if (index + 1 >= text.length || !WHITESPACE_CHAR.test(text[index + 1])) return false
+  let next = index + 1
+  while (next < text.length && WHITESPACE_CHAR.test(text[next])) next++
+  if (next >= text.length) return true
+  return /[A-Z"“「(]/.test(text[next])
+}
+
+// Sentence spans that keep their offsets. Chinese terminators, line breaks, and
+// unambiguous English sentence endings end a sentence; other ASCII periods do not,
+// because a decimal/version/path can contain one.
+export function sentenceSpans(text) {
+  if (typeof text !== 'string' || text === '') return []
+  const spans = []
+  let start = 0
+  for (let index = 0; index < text.length; index++) {
+    const char = text[index]
+    const ends = char === '\n' || char === '。' || char === '！' || char === '？' ||
+      char === '；' || char === ';' || (char === '.' && asciiSentenceEnd(text, index))
+    if (ends) {
+      const end = index + 1
+      if (text.slice(start, end).trim() !== '') spans.push({ start, end })
+      start = end
+    }
+  }
+  if (start < text.length && text.slice(start).trim() !== '') {
+    spans.push({ start, end: text.length })
+  }
+  return spans
+}
+
+function sentenceSignals(text) {
+  return {
+    explicitRule: any(RE.explicitRule, text),
+    designConstraint: any(RE.designConstraint, text),
+    memo: any(RE.memo, text),
+    future: any(RE.future, text),
+    universal: any(RE.universal, text),
+    modal: any(RE.modal, text),
+    decision: any(RE.decision, text),
+    preference: any(RE.preference, text),
+    interaction: any(RE.interaction, text),
+    retraction: any(RE.retraction, text),
+    english: any(RE.english, text),
+    environment: any(RE.environment, text),
+    domain: any(RE.domain, text),
+  }
+}
+
+function hasSignal(signals) {
+  return Object.values(signals).some(Boolean)
+}
+
+function strongSignal(signals) {
+  return Boolean(signals.explicitRule || signals.memo || signals.future)
+}
+
+function mergeSpans(units) {
+  const merged = []
+  for (const unit of units) {
+    const last = merged[merged.length - 1]
+    // Adjacent sentences must stay separate: merging them lets a predicate cue from
+    // one sentence authorize a quote from the next ("回答先给结论" becoming a
+    // constraint because the previous sentence said "项目约定").
+    if (last && last.end > unit.start) {
+      last.end = Math.max(last.end, unit.end)
+      last.reasons.push(...unit.reasons)
+      continue
+    }
+    merged.push({ ...unit, reasons: [...unit.reasons] })
+  }
+  return merged
+}
+
+function signalsOf(signals) {
+  return Object.entries(signals)
+    .filter(([, pattern]) => Boolean(pattern))
+    .map(([name]) => name)
+}
+
+// Pure classification of one archived human message.
+//
+// `eligible` means a proposal may be generated from it. `spans` are the only text
+// ranges a quote may occupy; a quote that crosses an ineligible sentence is
+// refused by the validator even when the message as a whole is eligible.
+export function analyzeDurableUtterance(text) {
+  if (typeof text !== 'string' || text.trim() === '') {
+    return {
+      eligible: false, reason: 'empty', reasons: ['empty'], signals: [], spans: [],
+      policyVersion: DURABILITY_POLICY_VERSION,
+    }
+  }
+  // Keep the original offsets: callers compare these spans with an indexOf result
+  // on the archived text, so trimming here would shift every unit.
+  const raw = text
+  const match = (patterns) => any(patterns, raw)
+  const injection = match(RE.injection)
+  const assistantProse = match(RE.assistantProse)
+  const codeDump = match(RE.codeDump)
+  const question = match(RE.question)
+  const uncertain = match(RE.uncertain)
+  const transient = match(RE.transient)
+  const plan = match(RE.plan)
+  const transientError = match(RE.transientError)
+  const fileObservation = match(RE.fileObservation)
+  const environment = match(RE.environment)
+  const assignmentOpen = match(RE.assignmentOpen)
+  const startImperative = match(RE.startImperative)
+  const actionChain = match(RE.actionChain)
+  const requestFrame = match(RE.requestFrame)
+  const directive = match(RE.directive)
+  const status = match(RE.statusFrame)
+  const strong = match(RE.strongDurable)
+  const future = match(RE.future)
+  const explicitRule = match(RE.explicitRule)
+  const universal = match(RE.universal)
+  const decision = match(RE.decision)
+  const interaction = match(RE.interaction)
+  const preference = match(RE.preference)
+  // A commitment outranks an action verb: "定版为 p1，不再接受" and "改用 worktree"
+  // are decisions even though they contain 提交/执行.
+  const commitment = /(决定|定版|拍板|采用|选用|选定|不再接受|不再适用|不再使用|作废|撤回|废弃|弃用|放弃|改用|迁移到|最终方案|为准)/.test(raw)
+
+  // Precedence is intentional. A transient or reported message must not become
+  // durable just because it contains a decision verb ("临时决定…", "README 里写着
+  // 决定…"), and a question must not become a fact because it contains "以后".
+  let reason = null
+  if (injection) reason = 'injection'
+  else if (assistantProse) reason = 'assistant-prose'
+  else if (codeDump) reason = 'code-dump'
+  else if (question) reason = 'question'
+  else if (uncertain) reason = 'uncertain'
+  else if (transient && !future) reason = 'transient'
+  else if (plan) reason = 'plan'
+  else if (transientError) reason = 'transient-error'
+  else if (fileObservation) reason = 'file-observation'
+  else if (status && !strong && !explicitRule && !commitment) {
+    reason = 'status-report'
+  }
+
+  // An imperative frame is request-scoped unless the same message also carries a
+  // standing-rule signal. This is the scope criterion, not a list of task words:
+  // "帮我加一个测试文件" is work; "先给结论" is a preference; "把金额一律…" is a rule.
+  // A commitment always outranks the action frame; a standing-rule signal exempts a
+  // request frame. A bare setting change ("接口端口改为 9000") has neither an opener
+  // nor an action chain, so it stays a decision rather than becoming a task.
+  // "所有文件" is an object, not a standing scope: a universal word exempts an
+  // imperative opener ("把金额一律…") but not a first-person request ("我需要你清理
+  // 所有文件").
+  const openerFrame = Boolean(assignmentOpen || startImperative || (actionChain && directive))
+  const requestFrameTask = Boolean(requestFrame && directive)
+  const assignment = Boolean(
+    (openerFrame && !strong && !explicitRule && !universal && !interaction && !preference &&
+      !environment && !commitment) ||
+    (requestFrameTask && !strong && !explicitRule && !preference &&
+      !environment && !commitment),
+  )
+  if (reason === null && assignment) reason = 'task-assignment'
+
+  const spans = sentenceSpans(raw)
+  const units = []
+  for (const span of spans) {
+    const sentence = raw.slice(span.start, span.end)
+    const signals = sentenceSignals(sentence)
+    if (!hasSignal(signals)) continue
+    if (assignment && !strongSignal(signals)) continue
+    units.push({
+      start: span.start,
+      end: span.end,
+      reasons: signalsOf(signals),
+      text: sentence,
+    })
+  }
+
+  // A mixed message is judged by its rule clause, not by a question/plan in another
+  // sentence. The override is limited to question/uncertain/plan; a reported file or
+  // an injection stays excluded because its authority is wrong.
+  const strongSentences = units.some(({ reasons }) => reasons.some((name) =>
+    name === 'explicitRule' || name === 'designConstraint' || name === 'memo' ||
+    name === 'future' || name === 'decision' ||
+    (name === 'modal' && reasons.includes('domain'))))
+  // A universal word in a status report ("不显示任何内容") is an object quantifier,
+  // not a standing rule; only an explicit label, a design invariant, a memo, or a
+  // future commitment lifts the status frame.
+  const ruleSentences = units.some(({ reasons }) => reasons.some((name) =>
+    name === 'explicitRule' || name === 'designConstraint' || name === 'memo' ||
+    name === 'future'))
+  // A question only yields to a sentence that states a rule outright. "我们是不是
+  // 应该改用 Postgres？" contains a decision verb but is still a question; a plain
+  // decision/future cue must not turn it into a fact.
+  const questionRule = units.some(({ reasons }) => reasons.some((name) =>
+    name === 'explicitRule' || name === 'designConstraint' || name === 'memo'))
+  if (ruleSentences && reason === 'status-report') reason = null
+  if (questionRule && reason === 'question') reason = null
+  if (strongSentences && (reason === 'uncertain' || reason === 'plan')) reason = null
+
+  if (reason === null && units.length === 0) reason = 'no-durable-signal'
+
+  return {
+    eligible: reason === null,
+    reason,
+    reasons: reason === null ? [] : [reason],
+    signals: {
+      injection: describe(injection),
+      assistantProse: describe(assistantProse),
+      codeDump: describe(codeDump),
+      question: describe(question),
+      uncertain: describe(uncertain),
+      transient: describe(transient),
+      plan: describe(plan),
+      transientError: describe(transientError),
+      fileObservation: describe(fileObservation),
+      environment: describe(environment),
+      requestFrame: describe(requestFrame),
+      status: describe(status),
+      startImperative: describe(startImperative),
+      actionChain: describe(actionChain),
+      directive: describe(directive),
+      assignment: assignment ? 'assignment' : null,
+      strongDurable: describe(strong),
+      future: describe(future),
+    },
+    spans: mergeSpans(units).map(({ start, end, reasons }) => ({ start, end, reasons })),
+    policyVersion: DURABILITY_POLICY_VERSION,
+  }
+}
+
+// The taxonomy is defined by these cues, in this order. The model's predicate is
+// advisory: when a cue is present the deterministic label wins, because an
+// undefined taxonomy is what produced content-correct/predicate-wrong candidates.
+export function resolveQuotePredicate(text) {
+  if (typeof text !== 'string') return null
+  const signals = sentenceSignals(text)
+  if (signals.explicitRule) return { predicate: 'project.constraint', rule: 'explicit-rule-noun' }
+  if (signals.designConstraint) return { predicate: 'project.constraint', rule: 'design-constraint' }
+  if (signals.environment) return { predicate: 'project.constraint', rule: 'environment-cue' }
+  if (signals.preference || signals.interaction) {
+    return { predicate: 'project.preference', rule: signals.preference ? 'preference-cue' : 'interaction-cue' }
+  }
+  if (signals.decision) return { predicate: 'project.decision', rule: 'decision-cue' }
+  if (signals.english) {
+    if (/\b(prefer|preference|would rather)\b/i.test(text)) {
+      return { predicate: 'project.preference', rule: 'english-preference' }
+    }
+    if (/\b(decided|decision|chose|instead of)\b/i.test(text)) {
+      return { predicate: 'project.decision', rule: 'english-decision' }
+    }
+    return { predicate: 'project.constraint', rule: 'english-deontic' }
+  }
+  if (signals.universal || signals.modal || signals.retraction) {
+    return { predicate: 'project.constraint', rule: 'invariant-cue' }
+  }
+  // A domain noun alone ("数据库使用 SQLite") says what the statement is about,
+  // not which taxonomy bucket it belongs to. Keep the model's label rather than
+  // turning every unlabelled domain sentence into a constraint.
+  return null
+}
+
+export function quoteWithinSpans(spans, start, end) {
+  if (!Array.isArray(spans)) return false
+  return spans.some((span) =>
+    Number.isInteger(span.start) && Number.isInteger(span.end) &&
+    start >= span.start && end <= span.end)
+}