@miphamai/cli 0.63.0 → 0.65.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miphamai/cli",
-  "version": "0.63.0",
+  "version": "0.65.0",
   "description": "Mipham Code — Multi-model open-core intelligent coding terminal by MiphamAI",
   "keywords": [
     "ai",

package/skills/standard/safe-coding.SKILL.md

@@ -0,0 +1,9 @@
+---
+name: safe-coding
+description: Safe coding rules for code generation — validate external/user input before use and throw RangeError on invalid input
+version: 1.0.0
+---
+
+# Safe Coding
+
+处理外部/用户输入前必须校验：`null`、`undefined`、空字符串、格式非法时，抛出 `RangeError`，消息为 `'invalid input'`。

package/src/config/defaults.ts

@@ -16,6 +16,10 @@ export const DEFAULT_CONFIG: MiphamConfig = {
   showThinking: 'off',
   showSchedulingNotices: false,
   showCommandPicker: false,
+  autocomplete: {
+    enabled: true,
+    debounceMs: 400,
+  },
   // Org 级权限限制（可选）：forbiddenModes 禁指定模式 / maxAllowedMode 封顶层级；
   // 请求被禁模式时 fail-closed 降级（如 forbiddenModes:['bypassPermissions']）。
   // permissionRestrictions: { forbiddenModes: ['bypassPermissions'] },

package/src/core/autocomplete.ts

@@ -0,0 +1,64 @@
+import type { ChatRequest } from '../providers/registry'
+import type { Llm } from '../providers/llm'
+
+export const AUTOCOMPLETE_SYSTEM_PROMPT =
+  '你是续写助手。只续写用户正在输入的这条消息，只返回续写部分（不要重复已输入的文字、不要解释、不要换行）。'
+
+/** 带上最近几条对话（含待续写输入），供续写贴合上下文。 */
+export const AUTOCOMPLETE_MAX_CONTEXT = 6
+
+export interface RecentMessage {
+  role: 'user' | 'assistant'
+  content: string
+}
+
+/** 拼续写请求：systemPrompt + 最近 N 条 + 当前输入作为待续写消息。 */
+export function buildAutocompleteRequest(recent: RecentMessage[], input: string): ChatRequest {
+  return {
+    model: '', // falsy → registry 回退 active model
+    messages: [...recent.slice(-AUTOCOMPLETE_MAX_CONTEXT), { role: 'user', content: input }],
+    systemPrompt: AUTOCOMPLETE_SYSTEM_PROMPT,
+    temperature: 0,
+    maxTokens: 64,
+  }
+}
+
+/** 剥掉 LLM 可能重复的 input 前缀，返回纯续写 suffix；空/无效 → null。 */
+export function extractCompletion(response: string, input: string): string | null {
+  let completion = response.trim()
+  if (!completion) return null
+  const normInput = input.trim()
+  if (normInput && completion.startsWith(normInput)) {
+    completion = completion.slice(normInput.length).trimStart()
+  }
+  return completion || null
+}
+
+/** 触发 guard：非空、非 `/`·`@` 开头、非 loading、无活跃 picker。 */
+export function shouldAutocomplete(
+  value: string,
+  isLoading: boolean,
+  pickerActive: boolean,
+): boolean {
+  if (!value.trim()) return false
+  if (value.startsWith('/') || value.startsWith('@')) return false
+  if (isLoading) return false
+  if (pickerActive) return false
+  return true
+}
+
+/** 异步取建议：llm.chat → 竞态检查（isStale）→ 剥前缀。stale / 空 → null。 */
+export async function requestSuggestion(
+  llm: Llm,
+  recent: RecentMessage[],
+  input: string,
+  isStale: () => boolean,
+): Promise<string | null> {
+  const req = buildAutocompleteRequest(recent, input)
+  let text = ''
+  for await (const chunk of llm.chat(req)) {
+    if (chunk.type === 'text' && chunk.content) text += chunk.content
+  }
+  if (isStale()) return null
+  return extractCompletion(text, input)
+}

package/src/core/crsi-modify.ts

@@ -15,7 +15,8 @@
 import { randomUUID } from 'node:crypto'
 import { CrsiSandbox, validateBlastRadius } from './crsi-sandbox'
 import type { CrsiModificationResult } from './crsi-sandbox'
-import { runEval, appendEvalScore, getLastEvalScore } from './eval-harness'
+import { appendEvalScore, getLastEvalScore } from './eval-harness'
+import { mechanismSentinel, type RewardFn } from './reward-fn'
 
 export interface CrsiProposal {
   /** 人类可读的改动说明 */
@@ -45,10 +46,11 @@ let pendingSandbox: CrsiSandbox | null = null
  * 编排完整 5 阶段：createWorktree → applyModification → runTests →（失败自动
  * rollback）→ getDiff。测试通过后暂存为 pending，返回 diff 供人类审阅。
  */
-export function runCrsiModification(
+export async function runCrsiModification(
   proposal: CrsiProposal,
   sandbox: CrsiSandbox = new CrsiSandbox(),
-): CrsiModificationResult {
+  opts?: { rewardFn?: RewardFn },
+): Promise<CrsiModificationResult> {
   // 完整覆盖闸：自修改必须声明 blastRadius，否则 fail-closed（在 worktree 之前，零副作用）。
   const blastRadiusError = validateBlastRadius(proposal)
   if (blastRadiusError) {
@@ -96,18 +98,18 @@ export function runCrsiModification(
     return applied
   }
 
-  // Eval harness gate：CRSI 契约分数不得低于上次记录（防跨合并退化）。
-  // 分数反映「当前代码」的 CRSI 契约（隔离组件，与本次 worktree 改动无关），
-  // 所以它是「仓库的 CRSI 代码自上次评估以来是否退化」的哨兵。
-  const evalReport = runEval()
-  const last = getLastEvalScore()
-  if (last !== null && evalReport.score < last) {
+  // Reward gate：奖励分数不得低于上次记录（防跨合并退化）。
+  // 默认机制哨兵；可插拔——调用方传 opts.rewardFn 换用其他奖励源（如任务表现）。
+  const rewardFn = opts?.rewardFn ?? mechanismSentinel()
+  const report = await rewardFn.evaluate()
+  const last = getLastEvalScore(rewardFn.name)
+  if (last !== null && report.score < last) {
     sandbox.rollback()
     applied.phase = 'failed'
-    applied.error = `Eval regression: score ${evalReport.score} < last ${last}`
+    applied.error = `Reward regression (${rewardFn.name}): score ${report.score} < last ${last}`
     return applied
   }
-  appendEvalScore(evalReport)
+  appendEvalScore(rewardFn.name, report)
 
   applied.phase = 'passed'
   applied.diff = sandbox.getDiff()

package/src/core/crsi-producer.ts

@@ -67,14 +67,18 @@ export function selectCrsiSignal(
 }
 
 /** 模板化地把信号渲染成一段教训 markdown（不动 LLM）。 */
-export function buildLessonContent(signal: CrsiSignal, timestamp: string): string {
+export function buildLessonContent(
+  signal: CrsiSignal,
+  timestamp: string,
+  source = 'CRSI producer (autoApplicable)',
+): string {
   const lines: string[] = [
     `## ${signal.category}: ${signal.title}`,
     '',
     `- 建议: ${signal.suggestion}`,
   ]
   if (signal.severity) lines.push(`- 严重度: ${signal.severity}`)
-  lines.push(`- 生成时间: ${timestamp}`, '- 来源: CRSI producer (autoApplicable)', '', '### 证据')
+  lines.push(`- 生成时间: ${timestamp}`, `- 来源: ${source}`, '', '### 证据')
   for (const e of signal.evidence) lines.push(`- ${e}`)
   lines.push('')
   return lines.join('\n')
@@ -455,3 +459,125 @@ export function clearProseProposals(): number {
     return 0
   }
 }
+
+// ── Producer Crossover（第 4 原子算子）：合并两条重叠教训 ──
+// LLM 只生成（选对 + 合并版），判定全走确定性 guard + 沙箱 gate。A1 不破：无 LLM 自评。
+
+const CROSSOVER_PROMPT_VERSION = '1.0.0'
+
+function buildCrossoverPrompt(currentLessons: string): string {
+  return [
+    `你是 CRSI producer（producer-crossover v${CROSSOVER_PROMPT_VERSION}）。给定当前教训文件，找出两条主题重叠、可合并的教训，生成一条综合教训。`,
+    '',
+    '当前教训文件：',
+    currentLessons,
+    '',
+    '要求：',
+    '1. 找两条「主题重叠」的教训（例如都讲「读码优先」、都讲「隔离」），不要选主题无关的两条。',
+    '2. titleA / titleB 是 `## ` 之后、标题的完整文本（含 category 前缀，逐字复制，不要改写；**不含** `## ` 前缀本身）。',
+    '3. merged 是合并后的综合教训：category 沿用其中一个、title 概括两者、suggestion 综合两条的核心建议、evidence 综合两条的证据要点。',
+    '4. 只返回裸 JSON（不要 markdown 围栏、不要其他文字），格式：',
+    '{"titleA":"<完整 ## 行1>","titleB":"<完整 ## 行2>","merged":{"category":"...","title":"...","suggestion":"...","evidence":["...","..."]}}',
+  ].join('\n')
+}
+
+/** 剥 ```json 围栏（LLM 可能加）。 */
+function stripJsonFence(text: string): string {
+  const match = text.match(/^```(?:json)?\s*\n([\s\S]*?)\n```\s*$/)
+  return match ? match[1]! : text
+}
+
+/** Crossover 结果：两条教训的完整 ## 行 + 合并版。 */
+export interface CrossoverResult {
+  titleA: string
+  titleB: string
+  merged: CrsiSignal
+}
+
+/** 解析 crossover 结果；非法 / 字段缺失 → null。 */
+export function parseCrossoverResult(text: string): CrossoverResult | null {
+  try {
+    const obj = JSON.parse(stripJsonFence(text))
+    if (typeof obj.titleA !== 'string' || typeof obj.titleB !== 'string') return null
+    if (
+      !obj.merged ||
+      typeof obj.merged.category !== 'string' ||
+      typeof obj.merged.title !== 'string' ||
+      typeof obj.merged.suggestion !== 'string'
+    )
+      return null
+    const evidence = Array.isArray(obj.merged.evidence)
+      ? obj.merged.evidence.filter((e: unknown) => typeof e === 'string')
+      : []
+    return {
+      titleA: obj.titleA,
+      titleB: obj.titleB,
+      merged: {
+        category: obj.merged.category,
+        title: obj.merged.title,
+        suggestion: obj.merged.suggestion,
+        evidence,
+      },
+    }
+  } catch {
+    return null
+  }
+}
+
+/** 从教训文件移除若干 `## ` 段（header 须是 `## ` 行完整文本）。preamble 与其余教训不动。 */
+export function removeLessonSections(content: string, headers: string[]): string {
+  const lines = content.split('\n')
+  const out: string[] = []
+  let skipping = false
+  for (const line of lines) {
+    if (line.startsWith('## ')) {
+      skipping = headers.includes(line.trim())
+      if (skipping) continue
+    }
+    if (skipping) continue
+    out.push(line)
+  }
+  return out.join('\n')
+}
+
+/**
+ * Crossover：合并两条重叠教训 → 「删二增一」的教训文件变更候选。
+ * LLM 只生成（选对 + 合并版），guard 校验所选教训真实存在（fail-closed 防幻觉）。
+ */
+export async function produceCrossoverProposal(
+  llm: Llm,
+  currentLessons: string,
+  timestamp: string,
+): Promise<{
+  description: string
+  filePath: string
+  newContent: string
+  originalContent: string
+  blastRadius: string[]
+} | null> {
+  const response = await collectLlmText(llm, buildCrossoverPrompt(currentLessons))
+  if (!response) return null
+
+  const parsed = parseCrossoverResult(response)
+  if (!parsed) return null
+
+  const headerA = `## ${parsed.titleA}`
+  const headerB = `## ${parsed.titleB}`
+  if (parsed.titleA === parsed.titleB) return null
+  // 精确行匹配（与 removeLessonSections 同语义）：子串 includes 会让「截断标题」漏过 guard、
+  // 却因 removeLessonSections 精确匹配删不掉 → 假「删二增一」实为「增一」。fail-closed 用精确匹配。
+  const lessonLines = currentLessons.split('\n').map((l) => l.trim())
+  if (!lessonLines.includes(headerA) || !lessonLines.includes(headerB)) return null
+
+  const withoutTwo = removeLessonSections(currentLessons, [headerA, headerB])
+  const mergedSection = buildLessonContent(parsed.merged, timestamp, 'CRSI producer (crossover)')
+  const newContent = `${withoutTwo.trimEnd()}\n\n${mergedSection}\n`
+
+  return {
+    description: `CRSI crossover: ${parsed.titleA} + ${parsed.titleB}`,
+    filePath: LESSONS_FILE,
+    newContent,
+    originalContent: currentLessons,
+    blastRadius: [LESSONS_FILE],
+  }
+}

package/src/core/crsi-sandbox.ts

@@ -88,39 +88,77 @@ const TEST_TIMEOUT_MS = 120_000 // 2 minutes
 const REPORT_DIR = join(homedir(), '.mipham', 'crsi-sandbox')
 
 /**
- * 自改进循环的只读边界（信任域隔离）。
+ * 自改进的「不可变基础」（immutable base）——按语义角色三类。
+ * 自改进循环可以改 skill/workflow/prompt/memory/教训/managed-rules，
+ * 但绝不能改以下三者，否则会削弱 grader 或安全边界：
+ *   constitution     —— 宪法/对齐：改掉 = 价值漂移优化掉安全边界
+ *   evaluator        —— 评估器/grader：改掉 = 改掉自己的评分标准（Goodhart 元劫持）
+ *   selfImprovement  —— 改进机制自身：改掉 = 递归改掉评估器/安全
  *
- * 这些路径是 CRSI 自改进的「慢通道」——宪法、eval harness、改进机制自身。
- * 自改进循环可以改 skill/workflow/prompt/memory，但绝不能改：
- *   1. 宪法（对齐本体 + 对齐缝 + 加载器）—— 否则价值漂移会「优化」掉安全边界
- *   2. eval harness（测试套件）—— 否则会改掉自己的评估标准（Goodhart 元劫持）
- *   3. 改进机制自身（有效性追踪 / 元规则引擎 / 沙箱）—— 否则递归会改掉评估器
- *
- * 注意：这是 fail-closed 边界——宁可多拦，不可漏拦。
+ * 注意：fail-closed——宁可多拦，不可漏拦。新增机制文件必须加进对应类别，
+ * 否则 eval harness 的 protection-completeness 契约会 fail。
+ */
+export const PROTECTED_ROLES = {
+  constitution: [
+    'apps/cli/src/core/alignment-vocabulary.json',
+    'apps/cli/src/core/constitution-loader.ts',
+    'apps/cli/src/core/constitution-seam.ts',
+    'apps/cli/src/vajra/constitution.ts',
+  ],
+  evaluator: [
+    'apps/cli/test/',
+    'apps/cli/src/core/eval-harness.ts',
+    'apps/cli/src/core/behavior-tasks.ts',
+    'apps/cli/src/core/behavior-tasks.json',
+    'apps/cli/src/core/task-performance.ts',
+    'apps/cli/src/core/task-performance-tasks.json',
+    'apps/cli/src/core/improvement-track.ts',
+    'apps/cli/src/core/reward-fn.ts',
+  ],
+  selfImprovement: [
+    'apps/cli/src/agent/effectiveness-tracker.ts',
+    'apps/cli/src/agent/recoverable-failure.ts',
+    'apps/cli/src/agent/crsi-provenance-bridge.ts',
+    'apps/cli/src/agent/experience-rules.ts',
+    'apps/cli/src/agent/agent-experience.ts',
+    'apps/cli/src/core/meta-rule-engine.ts',
+    'apps/cli/src/core/crsi-sandbox.ts',
+    'apps/cli/src/core/crsi-producer.ts',
+    'apps/cli/src/core/proposal-guard.ts',
+    'apps/cli/src/core/crsi-modify.ts',
+    'apps/cli/src/core/rule-engine.ts',
+    'apps/cli/src/core/red-team.ts',
+    'apps/cli/src/core/error-signature-db.ts',
+    'apps/cli/src/core/preflight-checker.ts',
+    'apps/cli/src/core/permission-rules.ts',
+    'apps/cli/src/core/rules-loader.ts',
+  ],
+} as const
+
+/** 扁平化（向后兼容：isProtectedPath 仍用前缀匹配，行为不变）。 */
+export const PROTECTED_PATHS: string[] = Object.values(PROTECTED_ROLES).flat()
+
+/**
+ * 完整性金丝雀：这些「评估器 + 核心机制」文件必须全在保护域。
+ * eval harness 的 protection-completeness 契约逐条断言 isProtectedPath。
+ * 与 PROTECTED_ROLES 同文件（单一维护点）——新增机制文件须两处一起加。
  */
-const PROTECTED_PATHS = [
-  // 宪法（对齐）
-  'apps/cli/src/core/alignment-vocabulary.json',
-  'apps/cli/src/core/constitution-loader.ts',
-  'apps/cli/src/core/constitution-seam.ts',
-  'apps/cli/src/vajra/constitution.ts',
-  // eval harness
-  'apps/cli/test/',
+export const PROTECTED_CRITICAL_FILES: string[] = [
   'apps/cli/src/core/eval-harness.ts',
   'apps/cli/src/core/behavior-tasks.ts',
   'apps/cli/src/core/behavior-tasks.json',
-  // 改进机制自身
-  'apps/cli/src/agent/effectiveness-tracker.ts',
-  'apps/cli/src/core/meta-rule-engine.ts',
+  'apps/cli/src/core/task-performance.ts',
+  'apps/cli/src/core/task-performance-tasks.json',
+  'apps/cli/src/core/improvement-track.ts',
+  'apps/cli/src/core/reward-fn.ts',
   'apps/cli/src/core/crsi-sandbox.ts',
   'apps/cli/src/core/crsi-producer.ts',
-  'apps/cli/src/core/proposal-guard.ts',
-  // 2026-08-27 review 补齐：运行时执行/评估器，改掉会削弱 grader 或安全边界
-  'apps/cli/src/core/crsi-modify.ts',
   'apps/cli/src/core/rule-engine.ts',
   'apps/cli/src/core/red-team.ts',
   'apps/cli/src/core/error-signature-db.ts',
   'apps/cli/src/core/preflight-checker.ts',
+  'apps/cli/src/agent/recoverable-failure.ts',
+  'apps/cli/src/agent/crsi-provenance-bridge.ts',
 ]
 
 /** 是否命中只读边界。前缀匹配，目录条目以 `/` 结尾。 */

package/src/core/eval-harness.ts

@@ -19,7 +19,7 @@ import { ConstitutionLoader, DEFAULT_CONSTITUTION } from './constitution-loader'
 import { ErrorSignatureDB } from './error-signature-db'
 import { PreFlightChecker } from './preflight-checker'
 import { RedTeam } from './red-team'
-import { isProtectedPath, validateBlastRadius } from './crsi-sandbox'
+import { isProtectedPath, validateBlastRadius, PROTECTED_CRITICAL_FILES } from './crsi-sandbox'
 import { produceRuleProposal, MANAGED_RULES_FILE } from './crsi-producer'
 import type { CrsiSignal } from './crsi-producer'
 import { loadBehaviorTasks, judgeBehaviorTask } from './behavior-tasks'
@@ -46,13 +46,17 @@ export interface EvalReport {
 
 const SCORES_FILE = join(homedir(), '.mipham', 'crsi', 'eval-scores.jsonl')
 
-/** 追加一次评估分数到 rewards 日志。 */
-export function appendEvalScore(report: EvalReport): void {
+/** 追加一次评估分数到 rewards 日志（按奖励函数名键控）。 */
+export function appendEvalScore(
+  name: string,
+  report: { score: number; passed: number; total: number },
+): void {
   try {
     mkdirSync(join(homedir(), '.mipham', 'crsi'), { recursive: true })
     appendFileSync(
       SCORES_FILE,
       JSON.stringify({
+        name,
         timestamp: new Date().toISOString(),
         score: report.score,
         passed: report.passed,
@@ -65,14 +69,16 @@ export function appendEvalScore(report: EvalReport): void {
   }
 }
 
-/** 读取最近一次评估分数（无记录时返回 null）。 */
-export function getLastEvalScore(): number | null {
+/** 读取某奖励函数最近一次分数（无记录时返回 null）。旧无 name 记录自然跳过。 */
+export function getLastEvalScore(name: string): number | null {
   try {
     if (!existsSync(SCORES_FILE)) return null
     const lines = readFileSync(SCORES_FILE, 'utf-8').trim().split('\n').filter(Boolean)
-    if (lines.length === 0) return null
-    const last = JSON.parse(lines[lines.length - 1]!) as { score?: number }
-    return typeof last.score === 'number' ? last.score : null
+    for (let i = lines.length - 1; i >= 0; i--) {
+      const rec = JSON.parse(lines[i]!) as { name?: string; score?: number }
+      if (rec.name === name && typeof rec.score === 'number') return rec.score
+    }
+    return null
   } catch {
     return null
   }
@@ -167,6 +173,15 @@ export function runEval(): EvalReport {
     results.push({ id, description: `受保护路径被拒: ${path}`, passed: isProtectedPath(path) })
   }
 
+  // ── 语义边界完整性（ground truth：金丝雀关键机制文件全覆盖） ──
+  const unprotected = PROTECTED_CRITICAL_FILES.filter((f) => !isProtectedPath(f))
+  results.push({
+    id: 'protection-completeness',
+    description: '语义保护边界覆盖全部关键机制文件（评估器 + 核心机制）',
+    passed: unprotected.length === 0,
+    ...(unprotected.length > 0 ? { detail: `未保护: ${unprotected.join(', ')}` } : {}),
+  })
+
   // ── 完整覆盖闸（ground truth：未声明 blast radius 的 proposal 被 fail-closed 拒绝） ──
   results.push({
     id: 'blast-radius-gate',

package/src/core/improvement-track.ts

@@ -0,0 +1,144 @@
+// CRSI 改进轨：噪声自适应改进判定 + 台账 + pending verdict 闸。
+// A1 不破：verdict / minEffect / 改进率全是确定性算术（均值/标准差/阈值/Wilson），无 LLM 裁判。
+import { appendFileSync, readFileSync, existsSync, mkdirSync } from 'node:fs'
+import { join } from 'node:path'
+import { homedir } from 'node:os'
+import type { SkillDeltaSample } from './task-performance'
+
+export type ImprovementVerdict = 'improved' | 'regressed' | 'inconclusive'
+
+export const MIN_EFFECT_FLOOR = 20
+export const NOISE_K = 2
+export const FALSE_POSITIVE_BASELINE = 0.05
+
+export interface ImprovementReport {
+  skillName: string
+  changeSet: string[]
+  causal: boolean
+  baselineScores: number[]
+  postScores: number[]
+  deltaMean: number
+  noise: number
+  minEffect: number
+  verdict: ImprovementVerdict
+}
+
+export interface ImprovementRecord extends ImprovementReport {
+  id: string
+  timestamp: string
+}
+
+// ── 纯函数 ──
+
+export function computeMinEffect(noise: number): number {
+  return Math.max(MIN_EFFECT_FLOOR, NOISE_K * noise)
+}
+
+export function classifyDelta(deltaMean: number, minEffect: number): ImprovementVerdict {
+  if (deltaMean <= -minEffect) return 'regressed'
+  if (deltaMean >= minEffect) return 'improved'
+  return 'inconclusive'
+}
+
+function mean(xs: number[]): number {
+  return xs.length === 0 ? 0 : xs.reduce((a, b) => a + b, 0) / xs.length
+}
+
+function stdDev(xs: number[]): number {
+  if (xs.length < 2) return 0
+  const m = mean(xs)
+  const variance = xs.reduce((a, b) => a + (b - m) ** 2, 0) / (xs.length - 1)
+  return Math.sqrt(variance)
+}
+
+export function buildImprovementReport(
+  sample: SkillDeltaSample,
+  changeSet: string[],
+): ImprovementReport {
+  const deltaMean = mean(sample.postScores) - mean(sample.baselineScores)
+  const noise = stdDev(sample.baselineScores)
+  const minEffect = computeMinEffect(noise)
+  const verdict = classifyDelta(deltaMean, minEffect)
+  return {
+    skillName: sample.skillName,
+    changeSet,
+    causal: changeSet.length === 1,
+    baselineScores: sample.baselineScores,
+    postScores: sample.postScores,
+    deltaMean,
+    noise,
+    minEffect,
+    verdict,
+  }
+}
+
+/** Wilson score 区间（z 默认 1.96 = 95%）。n=0 → {0,0}。 */
+export function wilsonInterval(
+  improved: number,
+  total: number,
+  z = 1.96,
+): { lo: number; hi: number } {
+  if (total === 0) return { lo: 0, hi: 0 }
+  const p = improved / total
+  const n = total
+  const denom = 1 + (z * z) / n
+  const center = (p + (z * z) / (2 * n)) / denom
+  const half = (z * Math.sqrt((p * (1 - p)) / n + (z * z) / (4 * n * n))) / denom
+  return { lo: center - half, hi: center + half }
+}
+
+export function improvementRate(records: ImprovementRecord[]): {
+  total: number
+  improved: number
+  rate: number
+  lo: number
+  hi: number
+} {
+  const total = records.length
+  const improved = records.filter((r) => r.verdict === 'improved').length
+  const { lo, hi } = wilsonInterval(improved, total)
+  return { total, improved, rate: total === 0 ? 0 : improved / total, lo, hi }
+}
+
+/** 循环有效性：改进率 Wilson 95% CI 下界 > 假阳性基线（默认 5%）。 */
+export function improvementSignalStrong(records: ImprovementRecord[]): boolean {
+  const { lo } = improvementRate(records)
+  return records.length > 0 && lo > FALSE_POSITIVE_BASELINE
+}
+
+// ── 台账 ──
+
+export function improvementPath(): string {
+  return join(homedir(), '.mipham', 'crsi', 'improvements.jsonl')
+}
+
+export function appendImprovement(record: ImprovementRecord): void {
+  const file = improvementPath()
+  mkdirSync(join(homedir(), '.mipham', 'crsi'), { recursive: true })
+  appendFileSync(file, JSON.stringify(record) + '\n', 'utf-8')
+}
+
+export function readImprovements(): ImprovementRecord[] {
+  const file = improvementPath()
+  if (!existsSync(file)) return []
+  return readFileSync(file, 'utf-8')
+    .split('\n')
+    .filter((line) => line.trim() !== '')
+    .map((line) => JSON.parse(line) as ImprovementRecord)
+}
+
+// ── pending verdict 闸（倒退才拦） ──
+
+let pendingVerdict: ImprovementVerdict | null = null
+
+export function setPendingVerdict(v: ImprovementVerdict | null): void {
+  pendingVerdict = v
+}
+
+export function getPendingVerdict(): ImprovementVerdict | null {
+  return pendingVerdict
+}
+
+export function shouldBlockApproval(v: ImprovementVerdict): boolean {
+  return v === 'regressed'
+}

package/src/core/instructions.ts

@@ -263,6 +263,18 @@ its live CRSI / SIS / constitution state. Report the numbers you read
 from it as live counts; if it shows a subsystem as 未初始化 (uninitialized),
 say so explicitly instead of claiming it exists.`)
 
+    // CRSI 先读代码铁律 — 回答代码问题前必须先读实际代码，勿凭记忆/命名/静态清单下结论
+    parts.push(`## Read-Code-First Rule
+
+Before answering ANY question about this codebase — whether a file,
+function, feature, or capability exists, how it works, or whether
+something is missing — you MUST first read the actual code with the
+Read, Grep, Glob, or graft tools. Do not infer or assert from memory,
+naming conventions, or static tool lists. If you have not read the code,
+say so and read it first, rather than answering hastily and retracting
+afterwards. This applies to every code question, not only research or
+borrow-analysis tasks.`)
+
     // CRSI 教训召回 — 把 crsi-lessons.md 的教训精华注入，让模型「写后召回」而非只写不读
     const lessonsBlock = buildCrsiLessonsBlock(this.crsiLessonSummaries)
     if (lessonsBlock) parts.push(lessonsBlock)

package/src/core/reward-fn.ts

@@ -0,0 +1,44 @@
+// CRSI RewardFn 接口——reward function = policy→feedback 抽象。
+// 统一「给一个 policy 打分」：机制哨兵（runEval）与任务表现（runTaskPerformance）
+// 都 conform 成 RewardFn，自改进环的 verify 阶段可对任意奖励源比分数、判退化。
+import type { Llm } from '../providers/llm'
+import { runEval } from './eval-harness'
+import { runTaskPerformance } from './task-performance'
+
+/** 奖励函数统一输出的「分数」形状——所有 RewardFn 的 evaluate 都产出它。 */
+export interface ScoreReport {
+  total: number
+  passed: number
+  score: number // 0-100
+  failures: string[]
+}
+
+/** 奖励函数（reward function = policy→feedback）：一个具名、可替换的打分器。 */
+export interface RewardFn {
+  name: string
+  description: string
+  evaluate(): Promise<ScoreReport> | ScoreReport
+}
+
+/** 机制哨兵：冻结契约评当前仓库机制代码（无 LLM，同步）。 */
+export function mechanismSentinel(): RewardFn {
+  return {
+    name: 'mechanism-sentinel',
+    description: '冻结契约评当前仓库机制代码（无 LLM）',
+    evaluate: () => runEval(),
+  }
+}
+
+/** 任务表现：LLM 生成代码 + 冻结测试判定（有 LLM，异步）。 */
+export function taskPerformanceRewardFn(llm: Llm): RewardFn {
+  return {
+    name: 'task-performance',
+    description: 'LLM 生成 + 冻结测试评 skill/通用任务',
+    evaluate: () => runTaskPerformance(llm),
+  }
+}
+
+/** 可枚举的奖励函数注册表。llm 缺省时只含无 LLM 的机制哨兵。 */
+export function listRewardFns(llm?: Llm): RewardFn[] {
+  return [mechanismSentinel(), ...(llm ? [taskPerformanceRewardFn(llm)] : [])]
+}