@miphamai/cli 0.49.0 → 0.51.0

package/src/core/crsi-producer.ts

@@ -12,6 +12,10 @@
 
 import type { CrsiInsight } from './auto-memory'
 import type { MetaRule } from './meta-rule-engine'
+import type { Llm } from '../providers/llm'
+import { readdirSync, appendFileSync, readFileSync, existsSync, mkdirSync, rmSync } from 'node:fs'
+import { join } from 'node:path'
+import { homedir } from 'node:os'
 
 /** 教训文件（相对仓库根）。预建，沙箱只能改已存在文件。 */
 export const LESSONS_FILE = 'apps/cli/crsi-lessons.md'
@@ -86,6 +90,9 @@ export function produceCrsiProposal(
   const signal = selectCrsiSignal(insights, metaRules)
   if (!signal) return null
 
+  // 幂等：同一信号的教训标题已在文件中，不再重复产出。
+  if (currentLessons.includes(`## ${signal.category}: ${signal.title}`)) return null
+
   const lesson = buildLessonContent(signal, timestamp)
   const newContent = currentLessons ? `${currentLessons.trimEnd()}\n\n${lesson}\n` : `${lesson}\n`
 
@@ -186,3 +193,210 @@ export function produceRuleProposal(
     originalContent: currentManagedRules,
   }
 }
+
+// ── Producer 散文提议（块 1）：从失败信号生成「改 skill 散文」提议 ──
+// A1 边界首次实演：LLM 只作「生成」（候选），判定仍走确定性（guard 预筛 / 行为效果 / 人审）。
+
+const PROSE_SELECT_PROMPT_VERSION = '1.0.0'
+
+function buildSelectSkillPrompt(signal: CrsiSignal, skillFiles: string[]): string {
+  return [
+    `你是 CRSI producer（producer-prose-select v${PROSE_SELECT_PROMPT_VERSION}）。给定失败信号，从候选 skill 文件列表中选出最相关的一个，返回其文件路径（只返回路径，一行，不要其他文字）。`,
+    '',
+    '失败信号：',
+    `- category: ${signal.category}`,
+    `- title: ${signal.title}`,
+    signal.severity ? `- severity: ${signal.severity}` : '',
+    `- suggestion: ${signal.suggestion}`,
+    `- evidence: ${signal.evidence.join(' | ')}`,
+    '',
+    '候选 skill 文件：',
+    ...skillFiles.map((f) => `- ${f}`),
+  ]
+    .filter(Boolean)
+    .join('\n')
+}
+
+async function collectLlmText(llm: Llm, prompt: string): Promise<string> {
+  let text = ''
+  const req = {
+    model: 'prose',
+    messages: [{ role: 'user' as const, content: prompt }],
+    systemPrompt: '',
+  }
+  for await (const chunk of llm.chat(req)) {
+    if (chunk.type === 'text' && chunk.content) text += chunk.content
+  }
+  return text.trim()
+}
+
+function extractFilePath(response: string, skillFiles: string[]): string | null {
+  for (const f of skillFiles) {
+    if (response.includes(f)) return f
+  }
+  return null
+}
+
+export async function selectTargetSkill(
+  signal: CrsiSignal,
+  llm: Llm,
+  skillFiles: string[],
+): Promise<string | null> {
+  if (skillFiles.length === 0) return null
+  const prompt = buildSelectSkillPrompt(signal, skillFiles)
+  const response = await collectLlmText(llm, prompt)
+  if (!response) return null
+  return extractFilePath(response, skillFiles)
+}
+
+const PROSE_GENERATE_PROMPT_VERSION = '1.0.0'
+
+function buildGenerateProsePrompt(
+  signal: CrsiSignal,
+  filePath: string,
+  originalContent: string,
+): string {
+  return [
+    `你是 CRSI producer（producer-prose-generate v${PROSE_GENERATE_PROMPT_VERSION}）。基于失败信号，改进目标 skill 的内容。`,
+    '',
+    '失败信号：',
+    `- category: ${signal.category}`,
+    `- title: ${signal.title}`,
+    `- suggestion: ${signal.suggestion}`,
+    `- evidence: ${signal.evidence.join(' | ')}`,
+    '',
+    `目标文件：${filePath}`,
+    '',
+    '当前内容：',
+    originalContent,
+    '',
+    '请返回改进后的完整 markdown（保持 YAML frontmatter 的 name/description 字段，正文针对失败信号做针对性改进）。只返回 markdown，不要额外说明。',
+  ].join('\n')
+}
+
+function stripMarkdownFence(text: string): string {
+  const match = text.match(/^```(?:markdown|md)?\s*\n([\s\S]*?)\n```\s*$/)
+  return match ? match[1]! : text
+}
+
+export async function generateProseContent(
+  signal: CrsiSignal,
+  llm: Llm,
+  filePath: string,
+  originalContent: string,
+): Promise<string | null> {
+  const prompt = buildGenerateProsePrompt(signal, filePath, originalContent)
+  const response = await collectLlmText(llm, prompt)
+  if (!response) return null
+  return stripMarkdownFence(response)
+}
+
+export interface ProseProposalResult {
+  filePath: string
+  newContent: string
+  originalContent: string
+  description: string
+}
+
+export async function produceProseProposal(
+  signal: CrsiSignal,
+  llm: Llm,
+  skillFiles: string[],
+  readSkill: (filePath: string) => string,
+): Promise<ProseProposalResult | null> {
+  const filePath = await selectTargetSkill(signal, llm, skillFiles)
+  if (!filePath) return null
+
+  let originalContent: string
+  try {
+    originalContent = readSkill(filePath)
+  } catch {
+    return null
+  }
+
+  const newContent = await generateProseContent(signal, llm, filePath, originalContent)
+  if (!newContent) return null
+
+  return { filePath, newContent, originalContent, description: signal.title }
+}
+
+const SKILL_DIRS: Array<[string, string]> = [
+  ['standard', '.SKILL.md'],
+  ['mipham', '.mipham-skill.md'],
+]
+
+/** 收集仓库内所有 skill 文件（相对仓库根的路径），供 produceProseProposal 选目标。 */
+export function collectSkillFiles(root: string): string[] {
+  const files: string[] = []
+  for (const [dir, ext] of SKILL_DIRS) {
+    let entries: string[] = []
+    try {
+      entries = readdirSync(join(root, 'apps', 'cli', 'skills', dir))
+    } catch {
+      continue
+    }
+    for (const entry of entries) {
+      if (entry.endsWith(ext)) files.push(`apps/cli/skills/${dir}/${entry}`)
+    }
+  }
+  return files
+}
+
+// ── 幂等去重（prose ledger） ──
+// 散文提议（块 1）的幂等：同一失败信号只生成一次提议。与 --rule 路径「目标文件内 id marker」去重不同，
+// 散文改的是 skill 内容（非追加 marker），故用 ~/.mipham 下的 append-only ledger 记录「已提议的信号」。
+
+/** 散文提议的稳定 id（同 category + 同 title → 同 id，同 managedRuleId 的 hash 语义）。 */
+export function proseProposalId(signal: CrsiSignal): string {
+  return `prose-${signal.category}-${stableHash(signal.title)}`
+}
+
+/** ledger 里的一条散文提议记录。 */
+export interface ProseProposalRecord {
+  id: string
+  filePath: string
+  timestamp: string
+}
+
+function proseLedgerFile(): string {
+  return join(homedir(), '.mipham', 'crsi', 'prose-proposals.jsonl')
+}
+
+/** 该信号是否已生成过散文提议。 */
+export function hasProposedProse(id: string): boolean {
+  try {
+    if (!existsSync(proseLedgerFile())) return false
+    const lines = readFileSync(proseLedgerFile(), 'utf-8').trim().split('\n').filter(Boolean)
+    return lines.some((line) => {
+      try {
+        return (JSON.parse(line) as { id?: string }).id === id
+      } catch {
+        return false
+      }
+    })
+  } catch {
+    return false
+  }
+}
+
+/** 追加一条散文提议记录（append-only，非关键——失败不影响提议本身）。 */
+export function appendProseProposal(record: ProseProposalRecord): void {
+  try {
+    mkdirSync(join(homedir(), '.mipham', 'crsi'), { recursive: true })
+    appendFileSync(proseLedgerFile(), JSON.stringify(record) + '\n', 'utf-8')
+  } catch {
+    // ledger 非关键，失败不影响提议本身
+  }
+}
+
+/** 清空散文提议 ledger，返回清除的记录数（无文件时返回 0）。 */
+export function clearProseProposals(): number {
+  try {
+    if (!existsSync(proseLedgerFile())) return 0
+    const lines = readFileSync(proseLedgerFile(), 'utf-8').trim().split('\n').filter(Boolean)
+    rmSync(proseLedgerFile(), { force: true })
+    return lines.length
+  } catch {
+    return 0
+  }
+}

package/src/core/crsi-sandbox.ts

@@ -84,7 +84,6 @@ export interface CrsiSessionReport {
 // ── Constants ──
 
 const WORKTREE_PREFIX = 'crsi-sandbox-'
-const MAX_WORKTREE_AGE_MS = 30 * 60 * 1000 // 30 minutes
 const TEST_TIMEOUT_MS = 120_000 // 2 minutes
 const REPORT_DIR = join(homedir(), '.mipham', 'crsi-sandbox')
 
@@ -107,10 +106,15 @@ const PROTECTED_PATHS = [
   'apps/cli/src/vajra/constitution.ts',
   // eval harness
   'apps/cli/test/',
+  '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/crsi-sandbox.ts',
+  'apps/cli/src/core/crsi-producer.ts',
+  'apps/cli/src/core/proposal-guard.ts',
 ]
 
 /** 是否命中只读边界。前缀匹配，目录条目以 `/` 结尾。 */

package/src/core/dream-engine.ts

@@ -469,7 +469,6 @@ export class DreamEngine {
       const lines = match[1]!.split('\n')
       const fm: Record<string, unknown> = {}
       let currentNested: Record<string, unknown> | null = null
-      let nestedKey = ''
       for (const line of lines) {
         // Handle nested YAML (e.g. "metadata:\n  type: feedback")
         const nested = line.match(/^  (\w[\w\s]*?):\s*(.+)$/)
@@ -485,7 +484,6 @@ export class DreamEngine {
           if (val === '') {
             // Empty value → start nested block
             currentNested = {}
-            nestedKey = key
             fm[key] = currentNested
           } else {
             currentNested = null

package/src/core/eval-harness.ts

@@ -22,6 +22,7 @@ import { RedTeam } from './red-team'
 import { isProtectedPath } from './crsi-sandbox'
 import { produceRuleProposal, MANAGED_RULES_FILE } from './crsi-producer'
 import type { CrsiSignal } from './crsi-producer'
+import { loadBehaviorTasks, judgeBehaviorTask } from './behavior-tasks'
 
 // ── Types ──
 
@@ -223,6 +224,12 @@ export function runEval(): EvalReport {
     })
   }
 
+  // ── 行为任务集（ground truth：约束行为效果，确定性无 LLM） ──
+  const behaviorTasks = loadBehaviorTasks()
+  for (const task of behaviorTasks) {
+    results.push(judgeBehaviorTask(task, ruleEngine))
+  }
+
   const passed = results.filter((r) => r.passed).length
   return {
     total: results.length,

package/src/core/instructions.ts

@@ -2,6 +2,7 @@ import { readFileSync, existsSync } from 'node:fs'
 import { join } from 'node:path'
 import { parse as parseYaml } from 'yaml'
 import type { InstructionFile } from '../shared/index.ts'
+import { COAUTHOR_TRAILER } from '../shared/index.ts'
 
 interface FrontmatterResult {
   data: Record<string, unknown>
@@ -163,6 +164,16 @@ 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.`)
 
+    // AI 署名披露：提交时附带 Co-Authored-By 署名（与 Undercover 式隐瞒相反）
+    parts.push(`## Commit Attribution
+
+When you create a git commit, always append this trailer on its own line
+at the end of the commit message, disclosing AI involvement:
+
+${COAUTHOR_TRAILER}
+
+Never omit it or present the work as purely human-authored.`)
+
     return parts.join('\n\n---\n\n')
   }
 

package/src/core/memory/memory-loader.ts

@@ -1,7 +1,9 @@
+import { join } from 'node:path'
+import { homedir } from 'node:os'
 import { MemoryManager } from './memory-manager'
 import type { MemoryManager as MemoryManagerType } from './memory-manager'
 
-const MEMORY_DIR = `${process.env.HOME || '~'}/.mipham/memory`
+const MEMORY_DIR = join(homedir(), '.mipham', 'memory')
 
 let instance: MemoryManagerType | null = null
 

package/src/core/permission-rules.ts

@@ -1,14 +1,16 @@
 import type { PermissionRuleEntry } from '../shared/index.ts'
+import { matchPath } from './credential-masker/matcher'
 
-/**
- * Match a Bash(command_pattern) rule against an actual tool call.
- *
- * Pattern formats:
- *   "Bash"              → matches any Bash call
- *   "Bash(git:*)"       → matches "git status", "git diff --cached", etc.
- *   "Bash(npm test:*)"  → matches "npm test -- --coverage"
- *   "Write(/etc/*)"     → matches Write to /etc/passwd, /etc/hosts, etc.
- */
+// Match a tool(parameter) rule against an actual tool call.
+//
+// Pattern formats:
+//   "Bash"              → matches any Bash call
+//   "Bash(git:*)"       → matches "git status", "git diff --cached", etc.
+//   "Bash(npm test:*)"  → matches "npm test -- --coverage"
+//   "Write(/etc/*)"     → matches Write to /etc/passwd, /etc/hosts, etc.
+//   "Read(**/.ssh/*)"   → matches Read of any path under .ssh
+//   "Grep(**/vendor)"   → matches Grep rooted under a vendor directory
+//   "Glob(**/.ssh)"     → matches Glob rooted under a .ssh directory
 export function matchBashRule(
   pattern: string,
   toolName: string,
@@ -31,10 +33,19 @@ export function matchBashRule(
     return wildcardMatch(subPattern!, cmd.trim())
   }
 
-  // For Write/Edit: match against the file_path
-  if (baseTool === 'Write' || baseTool === 'Edit') {
+  // For Write/Edit/Read: match against the file_path with path-glob semantics.
+  // Use matchPath (NOT wildcardMatch): wildcardMatch is tuned for Bash commands
+  // (`:` → colon-or-whitespace, `*` → `.*`), which is wrong for filesystem
+  // paths — `*` would cross `/` and Windows drive letters like `C:\` get mangled.
+  if (baseTool === 'Write' || baseTool === 'Edit' || baseTool === 'Read') {
     const path = String(toolInput.file_path || '')
-    return wildcardMatch(subPattern!, path)
+    return matchPath(path, subPattern!)
+  }
+
+  // For Grep/Glob: match against the base search path (a directory)
+  if (baseTool === 'Grep' || baseTool === 'Glob') {
+    const path = String(toolInput.path || '')
+    return matchPath(path, subPattern!)
   }
 
   return false

package/src/core/proposal-guard.ts

@@ -0,0 +1,100 @@
+// apps/cli/src/core/proposal-guard.ts
+// CRSI 提案预筛器（块 2 最小版）：结构确定性预筛，零 LLM。
+//
+// 这是 CRSI 自改进「producer 改散文」闭环的第一道闸（三层验证的第①层）：
+//   ① 结构确定性（本文件）：受保护路径 + 自引用封闭 + 目标范围 + 结构不变量
+//   ② 行为效果（M3/A：LLM 生成行为 → 测试判定）—— 留作 seam，未实现
+//   ③ 人类审批 —— CrsiSandbox 之后
+//
+// 诚实边界：本预筛只能证「结构合法」，不能证「散文更好」——
+// 后者是第②层（行为效果）的职责，需要 LLM 生成行为，属另一 A1 边界决策。
+
+import { parse as parseYaml } from 'yaml'
+import { isProtectedPath } from './crsi-sandbox'
+import { MANAGED_RULES_FILE, MANAGED_RULE_MARKER } from './crsi-producer'
+
+/** producer 产出的一条「改散文」提议（最小字段集，guard 阶段只消费这些）。 */
+export interface ProducerProposal {
+  id: string
+  /** 仓库根相对路径 */
+  filePath: string
+  /** 最小版只放行两类目标，收窄散文风险面 */
+  kind: 'skill' | 'managed-rule'
+  /** 改后的完整文件内容 */
+  newContent: string
+}
+
+export interface PrefilterVerdict {
+  pass: boolean
+  /** 拒绝理由；pass=true 时为空 */
+  reasons: string[]
+}
+
+const SKILL_DIR = 'apps/cli/skills/'
+
+function isSkillPath(filePath: string): boolean {
+  return (
+    filePath.startsWith(SKILL_DIR) &&
+    (filePath.endsWith('.SKILL.md') || filePath.endsWith('.mipham-skill.md'))
+  )
+}
+
+/**
+ * 更严、独立于 skills loader 的最小 frontmatter 解析。
+ *
+ * 与 loader 的宽松回退不同：frontmatter 缺失 / YAML 非法 / name 或 description
+ * 为空，都判失败（返回 null）——proposal 必须显式携带合法头部，不得依赖
+ * loader 的「文件名兜底 name」「空 description」等宽进逻辑。
+ */
+function parseStrictSkillFrontmatter(raw: string): { name: string; description: string } | null {
+  const match = raw.match(/^---\r?\n([\s\S]*?)\r?\n---\r?\n/)
+  if (!match) return null
+
+  let data: unknown
+  try {
+    data = parseYaml(match[1] || '')
+  } catch {
+    return null
+  }
+
+  if (typeof data !== 'object' || data === null || Array.isArray(data)) return null
+  const obj = data as Record<string, unknown>
+
+  const name = typeof obj.name === 'string' ? obj.name.trim() : ''
+  const description = typeof obj.description === 'string' ? obj.description.trim() : ''
+  if (!name || !description) return null
+  return { name, description }
+}
+
+export function prefilterProposal(p: ProducerProposal): PrefilterVerdict {
+  const reasons: string[] = []
+
+  // ① 受保护路径 + 自引用封闭（isProtectedPath 已含评估机制自身的 5 个补洞）
+  if (isProtectedPath(p.filePath)) {
+    reasons.push(`protected path: ${p.filePath} is read-only to the self-improvement loop`)
+  }
+
+  // ② 目标范围白名单
+  const inScope =
+    (p.kind === 'skill' && isSkillPath(p.filePath)) ||
+    (p.kind === 'managed-rule' && p.filePath === MANAGED_RULES_FILE)
+  if (!inScope) {
+    reasons.push(`out of scope: ${p.filePath} is not an allowed ${p.kind} target`)
+  }
+
+  // ③ 结构不变量
+  if (p.kind === 'skill') {
+    if (!parseStrictSkillFrontmatter(p.newContent)) {
+      reasons.push(
+        'skill frontmatter invalid: missing/illegal frontmatter or empty name/description',
+      )
+    }
+  } else if (!p.newContent.includes(MANAGED_RULE_MARKER)) {
+    reasons.push('managed rule missing the producer append marker')
+  }
+
+  // ②（行为效果 seam）：此处接入 M3/A 任务表现度量——用 LLM 生成行为 + 确定性
+  // 测试判定，过滤「行为效果退化」的提议。未实现，属 producer LLM 生成能力（块 1）。
+
+  return { pass: reasons.length === 0, reasons }
+}

package/src/core/rules-loader.ts

@@ -19,6 +19,7 @@
 
 import { readdirSync, readFileSync, existsSync } from 'node:fs'
 import { join } from 'node:path'
+import { globToRegexSource } from './credential-masker/matcher'
 
 interface RuleFile {
   name: string
@@ -117,18 +118,15 @@ export class RulesLoader {
   }
 
   /**
-   * Simple glob matching. Supports **, *, and exact file matching.
-   * Returns true if file matches pattern.
+   * Glob matching for path-scoped rules. Reuses the shared path-glob core
+   * (globToRegexSource) so `*`/`**`/`?` semantics match credential-file
+   * matching. Anchoring differs: rules match by *suffix* (a rule `*.ts`
+   * applies to any file ending in `.ts`), not full path.
    */
   private matchPath(file: string, pattern: string): boolean {
-    // Convert glob pattern to regex
-    let regexStr = pattern
-      .replace(/\./g, '\\.')
-      .replace(/\*\*/g, '<<GLOBSTAR>>')
-      .replace(/\*/g, '[^/]*')
-      .replace(/<<GLOBSTAR>>/g, '.*')
-
-    // If pattern doesn't start with ** or *, anchor to be a suffix match
+    let regexStr = globToRegexSource(pattern)
+
+    // If pattern doesn't start with ** or *, anchor to a suffix match
     if (!pattern.startsWith('**') && !pattern.startsWith('*')) {
       regexStr = regexStr + '$'
     }

package/src/core/task-runner-tasks.json

@@ -0,0 +1,14 @@
+{
+  "version": 1,
+  "tasks": [
+    {
+      "id": "task-answer-fn",
+      "instruction": "用 Write 工具在 <taskDir>/solution.ts 写入一个 TypeScript 文件，导出 `export function answer(): number { return 42 }`。",
+      "groundTruth": {
+        "kind": "file-contains",
+        "file": "solution.ts",
+        "contains": ["export function answer", "42"]
+      }
+    }
+  ]
+}