@miphamai/cli 0.68.0 → 0.70.0

package/src/core/post-flight-checker.ts

@@ -0,0 +1,102 @@
+/**
+ * PostFlightChecker — 事后检查器（Recuris C 组件）。
+ *
+ * Recuris M=(E,W,ρ,C) 的 C：检查器。核心原则「调用技能/尝试工具 ≠ 完成证据；
+ * 只有工具/env 的观察结果支撑才算」。`PreFlightChecker` 是事前拦截（ErrorSignatureDB
+ * + RuleEngine），本模块是事后验证：把「完成谓词 (observation) => boolean」抽象出来，
+ * 对工具结果核对，决策写进会话轨迹（供 ① 组件归因 + eval 打分）。
+ *
+ * 设计约束：
+ *   - 默认 no-checker 静默、rejected 只记录不阻塞（第一阶段只产证据，不强制拦截）。
+ *   - 确定性、无 LLM、可单测。
+ */
+
+import { existsSync, readFileSync, statSync } from 'node:fs'
+import type { ToolResult } from '../shared/types'
+
+/** 工具事后观察：工具实际用的参数 + 工具返回结果。 */
+export interface ToolObservation {
+  params: Record<string, unknown>
+  result: ToolResult
+}
+
+export type CheckerDecision =
+  | { verdict: 'supported'; checkerId: string }
+  | { verdict: 'rejected'; checkerId: string; reason: string }
+  | { verdict: 'no-checker' }
+
+/** 完成谓词：工具/env 观察是否支撑「该变更已生效」。绝不信任模型自称成功。 */
+export type Checker = (observation: ToolObservation) => boolean
+
+export class PostFlightChecker {
+  private checkers = new Map<string, { id: string; predicate: Checker }>()
+
+  register(checkerId: string, toolName: string, predicate: Checker): void {
+    this.checkers.set(toolName, { id: checkerId, predicate })
+  }
+
+  /** 事后核对一次工具结果。无匹配 checker 时返回 no-checker（不阻塞）。 */
+  check(toolName: string, observation: ToolObservation): CheckerDecision {
+    const entry = this.checkers.get(toolName)
+    if (!entry) return { verdict: 'no-checker' }
+    try {
+      return entry.predicate(observation)
+        ? { verdict: 'supported', checkerId: entry.id }
+        : {
+            verdict: 'rejected',
+            checkerId: entry.id,
+            reason: `post-condition failed (${entry.id})`,
+          }
+    } catch {
+      return { verdict: 'rejected', checkerId: entry.id, reason: `checker threw (${entry.id})` }
+    }
+  }
+}
+
+// ── 首批 checker（最小、确定性、可测） ──
+
+/** Bash → exit code 0（对齐 detectViolations 语义，不重复实现）。 */
+function bashExit(o: ToolObservation): boolean {
+  return o.result.success === true
+}
+
+/** Write → 文件存在且长度一致。 */
+function writeExists(o: ToolObservation): boolean {
+  if (!o.result.success) return false
+  const filePath = String(o.params.file_path ?? '')
+  if (!filePath) return false
+  try {
+    if (!existsSync(filePath)) return false
+    const expected = Buffer.byteLength(String(o.params.content ?? ''), 'utf-8')
+    return statSync(filePath).size === expected
+  } catch {
+    return false
+  }
+}
+
+/** Edit → 工具返回 applied 且 new_string 已生效（退化为「old_string 无残留」）。 */
+function editApplied(o: ToolObservation): boolean {
+  if (!o.result.success) return false
+  const filePath = String(o.params.file_path ?? '')
+  const newString = String(o.params.new_string ?? '')
+  const oldString = String(o.params.old_string ?? '')
+  if (!filePath) return false
+  try {
+    if (!existsSync(filePath)) return false
+    const text = readFileSync(filePath, 'utf-8')
+    if (newString) return text.includes(newString)
+    if (oldString) return !text.includes(oldString)
+    return true
+  } catch {
+    return false
+  }
+}
+
+/** 默认 checker：Bash/Write/Edit 三件套（工具作用域后置条件）。 */
+export function createDefaultPostFlightChecker(): PostFlightChecker {
+  const c = new PostFlightChecker()
+  c.register('bash-exit', 'Bash', bashExit)
+  c.register('write-exists', 'Write', writeExists)
+  c.register('edit-applied', 'Edit', editApplied)
+  return c
+}

package/src/core/reward-fn.ts

@@ -2,7 +2,7 @@
 // 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 { runEval, type EvalResult } 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。 */
@@ -11,6 +11,8 @@ export interface ScoreReport {
   passed: number
   score: number // 0-100
   failures: string[]
+  /** 逐契约结果（机制哨兵 runEval 天然携带；任务表现等 reward 可缺省）。 */
+  results?: EvalResult[]
 }
 
 /** 奖励函数（reward function = policy→feedback）：一个具名、可替换的打分器。 */

package/src/core/session-log.ts

@@ -2,6 +2,7 @@ import { appendFileSync, readFileSync, mkdirSync, existsSync } from 'node:fs'
 import { join } from 'node:path'
 import { createHash } from 'node:crypto'
 import type { Message, ToolUseContent, ToolResultContent, ToolResult } from '../shared/types'
+import type { CheckerDecision } from './post-flight-checker'
 
 export type SessionEvent =
   | {
@@ -20,6 +21,7 @@ export type SessionEvent =
   | { type: 'context/inject'; at: number; source: string; text: string }
   | { type: 'compaction/summary'; at: number; summary: string; replacedCount: number }
   | { type: 'compaction/rewrite'; at: number; messages: Message[] }
+  | { type: 'checker/decision'; at: number; toolName: string; decision: CheckerDecision }
 
 export function messageToEvents(msg: Message, at = 0): SessionEvent[] {
   if (msg.role === 'user') {
@@ -94,7 +96,7 @@ export function deriveMessages(events: SessionEvent[]): Message[] {
       // 快照替换：整个投影重建（微压缩/截断等结构性编辑的字节级复现）
       out = structuredClone(e.messages)
     }
-    // 'session/start' → 无消息
+    // 'session/start' / 'checker/decision' → 无消息（决策仅记录证据，不进投影，保字节级可逆）
   }
   return out
 }

package/src/core/working-memory.ts

@@ -0,0 +1,140 @@
+/**
+ * Working Memory — 用「已验证的当前任务状态」接地记忆召回（借鉴 Recuris 的 W，仅借概念）。
+ *
+ * 问题（治「前存后忘」）：MemoryManager.recall 用 TF-IDF 去和「刚说的 query」比，
+ * 历史越长，旧但相关的记忆越容易被埋掉。Recuris 的解法是把召回绑定到「当前还剩
+ * 什么没做」（pending/in_progress），而不是只绑定到「刚说了什么」。
+ *
+ * 本文件分两层：
+ *  - Phase 1（纯只读）：`renderWorkingState(tasks)` 把「现有任务追踪」的未完成任务
+ *    渲染成紧凑接地串，拼进 recall 的 query（状态接地召回）。
+ *  - Phase 2（证据接地状态机）：`WorkingMemory` 类复现 Recuris「w_t → ρ → E_t → (a,o)
+ *    → C → w_{t+1}」环——`done` 只能由 PostFlightChecker 的 `supported` 决策推进，
+ *    `blocked` 由 `rejected` 置位；**没有「模型自称 done」的入口**（不信任模型叙事，
+ *    只信任工具/env 观察）。
+ */
+
+import type { CheckerDecision } from './post-flight-checker'
+
+export interface WorkingGoal {
+  subject: string
+  status: string
+}
+
+/** 渲染「当前仍未完成」的任务主题为紧凑接地串；无未完成任务返回空串。 */
+export function renderWorkingState(tasks: ReadonlyArray<WorkingGoal>): string {
+  const open = tasks.filter((t) => t.status === 'pending' || t.status === 'in_progress')
+  if (open.length === 0) return ''
+  return open
+    .map((t) => t.subject.trim())
+    .filter(Boolean)
+    .join('\n')
+}
+
+// ── Phase 2：证据接地状态机 ──
+
+export type GoalStatus = 'pending' | 'done' | 'blocked'
+
+export interface GoalState {
+  id: string
+  content: string
+  status: GoalStatus
+  /** 推进/置位时固化的证据（checkerId + reason），供审计「谁证明了这个状态」。 */
+  evidence: string[]
+}
+
+/** 从现有任务追踪（Task 工具的 Task 列表）同步的最小结构，避免 import Task 造成循环依赖。 */
+export interface TaskLike {
+  id: string
+  subject: string
+  status: string
+}
+
+export class WorkingMemory {
+  private goals = new Map<string, GoalState>()
+
+  /** 注册一个待办目标（新目标默认 pending）。重复 id 只更新 content，不回退已验证的状态。 */
+  setGoal(id: string, content: string): void {
+    const existing = this.goals.get(id)
+    if (existing) {
+      existing.content = content
+    } else {
+      this.goals.set(id, { id, content, status: 'pending', evidence: [] })
+    }
+  }
+
+  /**
+   * 证据接地推进：`done` 只能由 checker `supported` 决策推进；`rejected` 置 `blocked`；
+   * `no-checker` 不改状态（模型自称不算）。已 `done` 的目标不再回退。
+   */
+  observe(goalId: string, decision: CheckerDecision): void {
+    const g = this.goals.get(goalId)
+    if (!g || g.status === 'done') return
+    if (decision.verdict === 'supported') {
+      g.status = 'done'
+      g.evidence.push(`supported:${decision.checkerId}`)
+    } else if (decision.verdict === 'rejected') {
+      g.status = 'blocked'
+      g.evidence.push(`rejected:${decision.checkerId}:${decision.reason}`)
+    }
+    // no-checker：证据不足，状态不变
+  }
+
+  getGoal(id: string): GoalState | undefined {
+    return this.goals.get(id)
+  }
+
+  /** 从现有任务追踪（Task 工具的 Task 列表）重建目标状态——TaskList 是唯一真源。
+   *  pending/in_progress → pending；completed → done；failed → blocked；deleted 跳过。 */
+  syncFromTasks(tasks: ReadonlyArray<TaskLike>): void {
+    this.goals.clear()
+    for (const t of tasks) {
+      if (t.status === 'deleted') continue
+      const status: GoalStatus =
+        t.status === 'completed' ? 'done' : t.status === 'failed' ? 'blocked' : 'pending'
+      this.goals.set(t.id, { id: t.id, content: t.subject, status, evidence: [] })
+    }
+  }
+
+  /** 紧凑三态块（检索/系统提示用）；无任何目标时返回空串。 */
+  renderWorkingState(): string {
+    if (this.goals.size === 0) return ''
+    const byStatus: Record<GoalStatus, string[]> = { pending: [], done: [], blocked: [] }
+    for (const g of this.goals.values()) {
+      byStatus[g.status].push(g.content.trim())
+    }
+    return (['pending', 'done', 'blocked'] as const)
+      .map((s) => {
+        const items = byStatus[s].filter(Boolean)
+        return `[WORKING] ${s}: ${items.length > 0 ? items.join(', ') : '(none)'}`
+      })
+      .join('\n')
+  }
+}
+
+// ── 证据账本（engine post-flight 后写入；Task 完成门读取） ──
+
+export interface ToolEvidence {
+  toolName: string
+  decision: CheckerDecision
+  at: number
+}
+
+/** 会话级证据账本（对齐 Task 工具的 module-level `tasks` Map，session 生命周期）。 */
+const evidenceLog: ToolEvidence[] = []
+
+/** 记录一次工具事后决策。no-checker 不产证据（忽略）。engine post-flight 后调用。 */
+export function recordToolEvidence(toolName: string, decision: CheckerDecision): void {
+  if (decision.verdict === 'no-checker') return
+  evidenceLog.push({ toolName, decision, at: Date.now() })
+}
+
+/** 某时间点之后是否有 supported 决策（用于任务完成证据门）。 */
+export function hasSupportedEvidenceSince(sinceAt: number): boolean {
+  return evidenceLog.some((e) => e.decision.verdict === 'supported' && e.at >= sinceAt)
+}
+
+/** 测试用：清空证据账本（避免跨测试污染）。 */
+export function clearEvidenceLog(): void {
+  evidenceLog.length = 0
+}

package/src/i18n-core/locales/en-US.json

@@ -407,7 +407,15 @@
       "error": "─ Memory ─\n\nCould not read memory directory.",
       "location": "Location: {path}",
       "total": "Total:    {count} memor{plural}",
-      "footer": "Memories are used by the AI to provide personalized context across sessions.\nEach file in ~/.mipham/memory/ represents one remembered fact or preference."
+      "footer": "Memories are used by the AI to provide personalized context across sessions.\nEach file in ~/.mipham/memory/ represents one remembered fact or preference.",
+      "gc_title": "─ Memory GC ─",
+      "gc_none": "Nothing to archive. All auto-* memories are recent or have been recalled.",
+      "gc_archived": "Archived ({count}):",
+      "gc_candidates": "Manual review candidates ({count}):",
+      "gc_done": "Done. Archived files moved to ~/.mipham/memory/archive/",
+      "consolidate_title": "─ Memory Consolidate ─",
+      "consolidate_none": "No auto-* memories to consolidate.",
+      "consolidate_result": "Consolidated: {removed} auto-* removed, {merged} lessons written."
     },
     "upgrade": {
       "uptodate": "── Upgrade Mipham Code ──\n\nCurrent version: v{current}\nLatest:          v{latest}\n\n✓ Already up to date.\n\nTo check manually: https://www.npmjs.com/package/@miphamai/cli",
@@ -571,7 +579,7 @@
     },
     "hooks": {
       "title": "── Hooks ──",
-      "no_hooks": "── Hooks ──\n\nNo hooks configured.\n\nCreate hook scripts in .mipham/hooks/:\n  pre-tool-use.sh   — runs before each tool\n  post-tool-use.sh  — runs after each tool\n  stop.sh           — runs when session ends\n\nUse /loop init to scaffold hooks directory.",
+      "no_hooks": "── Hooks ──\n\nNo hooks configured.\n\nConfigure hooks in .mipham/settings.json under the hooks section:\n  PreToolUse   — before each tool\n  PostToolUse  — after each tool\n  Stop         — on session end\n\nCommand hooks read JSON from stdin; exit 2 blocks (stderr as reason).\nUse /loop init to scaffold settings.json.",
       "location": "Location: {path}",
       "found": "{count} hook(s) found.",
       "error": "── Hooks ──\n\nCould not read hooks directory."
@@ -601,6 +609,26 @@
       "audit_clean": "  ✓ no code-derivable sections found",
       "audit_found": "Add the sections above to the `prompt-exclude` frontmatter to stop sending them to the model."
     },
+    "fix": {
+      "title": "── Fix ──",
+      "dryrun_banner": "[dry-run] preview — no changes written",
+      "doctor_clean": "✓ no code-derivable CLAUDE.md sections to fix",
+      "doctor_fixed": "✓ {path}: added prompt-exclude [{added}]",
+      "config_clean": "✓ config healthy, no disabled hooks",
+      "config_corrupt": "⚠ {path}: config corrupt",
+      "config_restored": "✓ restored config from backup: {path}",
+      "hook_disabled": "⚠ hook disabled: {key}",
+      "hook_reenabled": "✓ re-enabled hook: {key}",
+      "cache_clean": "✓ no corrupt cache lines",
+      "cache_found": "⚠ {path}: {count} corrupt line(s)",
+      "cache_cleaned": "✓ cleaned {path}: removed {count} line(s)",
+      "cache_hint": "Run /fix cache --apply to remove corrupt lines",
+      "test_usage": "Usage: /fix test <test-file> [--apply]\n\nLLM-driven repair of a failing test (test frozen, source only). Dry-run by default; --apply writes the fix.",
+      "test_no_llm": "No LLM available in this session — cannot generate a fix.",
+      "test_fixed": "✓ fixed {test} (source {source}, {attempts} attempt(s))",
+      "test_failed": "✗ could not fix {test}: {detail}",
+      "test_hint": "Run /fix test <file> --apply to write the fix."
+    },
     "plugins": {
       "unavailable": "PluginManager not available in this session.",
       "title": "── Installed Plugins ──",

package/src/i18n-core/locales/zh-CN.json

@@ -407,7 +407,15 @@
       "error": "─ 记忆 ─\n\n无法读取记忆目录。",
       "location": "位置: {path}",
       "total": "总计:    {count} 条记忆",
-      "footer": "AI 使用记忆提供跨会话的个性化上下文。\n~/.mipham/memory/ 中每个文件代表一条已记住的事实或偏好。"
+      "footer": "AI 使用记忆提供跨会话的个性化上下文。\n~/.mipham/memory/ 中每个文件代表一条已记住的事实或偏好。",
+      "gc_title": "─ 记忆 GC ─",
+      "gc_none": "无需归档。所有 auto-* 记忆要么较新、要么已被召回。",
+      "gc_archived": "已归档（{count} 条）:",
+      "gc_candidates": "待人工确认（{count} 条）:",
+      "gc_done": "完成。归档文件已移到 ~/.mipham/memory/archive/",
+      "consolidate_title": "─ 记忆合并 ─",
+      "consolidate_none": "没有需要合并的 auto-* 记忆。",
+      "consolidate_result": "合并完成：删除 {removed} 条 auto-*，写入 {merged} 条 lesson-*。"
     },
     "upgrade": {
       "uptodate": "── 升级 Mipham Code ──\n\n当前版本: v{current}\n最新版本:          v{latest}\n\n✓ 已是最新版本。\n\n手动检查: https://www.npmjs.com/package/@miphamai/cli",
@@ -571,7 +579,7 @@
     },
     "hooks": {
       "title": "── 钩子 ──",
-      "no_hooks": "── 钩子 ──\n\n未配置钩子。\n\n在 .mipham/hooks/ 中创建钩子脚本:\n  pre-tool-use.sh   — 在每次工具执行前运行\n  post-tool-use.sh  — 在每次工具执行后运行\n  stop.sh           — 会话结束时运行\n\n使用 /loop init 初始化钩子目录。",
+      "no_hooks": "── 钩子 ──\n\n未配置钩子。\n\n在 .mipham/settings.json 的 hooks 段配置钩子:\n  PreToolUse   — 工具执行前\n  PostToolUse  — 工具执行后\n  Stop         — 会话结束\n\n命令钩子从 stdin 读 JSON，exit 2 拦截（stderr 作理由）。\n使用 /loop init 初始化 settings.json。",
       "location": "位置: {path}",
       "found": "找到 {count} 个钩子。",
       "error": "── 钩子 ──\n\n无法读取钩子目录。"
@@ -601,6 +609,26 @@
       "audit_clean": "  ✓ 未发现可从代码推断的冗余章节",
       "audit_found": "将上述章节加入 `prompt-exclude` frontmatter，可停止发送给模型、节省 token。"
     },
+    "fix": {
+      "title": "── 修复 ──",
+      "dryrun_banner": "[dry-run] 预览模式 — 不落盘",
+      "doctor_clean": "✓ CLAUDE.md 无需要修复的冗余章节",
+      "doctor_fixed": "✓ {path}: 已加 prompt-exclude [{added}]",
+      "config_clean": "✓ 配置无损坏、hooks 无禁用",
+      "config_corrupt": "⚠ {path}: 配置损坏",
+      "config_restored": "✓ 已从备份恢复 config: {path}",
+      "hook_disabled": "⚠ hook 已禁用: {key}",
+      "hook_reenabled": "✓ 已重新启用 hook: {key}",
+      "cache_clean": "✓ 缓存无损坏行",
+      "cache_found": "⚠ {path}: {count} 行损坏",
+      "cache_cleaned": "✓ 已清理 {path}: 移除 {count} 行",
+      "cache_hint": "使用 /fix cache --apply 清理损坏行",
+      "test_usage": "用法: /fix test <测试文件> [--apply]\n\nLLM 自动修复失败测试（测试冻结，只改源码）。默认 dry-run，--apply 落盘。",
+      "test_no_llm": "当前会话无可用 LLM，无法生成修复。",
+      "test_fixed": "✓ 已修复 {test}（源码 {source}，{attempts} 次尝试）",
+      "test_failed": "✗ 未能修复 {test}: {detail}",
+      "test_hint": "使用 /fix test <文件> --apply 落盘修复。"
+    },
     "plugins": {
       "unavailable": "当前会话中 PluginManager 不可用。",
       "title": "── 已安装插件 ──",

package/src/index.tsx

@@ -10,6 +10,7 @@ import {
   loadInferenceHookConfig,
   loadCredentialMaskingConfig,
   loadCrossSessionConfig,
+  loadSettingsJson,
 } from './config/loader'
 import {
   registerActiveSession,
@@ -44,6 +45,7 @@ import { registerMcpServerTools, syncMcpToolsOnChange } from './mcp/registry'
 import { formatMcpConnectFailures, type McpConnectFailure } from './mcp/connect-failures'
 import { AgentRegistry } from './agent/agent-registry'
 import { HookEngine } from './core/hooks'
+import { loadHookConfigs } from './core/hooks-config'
 import { ArtifactServer } from './artifacts/server'
 import { getMetrics } from './core/metrics'
 import { getWorkspaceTrust } from './core/workspace-trust'
@@ -514,6 +516,12 @@ export async function runApp(options: RunOptions): Promise<void> {
     }
   }
 
+  // Register settings.json hooks (Claude Code convention — additive across levels)
+  const settingsJson = loadSettingsJson(process.cwd())
+  for (const def of loadHookConfigs(settingsJson.hooks)) {
+    hookEngine.register(def)
+  }
+
   // Start artifact server (lazy — first artifact creation triggers listening)
   const artifactsDir = join(process.cwd(), MIPHAM_DIR, ARTIFACTS_DIR)
   const artifactServer = new ArtifactServer(artifactsDir, ARTIFACT_PORT)
@@ -547,6 +555,10 @@ export async function runApp(options: RunOptions): Promise<void> {
     for (const rule of config.permissionRules.deny ?? []) engine.getPermission().deny(rule)
   }
 
+  // Apply settings.json permissions (allow/deny) — Claude Code convention
+  for (const rule of settingsJson.permissions.allow) engine.getPermission().allow(rule)
+  for (const rule of settingsJson.permissions.deny) engine.getPermission().deny(rule)
+
   // Initialize agent registry and load plugin agents/skills/MCP/hooks
   const agentRegistry = new AgentRegistry()
   agentRegistry.loadUserAgents()

package/src/shared/package-info.ts

@@ -9,7 +9,7 @@
 export const PACKAGE_NAME = '@miphamai/cli' as const
 
 /** 当前发布版本 */
-export const PACKAGE_VERSION = '0.68.0' as const
+export const PACKAGE_VERSION = '0.70.0' as const
 
 /** npm install 全局安装命令 */
 export const NPM_INSTALL_COMMAND = `npm install -g ${PACKAGE_NAME}` as const

package/src/shared/types.ts

@@ -234,6 +234,8 @@ export interface HookConfig {
   headers?: Record<string, string>
   mcpServer?: string
   mcpTool?: string
+  /** Command timeout in seconds (Claude Code default is 60). */
+  timeout?: number
   continueOnBlock?: boolean
 }
 

package/src/skills/community-registry.json

@@ -77,5 +77,21 @@
     "file": "frontend-design.SKILL.md",
     "category": "Design",
     "author": "Anthropic"
+  },
+  {
+    "name": "grill-me",
+    "description": "A relentless one-question-at-a-time interview that stress-tests your plan or design",
+    "url": "https://github.com/mattpocock/skills",
+    "file": "skills/productivity/grill-me/SKILL.md",
+    "category": "Development",
+    "author": "Matt Pocock"
+  },
+  {
+    "name": "eli5",
+    "description": "Explain any topic like the user knows nothing about it, using an HTML artifact with big pictures and few words",
+    "url": "https://github.com/anthropics/claude-plugins-community",
+    "file": "eli5/skills/eli5/SKILL.md",
+    "category": "Communication",
+    "author": "Anthropic"
   }
 ]