@miphamai/cli 0.63.0 → 0.64.0

package/src/ui/commands.ts

@@ -9,6 +9,7 @@ import type { MiphamConfig } from '../shared/index.ts'
 import type { SkillsLoader } from '../skills/loader'
 import type { PluginManager } from '../plugin/plugin-manager'
 import type { Message } from '../shared/types.js'
+import type { UpdateStatus } from '../shared/update'
 import { McpClient } from '../mcp/client'
 import { buildCapabilityReport } from '../core/capability-inventory'
 import { InstructionsLoader } from '../core/instructions'
@@ -18,6 +19,7 @@ import {
   produceCrsiProposal,
   produceRuleProposal,
   produceProseProposal,
+  produceCrossoverProposal,
   selectCrsiSignal,
   collectSkillFiles,
   proseProposalId,
@@ -29,6 +31,18 @@ import {
 } from '../core/crsi-producer'
 import { prefilterProposal } from '../core/proposal-guard'
 import { runEval, appendEvalScore } from '../core/eval-harness'
+import { listRewardFns } from '../core/reward-fn'
+import { runTaskPerformance, measureSkillDeltaRepeated } from '../core/task-performance'
+import { randomUUID } from 'node:crypto'
+import {
+  buildImprovementReport,
+  appendImprovement,
+  readImprovements,
+  improvementRate,
+  setPendingVerdict,
+  getPendingVerdict,
+  shouldBlockApproval,
+} from '../core/improvement-track'
 import { NPM_UPDATE_COMMAND, PACKAGE_VERSION, COAUTHOR_TRAILER } from '../shared/index.ts'
 import { getPreference } from '../config/preferences'
 import { loadCrossSessionConfig } from '../config/loader'
@@ -81,6 +95,7 @@ export interface CommandContext {
   setFocusMode: (on: boolean) => void
   setGoal: (text: string) => void
   setUltracodeMode: (on: boolean) => void
+  setUpdateStatus: (s: UpdateStatus) => void
   skillsLoader?: SkillsLoader
   pluginManager?: PluginManager
   /** i18n translate function — populated from React tree via useI18n().
@@ -761,11 +776,16 @@ const crsiInventoryCmd: CommandHandler = async (ctx) => {
 
 const crsiModifyCmd: CommandHandler = async (ctx, args) => {
   if (args[0] === '--approve') {
+    if (shouldBlockApproval(getPendingVerdict() ?? 'inconclusive')) {
+      return { content: '❌ 任务表现倒退，禁止固化。请 /crsi modify --reject 丢弃，或改进后再试。' }
+    }
     const r = approvePending()
+    setPendingVerdict(null)
     return { content: r.success ? `✅ ${r.message}` : `⚠️ ${r.message}` }
   }
   if (args[0] === '--reject') {
     const r = rejectPending()
+    setPendingVerdict(null)
     return { content: r.success ? `✅ ${r.message}` : `⚠️ ${r.message}` }
   }
   if (args.length < 3) {
@@ -794,17 +814,50 @@ const crsiModifyCmd: CommandHandler = async (ctx, args) => {
     // 文件不存在 → 宽松模式（originalContent 为空）
   }
 
-  const result = runCrsiModification({ description, filePath, newContent, originalContent })
+  const result = await runCrsiModification({
+    description,
+    filePath,
+    newContent,
+    originalContent,
+    blastRadius: [filePath],
+  })
   if (!result.applied || result.phase === 'failed') {
     return {
       content: `❌ 修改未通过（phase: ${result.phase}）。\n${result.error ?? ''}`,
     }
   }
 
+  // 测量在 runCrsiModification 成功后进行（避免 failed proposal 白跑 6 次 LLM 调用）。
+  const llm = ctx.engine.getLlm() ?? ctx.engine.getRegistry()
+  let improvementLine = ''
+  try {
+    const sample = await measureSkillDeltaRepeated(llm, { filePath, originalContent, newContent })
+    if (sample) {
+      const report = buildImprovementReport(sample, [filePath])
+      setPendingVerdict(report.verdict)
+      appendImprovement({ ...report, id: randomUUID(), timestamp: new Date().toISOString() })
+      const rate = improvementRate(readImprovements())
+      const label =
+        report.verdict === 'improved'
+          ? 'improved ✅'
+          : report.verdict === 'regressed'
+            ? 'regressed ⚠️'
+            : 'inconclusive'
+      const sign = report.deltaMean >= 0 ? '+' : ''
+      improvementLine =
+        `\n📊 改进判定: ${label} (delta ${sign}${report.deltaMean.toFixed(1)}, 噪声 ${report.noise.toFixed(1)}, 阈值 ${report.minEffect.toFixed(1)})` +
+        `\n   改进率: ${rate.improved}/${rate.total} (${(rate.rate * 100).toFixed(0)}%, Wilson 95% [${(rate.lo * 100).toFixed(0)}%, ${(rate.hi * 100).toFixed(0)}%])` +
+        (report.verdict === 'regressed' ? '\n   ⚠️ 任务表现倒退：--approve 将被拒绝。' : '')
+    }
+  } catch {
+    // 测量失败（LLM 不可用等）不阻断 modify 流程——改进信号是可选的。
+  }
+
   return {
     content:
-      `✅ 测试通过。审阅下方 diff：\n\n${result.diff}\n\n` +
-      '/crsi modify --approve  合并\n/crsi modify --reject   丢弃',
+      `✅ 测试通过。审阅下方 diff：\n\n${result.diff}\n` +
+      improvementLine +
+      '\n/crsi modify --approve  合并\n/crsi modify --reject   丢弃',
   }
 }
 
@@ -869,11 +922,12 @@ const crsiProposeCmd: CommandHandler = async (ctx, args) => {
       return { content: `❌ 提议未通过预筛：${verdict.reasons.join('; ')}` }
     }
 
-    const result = runCrsiModification({
+    const result = await runCrsiModification({
       description: proposal.description,
       filePath: proposal.filePath,
       newContent: proposal.newContent,
       originalContent: proposal.originalContent,
+      blastRadius: [proposal.filePath],
     })
     if (!result.applied || result.phase === 'failed') {
       return { content: `❌ 生成失败（phase: ${result.phase}）。\n${result.error ?? ''}` }
@@ -911,7 +965,7 @@ const crsiProposeCmd: CommandHandler = async (ctx, args) => {
       }
     }
 
-    const result = runCrsiModification(proposal)
+    const result = await runCrsiModification(proposal)
     if (!result.applied || result.phase === 'failed') {
       return { content: `❌ 固化失败（phase: ${result.phase}）。\n${result.error ?? ''}` }
     }
@@ -923,6 +977,38 @@ const crsiProposeCmd: CommandHandler = async (ctx, args) => {
     }
   }
 
+  // ── Crossover 路径：/crsi propose --crossover 合并两条重叠教训 ──
+  if (args[0] === '--crossover') {
+    const llm = ctx.engine.getLlm() ?? ctx.engine.getRegistry()
+    let current = ''
+    try {
+      const { readFileSync } = await import('node:fs')
+      const { join } = await import('node:path')
+      current = readFileSync(join(root, LESSONS_FILE), 'utf-8')
+    } catch {
+      current = ''
+    }
+    if (!current) {
+      return { content: '教训文件为空，无可合并。' }
+    }
+
+    const proposal = await produceCrossoverProposal(llm, current, new Date().toISOString())
+    if (!proposal) {
+      return { content: '没有找到可合并的重叠教训对。' }
+    }
+
+    const result = await runCrsiModification(proposal)
+    if (!result.applied || result.phase === 'failed') {
+      return { content: `❌ 合并失败（phase: ${result.phase}）。\n${result.error ?? ''}` }
+    }
+
+    return {
+      content:
+        `✅ 已生成合并教训并跑过测试。审阅 diff：\n\n${result.diff}\n\n` +
+        '/crsi modify --approve 合并 | /crsi modify --reject 丢弃',
+    }
+  }
+
   // ── 教训路径（默认）：/crsi propose 追加教训 ──
   let current = ''
   try {
@@ -939,7 +1025,7 @@ const crsiProposeCmd: CommandHandler = async (ctx, args) => {
     return { content: '没有足够的失败信号（autoApplicable insight 或高置信元规则）来生成教训。' }
   }
 
-  const result = runCrsiModification(proposal)
+  const result = await runCrsiModification(proposal)
   if (!result.applied || result.phase === 'failed') {
     return { content: `❌ 生成失败（phase: ${result.phase}）。\n${result.error ?? ''}` }
   }
@@ -951,9 +1037,28 @@ const crsiProposeCmd: CommandHandler = async (ctx, args) => {
   }
 }
 
-const crsiEvalCmd: CommandHandler = async () => {
+const crsiEvalCmd: CommandHandler = async (ctx, args) => {
+  const rewardIdx = args.indexOf('--reward')
+  const rewardName = rewardIdx >= 0 ? args[rewardIdx + 1] : undefined
+
+  if (rewardName) {
+    const llm = ctx.engine.getLlm() ?? ctx.engine.getRegistry()
+    const fns = listRewardFns(llm)
+    const fn = fns.find((f) => f.name === rewardName)
+    if (!fn) {
+      return {
+        content: `❌ 未知 reward: ${rewardName}。可用: ${fns.map((f) => f.name).join(', ')}`,
+      }
+    }
+    const report = await fn.evaluate()
+    appendEvalScore(fn.name, report)
+    return {
+      content: `得分 **${report.score}/100** (${report.passed}/${report.total})\n失败: ${report.failures.join(', ') || '无'}`,
+    }
+  }
+
   const report = runEval()
-  appendEvalScore(report)
+  appendEvalScore('mechanism-sentinel', report)
 
   const lines: string[] = ['## 🧪 CRSI Eval Harness', '']
   lines.push(`得分: **${report.score}/100** (${report.passed}/${report.total})`, '')
@@ -967,6 +1072,51 @@ const crsiEvalCmd: CommandHandler = async () => {
   if (report.failures.length > 0) {
     lines.push('', `❌ 失败任务: ${report.failures.join(', ')}`)
   }
+
+  // 奖励函数注册表（reward function = policy→feedback 抽象可见）
+  // 传 llm 列出完整注册表（task-performance 需 llm 才能跑，但构造它零 LLM 调用）。
+  const llm = ctx.engine.getLlm() ?? ctx.engine.getRegistry()
+  const fns = listRewardFns(llm)
+  lines.push('', '## 🎁 奖励函数注册表', '')
+  for (const f of fns) {
+    lines.push(`- **${f.name}** — ${f.description}`)
+  }
+  lines.push('', '`/crsi eval --reward <name>` 跑指定奖励函数')
+
+  return { content: lines.join('\n') }
+}
+
+const crsiBenchCmd: CommandHandler = async (ctx, args) => {
+  const llm = ctx.engine.getLlm() ?? ctx.engine.getRegistry()
+
+  const skillIdx = args.indexOf('--skill')
+  const skillName = skillIdx >= 0 ? args[skillIdx + 1] : undefined
+  let skill: { name: string; text: string } | undefined
+  if (skillName) {
+    const body = ctx.skillsLoader?.get(skillName)?.body
+    if (!body) {
+      return { content: `❌ 未找到 skill: ${skillName}` }
+    }
+    skill = { name: skillName, text: body }
+  }
+
+  const report = await runTaskPerformance(llm, skill ? { skill } : undefined)
+
+  const lines: string[] = [
+    '## 🎯 CRSI 任务表现基准' + (skill ? `（skill: ${skill.name}）` : ''),
+    '',
+  ]
+  lines.push(`得分: **${report.score}/100** (${report.passed}/${report.total})`, '')
+  lines.push('| 任务 | 结果 |')
+  lines.push('|------|------|')
+  for (const r of report.results) {
+    lines.push(
+      `| ${r.description.slice(0, 60)} | ${r.passed ? '✅' : '❌'}${r.detail ? ` — ${r.detail.slice(0, 80)}` : ''} |`,
+    )
+  }
+  if (report.failures.length > 0) {
+    lines.push('', `❌ 失败任务: ${report.failures.join(', ')}`)
+  }
   return { content: lines.join('\n') }
 }
 
@@ -3709,6 +3859,7 @@ const upgradeCmd: CommandHandler = async (ctx) => {
   if (ok) {
     const configPath = getConfigPath()
     const { existsSync } = await import('node:fs')
+    ctx.setUpdateStatus({ state: 'installed', latest: update.latest })
     lines.push('')
     lines.push(t('commands.upgrade.updated', { version: update.latest }))
 
@@ -4695,6 +4846,7 @@ const commandsListCmd: CommandHandler = () => {
     '/crsi propose': 'Tools & Skills',
     '/crsi prose-clear': 'Tools & Skills',
     '/crsi eval': 'Tools & Skills',
+    '/crsi bench': 'Tools & Skills',
     '/crsi meta': 'Tools & Skills',
     '/crsi interpret': 'Tools & Skills',
     '/crsi critique': 'Tools & Skills',
@@ -4853,6 +5005,7 @@ registry.set('/crsi modify', crsiModifyCmd)
 registry.set('/crsi propose', crsiProposeCmd)
 registry.set('/crsi prose-clear', crsiProseClearCmd)
 registry.set('/crsi eval', crsiEvalCmd)
+registry.set('/crsi bench', crsiBenchCmd)
 registry.set('/crsi meta', crsiMetaCmd)
 registry.set('/crsi interpret', crsiInterpretCmd)
 registry.set('/crsi critique', crsiCritiqueCmd)
@@ -5037,6 +5190,7 @@ const COMMAND_DESCRIPTIONS: Record<string, string> = {
     '固化 CRSI 失败信号（默认教训 / --rule 受管理规则 / --prose 改 skill 散文），沙箱 + 人批准门控',
   '/crsi prose-clear': '清空散文提议去重 ledger（~/.mipham/crsi/prose-proposals.jsonl）',
   '/crsi eval': 'Run the ground-truth CRSI eval harness and record the score',
+  '/crsi bench': 'Run the LLM task-performance benchmark and report the score',
   '/crsi meta': 'RSI Level 3 meta-rule analysis — rules that improve the rules',
   '/crsi interpret': 'Tool-call behavior dashboard — error patterns, usage, health',
   '/crsi critique': 'Enable/disable RLAIF self-critique on tool calls',

package/src/ui/input.tsx

@@ -5,6 +5,8 @@ import { getCommandList } from './commands.js'
 import { CommandPicker } from './command-picker.js'
 import { useI18n } from '../i18n-context'
 import { discoverSessions } from '../agent/cross-session/discovery'
+import { requestSuggestion, shouldAutocomplete, type RecentMessage } from '../core/autocomplete'
+import type { Llm } from '../providers/llm'
 
 interface InputBarProps {
   onSubmit: (input: string) => void
@@ -23,6 +25,14 @@ interface InputBarProps {
   onCancel?: () => void
   /** When false, don't auto-open the slash-command picker when typing `/`. */
   showCommandPicker?: boolean
+  /** LLM 续写建议所需的模型（app.tsx 传；RemoteEngine 下 undefined → 补全禁用）。 */
+  llm?: Llm
+  /** 最近对话上下文（供续写贴合）。 */
+  recentMessages?: RecentMessage[]
+  /** 默认 true；app.tsx 传 config.autocomplete?.enabled ?? true。 */
+  autocompleteEnabled?: boolean
+  /** 默认 400ms；app.tsx 传 config.autocomplete?.debounceMs ?? 400。 */
+  autocompleteDebounceMs?: number
 }
 
 // ── Loading verb keys (i18n) ──
@@ -92,6 +102,10 @@ export function InputBar({
   onCyclePermission,
   onCancel,
   showCommandPicker = true,
+  llm,
+  recentMessages,
+  autocompleteEnabled = true,
+  autocompleteDebounceMs = 400,
 }: InputBarProps) {
   const { t } = useI18n()
   const [value, setValue] = useState('')
@@ -109,6 +123,22 @@ export function InputBar({
   const historyIndexRef = useRef(-1) // -1 = not browsing history
   const savedDraftRef = useRef('') // saved user draft before browsing history
 
+  // ── Ghost-text 自动补全 ──
+  const [suggestion, setSuggestion] = useState<string | null>(null)
+  const suggestionTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null)
+  const suggestionReqIdRef = useRef(0)
+
+  // 统一清理：清 suggestion + 使在途/待发请求失效 + 取消防抖定时器。
+  // Escape / handleSubmit / 翻历史三处共用，避免各自手写漏掉某一环（rule of three）。
+  const clearSuggestion = () => {
+    setSuggestion(null)
+    suggestionReqIdRef.current++
+    if (suggestionTimerRef.current) {
+      clearTimeout(suggestionTimerRef.current)
+      suggestionTimerRef.current = null
+    }
+  }
+
   // Stabilize t ref — prevents stale closures in intervals and avoids
   // unnecessary effect re-runs when the i18n context value object changes.
   const tRef = useRef(t)
@@ -180,6 +210,7 @@ export function InputBar({
       // Idle → clear the draft (the intuitive "cancel")
       setValue('')
       valueRef.current = ''
+      clearSuggestion()
       return
     }
 
@@ -189,6 +220,14 @@ export function InputBar({
       onCyclePermission?.()
       return
     }
+    // Tab → 接受 ghost-text 建议（复用 Ctrl-key 的 revert 手法）
+    if (key.tab && !key.shift && suggestion) {
+      const next = valueBeforeShortcut.current + suggestion
+      setValue(next)
+      valueRef.current = next
+      setSuggestion(null)
+      return
+    }
     // Ctrl+P → toggle model picker
     // NOTE: Ink passes input=keypress.name (just 'p') when ctrl is true, not raw \x10.
     // ink-text-input inserts 'p' as literal text — revert it.
@@ -218,6 +257,9 @@ export function InputBar({
 
     // ── Arrow-key history navigation (Claude Code parity) ──
     if (key.upArrow || key.downArrow) {
+      // History navigation changes value via setValue (no onChange) — clear any
+      // ghost suggestion so a stale one isn't Tab-accepted onto a recalled entry.
+      clearSuggestion()
       // Ignore if picker is active (command picker handles its own arrows)
       if (value.startsWith('/')) return
 
@@ -299,6 +341,7 @@ export function InputBar({
     setValue('')
     valueRef.current = ''
     setPickerActive(false)
+    clearSuggestion()
   }
 
   // ── Picker mode: CommandPicker overlay ──
@@ -339,6 +382,33 @@ export function InputBar({
             // Normalize newlines → spaces. ink-text-input is single-line; multi-line
             // paste would trap arrow-key navigation on the first line.
             const normalized = val.replace(/\n/g, ' ')
+            // ── Ghost-text 自动补全：每次输入清 suggestion + 重排防抖 ──
+            setSuggestion(null)
+            const suggestionReqId = ++suggestionReqIdRef.current
+            if (suggestionTimerRef.current) {
+              clearTimeout(suggestionTimerRef.current)
+              suggestionTimerRef.current = null
+            }
+            if (
+              llm &&
+              autocompleteEnabled &&
+              shouldAutocomplete(normalized, isLoading, pickerActive)
+            ) {
+              suggestionTimerRef.current = setTimeout(() => {
+                requestSuggestion(
+                  llm,
+                  recentMessages ?? [],
+                  normalized,
+                  () => suggestionReqId !== suggestionReqIdRef.current,
+                )
+                  .then((completion) => {
+                    if (completion) setSuggestion(completion)
+                  })
+                  .catch(() => {
+                    // 补全失败非关键——静默忽略
+                  })
+              }, autocompleteDebounceMs)
+            }
             // 批量输入（paste/IME 替换）节流防渲染风暴；普通单字符输入立即显示，
             // 避免 33ms trailing 的「慢半拍」尾巴。
             const bulk = isBulkInput(valueRef.current, normalized)
@@ -363,6 +433,7 @@ export function InputBar({
             isLoading ? `${verb}...` : completionVerb ? completionVerb : t('ui.input.placeholder')
           }
         />
+        {suggestion && <Text dimColor>{suggestion}</Text>}
       </Box>
       {/* Slash command hints — shown when typing / (only when picker is NOT active) */}
       {slashHints.length > 0 && !pickerActive && (