mio-agent-runtime 0.16.1 → 0.17.0

package/README.md

@@ -58,6 +58,7 @@ mio creativity list          List creativity hypotheses (--status active|validat
 mio creativity generate      Generate hypotheses from 2+ --source "name|content" (or --from-insights; calls an LLM)
 mio creativity ferment       Review and refine active hypotheses (calls an LLM)
 mio creativity adopt         Record an adoption event for a stored hypothesis (writes traces only)
+mio creativity feedback      Show feedback runtime mode, snapshot metrics, and rollout decision (read-only)
 mio idea generate --goal "..."   Grounded idea pipeline: auto sources + novelty gate, persisted (calls an LLM)
 mio insight status           Insight counts: total, reported, unreported, high-value (needs @akemi-mio/insight)
 mio insight list             List insights (--unreported, --min-score N, --detector X, --limit N)
@@ -384,6 +385,7 @@ Historically risky: ...
 mio creativity status
 mio creativity list --status rejected --limit 10
 mio creativity list --json
+mio creativity feedback
 ```
 
 ```text
@@ -404,6 +406,7 @@ Recent top ideas:
 - **`status` 计入 `draft`。** `draft` 是发酵中的真实状态（`ferment` 会同时处理 `draft` 与 `active`），此前漏计导致 `status` 与 `list --status draft` 对不上。
 - **空存储是正常状态。** 引擎运行之前，`status` 报告全零计数，`list` 显示 "No hypotheses match." 两者都是有效输出，不是错误。
 - **缓存会在文件变化时失效。** 三个 JSONL 的读取按 `(size, mtime)` 校验，因此另一个进程写入后长驻的 MCP 服务端无需重启就能看到——与 `server/insight-store.js` 同一套守卫。
+- **`feedback` 只读展示正向反馈闭环状态**（`mio.creativity.feedback` 同面）：当前运行时模式（`off`/`shadow`/`canary`/`full`，默认 `off` = 逐字节原行为）、不可变快照指标（概念对数、归因覆盖率、重复加权率）、观测运行摘要（`feedback-runtime-runs.jsonl`，`shadow`/`canary`/`full` 模式下 generate 自动落盘），以及按预固定阈值算出的放量判定（`rollback`/`advance`/`hold`）。阈值走 `MIO_FEEDBACK_MIN_STAGE_RUNS`/`MIO_FEEDBACK_MIN_COVERAGE`/`MIO_FEEDBACK_MAX_NEGATIVE_RATE`，fail-closed 解析。反馈只影响生成，**永不进 `checkPromotionGate`/`checkSubmitGate`**。
 
 ### 采用证据链（adopt / status / list）
 

package/bin/mio.js

@@ -2060,6 +2060,7 @@ function creativityUsage(write = console.log) {
   write(`Usage:
   mio creativity status              Show hypothesis counts, recent top ideas, adoption join
   mio creativity list                List hypotheses (--status active|validated|rejected|draft, --sort novelty, --limit N)
+  mio creativity feedback            Show feedback runtime mode, snapshot metrics, and rollout decision (read-only)
 
   mio creativity generate        Generate hypotheses by combining sources (needs 2+ --source, calls an LLM)
   mio creativity ferment         Review/refine active hypotheses (calls an LLM)
@@ -2078,6 +2079,7 @@ Examples:
   mio creativity status
   mio creativity list --status rejected --limit 10
   mio creativity list --sort novelty
+  mio creativity feedback
   mio creativity generate --source "auth|token rotation" --source "cache|write-through"
   mio creativity generate --from-insights --source "auth|token rotation"
 
@@ -2428,7 +2430,7 @@ function creativityCommand(args, useJson) {
   if (sub === 'ferment') return creativityFermentCommand(args, useJson)
   if (sub === 'adopt') return creativityAdoptCommand(args, useJson)
 
-  if (!['status', 'list'].includes(sub)) {
+  if (!['status', 'list', 'feedback'].includes(sub)) {
     console.error(`Unknown creativity subcommand: ${sub}`)
     creativityUsage(console.error)
     process.exitCode = 1
@@ -2442,6 +2444,8 @@ function creativityCommand(args, useJson) {
   try {
     if (sub === 'status') {
       result = engine.status()
+    } else if (sub === 'feedback') {
+      result = engine.feedbackStatus()
     } else {
       const sort = optionValue(flags, '--sort')
       if (sort !== undefined && sort !== 'novelty') {
@@ -2463,9 +2467,38 @@ function creativityCommand(args, useJson) {
 
   if (useJson) return jsonOrText(result, true)
   if (sub === 'status') printCreativityStatus(result)
+  else if (sub === 'feedback') printCreativityFeedback(result)
   else printCreativityList(result)
 }
 
+function printCreativityFeedback(status) {
+  console.log('Feedback runtime (positive feedback loop):')
+  console.log(`  mode: ${status.mode}`)
+  const snap = status.snapshot
+  if (snap) {
+    console.log(
+      `  snapshot: v${snap.version}  pairs=${snap.pairs}  events=${snap.eventCount}  coverage=${snap.attributionCoverage.toFixed(3)}  dupRate=${snap.duplicateApplicationRate.toFixed(3)}`,
+    )
+  } else {
+    console.log('  snapshot: unavailable')
+  }
+  const t = status.thresholds
+  const r = status.rollout
+  console.log(
+    `  runs: shadow=${r.metrics.shadowRuns} canary=${r.metrics.canaryRuns} (min per stage ${t.minStageRuns})`,
+  )
+  console.log(`  rollout: ${r.decision}${r.to ? ` → ${r.to}` : ''} (from ${r.from})`)
+  for (const reason of r.reasons) console.log(`    - ${reason}`)
+  if (status.recentRuns.length > 0) {
+    console.log('  recent runs:')
+    for (const run of status.recentRuns) {
+      console.log(
+        `    ${new Date(run.at).toISOString().slice(0, 19)}  ${run.mode}  v${run.snapshotVersion || '-'}  pairs=${run.pairCount}  applied=${run.applied}`,
+      )
+    }
+  }
+}
+
 function mechanismUsage(write = console.log) {
   write(`mio mechanism <subcommand> — mechanism registry (promotion/rollback)
 
@@ -3582,6 +3615,7 @@ Usage:
   mio experience reuse --source-agent A --target-agent B --experience-id X   Record a reuse
   mio creativity status        Show creativity hypothesis counts, recent top ideas, adoption join
   mio creativity list          List creativity hypotheses (--status active|validated|rejected|draft, --sort novelty, --limit N)
+  mio creativity feedback      Show feedback runtime mode, snapshot metrics, and rollout decision (read-only)
   mio creativity generate      Generate hypotheses from 2+ --source "name|content" (or --from-insights; calls an LLM)
   mio creativity ferment       Review and refine active hypotheses (calls an LLM)
   mio creativity adopt         Record an adoption event for a stored hypothesis (writes traces only)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mio-agent-runtime",
-  "version": "0.16.1",
+  "version": "0.17.0",
   "description": "Mio Agent Runtime installer and MCP intelligence layer",
   "license": "MIT",
   "type": "commonjs",
@@ -36,7 +36,7 @@
     "@akemi-mio/evolution-scheduler": "^0.1.0",
     "@akemi-mio/observer": "^0.1.5",
     "@akemi-mio/insight": "^0.1.1",
-    "@akemi-mio/creativity": "^0.1.1"
+    "@akemi-mio/creativity": "^0.2.0"
   },
   "main": "index.js",
   "directories": {

package/server/creativity-engine.js

@@ -5,6 +5,24 @@ const path = require('path')
 const crypto = require('crypto')
 const { readJsonl } = require('./memory-store.js')
 const { evaluateNovelty } = require('@akemi-mio/creativity/NoveltyScorer')
+const { JsonlFeedbackEventStore } = require('@akemi-mio/creativity/JsonlFeedbackEventStore')
+const {
+  buildFeedbackSnapshot,
+  buildFeedbackPromptBlock,
+  replayFeedbackAdjustment,
+  DEFAULT_FEEDBACK_LEARNING_CONFIG,
+} = require('@akemi-mio/creativity/FeedbackLearning')
+const {
+  resolveFeedbackRuntime,
+  isFeedbackApplied,
+  isFeedbackObserved,
+  DEFAULT_FEEDBACK_RUNTIME,
+  FEEDBACK_MODES,
+} = require('@akemi-mio/creativity/FeedbackRuntime')
+const {
+  resolveFeedbackRolloutThresholds,
+  buildFeedbackRolloutReport,
+} = require('@akemi-mio/creativity/FeedbackRollout')
 
 // Cache validity key for a JSONL file: a (size, mtimeMs) pair, the same
 // guard server/insight-store.js uses (commit 0ba7fa8). Every getter below used
@@ -172,7 +190,7 @@ function pairSourceKey(name) {
   return String(name == null ? '' : name).trim().slice(0, 120)
 }
 
-function mixConcepts(sources, exploredPairs, rng, strategy) {
+function mixConcepts(sources, exploredPairs, rng, strategy, pairAdjustments) {
   if (sources.length < 2) return []
   // Dedup identity is (pair, strategy): `strategy` picks the prompt and the
   // temperature, so the documented fallback "hit 409, retry with another
@@ -199,6 +217,16 @@ function mixConcepts(sources, exploredPairs, rng, strategy) {
     const j = Math.floor(rng() * (i + 1))
     ;[pairs[i], pairs[j]] = [pairs[j], pairs[i]]
   }
+  // 正向反馈闭环 P3：按概念对反馈调整重排（缺省 pairAdjustments = 原顺序不变）。
+  // key 与快照 conceptPairKey 对齐（origin 优先，name 回退）。
+  if (pairAdjustments && pairAdjustments.size > 0) {
+    const fbId = (s) => (s.origin ? s.origin : `name:${s.name}`)
+    const adj = ([a, b]) => {
+      const k = [fbId(a), fbId(b)].sort().join('|')
+      return pairAdjustments.get(k) || 0
+    }
+    pairs.sort((x, y) => adj(y) - adj(x))
+  }
   return pairs.slice(0, Math.min(5, pairs.length))
 }
 
@@ -312,9 +340,109 @@ class CreativityEngine {
     this.strategyIndex = 0
     this.exploredPairs = new Set()
     this.rng = mulberry32(Date.now())
+    // 正向反馈闭环 P0：CLI 侧证据事件只追加 feedback-events.jsonl（交换格式），
+    // 权威真相源在主进程 SQLite，由 ingest 单向导入。
+    this.feedbackEvents = new JsonlFeedbackEventStore(dataDir)
+    // 正向反馈闭环 P3：受控启用运行时（默认 off，行为不变）。
+    this.feedbackRuntime = resolveFeedbackRuntime()
+  }
+
+  /** P3：热切换反馈模式（回滚 = 切回 off）。 */
+  setFeedbackRuntime(cfg) {
+    this.feedbackRuntime = cfg || DEFAULT_FEEDBACK_RUNTIME
+    return this.feedbackRuntime
+  }
+
+  /** P3：用本机 JSONL 事件 + 假设 provenance 构建不可变快照。 */
+  buildFeedbackSnapshot() {
+    const events = this.feedbackEvents.list()
+    const hypotheses = {}
+    for (const h of this.store.getHypotheses()) {
+      hypotheses[h.id] = h.provenance ? { sources: h.provenance.sources } : undefined
+    }
+    return buildFeedbackSnapshot(events, hypotheses, DEFAULT_FEEDBACK_LEARNING_CONFIG)
+  }
+
+  /** P3：把快照转成 mixer 调整表（key = 排序 conceptId 对）。 */
+  buildPairAdjustments(snapshot) {
+    const map = new Map()
+    for (const pair of snapshot.pairs) {
+      map.set(pair.pairKey, replayFeedbackAdjustment(pair, DEFAULT_FEEDBACK_LEARNING_CONFIG).clampedAdjustment)
+    }
+    return map
+  }
+
+  /** P3 收尾：观测运行摘要落 feedback-runtime-runs.jsonl（append-only，供放量报告）。 */
+  recordRuntimeRun(run) {
+    if (!fs.existsSync(this.dataDir)) fs.mkdirSync(this.dataDir, { recursive: true })
+    fs.appendFileSync(path.join(this.dataDir, 'feedback-runtime-runs.jsonl'), JSON.stringify(run) + '\n', 'utf8')
   }
 
-  async generate(sourceLabels, strategy) {
+  readRuntimeRuns() {
+    const file = path.join(this.dataDir, 'feedback-runtime-runs.jsonl')
+    if (!fs.existsSync(file)) return []
+    const out = []
+    for (const line of fs.readFileSync(file, 'utf8').split('\n')) {
+      if (!line.trim()) continue
+      try {
+        const row = JSON.parse(line)
+        if (
+          row &&
+          typeof row === 'object' &&
+          typeof row.mode === 'string' &&
+          row.mode !== 'off' &&
+          FEEDBACK_MODES.includes(row.mode)
+        ) {
+          out.push({
+            at: Number(row.at) || 0,
+            mode: row.mode,
+            snapshotVersion: typeof row.snapshotVersion === 'string' ? row.snapshotVersion : null,
+            pairCount: Number(row.pairCount) || 0,
+            applied: !!row.applied,
+          })
+        }
+      } catch (_) {
+        // 损坏行跳过（同 JsonlFeedbackEventStore 策略）
+      }
+    }
+    return out
+  }
+
+  /** P3 收尾：放量状态 + 判定报告（只读；阈值走 env 配置，缺省预固定默认）。 */
+  feedbackStatus() {
+    let snapshot = null
+    try {
+      snapshot = this.buildFeedbackSnapshot()
+    } catch (_) {
+      snapshot = null
+    }
+    const runs = this.readRuntimeRuns()
+    const thresholds = resolveFeedbackRolloutThresholds()
+    const rollout = buildFeedbackRolloutReport({
+      snapshot,
+      runs,
+      currentMode: this.feedbackRuntime.mode,
+      thresholds,
+    })
+    return {
+      mode: this.feedbackRuntime.mode,
+      config: this.feedbackRuntime,
+      snapshot: snapshot
+        ? {
+            version: snapshot.version,
+            pairs: snapshot.pairs.length,
+            eventCount: snapshot.eventCount,
+            attributionCoverage: snapshot.attributionCoverage,
+            duplicateApplicationRate: snapshot.duplicateApplicationRate,
+          }
+        : null,
+      recentRuns: runs.slice(-5),
+      thresholds,
+      rollout,
+    }
+  }
+
+  async generate(sourceLabels, strategy, feedbackBlockArg) {
     // A missing sources array used to reach .map() and surface as
     // "Cannot read properties of undefined", which tells the caller nothing.
     if (!Array.isArray(sourceLabels)) {
@@ -325,6 +453,8 @@ class CreativityEngine {
       content: s.content || '',
       type: s.type || 'knowledge',
       weight: s.weight || 0.5,
+      // 正向反馈闭环 P2：保留 origin 供快照 pair 对齐（缺省 = 不影响既有行为）
+      ...(s.origin ? { origin: s.origin } : {}),
     }))
 
     if (sources.length < 2) {
@@ -349,7 +479,53 @@ class CreativityEngine {
         })
       )
     )
-    const pairs = mixConcepts(sources, exploredPairs, this.rng.bind(this), strat)
+    // 正向反馈闭环 P3：受控启用。
+    //   off     = 原行为（缺省，逐字节不变）
+    //   shadow  = 只观测：计算快照与调整，记录版本，不应用到排序/prompt
+    //   canary  = 按比例应用（canaryFraction 缩放 λ）
+    //   full    = 全部应用
+    let feedbackBlock = feedbackBlockArg
+    let pairAdjustments = null
+    const feedbackInfo = { mode: this.feedbackRuntime.mode, version: null, pairs: 0, applied: false }
+    if (isFeedbackObserved(this.feedbackRuntime.mode) && !feedbackBlock) {
+      try {
+        const snapshot = this.buildFeedbackSnapshot()
+        feedbackInfo.version = snapshot.version
+        feedbackInfo.pairs = snapshot.pairs.length
+        if (snapshot.pairs.length > 0) {
+          const block = buildFeedbackPromptBlock(snapshot)
+          const adjustments = this.buildPairAdjustments(snapshot)
+          if (this.feedbackRuntime.mode === 'canary') {
+            const scale = this.feedbackRuntime.canaryFraction
+            for (const [k, v] of adjustments) adjustments.set(k, v * scale)
+          }
+          if (isFeedbackApplied(this.feedbackRuntime.mode)) {
+            feedbackBlock = block
+            pairAdjustments = adjustments
+            feedbackInfo.applied = true
+          }
+          // shadow：保持 block=null / adjustments=null，仅观测
+        }
+      } catch (_) {
+        // best-effort：反馈构建失败不得阻断生成
+      }
+    }
+    // P3 收尾：观测模式下落运行摘要，供放量报告统计阶段门槛。
+    if (isFeedbackObserved(this.feedbackRuntime.mode)) {
+      try {
+        this.recordRuntimeRun({
+          at: Date.now(),
+          mode: this.feedbackRuntime.mode,
+          snapshotVersion: feedbackInfo.version,
+          pairCount: feedbackInfo.pairs,
+          applied: feedbackInfo.applied,
+        })
+      } catch (_) {
+        // best-effort：观测落盘失败不得阻断生成
+      }
+    }
+
+    const pairs = mixConcepts(sources, exploredPairs, this.rng.bind(this), strat, pairAdjustments)
 
     if (pairs.length === 0) {
       // The bare string made a deterministic failure look like "this strategy
@@ -362,6 +538,7 @@ class CreativityEngine {
         reason: `all pairs already explored for strategy "${strat}" (${sources.length} sources, ${candidatePairs} candidate pairs)`,
         strategy: strat,
         pairsAvailable: 0,
+        feedback: feedbackInfo,
       }
     }
 
@@ -373,7 +550,7 @@ class CreativityEngine {
     // down, the JSON was unparseable, or every pair really was a duplicate.
     const errors = []
     for (const [a, b] of pairs) {
-      const prompt = this._buildGenerationPrompt(a, b, strat, recent.slice(-10), rejected.slice(-5))
+      const prompt = this._buildGenerationPrompt(a, b, strat, recent.slice(-10), rejected.slice(-5), feedbackBlock)
       try {
         const result = await this.chatJson(prompt, {
           system: [
@@ -444,7 +621,7 @@ class CreativityEngine {
       }
     }
 
-    return { ideas, strategy: strat, pairsAttempted: pairs.length, errors }
+    return { ideas, strategy: strat, pairsAttempted: pairs.length, errors, feedback: feedbackInfo }
   }
 
   async ferment(limit = 5) {
@@ -549,6 +726,23 @@ Output JSON with: title, idea, expectedBenefit, risk, novelty (0-100), feasibili
         }
 
         this.store.updateHypothesis(h.id, patch)
+
+        // 正向反馈闭环 P0：ferment 终局追加证据事件（validated/rejected）。
+        // sourceId 按 ferment 轮次确定 → 同轮重放幂等；keep/merge 不产事件。
+        // 只写 JSONL 交换层，不改假设状态、不进晋级门。
+        if (patch.status === 'validated' || patch.status === 'rejected') {
+          const feedback = this.feedbackEvents.append({
+            hypothesisId: h.id,
+            eventType: patch.status,
+            sourceType: 'experiment',
+            sourceId: `ferment:${h.id}:${patch.fermentCount}`,
+            evidenceRef: typeof result.data.reason === 'string' ? result.data.reason.slice(0, 200) : undefined,
+          })
+          if (!feedback.appended && !feedback.deduped && feedback.quarantine) {
+            errors.push({ id: h.id, title: h.title, error: `feedback quarantine: ${feedback.quarantine.reason}` })
+          }
+        }
+
         results.push({ id: h.id, title: h.title, verdict, reason: result.data.reason })
       } catch (err) {
         errors.push({ id: h.id, title: h.title, error: String((err && err.message) || err) })
@@ -644,7 +838,7 @@ Output JSON with: title, idea, expectedBenefit, risk, novelty (0-100), feasibili
       .map(decorate)
   }
 
-  _buildGenerationPrompt(a, b, strategy, recentHypotheses, rejectedHypotheses) {
+  _buildGenerationPrompt(a, b, strategy, recentHypotheses, rejectedHypotheses, feedbackBlock) {
     // Long sources used to be pasted unbounded, crowding the instructions out
     // of the context window; 600 chars keeps the pairing signal without the
     // flood.
@@ -685,6 +879,12 @@ Output JSON with: title, idea, expectedBenefit, risk, novelty (0-100), feasibili
       }
     }
 
+    // 正向反馈闭环 P2：结构化快照块（已验证来源 + 失败边界）。缺省 = 提示词不变。
+    if (feedbackBlock) {
+      const { buildFeedbackBlockSection } = require('@akemi-mio/creativity/CreativityPrompt')
+      prompt += `\n${buildFeedbackBlockSection(feedbackBlock)}`
+    }
+
     prompt += `\nOutput JSON: {"title": "...", "idea": "...", "expectedBenefit": "...", "risk": "...", "novelty": 0-100, "feasibility": 0-100, "impact": 0-100, "logic": 0-100}`
     return prompt
   }
@@ -737,8 +937,9 @@ function adoptHypothesis(engine, args = {}, ingest, dataRoot) {
   }
   const taskId = String(args.taskId || '').trim()
   const note = String(args.note || '').trim()
+  const occurredAt = Date.now()
   const result = ingest({
-    trace_id: `creativity-adopt:${hypothesisId}:${Date.now()}`,
+    trace_id: `creativity-adopt:${hypothesisId}:${occurredAt}`,
     event_type: 'creativity.adopt',
     outcome: 'success',
     payload: {
@@ -748,10 +949,22 @@ function adoptHypothesis(engine, args = {}, ingest, dataRoot) {
       ...(note ? { note } : {}),
     },
   })
+  // 正向反馈闭环 P0：adopt 是选择证据（weight 0.2 档），追加 adopted 事件到
+  // feedback-events.jsonl 交换层。store 状态不变（D1），不写 Memory（D4）。
+  // 每次有意的 adopt 调用各记一条事件；P1 聚合时在假设级归并，不虚增样本。
+  const feedback = engine.feedbackEvents.append({
+    hypothesisId,
+    eventType: 'adopted',
+    sourceType: 'adoption',
+    sourceId: `adopt:${hypothesisId}:${occurredAt}`,
+    ...(memoryId ? { evidenceRef: `memory:${memoryId}` } : {}),
+    occurredAt,
+  })
   return {
     recorded: true,
     hypothesisId,
     event: (result && result.event) || result,
+    feedback: { appended: feedback.appended, deduped: feedback.deduped },
   }
 }
 

package/server/mio-intelligence-mcp/index.js

@@ -803,6 +803,12 @@ const TOOLS = [
       required: ['hypothesisId'],
     },
   },
+  {
+    name: 'mio.creativity.feedback',
+    description:
+      'Read-only positive-feedback-loop status: runtime mode (off/shadow/canary/full), immutable snapshot metrics (pairs, attribution coverage, duplicate application rate), recent observed runs, and the pre-fixed rollout decision (rollback/advance/hold with reasons). Rollout thresholds come from env (MIO_FEEDBACK_MIN_STAGE_RUNS / MIO_FEEDBACK_MIN_COVERAGE / MIO_FEEDBACK_MAX_NEGATIVE_RATE). Feedback never gates promotion or submit.',
+    inputSchema: { type: 'object', properties: {} },
+  },
   {
     name: 'mio.mechanism.registry',
     description:
@@ -1234,6 +1240,8 @@ async function callTool(name, args = {}) {
       return creativityEngine.ferment(args.limit)
     case 'mio.creativity.adopt':
       return adoptHypothesis(creativityEngine, args, ingestObservation, dataDir)
+    case 'mio.creativity.feedback':
+      return creativityEngine.feedbackStatus()
     case 'mio.mechanism.registry':
       return mechanismRegistry(args, dataDir)
     case 'mio.mechanism.promote':