oh-my-knowledge 0.39.0 → 0.41.0

package/README.md

@@ -105,8 +105,7 @@ RAG-specific evals: see RAGAS (separate niche, complementary to omk). Full compa
 | **Git & remote sources** | install / eval from a local git ref or a remote git URL (`--git-url`); directory-skills run in a content-addressed **isolated copy** so `references/` assets are real measured input, not just `SKILL.md` |
 | **Evidence-gated management** | `omk install` registers a managed record; `omk eval` auto-writes evidence bound by content fingerprint, moving a skill `installed → measurable`; `omk list` surfaces each managed skill's status (installed / measurable / promoted / stale); `omk promote` accepts a version once its evidence passes the gate (default PROGRESS only); `omk rollback` revokes that acceptance, returning the skill to `measurable`. [spec →](docs/specs/evidence-gated-management.md) |
 | **Sample design science** | sample schema with `capability` / `difficulty` / `construct` / `provenance` metadata (HF Dataset Cards style); studio surfaces coverage breakdown plus `rubric_clarity_low` / `capability_thin` flags. [docs/specs/sample-design-spec.md](docs/specs/sample-design-spec.md) |
-| **Multi-judge ensemble** | `--judge-models claude:opus,openai:gpt-4o` cross-vendor scoring + agreement metrics |
-| **Blind A/B** | `--blind` hides variant names; HTML report has a reveal button |
+| **Multi-judge ensemble** | `--judge-models claude:opus,openai-api:gpt-4o` cross-vendor scoring + agreement metrics |
 | **Multi-run variance** | `--repeat N` repeats the eval and computes mean / SD / CI / t-test |
 | **MCP URL fetching** | pull content from private-doc URLs via an MCP server (SSO-protected knowledge bases, etc.) |
 | **Auto analysis** | detects low-discrimination assertions, flat scores, all-pass / all-fail, expensive samples |

package/README.zh.md

@@ -105,8 +105,7 @@ RAG 专项评测请看 RAGAS（独立 niche，跟 omk 互补）。完整对比
 | **Git / 远端源** | install / eval 支持本地 git ref 或远端 git URL（`--git-url`）；目录-skill 在内容寻址**隔离副本**里执行，`references/` 资产是真实测量输入，不只是 `SKILL.md` |
 | **证据门控管理** | `omk install` 登记受管记录；`omk eval` 按内容指纹自动写入证据，把 skill 从 `installed` 推到 `measurable`；`omk list` 查看各受管 skill 的状态（installed / measurable / promoted / stale）；`omk promote` 在证据过门禁（默认仅 PROGRESS）后把该版本接受为当前版本；`omk rollback` 撤销这次接受，让 skill 回到 `measurable`。[规范 →](docs/zh/specs/evidence-gated-management.md) |
 | **用例设计科学性** | Sample schema 加 `capability` / `difficulty` / `construct` / `provenance` 元数据字段（HF Dataset Cards 风），studio 输出 coverage 分桶 + `rubric_clarity_low` / `capability_thin` issue。[docs/zh/specs/sample-design-spec.md](docs/zh/specs/sample-design-spec.md) |
-| **多评委 ensemble** | `--judge-models claude:opus,openai:gpt-4o` 跨厂商评分 + agreement 度量 |
-| **盲测 A/B** | `--blind` 隐藏变体名称，HTML 报告有揭晓按钮 |
+| **多评委 ensemble** | `--judge-models claude:opus,openai-api:gpt-4o` 跨厂商评分 + agreement 度量 |
 | **多轮方差分析** | `--repeat N` 重复 N 次，计算均值/标准差/置信区间/t 检验 |
 | **MCP URL 获取** | 通过 MCP Server 获取私有文档 URL 内容（SSO 保护的知识库等） |
 | **自动分析** | 检测低区分度断言、均匀分数、全通过/全失败、高成本用例 |

package/dist/analysis/report-diagnostics.js

@@ -2,6 +2,7 @@
  * Auto-analysis: detect patterns and generate insights from evaluation results.
  */
 import { normalizeCapability } from './sample-diagnostics.js';
+import { analyzeJudgeIndependence } from '../eval-core/judge-independence.js';
 /**
  * Analyze an evaluation report and produce structured insights.
  */
@@ -15,6 +16,10 @@ export function analyzeResults(report, opts = {}) {
     const sampleQuality = opts.samples
         ? buildSampleQualityAggregate(opts.samples)
         : undefined;
+    // 评委独立性对**单变体**报告同样适用(SOLO 只有绝对分、没有 A/B 差值去抵消自我偏好,
+    // 反而更该报),故放在 variants<2 早退之前;仅 results 为空(dry-run)时不评估。
+    if (results.length > 0)
+        detectJudgeIndependence(report, insights);
     if (results.length === 0 || variants.length < 2) {
         return { insights, ...(sampleQuality && { sampleQuality }) };
     }
@@ -38,11 +43,33 @@ export function analyzeResults(report, opts = {}) {
     detectAgentAssertionDiscrimination(results, variants, insights);
     // 10. Suggest --repeat when score variance is high and no repeat data
     detectNeedRepeat(report, results, variants, insights);
+    // 11. Judge independence — 已在 variants<2 早退前调过(单变体也适用),此处不重复。
     return {
         insights,
         ...(sampleQuality && { sampleQuality }),
     };
 }
+/**
+ * 评委独立性诊断:同厂商评委(自我偏好敞口)/ 单厂商 ensemble(一致性不反驳共有偏置)。
+ * 判定走单一来源 `analyzeJudgeIndependence`;挂了 gold 校准则软化为 info。
+ */
+function detectJudgeIndependence(report, insights) {
+    const ind = analyzeJudgeIndependence(report);
+    if (ind.sameVendorJudge) {
+        insights.push({
+            type: 'judge_self_preference',
+            severity: ind.goldCalibrated ? 'info' : 'warning',
+            details: { judgeVendors: ind.judgeVendors, outputVendors: ind.outputVendors, goldCalibrated: ind.goldCalibrated },
+        });
+    }
+    if (ind.singleVendorEnsemble) {
+        insights.push({
+            type: 'single_vendor_ensemble',
+            severity: ind.goldCalibrated ? 'info' : 'warning',
+            details: { judgeVendors: ind.judgeVendors, goldCalibrated: ind.goldCalibrated },
+        });
+    }
+}
 /**
  * Build sample design science aggregate from sample metadata.
  *

package/dist/assets/agent-skills/omk/references/commands.md

@@ -67,7 +67,6 @@ omk eval [flags]
 **Flags:**
 
 - `--batch` `boolean`:batch 模式:baseline vs 每个 skill
-- `--blind` `boolean`:judge blind 模式
 - `--bootstrap` `boolean`:加 bootstrap CI
 - `--bootstrap-samples` `option`:bootstrap 重采样次数，默认 1000
 - `--budget-per-sample-ms` `option`:单 sample 时长上限 ms（必须 > 0，不传则无上限）
@@ -82,7 +81,7 @@ omk eval [flags]
 - `--executor` `option`:执行器:claude / claude-sdk / codex / codex-sdk / openai-api / gemini / 自定义命令（默认 claude）。
 - `--global` `boolean`:报告写全局 ~/.oh-my-knowledge/reports，而非项目 .omk/
 - `--gold-dir` `option`:gold dataset 目录
-- `--judge-models` `option`:评委配置，格式 executor:model[,...]，例 claude:haiku 或 claude:opus,openai:gpt-4o(≥ 2 个 = ensemble）。默认 <executor>:haiku。
+- `--judge-models` `option`:评委配置，格式 executor:model[,...]，例 claude:haiku 或 claude:opus,openai-api:gpt-4o(≥ 2 个 = ensemble）。默认 <executor>:haiku。
 - `--judge-repeat` `option`:每个 dim 评 N 次
 - `--lang` `option` (默认 `zh`):输出语言 zh|en，优先级 CLI > OMK_LANG env > zh。
 - `--layered-stats` `boolean`:输出分层统计
@@ -403,6 +402,7 @@ omk observe [sessionsDir] [flags]
 
 **Flags:**
 
+- `--feedback` `boolean`:把生产健康观测反哺已纳管的同名 skill（--no-feedback 关闭）
 - `--from` `option`:起始时间 ISO，优先级高于 --last
 - `--global` `boolean`:写全局 ~/.oh-my-knowledge/observe-health，而非项目 .omk/observe-health
 - `--kb` `option`:知识库 root，启用 KB-aware 分析

package/dist/authoring/evolver.js

@@ -868,11 +868,15 @@ export async function evolveSkill({ skillPath, samplesPath, rounds = 5, target =
         // Significance accept gate: accept only when the candidate is *significantly*
         // above the current best on the decision (val) set, not merely numerically higher
         // — rejecting gains indistinguishable from judge noise. `lastReport` is the current
-        // best's fresh eval and `candidateReport` the candidate's, over the same samples;
-        // bootstrapDiffCI resamples the two arrays independently (conservative — not a paired
-        // bootstrap). Under-powered decision sets degrade to the legacy point-estimate accept
-        // (note: that path compares the prior-round best scalar, not this fresh re-eval) and
-        // flag `gate.underpowered`.
+        // best's fresh eval and `candidateReport` the candidate's, over the same samples.
+        // **Deliberately UNPAIRED here** (independent resampling, `decideAccept` → `bootstrapDiffCI`),
+        // unlike the report verdict / debias which switched to `bootstrapPairedDiffCI`: this is an
+        // *optimization accept-gate*, and the unpaired bootstrap's wider, conservative CI is the
+        // wanted bias — it raises the bar to accept, so evolve does not chase paired-tightened,
+        // marginally-significant gains that risk overfitting the decision set. Power is not the goal
+        // for de-/escalation gates; not accepting noise is. Under-powered decision sets degrade to the
+        // legacy point-estimate accept (note: that path compares the prior-round best scalar, not this
+        // fresh re-eval) and flag `gate.underpowered`.
         const valIds = split ? split.valIds : new Set(allSampleIds);
         const bestScores = perSampleComposite(lastReport, lastVariantKey, valIds);
         const candScores = perSampleComposite(candidateReport, candidateVariantKey, valIds);

package/dist/authoring/generator.js

@@ -1,4 +1,5 @@
 import { createExecutor } from '../executors/index.js';
+import { DEFAULT_GATE_THRESHOLD } from '../eval-core/verdict.js';
 /**
  * Generator 默认模型 'opus' (跟 eval 默认对齐)。
  * lean=true 路径会自动追加 `--effort low`,关掉 opus 默认的扩展思考,
@@ -76,7 +77,7 @@ const SYSTEM_PROMPT = `你是一个评测用例生成器。你的任务是根据
   来涨数量)。omk 的评分体系是 layered scoring: **fact 层**(deterministic 字面/工具断言)
   + **behavior 层**(代价指标如 turn 数 / 工具失败率) + **judge 层**(主观语义评分,从
   sample.rubric 派生维度,judge LLM 看 trace 评 1-5)三层独立计分,verdict 是三层
-  独立过 threshold(默认 3.5)。**fact 层的本职是测 deterministic 端点,不是测轨迹**。
+  独立过 threshold(默认 ${DEFAULT_GATE_THRESHOLD})。**fact 层的本职是测 deterministic 端点,不是测轨迹**。
 
   **断言哲学(关键):fact 测结果+里程碑,过程质量交 judge**
   ─────────────────────────────────────────────────────────────
@@ -109,7 +110,7 @@ const SYSTEM_PROMPT = `你是一个评测用例生成器。你的任务是根据
     - tools_not_called 反模式断言 0-1 条(禁止接触某禁忌工具,如 tripwire sample)
     - rubric 3-5 个判分维度(细致写明 judge 该看什么),由 sample.rubric 字段承载
 
-  *测量学背景:* 当前 omk verdict 三层独立 threshold(默认 3.5),fact 条目少之后单条
+  *测量学背景:* 当前 omk verdict 三层独立 threshold(默认 ${DEFAULT_GATE_THRESHOLD}),fact 条目少之后单条
   权重大、单次评测方差大,**强烈建议** 评测时带 \`--repeat 2\` 或更大测稳定性(coefficient
   of variation),并参考 bootstrap CI 而非点估计。这是 fact 层稀疏化的代价,换来的是
   fact 信号干净(不被 trajectory 字面噪音污染)。

package/dist/cli/commands/eval/index.d.ts

@@ -37,7 +37,6 @@ export default class Eval extends BaseCommand {
         'no-strict-baseline': import("@oclif/core/interfaces").BooleanFlag<boolean>;
         effort: import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
         'no-diagnostic': import("@oclif/core/interfaces").BooleanFlag<boolean>;
-        blind: import("@oclif/core/interfaces").BooleanFlag<boolean>;
         repeat: import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
         'judge-repeat': import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
         bootstrap: import("@oclif/core/interfaces").BooleanFlag<boolean>;

package/dist/cli/commands/eval/index.js

@@ -8,6 +8,7 @@ import { parseRunConfig } from '../../lib/parse-run-config.js';
 import { makeOnProgress } from '../../lib/progress.js';
 import { computeRunTally } from '../../lib/run-tally.js';
 import { DEFAULT_BOOTSTRAP_SAMPLES } from '../../../eval-core/bootstrap.js';
+import { DEFAULT_GATE_THRESHOLD } from '../../../eval-core/verdict.js';
 import { EVALUATION_REPORT_SCHEMA_VERSION } from '../../../eval-core/evaluation-reporting.js';
 function isDryRunReport(report) {
     return Boolean(report && typeof report === 'object' && report.dryRun === true);
@@ -17,9 +18,11 @@ function isDryRunBatchReport(report) {
 }
 function verdictOptions(values) {
     const rawThreshold = values.threshold;
+    // 不传 --threshold 时返回 undefined,由 computeVerdict 应用 DEFAULT_GATE_THRESHOLD ——
+    // 避免在此再硬编码一份 3.5(单一来源在 verdict.ts)。
     const gateThreshold = rawThreshold !== undefined && Number.isFinite(Number(rawThreshold))
         ? Number(rawThreshold)
-        : 3.5;
+        : undefined;
     const rawTrivial = values['trivial-diff'];
     const triviallySmallDiff = rawTrivial !== undefined && Number.isFinite(Number(rawTrivial))
         ? Number(rawTrivial)
@@ -158,9 +161,6 @@ async function announceSavedReport({ report, filePath, reportsDir, values, lang,
 async function runEval(_args, flags, lang) {
     const { values, config, evalConfig } = parseRunConfig({ ...flags });
     const { runEvaluation, runMultiple, runBatchEvaluation } = await import('../../../eval-workflows/run-evaluation.js');
-    if (values.blind !== undefined) {
-        config.blind = values.blind;
-    }
     config.onProgress = makeOnProgress(lang);
     const repeatRaw = values.repeat;
     const parsedRepeat = repeatRaw !== undefined ? Number(repeatRaw) : (evalConfig?.repeat ?? 1);
@@ -368,8 +368,8 @@ export default class Eval extends BaseCommand {
         }),
         'judge-models': Flags.string({
             description: bilingual({
-                zh: '评委配置，格式 executor:model[,...]，例 claude:haiku 或 claude:opus,openai:gpt-4o(≥ 2 个 = ensemble）。默认 <executor>:haiku。',
-                en: 'Judge config: executor:model[,...]. e.g. claude:haiku or claude:opus,openai:gpt-4o (≥ 2 = ensemble). Default <executor>:haiku.',
+                zh: '评委配置，格式 executor:model[,...]，例 claude:haiku 或 claude:opus,openai-api:gpt-4o(≥ 2 个 = ensemble）。默认 <executor>:haiku。',
+                en: 'Judge config: executor:model[,...]. e.g. claude:haiku or claude:opus,openai-api:gpt-4o (≥ 2 = ensemble). Default <executor>:haiku.',
             }),
         }),
         'output-dir': Flags.string({
@@ -453,9 +453,6 @@ export default class Eval extends BaseCommand {
             }),
         }),
         // ── eval-runner extra ──
-        blind: Flags.boolean({
-            description: bilingual({ zh: 'judge blind 模式', en: 'Blind judge mode' }),
-        }),
         repeat: Flags.string({
             description: bilingual({ zh: '每个 sample 重复跑 N 次', en: 'Repeat each sample N times' }),
             parse: integerStringParser('--repeat', { min: 1 }),
@@ -490,7 +487,10 @@ export default class Eval extends BaseCommand {
             parse: numberStringParser('--budget-per-sample-ms', { minExclusive: 0 }),
         }),
         threshold: Flags.string({
-            description: bilingual({ zh: 'verdict 阈值，默认 3.5', en: 'Verdict threshold, default 3.5' }),
+            description: bilingual({
+                zh: `verdict 阈值，默认 ${DEFAULT_GATE_THRESHOLD}`,
+                en: `Verdict threshold, default ${DEFAULT_GATE_THRESHOLD}`,
+            }),
             parse: numberStringParser('--threshold'),
         }),
         'trivial-diff': Flags.string({

package/dist/cli/commands/list.d.ts

@@ -2,7 +2,10 @@ import { BaseCommand } from '../oclif/base-command.js';
 import { sanitizeCell } from '../lib/cell-format.js';
 import { type ManagedListRow } from '../../managed/index.js';
 import type { CliLang } from '../lib/i18n.js';
-/** CJK 全角字符按 2 列计宽,使含中文表头的列也能对齐。 */
+/** CJK 全角字符 + 星平面 emoji 按 2 列计宽,使含中文表头 / emoji 标记(如 🔬)的列也能对齐。
+ *  逐**码点**迭代(for...of),星平面 emoji 是单码点,加 `\u{1F300}-\u{1FAFF}`(含 🔬 U+1F52C)判 2 列。
+ *  注:`⚠️` 是「U+26A0 + U+FE0F」两码点、各 1 列合计 2,本就与 2 列渲染对齐,不另判(避免重复计数);
+ *  `✓`(U+2713)终端按 1 列渲染、保持 1。 */
 export declare function dispWidth(s: string): number;
 /** 按**显示宽度**截断(不是 code unit):逐码点累加 dispWidth,绝不切断 surrogate 对、CJK 也不溢出列。 */
 export declare function truncate(s: string, max: number): string;

package/dist/cli/commands/list.js

@@ -4,11 +4,14 @@ import { BaseCommand } from '../oclif/base-command.js';
 import { tCli } from '../lib/i18n.js';
 import { sanitizeCell } from '../lib/cell-format.js';
 import { globalManagedDir, listManagedRows, managedDir, resolveManagedDir, } from '../../managed/index.js';
-/** CJK 全角字符按 2 列计宽,使含中文表头的列也能对齐。 */
+/** CJK 全角字符 + 星平面 emoji 按 2 列计宽,使含中文表头 / emoji 标记(如 🔬)的列也能对齐。
+ *  逐**码点**迭代(for...of),星平面 emoji 是单码点,加 `\u{1F300}-\u{1FAFF}`(含 🔬 U+1F52C)判 2 列。
+ *  注:`⚠️` 是「U+26A0 + U+FE0F」两码点、各 1 列合计 2,本就与 2 列渲染对齐,不另判(避免重复计数);
+ *  `✓`(U+2713)终端按 1 列渲染、保持 1。 */
 export function dispWidth(s) {
     let w = 0;
     for (const ch of s)
-        w += /[ᄀ-ᅟ⺀-꓏가-힣豈-﫿︰-﹏＀-｠￠-￦]/.test(ch) ? 2 : 1;
+        w += /[ᄀ-ᅟ⺀-꓏가-힣豈-﫿︰-﹏＀-｠￠-￦\u{1F300}-\u{1FAFF}]/u.test(ch) ? 2 : 1;
     return w;
 }
 function pad(s, width) {
@@ -45,7 +48,9 @@ export function renderTable(rows, lang) {
         truncate(sanitizeCell(r.name), 40), // name 与 source 同为用户可控、可超长 → 同样按显示宽度截断,防撑爆表宽
         r.kind, // ArtifactKind 枚举(validator 已收窄),无需洗
         // 不可达 → 标「?」(drift 未核),绝不冒充 stale;reachable 且漂移才 stale ⚠️;已人工接受标 promoted ✓。
-        !r.reachable ? `${r.state} ?` : r.state === 'stale' ? 'stale ⚠️' : r.state === 'promoted' ? 'promoted ✓' : r.state,
+        // 生产盲区 🔬 与生命周期**正交**(observe 量线上部署版),叠加在 state token 之后,不替换 state。
+        (!r.reachable ? `${r.state} ?` : r.state === 'stale' ? 'stale ⚠️' : r.state === 'promoted' ? 'promoted ✓' : r.state)
+            + (r.productionGap ? ' 🔬' : ''),
         r.latestVerdict ? sanitizeCell(r.latestVerdict) : '—',
         `${r.currentEvidenceCount}/${r.totalEvidenceCount}`,
         truncate(sanitizeCell(r.sourceLabel), 48),
@@ -99,7 +104,8 @@ export default class List extends BaseCommand {
             const hasDrift = rows.some((r) => r.drifted);
             const hasUnreachable = rows.some((r) => !r.reachable);
             const hasPromoted = rows.some((r) => r.state === 'promoted');
-            if (hasDrift || hasUnreachable || hasPromoted)
+            const hasProductionGap = rows.some((r) => r.productionGap);
+            if (hasDrift || hasUnreachable || hasPromoted || hasProductionGap)
                 process.stderr.write('\n');
             if (hasPromoted)
                 process.stderr.write(tCli('cli.list.promoted_note', lang));
@@ -107,6 +113,8 @@ export default class List extends BaseCommand {
                 process.stderr.write(tCli('cli.list.drift_note', lang));
             if (hasUnreachable)
                 process.stderr.write(tCli('cli.list.unreachable_note', lang));
+            if (hasProductionGap)
+                process.stderr.write(tCli('cli.list.production_gap_note', lang));
             process.stderr.write(tCli('cli.list.legend', lang));
         });
     }

package/dist/cli/commands/observe/index.d.ts

@@ -9,6 +9,14 @@ export declare function persistObserveHealthReport(report: SkillHealthReport, ou
     id: string;
     jsonPath: string;
 };
+/**
+ * SkillHealthReport → managed 反哺的结构化最小入参(#235)。纯映射、可单测 —— 把「observe 报告 →
+ * ObserveReportView」这段层间胶水从 CLI 副作用里拆出来,免得 healthBand 取错字段 / observedAt 取错时刻
+ * 这类映射 bug 无人验。`observedAt` 取**流量窗口结束时刻**(timeRange.to,空则退 generatedAt),不是「此刻」
+ * 的 generatedAt —— 否则 latest-wins 会把所有观测当成一样新(见 ManagedObservation.observedAt)。
+ * `healthBand` 由 observability 的 `healthBandOf` 逐 skill 算(阈值单一来源,注入以保可测)。
+ */
+export declare function buildObserveReportView(report: SkillHealthReport, reportId: string, healthBandOf: (weightedGapRate: number) => 'green' | 'yellow' | 'red'): import('../../../managed/index.js').ObserveReportView;
 export default class Observe extends BaseCommand {
     static description: string;
     static examples: {
@@ -27,6 +35,7 @@ export default class Observe extends BaseCommand {
         skills: import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
         'output-dir': import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
         global: import("@oclif/core/interfaces").BooleanFlag<boolean>;
+        feedback: import("@oclif/core/interfaces").BooleanFlag<boolean>;
     };
     run(): Promise<void>;
 }

package/dist/cli/commands/observe/index.js

@@ -23,6 +23,68 @@ export function persistObserveHealthReport(report, outDir) {
     indexObserveWrite(report, jsonPath, outDir, id);
     return { id, jsonPath };
 }
+// 盲区信号类型 → 人话标签(建议补样本提示用)。技术枚举键的展示名,zh/en 分列。
+const GAP_AREA_LABELS = {
+    failed_search: { zh: '检索失败', en: 'failed search' },
+    explicit_marker: { zh: '显式缺口', en: 'explicit gap' },
+    hedging: { zh: '含糊回避', en: 'hedging' },
+    repeated_failure: { zh: '反复失败', en: 'repeated failure' },
+};
+/** 取盲区计数最高的前几类,组成「建议补哪类用例」的人话区域串(只展示信号所在,不生成具体用例)。
+ *  只迭代**四个已知盲区类型**(GAP_AREA_LABELS 的键),记录 / 报告里若混入额外键一律忽略,不进展示。 */
+function topGapAreas(gapByType, lang) {
+    const sep = lang === 'zh' ? '、' : ', ';
+    const areas = Object.keys(GAP_AREA_LABELS)
+        .map((k) => [k, gapByType[k] ?? 0])
+        .filter(([, n]) => n > 0)
+        .sort((a, b) => b[1] - a[1])
+        .slice(0, 3)
+        .map(([k]) => GAP_AREA_LABELS[k][lang])
+        .join(sep);
+    return areas || (lang === 'zh' ? '未归类盲区' : 'uncategorized gaps');
+}
+/**
+ * SkillHealthReport → managed 反哺的结构化最小入参(#235)。纯映射、可单测 —— 把「observe 报告 →
+ * ObserveReportView」这段层间胶水从 CLI 副作用里拆出来,免得 healthBand 取错字段 / observedAt 取错时刻
+ * 这类映射 bug 无人验。`observedAt` 取**流量窗口结束时刻**(timeRange.to,空则退 generatedAt),不是「此刻」
+ * 的 generatedAt —— 否则 latest-wins 会把所有观测当成一样新(见 ManagedObservation.observedAt)。
+ * `healthBand` 由 observability 的 `healthBandOf` 逐 skill 算(阈值单一来源,注入以保可测)。
+ */
+export function buildObserveReportView(report, reportId, healthBandOf) {
+    return {
+        reportId,
+        observedAt: report.meta.timeRange?.to || report.meta.generatedAt,
+        skills: Object.values(report.bySkill).map((s) => ({
+            skillName: s.skillName,
+            segmentCount: s.segmentCount,
+            gapRate: s.gap.gapRate,
+            weightedGapRate: s.gap.weightedGapRate,
+            confidence: s.confidence,
+            healthBand: healthBandOf(s.gap.weightedGapRate),
+            gapByType: s.gap.byType,
+        })),
+    };
+}
+/**
+ * observe → 管理支柱反哺(#235):把每个 skill 的生产健康落成观测追加进同名受管记录,并打印「已记录 / 生产
+ * 盲区警示」。**非致命**:管理是 observe 旁路,任何异常都不该让 observe 失败(try/catch 吞掉)。observability /
+ * managed 运行时函数动态 import,与 observe 主体一致、不拖累 CLI 启动。
+ */
+async function recordObserveFeedback(report, reportId, lang) {
+    try {
+        const { healthBandOf } = await import('../../../observability/skill-health-analyzer.js');
+        const { recordObserveHealth } = await import('../../../managed/index.js');
+        const written = recordObserveHealth(buildObserveReportView(report, reportId, healthBandOf));
+        for (const w of written) {
+            process.stdout.write(w.isProductionGap
+                ? tCli('cli.observe.production_gap', lang, { name: w.name, areas: topGapAreas(w.gapByType, lang) })
+                : tCli('cli.observe.observation_recorded', lang, { name: w.name }));
+        }
+    }
+    catch {
+        // 反哺是 observe 旁路,任何异常都不该让 observe 失败。
+    }
+}
 // `omk observe <sessions-dir>` 是默认命令 —— 分析 sessions 目录的 skill 调用健康度，产出 observe-health 报告(JSON)，
 // 由 Studio 健康报告页按需渲染。observe 这条线的另一条产物是观测收件箱(observe-inbox)，走子命令 ingest / inbox / show。
 export default class Observe extends BaseCommand {
@@ -80,6 +142,14 @@ export default class Observe extends BaseCommand {
                 en: 'Write to global ~/.oh-my-knowledge/observe-health instead of project .omk/observe-health',
             }),
         }),
+        feedback: Flags.boolean({
+            default: true,
+            allowNo: true,
+            description: bilingual({
+                zh: '把生产健康观测反哺已纳管的同名 skill（--no-feedback 关闭）',
+                en: 'Feed production-health observations back to managed skills of the same name (--no-feedback to disable)',
+            }),
+        }),
     };
     async run() {
         const { args, flags } = await this.parse(Observe);
@@ -120,7 +190,7 @@ export default class Observe extends BaseCommand {
             const outDir = flags['output-dir']
                 ? resolve(flags['output-dir'])
                 : (flags.global ? globalObserveHealthDir() : projectObserveHealthDir());
-            const { jsonPath } = persistObserveHealthReport(report, outDir);
+            const { id, jsonPath } = persistObserveHealthReport(report, outDir);
             const { sessionCount, segmentCount, toolCallCount, toolFailureRate } = report.meta;
             console.log('');
             console.log(`sessions: ${sessionCount} · segments: ${segmentCount} · tool calls: ${toolCallCount} · fail rate: ${(toolFailureRate * 100).toFixed(1)}%`);
@@ -139,6 +209,9 @@ export default class Observe extends BaseCommand {
             console.log('');
             console.log(`report written to: ${jsonPath}`);
             console.log(tCli('cli.observe.view_hint', lang));
+            // #235 受管反哺:把生产健康观测落进同名受管 skill(--no-feedback 关)。非致命旁路。
+            if (flags.feedback)
+                await recordObserveFeedback(report, id, lang);
         });
     }
 }

package/dist/cli/lib/cmd-flags.d.ts

@@ -96,7 +96,6 @@ export interface EvalFlags {
     'no-strict-baseline'?: boolean;
     effort?: string;
     'no-diagnostic'?: boolean;
-    blind?: boolean;
     repeat?: string;
     'judge-repeat'?: string;
     bootstrap?: boolean;

package/dist/cli/lib/i18n-dict/common.d.ts

@@ -1,3 +1,3 @@
 import type { CliMessage } from './types.js';
-export type CommonMessageKey = 'cli.common.unknown_domain' | 'cli.common.error_prefix' | 'cli.common.skill_dir_not_found' | 'cli.common.skill_file_not_found' | 'cli.common.skill_dir_no_skill_md' | 'cli.common.report_not_found' | 'cli.common.no_judge_model' | 'cli.common.judge_models_single_only' | 'cli.common.warn_load_samples_failed' | 'cli.update.new_version_available' | 'cli.update.box_title' | 'cli.update.box_version_line' | 'cli.update.box_upgrade_line' | 'cli.update.box_silence_line' | 'cli.observe.view_hint' | 'cli.studio.started' | 'cli.studio.stop_hint' | 'cli.studio.open_failed' | 'cli.doctor.no_skill_found' | 'cli.doctor.samples_detected' | 'cli.doctor.progress_skill_start' | 'cli.doctor.progress_skill_done';
+export type CommonMessageKey = 'cli.common.unknown_domain' | 'cli.common.error_prefix' | 'cli.common.skill_dir_not_found' | 'cli.common.skill_file_not_found' | 'cli.common.skill_dir_no_skill_md' | 'cli.common.report_not_found' | 'cli.common.no_judge_model' | 'cli.common.judge_models_single_only' | 'cli.common.warn_load_samples_failed' | 'cli.update.new_version_available' | 'cli.update.box_title' | 'cli.update.box_version_line' | 'cli.update.box_upgrade_line' | 'cli.update.box_silence_line' | 'cli.observe.view_hint' | 'cli.observe.observation_recorded' | 'cli.observe.production_gap' | 'cli.studio.started' | 'cli.studio.stop_hint' | 'cli.studio.open_failed' | 'cli.doctor.no_skill_found' | 'cli.doctor.samples_detected' | 'cli.doctor.progress_skill_start' | 'cli.doctor.progress_skill_done';
 export declare const commonDict: Record<CommonMessageKey, CliMessage>;

package/dist/cli/lib/i18n-dict/common.js

@@ -59,6 +59,14 @@ export const commonDict = {
         zh: '分析 JSON 已写入 output-dir；后续可用 omk observe 持续生成健康报告。',
         en: 'Analysis JSON written to output-dir; use omk observe to keep producing health reports.',
     },
+    'cli.observe.observation_recorded': {
+        zh: '🔭 已为受管 skill「{name}」记录生产健康观测。\n',
+        en: '🔭 Recorded a production-health observation for managed skill "{name}".\n',
+    },
+    'cli.observe.production_gap': {
+        zh: '⚠️ 受管 skill「{name}」线上检测到生产盲区（{areas}）。建议补充对应用例后重跑 omk eval —— observe 只提示，不改样本集。\n',
+        en: '⚠️ Managed skill "{name}" shows production gaps in real traffic ({areas}). Consider adding matching samples and re-running omk eval — observe only suggests, it does not edit the sample set.\n',
+    },
     'cli.studio.started': {
         zh: 'studio 已启动：{url}',
         en: 'Studio running at {url}',

package/dist/cli/lib/i18n-dict/list.d.ts

@@ -1,3 +1,3 @@
 import type { CliMessage } from './types.js';
-export type ListMessageKey = 'cli.list.header' | 'cli.list.empty' | 'cli.list.empty_hint' | 'cli.list.col_name' | 'cli.list.col_kind' | 'cli.list.col_state' | 'cli.list.col_verdict' | 'cli.list.col_evidence' | 'cli.list.col_source' | 'cli.list.drift_note' | 'cli.list.unreachable_note' | 'cli.list.promoted_note' | 'cli.list.legend';
+export type ListMessageKey = 'cli.list.header' | 'cli.list.empty' | 'cli.list.empty_hint' | 'cli.list.col_name' | 'cli.list.col_kind' | 'cli.list.col_state' | 'cli.list.col_verdict' | 'cli.list.col_evidence' | 'cli.list.col_source' | 'cli.list.drift_note' | 'cli.list.unreachable_note' | 'cli.list.promoted_note' | 'cli.list.production_gap_note' | 'cli.list.legend';
 export declare const listDict: Record<ListMessageKey, CliMessage>;

package/dist/cli/lib/i18n-dict/list.js

@@ -29,6 +29,10 @@ export const listDict = {
         zh: '✓ = 当前版本已按证据人工接受为 promoted（omk promote）。\n',
         en: '✓ = current version accepted as promoted on evidence (omk promote).\n',
     },
+    'cli.list.production_gap_note': {
+        zh: '🔬 = observe 在线上检测到生产盲区（与生命周期无关的版本无关信号）；建议补对应用例后重跑 omk eval。\n',
+        en: '🔬 = observe detected a production gap in real traffic (a version-agnostic signal, orthogonal to lifecycle); add matching samples and re-run omk eval.\n',
+    },
     'cli.list.legend': {
         zh: '证据列 = 当前有效 / 全部（历史含旧内容证据，供回滚）。\n',
         en: 'EVIDENCE column = current / total (history keeps old-content evidence for rollback).\n',

package/dist/cli/lib/i18n-dict/run.js

@@ -52,8 +52,8 @@ export const runDict = {
         en: '⚠ --judge-repeat "{value}" is invalid (expected an integer ≥ 1), falling back to 1 judge call\n',
     },
     'cli.run.no_debias_length_active': {
-        zh: 'ℹ --no-debias-length 已生效: judge prompt 退回 v2-cot, 与 < v0.21 报告 hash 一致。\n',
-        en: 'ℹ --no-debias-length is active: judge prompt reverts to v2-cot, matching < v0.21 report hashes.\n',
+        zh: 'ℹ --no-debias-length 已生效：judge prompt 去掉长度去偏指令（debias-off 变体），hash 与默认开启时不同。\n',
+        en: 'ℹ --no-debias-length is active: the judge prompt drops the length-debias instruction (debias-off variant); its hash differs from the default.\n',
     },
     'cli.run.invalid_bootstrap_samples': {
         zh: '⚠ --bootstrap-samples "{value}" 无效 (期望 ≥ 100 的整数), 已按 1000 执行\n',

package/dist/cli/lib/parse-run-config.d.ts

@@ -42,7 +42,6 @@ export interface RunConfig {
     lang: 'zh' | 'en' | undefined;
     mcpConfig: string | undefined;
     verbose: boolean | undefined;
-    blind?: boolean | undefined;
     retry?: number;
     resume?: string;
     layeredStats?: boolean;

package/dist/cli/lib/parse-run-config.js

@@ -94,7 +94,6 @@ export function parseRunConfig(values) {
     const verbose = values.verbose ?? false;
     const retry = Math.max(0, Number(values.retry ?? 0) || 0);
     const resume = values.resume;
-    const blind = values.blind ?? evalConfig?.blind ?? false;
     const layeredStats = values['layered-stats'] ?? false;
     // strict-baseline default true. Reconcile both flag forms with eval.yaml fallback.
     // Priority: --no-strict-baseline > --strict-baseline > eval.yaml strictBaseline > true。
@@ -144,7 +143,6 @@ export function parseRunConfig(values) {
             verbose,
             retry,
             resume,
-            blind,
             layeredStats,
             budget: evalConfig?.budget,
             strictBaseline,

package/dist/cli/lib/record-evolve-outcome.js

@@ -2,26 +2,35 @@ import { resolve } from 'node:path';
 import { createOverlayReportStore } from '../../server/report-store.js';
 import { projectReportsDir, globalReportsDir } from '../../eval-core/measurement-dirs.js';
 import { computeVerdict } from '../../eval-core/verdict.js';
-import { bootstrapDiffCI, DEFAULT_BOOTSTRAP_ALPHA, DEFAULT_BOOTSTRAP_SAMPLES } from '../../eval-core/bootstrap.js';
+import { bootstrapPairedDiffCI, DEFAULT_BOOTSTRAP_ALPHA, DEFAULT_BOOTSTRAP_SAMPLES } from '../../eval-core/bootstrap.js';
 import { resolveManagedDir, managedDir, loadAllManagedRecords, appendManagedEvidence, rebaselineManagedContentHash, buildEvidenceRef, probeSourceState, } from '../../managed/index.js';
 /**
  * round-bestRound vs round-0 的忠实 verdict：复刻 eval 管线（evaluation-reporting.ts）对两变体抽
- * per-sample composite → bootstrapDiffCI → computeVerdict，与 `omk eval --bootstrap` 同口径、同 α / 重采样数。
+ * per-sample composite → 按 sample 配对 bootstrapPairedDiffCI → computeVerdict，与 `omk eval --bootstrap`
+ * 同口径、同 α / 重采样数。
  * 不自造门限：评委是否显著、是否 PROGRESS 全交给既有 computeVerdict。
  */
 function winnerVerdict(report, winnerVariant) {
     const baseline = 'round-0';
-    const scoresOf = (v) => report.results
-        .map((r) => r.variants[v])
-        .filter((e) => !!e && typeof e.compositeScore === 'number' && e.compositeScore > 0)
-        .map((e) => e.compositeScore);
-    const ctrl = scoresOf(baseline);
-    const treat = scoresOf(winnerVariant);
-    const pairComparisons = ctrl.length >= 2 && treat.length >= 2
+    // 按 sample **配对**(与 evaluation-reporting 主 A/B 同口径 —— 本函数职责就是复刻 eval 管线):baseline 与
+    // winner 在同一 sample 上都可测(composite > 0)才入对。同一 sample 两版分数正相关,配对 bootstrap 收紧 diff
+    // CI;独立重采样会高估方差、保守失功效。diff = b − a = winner − baseline。
+    const compositeOf = (v, r) => {
+        const e = r.variants[v];
+        return e && typeof e.compositeScore === 'number' && e.compositeScore > 0 ? e.compositeScore : undefined;
+    };
+    const pairs = [];
+    for (const r of report.results) {
+        const a = compositeOf(baseline, r);
+        const b = compositeOf(winnerVariant, r);
+        if (a !== undefined && b !== undefined)
+            pairs.push({ a, b });
+    }
+    const pairComparisons = pairs.length >= 2
         ? [{
                 control: baseline,
                 treatment: winnerVariant,
-                diffBootstrapCI: bootstrapDiffCI(ctrl, treat, DEFAULT_BOOTSTRAP_ALPHA, DEFAULT_BOOTSTRAP_SAMPLES),
+                diffBootstrapCI: bootstrapPairedDiffCI(pairs, DEFAULT_BOOTSTRAP_ALPHA, DEFAULT_BOOTSTRAP_SAMPLES),
             }]
         : undefined;
     const slice = {