oh-my-knowledge 0.20.1 → 0.22.0

package/dist/src/cli.js

@@ -4,6 +4,7 @@ import { resolve } from 'node:path';
 import { homedir } from 'node:os';
 import { join } from 'node:path';
 import { existsSync } from 'node:fs';
+import { tCli, getCliLang, parseLangFromArgv, langFromArgv } from './cli/i18n.js';
 import { discoverVariants, parseVariantCwd } from './inputs/skill-loader.js';
 import { loadEvalConfig, configVariantsToSpecs } from './inputs/eval-config.js';
 // ---------------------------------------------------------------------------
@@ -14,7 +15,15 @@ const DEFAULT_REPORTS_DIR = join(homedir(), '.oh-my-knowledge', 'reports');
 // Defaults are applied inside parseRunConfig (after config-file merge) so that
 // CLI `undefined` can be reliably distinguished from "user passed the default value".
 // Priority order resolved in parseRunConfig: CLI arg > --config file > hard-coded default.
+/**
+ * 所有子命令都接受的通用 flag。新增 --lang 让 parseArgs strict:false 模式下
+ * 仍能把值类型化到 values.lang 上(否则未声明的 flag 会被丢弃)。
+ */
+const COMMON_OPTIONS = {
+    lang: { type: 'string' },
+};
 const RUN_OPTIONS = {
+    ...COMMON_OPTIONS,
     samples: { type: 'string' },
     'skill-dir': { type: 'string' },
     control: { type: 'string' },
@@ -38,6 +47,10 @@ const RUN_OPTIONS = {
     retry: { type: 'string' },
     resume: { type: 'string' },
     'layered-stats': { type: 'boolean' },
+    // v0.22 — strict-baseline default true. Declare both forms; reconcile in
+    // parseRunConfig (后者赢)。strict-baseline 没传 + no-strict-baseline 没传 = default true。
+    'strict-baseline': { type: 'boolean' },
+    'no-strict-baseline': { type: 'boolean' },
 };
 // ---------------------------------------------------------------------------
 // parseRunConfig
@@ -144,6 +157,21 @@ function parseRunConfig(argv, extraOptions = {}) {
     const resume = values.resume;
     const blind = values.blind ?? evalConfig?.blind ?? false;
     const layeredStats = values['layered-stats'] ?? false;
+    // v0.22 — strict-baseline default true. Reconcile both flag forms.
+    // Priority: --no-strict-baseline > --strict-baseline > undefined(=true).
+    const noStrictFlag = values['no-strict-baseline'];
+    const strictFlag = values['strict-baseline'];
+    const strictBaseline = noStrictFlag === true ? false : (strictFlag ?? true);
+    // v0.22 — extract eval.yaml variant.allowedSkills overrides (per-variant). Always
+    // wins over strictBaseline default. Empty object when no eval.yaml or no overrides.
+    const variantAllowedSkills = {};
+    if (evalConfig?.variants) {
+        for (const v of evalConfig.variants) {
+            if (v.allowedSkills !== undefined) {
+                variantAllowedSkills[v.name] = v.allowedSkills;
+            }
+        }
+    }
     return {
         values,
         config: {
@@ -168,178 +196,56 @@ function parseRunConfig(argv, extraOptions = {}) {
             blind,
             layeredStats,
             budget: evalConfig?.budget,
+            strictBaseline,
+            ...(Object.keys(variantAllowedSkills).length > 0 && { variantAllowedSkills }),
         },
     };
 }
 // ---------------------------------------------------------------------------
-// Help text
-// ---------------------------------------------------------------------------
-const HELP = `
-oh-my-knowledge — Knowledge artifact evaluation toolkit
-
-Usage:
-  omk bench run [options]     Run an evaluation
-  omk bench report [options]  Start the report server
-  omk bench ci [options]      Run evaluation and exit with pass/fail code
-  omk bench init [dir]        Scaffold a new eval project
-  omk bench gen-samples [skill]  Generate eval-samples from skill content
-  omk bench diff <id1> <id2>      Compare two evaluation reports
-  omk bench evolve <skill>       Self-improve a skill through iterative evaluation
-
-  omk analyze <dir>           Analyze cc session trace(s), produce skill 健康度日报 (v0.18)
-
-Options for "bench run":
-
-  --samples <path>       Sample file (default: eval-samples.json)
-  --skill-dir <path>     Skill definitions directory (default: skills)
-  --control <expr>       Control-group variant expression (experiment role = control)
-  --treatment <v1,v2>    Treatment-group variant expressions (comma-separated; role = treatment)
-                         Each variant expression resolves to an artifact and optional runtime context:
-                           "baseline"       — bare model, no artifact injected
-                           "git:name"       — artifact from last commit
-                           "git:ref:name"   — artifact from specific commit
-                           path with "/"    — artifact from file directly (e.g. ./v1.md)
-                           "name@/cwd"      — attach runtime context / cwd
-                         At least one of --control / --treatment must be provided.
-  --config <path>        YAML/JSON config file (evaluation-as-code).
-                         Declares samples + variants + model + executor in one file.
-                         CLI flags override config fields when both are provided.
-                         Relative paths inside the config are resolved against its directory.
-  --model <name>         Model under test (default: sonnet)
-  --judge-model <name>   Judge model (default: haiku)
-  --output-dir <path>    Report output directory (default: ~/.oh-my-knowledge/reports/)
-  --no-judge             Skip LLM judging
-  --no-cache             Disable result caching
-  --dry-run              Preview tasks without executing
-  --blind                Blind A/B mode: hide variant names in report
-  --concurrency <n>      Number of parallel tasks (default: 1)
-  --timeout <seconds>    Executor timeout per task in seconds (default: 120)
-  --repeat <n>           Run evaluation N times for variance analysis (default: 1)
-  --judge-repeat <n>     Call LLM judge N times per (sample × dimension) for self-
-                         consistency (default: 1). High stddev across runs = the
-                         judge is unstable on this rubric and the score is noisy.
-  --judge-models <list>  Multi-judge ensemble. Comma-separated executor:model pairs,
-                         e.g. claude:opus,openai:gpt-4o,gemini:pro. Each judge scores
-                         every (sample × dimension); report includes per-judge break-
-                         down + Pearson/MAD inter-judge agreement. Refutes "Claude
-                         judge Claude same-modality bias" critique. Combines with
-                         --judge-repeat. Cost ~ N_judges × N_repeat × N_samples.
-  --bootstrap            Compute bootstrap confidence intervals (distribution-free,
-                         preferred over t-interval for ordinal LLM scores). Adds
-                         per-variant CI on the mean + pairwise CI on treatment-vs-
-                         control difference (significant=0 outside CI). Reports both
-                         t-interval and bootstrap so old tooling still works.
-  --bootstrap-samples <n>  Number of bootstrap resamples (default 1000). N>10000
-                         triggers a stderr warning about runtime cost.
-  --retry <n>            Retry failed tasks up to N times with exponential backoff (default: 0)
-  --resume <report-id>   Resume from a previous report, skipping completed tasks
-  --executor <name>      Executor: claude, openai, gemini, anthropic-api, openai-api,
-                         or any shell command (e.g. "python my_provider.py")
-  --judge-executor <name> Executor for LLM judge (default: same as --executor)
-  --each                 Evaluate each skill independently against baseline
-                         Requires {name}.eval-samples.json paired with each skill
-  --skip-preflight       Skip model connectivity check before evaluation
-  --mcp-config <path>    MCP config file for URL fetching via MCP servers
-                         (default: .mcp.json in current directory)
-  --no-serve             Skip auto-starting report server after evaluation
-  --verbose              Print detailed progress for each sample (exec result, grading phases)
-  --layered-stats        Expand the three-layer (fact/behavior/judge) independent
-                         significance breakdown in the HTML report by default.
-                         Without this flag, the breakdown is collapsed behind a
-                         click-to-expand summary under each comparison.
-
-Options for "bench ci":
-  (same as "bench run", plus:)
-  --threshold <number>   Minimum score to pass, applied INDEPENDENTLY to each of
-                         the three layers (fact / behavior / LLM judge). ANY
-                         layer below threshold fails the gate — this prevents
-                         composite averaging from masking a single-layer collapse.
-                         Default: 3.5. If all three layers are absent (no
-                         assertions and no rubric defined in eval-samples), the
-                         gate FAILS with a configuration hint — no composite fallback.
-
-Options for "bench report":
-  --port <number>        Server port (default: 7799)
-  --reports-dir <path>   Reports directory (default: ~/.oh-my-knowledge/reports/)
-  --export <id>          Export report as standalone HTML file
-  --dev                  Dev mode: auto-restart on lib/ file changes
-
-Options for "bench gen-samples":
-  --each                 Generate for all skills missing eval-samples
-  --count <n>            Number of samples to generate per skill (default: 5)
-  --model <name>         Model for generation (default: sonnet)
-  --skill-dir <path>     Skill directory (default: skills), used with --each
-
-Options for "analyze":
-  <dir>                  Input: cc session JSONL file / dir (e.g. ~/.claude/projects/<slug>)
-  --kb <path>            Knowledge base root (default: auto-infer from trace cwd)
-  --last <duration>      Time window like "7d" / "30d" (default: all)
-  --from <iso>           Window start (ISO8601), takes precedence over --last
-  --to <iso>             Window end (ISO8601), takes precedence over --last
-  --skills <n1,n2,...>   Whitelist skills to analyze (default: all)
-  --output-dir <path>    Output dir (default: ~/.oh-my-knowledge/analyses/)
-
-Options for "bench evolve":
-  --rounds <n>           Maximum evolution rounds (default: 5)
-  --target <score>       Stop early when score reaches this threshold
-  --samples <path>       Sample file (default: eval-samples.json)
-  --model <name>         Model under test (default: sonnet)
-  --judge-model <name>   Judge model (default: haiku)
-  --improve-model <name> Model for generating improvements (default: sonnet)
-  --concurrency <n>      Parallel eval tasks (default: 1)
-  --timeout <seconds>    Executor timeout per task in seconds (default: 120)
-  --executor <name>      Executor to use (default: claude)
-
-Examples:
-  omk bench run --control v1 --treatment v2
-  omk bench run --control baseline --treatment my-skill
-  omk bench run --control git:my-skill --treatment my-skill
-  omk bench run --control ./old-skill.md --treatment ./new-skill.md
-  omk bench run --control baseline --treatment v1,v2,v3
-  omk bench run --config eval.yaml
-  omk bench run --config eval.yaml --model sonnet-4.6   # CLI overrides config
-  omk bench run --each
-  omk bench run --dry-run
-  omk bench report --port 8080
-  omk bench report --export v1-vs-v2-20260326-1832
-  omk bench init my-eval
-  omk bench gen-samples skills/my-skill.md
-  omk bench gen-samples --each
-  omk bench diff <report-id-1> <report-id-2>
-  omk bench evolve skills/my-skill.md --rounds 5
-  omk analyze ~/.claude/projects/-Users-lizhiyao-Documents-oh-my-knowledge
-  omk analyze ~/.claude/projects/my-project --last 7d --kb /path/to/project
-  omk analyze ~/.claude/projects/my-project --skills audit,polish
-`.trim();
-// ---------------------------------------------------------------------------
 // Update check
 // ---------------------------------------------------------------------------
-async function checkUpdate() {
+async function checkUpdate(lang) {
     try {
         const { readFileSync } = await import('node:fs');
         const { fileURLToPath } = await import('node:url');
         const { dirname, join } = await import('node:path');
         const __dirname = dirname(fileURLToPath(import.meta.url));
-        const pkg = JSON.parse(readFileSync(join(__dirname, 'package.json'), 'utf-8'));
+        const findPackageJson = (startDir) => {
+            let dir = startDir;
+            for (let i = 0; i < 5; i++) {
+                const candidate = join(dir, 'package.json');
+                if (existsSync(candidate))
+                    return candidate;
+                dir = dirname(dir);
+            }
+            return null;
+        };
+        const pkgPath = findPackageJson(__dirname);
+        if (!pkgPath)
+            return;
+        const pkg = JSON.parse(readFileSync(pkgPath, 'utf-8'));
         const registry = pkg.publishConfig?.registry || 'https://registry.npmjs.org';
         const res = await fetch(`${registry}/${pkg.name}/latest`, { signal: AbortSignal.timeout(3000) });
         if (!res.ok)
             return;
         const data = await res.json();
         if (data.version && data.version !== pkg.version) {
-            process.stderr.write(`\n💡 新版本可用: ${pkg.version} → ${data.version}，运行 npm update ${pkg.name} -g 更新\n\n`);
+            process.stderr.write(tCli('cli.update.new_version_available', lang, {
+                old: pkg.version, new: data.version, pkg: pkg.name,
+            }));
         }
     }
-    catch { /* 静默失败，不影响正常使用 */ }
+    catch { /* 静默失败,不影响正常使用 */ }
 }
 // ---------------------------------------------------------------------------
 // Main
 // ---------------------------------------------------------------------------
 async function main() {
-    checkUpdate();
+    const lang = getCliLang(parseLangFromArgv(process.argv));
+    checkUpdate(lang);
     const [domain, command, ...rest] = process.argv.slice(2);
     if (!domain || domain === '--help' || domain === '-h') {
-        console.log(HELP);
+        console.log(tCli('cli.help.main', lang).trim());
         process.exit(0);
     }
     if (domain === 'analyze') {
@@ -348,11 +254,11 @@ async function main() {
         return;
     }
     if (domain !== 'bench') {
-        console.error(`Unknown domain: ${domain}. Use "omk bench <command>" or "omk analyze <dir>".`);
+        console.error(tCli('cli.common.unknown_domain', lang, { domain }));
         process.exit(1);
     }
     if (!command || command === '--help' || command === '-h') {
-        console.log(HELP);
+        console.log(tCli('cli.help.main', lang).trim());
         process.exit(0);
     }
     switch (command) {
@@ -365,8 +271,8 @@ async function main() {
         case 'init':
             await handleInit(rest);
             break;
-        case 'ci':
-            await handleCi(rest);
+        case 'gate':
+            await handleGate(rest);
             break;
         case 'gen-samples':
             await handleGenSamples(rest);
@@ -396,60 +302,77 @@ async function main() {
             await handleFailures(rest);
             break;
         default:
-            console.error(`Unknown command: bench ${command}. Use "run", "report", "ci", "init", "gen-samples", "evolve", "diff", "gold", "debias-validate", "saturation", "verdict", "diagnose", or "failures".`);
+            console.error(tCli('cli.common.unknown_bench_command', lang, { command }));
             process.exit(1);
     }
 }
 // ---------------------------------------------------------------------------
 // Progress callback
 // ---------------------------------------------------------------------------
-function defaultOnProgress({ phase, completed, total, sample_id, variant, durationMs, inputTokens, outputTokens, costUSD, score, outputPreview, judgePhase: _judgePhase, judgeDim, skipped, attempt, maxAttempts, error, }) {
-    if (phase === 'preflight') {
-        process.stderr.write('⏳ 预检模型连通性...\n');
-        return;
-    }
-    if (phase === 'retry') {
-        process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} 🔄 重试 ${attempt}/${maxAttempts}...\n`);
-        return;
-    }
-    if (phase === 'error') {
-        process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} ❌ ${error}\n`);
-        return;
-    }
-    if (phase === 'start') {
-        process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} ⏳ 执行中...\n`);
-    }
-    else if (phase === 'exec_done') {
-        const costInfo = costUSD != null && costUSD > 0 ? ` $${costUSD.toFixed(4)}` : '';
-        process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} 执行完成 ${durationMs}ms ${inputTokens}+${outputTokens} tokens${costInfo}\n`);
-        if (outputPreview) {
-            process.stderr.write(`  输出预览: ${outputPreview.slice(0, 150).replace(/\n/g, ' ')}\n`);
+/**
+ * Factory: 闭住 lang, 返回 onProgress callback。evaluation engine 回调时不传
+ * 上下文, 所以 lang 必须在 handler 入口处通过 closure 传进来。
+ */
+function makeOnProgress(lang) {
+    return ({ phase, completed, total, sample_id, variant, durationMs, inputTokens, outputTokens, costUSD, score, outputPreview, judgePhase: _judgePhase, judgeDim, skipped, attempt, maxAttempts, error, }) => {
+        const ctx = { i: completed ?? '', n: total ?? '', sample: sample_id ?? '', variant: variant ?? '' };
+        if (phase === 'preflight') {
+            process.stderr.write(tCli('cli.progress.preflight_starting', lang));
+            return;
         }
-    }
-    else if (phase === 'grading') {
-        const dimInfo = judgeDim ? ` [${judgeDim}]` : '';
-        process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} 评审中${dimInfo}...\n`);
-    }
-    else if (phase === 'judge_done') {
-        const dimInfo = judgeDim ? ` [${judgeDim}]` : '';
-        process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} 评审完成${dimInfo} score=${score}\n`);
-    }
-    else if (phase === 'done' && skipped) {
-        if (sample_id)
-            process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} ⏭ 已跳过（已有结果）\n`);
-    }
-    else {
-        const costInfo = costUSD != null && costUSD > 0 ? ` $${costUSD.toFixed(4)}` : '';
-        const scoreInfo = typeof score === 'number' ? ` score=${score}` : '';
-        process.stderr.write(`[${completed}/${total}] ${sample_id}/${variant} ✓ ${durationMs}ms ${inputTokens}+${outputTokens} tokens${costInfo}${scoreInfo}\n`);
-    }
+        if (phase === 'retry') {
+            process.stderr.write(tCli('cli.progress.sample_retry', lang, {
+                ...ctx, attempt: attempt ?? '', max: maxAttempts ?? '',
+            }));
+            return;
+        }
+        if (phase === 'error') {
+            process.stderr.write(tCli('cli.progress.sample_error', lang, { ...ctx, error: error ?? '' }));
+            return;
+        }
+        if (phase === 'start') {
+            process.stderr.write(tCli('cli.progress.sample_executing', lang, ctx));
+        }
+        else if (phase === 'exec_done') {
+            const cost = costUSD != null && costUSD > 0 ? ` $${costUSD.toFixed(4)}` : '';
+            process.stderr.write(tCli('cli.progress.sample_exec_done', lang, {
+                ...ctx, ms: durationMs ?? '', input: inputTokens ?? '', output: outputTokens ?? '', cost,
+            }));
+            if (outputPreview) {
+                process.stderr.write(tCli('cli.progress.output_preview', lang, {
+                    preview: outputPreview.slice(0, 150).replace(/\n/g, ' '),
+                }));
+            }
+        }
+        else if (phase === 'grading') {
+            const dim = judgeDim ? ` [${judgeDim}]` : '';
+            process.stderr.write(tCli('cli.progress.judging', lang, { ...ctx, dim }));
+        }
+        else if (phase === 'judge_done') {
+            const dim = judgeDim ? ` [${judgeDim}]` : '';
+            process.stderr.write(tCli('cli.progress.judged', lang, { ...ctx, dim, score: score ?? '' }));
+        }
+        else if (phase === 'done' && skipped) {
+            if (sample_id)
+                process.stderr.write(tCli('cli.progress.skipped', lang, ctx));
+        }
+        else {
+            const cost = costUSD != null && costUSD > 0 ? ` $${costUSD.toFixed(4)}` : '';
+            const scoreInfo = typeof score === 'number' ? ` score=${score}` : '';
+            process.stderr.write(tCli('cli.progress.sample_done', lang, {
+                ...ctx, ms: durationMs ?? '', input: inputTokens ?? '', output: outputTokens ?? '',
+                cost, score: scoreInfo,
+            }));
+        }
+    };
 }
 // ---------------------------------------------------------------------------
 // handleRun
 // ---------------------------------------------------------------------------
 async function handleRun(argv) {
+    const lang = langFromArgv(argv);
     const { values, config } = parseRunConfig(argv, {
-        blind: { type: 'boolean', default: false },
+        blind: { type: 'boolean' },
         repeat: { type: 'string', default: '1' },
         'judge-repeat': { type: 'string', default: '1' },
         'judge-models': { type: 'string' },
@@ -462,27 +385,29 @@ async function handleRun(argv) {
         'budget-per-sample-ms': { type: 'string' },
     });
     const { runEvaluation, runMultiple, runEachEvaluation } = await import('./eval-workflows/run-evaluation.js');
-    config.blind = values.blind;
-    config.onProgress = defaultOnProgress;
-    // --repeat 诚实输入校验:非 ≥1 整数时提示并钳到 1,不静默掩盖用户错字/极端输入
-    // 提前到 --each 分支之前,保证 each 模式也能读到 repeat (曾经 bug: --each 吞 --repeat)
+    if (values.blind !== undefined) {
+        config.blind = values.blind;
+    }
+    config.onProgress = makeOnProgress(lang);
+    // --repeat 输入校验: 非 ≥1 整数时提示并钳到 1, 不静默掩盖用户错字 / 极端输入。
+    // 提前到 --each 分支之前, 保证 each 模式也能读到 repeat (曾经 bug: --each 吞 --repeat)。
     const repeatRaw = values.repeat;
     const parsedRepeat = repeatRaw !== undefined ? Number(repeatRaw) : 1;
     if (repeatRaw !== undefined && (!Number.isFinite(parsedRepeat) || parsedRepeat < 1)) {
-        process.stderr.write(`⚠ --repeat "${repeatRaw}" 无效(期望 ≥ 1 的整数),已按 1 次评测执行\n`);
+        process.stderr.write(tCli('cli.run.invalid_repeat', lang, { value: repeatRaw }));
     }
     const repeatCount = Math.max(1, Math.floor(parsedRepeat) || 1);
-    // --judge-repeat 同样的诚实校验:非 ≥1 整数时钳到 1
+    // --judge-repeat 同样校验: 非 ≥1 整数时钳到 1
     const judgeRepeatRaw = values['judge-repeat'];
     const parsedJudgeRepeat = judgeRepeatRaw !== undefined ? Number(judgeRepeatRaw) : 1;
     if (judgeRepeatRaw !== undefined && (!Number.isFinite(parsedJudgeRepeat) || parsedJudgeRepeat < 1)) {
-        process.stderr.write(`⚠ --judge-repeat "${judgeRepeatRaw}" 无效(期望 ≥ 1 的整数),已按 1 次 judge 执行\n`);
+        process.stderr.write(tCli('cli.run.invalid_judge_repeat', lang, { value: judgeRepeatRaw }));
     }
     const judgeRepeatCount = Math.max(1, Math.floor(parsedJudgeRepeat) || 1);
     if (judgeRepeatCount > 1)
         config.judgeRepeat = judgeRepeatCount;
     // --judge-models executor:model,executor:model,... -> JudgeConfig[]
-    // 至少 2 个才进 ensemble 模式,1 个等同于 --judge-model
+    // 至少 2 个才进 ensemble 模式, 1 个等同于 --judge-model
     const judgeModelsRaw = values['judge-models'];
     if (judgeModelsRaw) {
         const parts = judgeModelsRaw.split(',').map((s) => s.trim()).filter(Boolean);
@@ -490,7 +415,7 @@ async function handleRun(argv) {
             const [executor, ...modelParts] = p.split(':');
             const model = modelParts.join(':');
             if (!executor || !model) {
-                throw new Error(`--judge-models 格式错误: "${p}",应为 "executor:model" (如 claude:opus)`);
+                throw new Error(tCli('cli.run.invalid_judge_models_format', lang, { part: p }));
             }
             return { executor, model };
         });
@@ -498,8 +423,10 @@ async function handleRun(argv) {
             config.judgeModels = judges;
         }
         else if (judges.length === 1) {
-            // 单 judge 不走 ensemble,但允许这样写,等同于 --judge-model + --executor
-            process.stderr.write(`ℹ --judge-models 只指定 1 个 judge (${judges[0].executor}:${judges[0].model}),不触发 ensemble。如需 ensemble 至少给 2 个。\n`);
+            // 单 judge 不走 ensemble, 但允许这样写, 等同于 --judge-model + --executor
+            process.stderr.write(tCli('cli.run.judge_models_single_warning', lang, {
+                executor: judges[0].executor, model: judges[0].model,
+            }));
         }
     }
     // --budget-usd / --budget-per-sample-usd / --budget-per-sample-ms:
@@ -521,7 +448,7 @@ async function handleRun(argv) {
     // off so historical reports (judgePromptHash from v2-cot era) can be reproduced.
     if (values['no-debias-length']) {
         config.lengthDebias = false;
-        process.stderr.write('ℹ --no-debias-length 已生效:judge prompt 退回 v2-cot,与 < v0.21 报告 hash 一致。\n');
+        process.stderr.write(tCli('cli.run.no_debias_length_active', lang));
     }
     // --bootstrap / --bootstrap-samples
     if (values.bootstrap) {
@@ -529,11 +456,11 @@ async function handleRun(argv) {
         const bsRaw = values['bootstrap-samples'];
         const parsedBs = bsRaw !== undefined ? Number(bsRaw) : 1000;
         if (bsRaw !== undefined && (!Number.isFinite(parsedBs) || parsedBs < 100)) {
-            process.stderr.write(`⚠ --bootstrap-samples "${bsRaw}" 无效(期望 ≥ 100 的整数),已按 1000 执行\n`);
+            process.stderr.write(tCli('cli.run.invalid_bootstrap_samples', lang, { value: bsRaw }));
         }
         const bsCount = Math.max(100, Math.floor(parsedBs) || 1000);
         if (bsCount > 10000) {
-            process.stderr.write(`⚠ --bootstrap-samples ${bsCount} 较大,可能耗时数秒。1000 是业内标准,通常已够用。\n`);
+            process.stderr.write(tCli('cli.run.bootstrap_samples_too_large', lang, { n: bsCount }));
         }
         config.bootstrapSamples = bsCount;
     }
@@ -545,30 +472,32 @@ async function handleRun(argv) {
                 repeat: repeatCount,
                 onSkillProgress({ phase, skill, current, total }) {
                     if (phase === 'start') {
-                        process.stderr.write(`\n=== [${current}/${total}] Skill: ${skill} ===\n`);
+                        process.stderr.write(tCli('cli.run.skill_section', lang, {
+                            i: current ?? '', n: total ?? '', skill: skill ?? '',
+                        }));
                     }
                 },
             });
             console.log(JSON.stringify(report, null, 2));
             if (filePath) {
-                process.stderr.write('\n✅ 批量评测完成\n');
-                process.stderr.write(`📄 Report saved to: ${filePath}\n`);
+                process.stderr.write(tCli('cli.run.batch_complete', lang));
+                process.stderr.write(tCli('cli.run.report_saved', lang, { path: filePath }));
                 if (!values['no-serve'] && process.stdout.isTTY) {
                     const { createReportServer } = await import('./server/report-server.js');
                     const server = createReportServer({ reportsDir: config.outputDir });
                     const serverUrl = await server.start();
-                    const reportUrl = `${serverUrl}/run/${report.id}`;
-                    process.stderr.write(`\n📊 Report server running at ${serverUrl}\n`);
-                    process.stderr.write(`👉 View report: ${reportUrl}\n`);
-                    process.stderr.write('\nPress Ctrl+C to stop the server\n');
+                    const reportUrl = `${serverUrl}/reports/${report.id}`;
+                    process.stderr.write(tCli('cli.run.report_server_running', lang, { url: serverUrl }));
+                    process.stderr.write(tCli('cli.run.report_server_view', lang, { url: reportUrl }));
+                    process.stderr.write(tCli('cli.run.report_server_stop', lang));
                     const { platform } = await import('node:os');
                     const openCmd = platform() === 'darwin' ? 'open' : platform() === 'win32' ? 'start' : 'xdg-open';
                     const { execFile: execFileCb } = await import('node:child_process');
                     execFileCb(openCmd, [reportUrl], () => { });
                 }
                 else if (!values['no-serve']) {
-                    process.stderr.write('\n💡 非交互环境，已跳过 report server\n');
-                    process.stderr.write(`   查看报告: omk bench report --reports-dir ${config.outputDir}\n`);
+                    process.stderr.write(tCli('cli.run.no_serve_in_non_tty', lang));
+                    process.stderr.write(tCli('cli.run.no_serve_view_hint', lang, { dir: config.outputDir }));
                 }
             }
             return;
@@ -580,7 +509,7 @@ async function handleRun(argv) {
                 ...config,
                 repeat: repeatCount,
                 onRepeatProgress({ run, total }) {
-                    process.stderr.write(`\n=== Run ${run}/${total} ===\n`);
+                    process.stderr.write(tCli('cli.run.run_section', lang, { i: run, n: total }));
                 },
             });
             report = result.report;
@@ -604,19 +533,22 @@ async function handleRun(argv) {
             if (out.result && out.gold) {
                 process.stderr.write(formatGoldCompare(out.result, out.gold));
                 if (out.result.contaminationWarning) {
-                    process.stderr.write(`\n⚠ ${out.result.contaminationWarning}\n`);
+                    process.stderr.write(tCli('cli.run.contamination_warning', lang, {
+                        warning: out.result.contaminationWarning,
+                    }));
                 }
             }
             else {
-                process.stderr.write(`\n⚠ gold dataset 加载失败 (${goldDir}):\n`);
-                for (const m of out.loadIssues)
-                    process.stderr.write(`  - ${m}\n`);
+                process.stderr.write(tCli('cli.run.gold_load_failed', lang, { dir: goldDir }));
+                for (const m of out.loadIssues) {
+                    process.stderr.write(tCli('cli.run.gold_load_issue', lang, { message: m }));
+                }
             }
         }
         console.log(JSON.stringify(report, null, 2));
         if (filePath) {
-            process.stderr.write('\n✅ 评测完成\n');
-            process.stderr.write(`📄 Report saved to: ${filePath}\n`);
+            process.stderr.write(tCli('cli.run.eval_complete', lang));
+            process.stderr.write(tCli('cli.run.report_saved', lang, { path: filePath }));
             if (!values['no-serve'] && process.stdout.isTTY) {
                 // Auto-start report server
                 const { createReportServer } = await import('./server/report-server.js');
@@ -624,10 +556,10 @@ async function handleRun(argv) {
                     reportsDir: config.outputDir,
                 });
                 const serverUrl = await server.start();
-                const reportUrl = `${serverUrl}/run/${report.id}`;
-                process.stderr.write(`\n📊 Report server running at ${serverUrl}\n`);
-                process.stderr.write(`👉 View report: ${reportUrl}\n`);
-                process.stderr.write('\nPress Ctrl+C to stop the server\n');
+                const reportUrl = `${serverUrl}/reports/${report.id}`;
+                process.stderr.write(tCli('cli.run.report_server_running', lang, { url: serverUrl }));
+                process.stderr.write(tCli('cli.run.report_server_view', lang, { url: reportUrl }));
+                process.stderr.write(tCli('cli.run.report_server_stop', lang));
                 // Auto-open report in browser
                 const { platform } = await import('node:os');
                 const openCmd = platform() === 'darwin' ? 'open' : platform() === 'win32' ? 'start' : 'xdg-open';
@@ -635,13 +567,13 @@ async function handleRun(argv) {
                 execFileCb(openCmd, [reportUrl], () => { });
             }
             else if (!values['no-serve']) {
-                process.stderr.write('\n💡 非交互环境，已跳过 report server\n');
-                process.stderr.write(`   查看报告: omk bench report --reports-dir ${config.outputDir}\n`);
+                process.stderr.write(tCli('cli.run.no_serve_in_non_tty', lang));
+                process.stderr.write(tCli('cli.run.no_serve_view_hint', lang, { dir: config.outputDir }));
             }
         }
     }
     catch (err) {
-        console.error(`Error: ${err.message}`);
+        console.error(tCli('cli.common.error_prefix', lang, { message: err.message }));
         process.exit(1);
     }
 }
@@ -649,9 +581,11 @@ async function handleRun(argv) {
 // handleReport
 // ---------------------------------------------------------------------------
 async function handleReport(argv) {
+    const lang = langFromArgv(argv);
     const { values } = parseArgs({
         args: argv,
         options: {
+            ...COMMON_OPTIONS,
             port: { type: 'string', default: '7799' },
             'reports-dir': { type: 'string', default: DEFAULT_REPORTS_DIR },
             export: { type: 'string' },
@@ -684,7 +618,7 @@ async function handleReport(argv) {
         const store = createFileStore(resolve(values['reports-dir']));
         const report = await store.get(values.export);
         if (!report) {
-            console.error(`Report not found: ${values.export}`);
+            console.error(tCli('cli.common.report_not_found', lang, { id: values.export }));
             process.exit(1);
         }
         const html = report.each ? renderEachRunDetail(report) : renderRunDetail(report);
@@ -779,10 +713,12 @@ function parseLastWindow(spec) {
     return new Date(Date.now() - ms).toISOString();
 }
 async function handleAnalyze(argv) {
+    const lang = langFromArgv(argv);
     const { values, positionals } = parseArgs({
         args: argv,
         allowPositionals: true,
         options: {
+            ...COMMON_OPTIONS,
             kb: { type: 'string' },
             last: { type: 'string' },
             from: { type: 'string' },
@@ -793,7 +729,7 @@ async function handleAnalyze(argv) {
     });
     const dir = positionals[0];
     if (!dir) {
-        console.error('Usage: omk analyze <dir> [--kb <path>] [--last 7d] [--from ISO] [--to ISO] [--skills name1,name2]');
+        console.error(tCli('cli.help.analyze_usage', lang));
         process.exit(1);
     }
     const tracePath = resolve(dir);
@@ -842,29 +778,32 @@ async function handleAnalyze(argv) {
     console.log(skillRows.join('\n'));
     console.log('');
     console.log(`report written to: ${jsonPath}`);
-    console.log(`view in browser: omk bench report  # 打开后点首页的 "📊 Skill 健康度日报"`);
+    console.log(tCli('cli.analyze.view_in_browser', lang));
 }
 async function handleInit(argv) {
+    const lang = langFromArgv(argv);
     const targetDir = resolve(argv[0] || '.');
     const { writeFileSync, mkdirSync } = await import('node:fs');
     mkdirSync(join(targetDir, 'skills'), { recursive: true });
     writeFileSync(join(targetDir, 'eval-samples.json'), INIT_SAMPLES);
     writeFileSync(join(targetDir, 'skills', 'v1.md'), INIT_SKILL_V1);
     writeFileSync(join(targetDir, 'skills', 'v2.md'), INIT_SKILL_V2);
-    console.log(`Eval project scaffolded at: ${targetDir}`);
+    console.log(tCli('cli.init.scaffolded', lang, { dir: targetDir }));
     console.log('');
-    console.log('Next steps:');
-    console.log('  1. Edit eval-samples.json to add your test cases');
-    console.log('  2. Edit skills/v1.md and skills/v2.md with your skill versions');
-    console.log('  3. Run: omk bench run --control v1 --treatment v2');
+    console.log(tCli('cli.init.next_steps_title', lang));
+    console.log(tCli('cli.init.next_step_edit_samples', lang));
+    console.log(tCli('cli.init.next_step_edit_skills', lang));
+    console.log(tCli('cli.init.next_step_run', lang));
 }
 // ---------------------------------------------------------------------------
 // handleGenSamples
 // ---------------------------------------------------------------------------
 async function handleGenSamples(argv) {
+    const lang = langFromArgv(argv);
     const { values } = parseArgs({
         args: argv,
         options: {
+            ...COMMON_OPTIONS,
             each: { type: 'boolean', default: false },
             count: { type: 'string', default: '5' },
             model: { type: 'string', default: 'sonnet' },
@@ -881,7 +820,7 @@ async function handleGenSamples(argv) {
         // Batch mode: generate for all skills missing eval-samples
         const skillDir = resolve(values['skill-dir']);
         if (!existsSync(skillDir)) {
-            console.error(`Skill directory not found: ${skillDir}`);
+            console.error(tCli('cli.common.skill_dir_not_found', lang, { path: skillDir }));
             process.exit(1);
         }
         const { readdirSync, statSync } = await import('node:fs');
@@ -909,55 +848,63 @@ async function handleGenSamples(argv) {
                 continue;
             }
             if (existsSync(samplesPath)) {
-                process.stderr.write(`⏭️  ${name}: eval-samples 已存在，跳过\n`);
+                process.stderr.write(tCli('cli.gen.skill_skipped_existing', lang, { name }));
                 continue;
             }
-            process.stderr.write(`🔄 ${name}: 正在生成 ${count} 个测试样本...\n`);
+            process.stderr.write(tCli('cli.gen.skill_generating', lang, { name, count }));
             try {
                 const skillContent = readFileSync(skillPath, 'utf-8');
                 const { samples, costUSD } = await generateSamples({ skillContent, count, model });
                 writeFileSync(samplesPath, JSON.stringify(samples, null, 2));
-                process.stderr.write(`✅ ${name}: 已生成 ${samples.length} 个样本 → ${samplesPath} (${costUSD > 0 ? `$${costUSD.toFixed(4)}` : ''})\n`);
+                const cost = costUSD > 0 ? ` $${costUSD.toFixed(4)}` : '';
+                process.stderr.write(tCli('cli.gen.skill_done', lang, {
+                    name, n: samples.length, path: samplesPath, cost,
+                }));
                 generated++;
             }
             catch (err) {
-                process.stderr.write(`❌ ${name}: ${err.message}\n`);
+                process.stderr.write(tCli('cli.gen.skill_failed', lang, {
+                    name, message: err.message,
+                }));
             }
         }
         if (generated === 0) {
-            console.log('没有需要生成的 eval-samples（所有 skill 已有配对文件）');
+            console.log(tCli('cli.gen.batch_none_needed', lang));
         }
         else {
-            console.log(`\n共生成 ${generated} 份 eval-samples，请审查后运行: omk bench run --each`);
+            console.log(tCli('cli.gen.batch_summary', lang, { n: generated }));
         }
     }
     else {
         // Single skill mode
         const skillPath = argv.find((a) => !a.startsWith('-'));
         if (!skillPath) {
-            console.error('请指定 skill 文件路径，例如: omk bench gen-samples skills/my-skill.md');
+            console.error(tCli('cli.gen.specify_skill_path', lang));
             process.exit(1);
         }
         const resolvedPath = resolve(skillPath);
         if (!existsSync(resolvedPath)) {
-            console.error(`Skill file not found: ${resolvedPath}`);
+            console.error(tCli('cli.common.skill_file_not_found', lang, { path: resolvedPath }));
             process.exit(1);
         }
         const skillContent = readFileSync(resolvedPath, 'utf-8');
         const outputPath = resolve('eval-samples.json');
         if (existsSync(outputPath)) {
-            console.error(`eval-samples.json 已存在。如需覆盖请先删除。`);
+            console.error(tCli('cli.gen.samples_already_exists', lang));
             process.exit(1);
         }
-        process.stderr.write(`🔄 正在生成 ${count} 个测试样本...\n`);
+        process.stderr.write(tCli('cli.gen.single_generating', lang, { count }));
         try {
             const { samples, costUSD } = await generateSamples({ skillContent, count, model });
             writeFileSync(outputPath, JSON.stringify(samples, null, 2));
-            process.stderr.write(`✅ 已生成 ${samples.length} 个样本 → ${outputPath} (${costUSD > 0 ? `$${costUSD.toFixed(4)}` : ''})\n`);
-            console.log('\n请审查生成的测试样本后运行: omk bench run');
+            const cost = costUSD > 0 ? ` $${costUSD.toFixed(4)}` : '';
+            process.stderr.write(tCli('cli.gen.single_done', lang, {
+                n: samples.length, path: outputPath, cost,
+            }));
+            console.log(tCli('cli.gen.review_hint', lang));
         }
         catch (err) {
-            console.error(`生成失败: ${err.message}`);
+            console.error(tCli('cli.gen.failed', lang, { message: err.message }));
             process.exit(1);
         }
     }
@@ -966,9 +913,11 @@ async function handleGenSamples(argv) {
 // handleEvolve
 // ---------------------------------------------------------------------------
 async function handleEvolve(argv) {
+    const lang = langFromArgv(argv);
     const { values } = parseArgs({
         args: argv,
         options: {
+            ...COMMON_OPTIONS,
             rounds: { type: 'string', default: '5' },
             target: { type: 'string' },
             samples: { type: 'string', default: 'eval-samples.json' },
@@ -985,7 +934,7 @@ async function handleEvolve(argv) {
     });
     const skillPath = argv.find((a) => !a.startsWith('-'));
     if (!skillPath) {
-        console.error('请指定 skill 文件路径，例如: omk bench evolve skills/my-skill.md');
+        console.error(tCli('cli.evolve.specify_skill_path', lang));
         process.exit(1);
     }
     let samplesFile = values.samples ?? 'eval-samples.json';
@@ -996,7 +945,7 @@ async function handleEvolve(argv) {
             samplesFile = 'eval-samples.yml';
     }
     const { evolveSkill } = await import('./authoring/evolver.js');
-    process.stderr.write(`\n=== Evolution: ${skillPath} ===\n`);
+    process.stderr.write(tCli('cli.evolve.section_header', lang, { path: skillPath }));
     try {
         const result = await evolveSkill({
             skillPath: resolve(skillPath),
@@ -1010,59 +959,93 @@ async function handleEvolve(argv) {
             concurrency: Math.max(1, Number(values.concurrency) || 1),
             timeoutMs: Math.max(1, Number(values.timeout) || 120) * 1000,
             skipPreflight: values['skip-preflight'],
-            onProgress: defaultOnProgress,
+            onProgress: makeOnProgress(lang),
             onRoundProgress({ round, totalRounds: _totalRounds, phase, score, delta, accepted, costUSD, error }) {
                 if (phase === 'baseline') {
-                    process.stderr.write(`Round 0 (baseline): score=${score.toFixed(2)} ($${costUSD.toFixed(4)})\n`);
+                    process.stderr.write(tCli('cli.evolve.round_baseline', lang, {
+                        score: score.toFixed(2), cost: costUSD.toFixed(4),
+                    }));
                 }
                 else if (phase === 'error') {
-                    process.stderr.write(`Round ${round}: ✗ 改进生成失败: ${error}\n`);
+                    process.stderr.write(tCli('cli.evolve.round_error', lang, {
+                        round, error: String(error ?? ''),
+                    }));
                 }
                 else if (phase === 'done') {
-                    const deltaStr = delta >= 0 ? `+${delta.toFixed(2)}` : delta.toFixed(2);
+                    const delta_ = delta >= 0 ? `+${delta.toFixed(2)}` : delta.toFixed(2);
                     const status = accepted ? '✓ ACCEPT' : '✗ REJECT';
-                    process.stderr.write(`Round ${round}: score=${score.toFixed(2)} (${deltaStr}) ${status} ($${costUSD.toFixed(4)})\n`);
+                    process.stderr.write(tCli('cli.evolve.round_done', lang, {
+                        round, score: score.toFixed(2), delta: delta_, status, cost: costUSD.toFixed(4),
+                    }));
                 }
             },
         });
         const improvement = result.startScore > 0
             ? ((result.finalScore - result.startScore) / result.startScore * 100).toFixed(1)
             : '0';
-        process.stderr.write(`\n✅ ${result.startScore.toFixed(2)} → ${result.finalScore.toFixed(2)} (+${improvement}%) | ${result.totalRounds} 轮 | $${result.totalCostUSD.toFixed(4)}\n`);
-        process.stderr.write(`Best: ${result.bestSkillPath} → ${resolve(skillPath)}\n`);
-        process.stderr.write(`所有版本保存在: ${join(resolve(skillPath, '..'), 'evolve')}/\n`);
+        process.stderr.write(tCli('cli.evolve.summary', lang, {
+            start: result.startScore.toFixed(2), final: result.finalScore.toFixed(2),
+            percent: improvement, rounds: result.totalRounds, cost: result.totalCostUSD.toFixed(4),
+        }));
+        process.stderr.write(tCli('cli.evolve.best_path', lang, {
+            best: result.bestSkillPath, target: resolve(skillPath),
+        }));
+        process.stderr.write(tCli('cli.evolve.versions_saved', lang, {
+            dir: join(resolve(skillPath, '..'), 'evolve'),
+        }));
         if (result.reportId) {
-            process.stderr.write(`📊 评测报告: omk bench report (ID: ${result.reportId})\n`);
+            process.stderr.write(tCli('cli.evolve.report_link', lang, { id: result.reportId }));
         }
         console.log(JSON.stringify(result, null, 2));
     }
     catch (err) {
-        console.error(`Error: ${err.message}`);
+        console.error(tCli('cli.common.error_prefix', lang, { message: err.message }));
         process.exit(1);
     }
 }
 // ---------------------------------------------------------------------------
-// handleCi
+// handleGate — 跑评测 + 应用 gate, exit code 0/1 适合 CI/CD pipeline 调用。
+// 内部 = runEvaluation + computeVerdict + formatVerdictText, 与 bench verdict
+// 共用决策内核(只是 verdict 读已有报告, gate 跑完再判)。
 // ---------------------------------------------------------------------------
-async function handleCi(argv) {
+async function handleGate(argv) {
+    const lang = langFromArgv(argv);
+    if (argv[0] === '--help' || argv[0] === '-h') {
+        console.log(tCli('cli.help.main', lang).trim());
+        process.exit(0);
+    }
     const { values, config } = parseRunConfig(argv, {
         threshold: { type: 'string', default: '3.5' },
+        'trivial-diff': { type: 'string' },
     });
     const { runEvaluation } = await import('./eval-workflows/run-evaluation.js');
-    config.onProgress = defaultOnProgress;
+    config.onProgress = makeOnProgress(lang);
     try {
         const { report } = (await runEvaluation(config));
-        const threshold = Number(values.threshold);
         if (report.dryRun) {
-            console.log('CI dry-run: no scores to check');
+            console.log('Gate dry-run: no scores to check');
+            process.exit(0);
+        }
+        // gate 内核 = run + verdict, 自动覆盖 omk 全部决策维度(三层 layer-gate /
+        // bootstrap diff CI / saturation / Krippendorff α)。computeVerdict 是单一
+        // 决策源, exit code 跟 verdict.level 走 — 数据 underpowered 直接 FAIL,
+        // 堵住"过 PASS 就 deploy"的漏洞。
+        const { computeVerdict, formatVerdictText } = await import('./eval-core/verdict.js');
+        const result = computeVerdict(report, {
+            gateThreshold: Number(values.threshold),
+            triviallySmallDiff: values['trivial-diff'] != null ? Number(values['trivial-diff']) : undefined,
+        });
+        console.log(formatVerdictText(result, { verbose: true }));
+        // exit code 与 handleVerdict 对齐:只有 PROGRESS / SOLO-pass 才 0,
+        // NOISE / UNDERPOWERED / CAUTIOUS / REGRESS 全 1。pipeline `omk bench gate
+        // && deploy` 数据不显著就不会误 deploy。
+        if (result.level === 'PROGRESS') {
             process.exit(0);
         }
-        // three-gate 逻辑抽到 src/eval-core/ci-gates.ts 作纯函数,便于测试;此处只做 IO。
-        const { evaluateCiGates } = await import('./eval-core/ci-gates.js');
-        const { allPass, lines } = evaluateCiGates(report.summary || {}, threshold);
-        for (const line of lines)
-            console.log(line);
-        process.exit(allPass ? 0 : 1);
+        if (result.level === 'SOLO' && result.headline.includes('PASS')) {
+            process.exit(0);
+        }
+        process.exit(1);
     }
     catch (err) {
         console.error(`Error: ${err.message}`);
@@ -1073,6 +1056,7 @@ async function handleCi(argv) {
 // handleDiff
 // ---------------------------------------------------------------------------
 async function handleDiff(argv) {
+    const lang = langFromArgv(argv);
     // Flag-aware split: separate positional report IDs from flags so we can support
     //   omk bench diff <id>                      — within-report sample-level (v0.22)
     //   omk bench diff <id1> <id2>               — cross-report variant-level (legacy)
@@ -1094,22 +1078,13 @@ async function handleDiff(argv) {
         }
     }
     if (positional.length === 0) {
-        console.error([
-            'Usage:',
-            '  omk bench diff <reportId>                     within-report per-sample diff (v0.22)',
-            '  omk bench diff <reportId1> <reportId2>        cross-report variant-level diff',
-            '',
-            'Options:',
-            '  --regressions-only          只列 treatment < control 的样本',
-            '  --threshold <num>           regression 阈值 (default 0,即任一负 Δ 算回退)',
-            '  --variant <name>            within-report 模式下指定要钻取的 variant (default: variants[1])',
-            '  --top <n>                   只列差距最大的前 N 个样本',
-        ].join('\n'));
+        console.error(tCli('cli.help.diff_usage', lang));
         process.exit(positional.length === 0 ? 1 : 0);
     }
     const { values } = parseArgs({
         args: flagArgs,
         options: {
+            ...COMMON_OPTIONS,
             'regressions-only': { type: 'boolean', default: false },
             threshold: { type: 'string' },
             variant: { type: 'string' },
@@ -1120,18 +1095,18 @@ async function handleDiff(argv) {
     const { createFileStore } = await import('./server/report-store.js');
     const store = createFileStore(resolve(DEFAULT_REPORTS_DIR));
     if (positional.length === 1) {
-        await runSampleLevelDiff(positional[0], store, values);
+        await runSampleLevelDiff(positional[0], store, values, lang);
         return;
     }
     const [id1, id2] = positional;
     const r1 = await store.get(id1);
     const r2 = await store.get(id2);
     if (!r1) {
-        console.error(`Report not found: ${id1}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: id1 }));
         process.exit(1);
     }
     if (!r2) {
-        console.error(`Report not found: ${id2}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: id2 }));
         process.exit(1);
     }
     console.log(`\n  Diff: ${id1} → ${id2}\n`);
@@ -1187,10 +1162,10 @@ async function handleDiff(argv) {
  * Default focus is variants[0] (control) vs variants[1] (treatment), but
  * `--variant` overrides which variant is the "treatment" side.
  */
-async function runSampleLevelDiff(reportId, store, flags) {
+async function runSampleLevelDiff(reportId, store, flags, lang) {
     const report = await store.get(reportId);
     if (!report) {
-        console.error(`Report not found: ${reportId}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: reportId }));
         process.exit(1);
     }
     const variants = report.meta?.variants ?? [];
@@ -1267,27 +1242,18 @@ async function runSampleLevelDiff(reportId, store, flags) {
 // handleGold — gold dataset workflow (init / validate / compare)
 // ---------------------------------------------------------------------------
 async function handleGold(argv) {
+    const lang = langFromArgv(argv);
     const sub = argv[0];
     const rest = argv.slice(1);
     if (!sub || sub === '--help' || sub === '-h') {
-        console.log([
-            '',
-            'Usage: omk bench gold <subcommand>',
-            '',
-            'Subcommands:',
-            '  init [--out <dir>] [--annotator <id>]    生成空白 gold dataset 模板',
-            '  validate <dir>                           校验数据集结构',
-            '  compare <reportId> --gold-dir <dir>      与已有 report 计算 α/κ/Pearson',
-            '    [--variant <name>] [--reports-dir <d>]',
-            '    [--bootstrap-samples N] [--seed N]',
-            '',
-        ].join('\n'));
+        console.log(tCli('cli.help.gold', lang));
         process.exit(sub ? 0 : 1);
     }
     if (sub === 'init') {
         const { values } = parseArgs({
             args: rest,
             options: {
+                ...COMMON_OPTIONS,
                 out: { type: 'string', default: './gold-dataset' },
                 annotator: { type: 'string' },
             },
@@ -1298,10 +1264,12 @@ async function handleGold(argv) {
             const written = initGoldDataset(values.out, {
                 annotator: values.annotator,
             });
-            console.log(`Created ${written.length} files in ${values.out}:`);
+            console.log(tCli('cli.gold.created_files', lang, {
+                n: written.length, dir: values.out,
+            }));
             for (const p of written)
                 console.log(`  ${p}`);
-            console.log('\n下一步: 编辑 annotations.yaml 加入真实标注 → 跑 omk bench gold validate');
+            console.log(tCli('cli.gold.next_step_edit_annotations', lang));
         }
         catch (err) {
             console.error(err.message);
@@ -1312,13 +1280,13 @@ async function handleGold(argv) {
     if (sub === 'validate') {
         const dir = rest[0];
         if (!dir) {
-            console.error('Usage: omk bench gold validate <dir>');
+            console.error(tCli('cli.common.usage_gold_validate', lang));
             process.exit(1);
         }
         const { validateGoldDataset } = await import('./grading/gold-cli.js');
         const result = validateGoldDataset(dir);
         if (result.ok) {
-            console.log(`✓ gold dataset OK — ${result.sampleCount} 条标注`);
+            console.log(tCli('cli.gold.validate_ok', lang, { n: result.sampleCount }));
             return;
         }
         console.error(`✗ gold dataset has ${result.issues.length} issue(s):`);
@@ -1335,6 +1303,7 @@ async function handleGold(argv) {
         const { values } = parseArgs({
             args: rest.slice(1),
             options: {
+                ...COMMON_OPTIONS,
                 'gold-dir': { type: 'string' },
                 variant: { type: 'string' },
                 'reports-dir': { type: 'string', default: DEFAULT_REPORTS_DIR },
@@ -1366,7 +1335,7 @@ async function handleGold(argv) {
         const store = createFileStore(resolve(values['reports-dir']));
         const report = await store.get(reportId);
         if (!report) {
-            console.error(`Report not found: ${reportId}`);
+            console.error(tCli('cli.common.report_not_found', lang, { id: reportId }));
             process.exit(1);
         }
         const samples = Math.max(100, Number(values['bootstrap-samples']) || 1000);
@@ -1388,27 +1357,11 @@ async function handleGold(argv) {
 // handleDebiasValidate — measure length-debias prompt sensitivity (Phase 3a)
 // ---------------------------------------------------------------------------
 async function handleDebiasValidate(argv) {
+    const lang = langFromArgv(argv);
     const sub = argv[0];
     const rest = argv.slice(1);
     if (!sub || sub === '--help' || sub === '-h') {
-        console.log([
-            '',
-            'Usage: omk bench debias-validate <kind> <reportId> [options]',
-            '',
-            'Kinds:',
-            '  length    re-judge with the opposite length-debias setting and bootstrap CI',
-            '            on the score diff. Cost ~doubles vs the original judge pass.',
-            '',
-            'Options:',
-            '  --reports-dir <dir>          report store dir (default: ~/.oh-my-knowledge/reports)',
-            '  --samples <path>             override samples file (default: from report.meta.request)',
-            '  --variant <name>             which variant to validate (default: first)',
-            '  --judge-executor <name>      executor for judge calls (default: claude)',
-            '  --judge-model <model>        judge model id (default: from report)',
-            '  --bootstrap-samples N        bootstrap iterations (default 1000)',
-            '  --seed N                     deterministic CI seed',
-            '',
-        ].join('\n'));
+        console.log(tCli('cli.help.debias_validate', lang));
         process.exit(sub ? 0 : 1);
     }
     if (sub !== 'length') {
@@ -1423,6 +1376,7 @@ async function handleDebiasValidate(argv) {
     const { values } = parseArgs({
         args: rest.slice(1),
         options: {
+            ...COMMON_OPTIONS,
             'reports-dir': { type: 'string', default: DEFAULT_REPORTS_DIR },
             samples: { type: 'string' },
             variant: { type: 'string' },
@@ -1437,7 +1391,7 @@ async function handleDebiasValidate(argv) {
     const store = createFileStore(resolve(values['reports-dir']));
     const report = await store.get(reportId);
     if (!report) {
-        console.error(`Report not found: ${reportId}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: reportId }));
         process.exit(1);
     }
     // Resolve samples path: --samples overrides; otherwise read from report.meta.request.
@@ -1452,10 +1406,10 @@ async function handleDebiasValidate(argv) {
     const judgeModel = values['judge-model']
         ?? report.meta?.judgeModel;
     if (!judgeModel) {
-        console.error('No judge model. Pass --judge-model <id> or ensure report has meta.judgeModel.');
+        console.error(tCli('cli.common.no_judge_model', lang));
         process.exit(1);
     }
-    process.stderr.write('\n⚠ debias-validate 会重判所有 (sample × variant),judge cost 大致翻倍。\n');
+    process.stderr.write(tCli('cli.debias.warn_cost_doubles', lang));
     const { createExecutor } = await import('./executors/index.js');
     const judgeExecutor = createExecutor(values['judge-executor']);
     const { validateLengthDebias, formatDebiasValidate } = await import('./grading/debias-validate.js');
@@ -1479,29 +1433,16 @@ async function handleDebiasValidate(argv) {
 // handleSaturation — re-compute saturation verdict from a finished report
 // ---------------------------------------------------------------------------
 async function handleSaturation(argv) {
+    const lang = langFromArgv(argv);
     const reportId = argv[0];
     if (!reportId || reportId === '--help' || reportId === '-h') {
-        console.log([
-            '',
-            'Usage: omk bench saturation <reportId> [options]',
-            '',
-            '回答"我跑够样本了吗?"。复述已有 report 中持久化的饱和判定。',
-            '',
-            '注:本命令读取 run 时跑出的 verdict(运行时已用 method=bootstrap-ci-width',
-            '默认阈值 + 3 窗口 持续条件)。如要换 method/threshold 重新计算,需要重跑',
-            '`omk bench run --repeat ≥ 5`(运行时持久化的 trace 不含原始分数,无法',
-            '在事后用其他参数复算)。',
-            '',
-            'Options:',
-            '  --reports-dir <dir>   report store dir (default: ~/.oh-my-knowledge/reports)',
-            '  --variant <name>      只看一个 variant (default: all)',
-            '',
-        ].join('\n'));
+        console.log(tCli('cli.help.saturation', lang));
         process.exit(reportId ? 0 : 1);
     }
     const { values } = parseArgs({
         args: argv.slice(1),
         options: {
+            ...COMMON_OPTIONS,
             'reports-dir': { type: 'string', default: DEFAULT_REPORTS_DIR },
             variant: { type: 'string' },
         },
@@ -1511,12 +1452,12 @@ async function handleSaturation(argv) {
     const store = createFileStore(resolve(values['reports-dir']));
     const report = await store.get(reportId);
     if (!report) {
-        console.error(`Report not found: ${reportId}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: reportId }));
         process.exit(1);
     }
     const saturation = report.variance?.saturation;
     if (!saturation) {
-        console.error('该 report 无 saturation 数据 (需要 --repeat ≥ 2 才会记录)。');
+        console.error(tCli('cli.saturation.no_data', lang));
         process.exit(1);
     }
     // Print the persisted verdict from the original run. The trace stores
@@ -1527,22 +1468,32 @@ async function handleSaturation(argv) {
     // run time to enable post-hoc parameter sweeps.
     const variants = report.meta.variants ?? [];
     const targetVariants = values.variant ? [values.variant] : variants;
-    console.log(`\n  Saturation verdict (复述持久化结果)\n`);
+    console.log(tCli('cli.saturation.verdict_header', lang));
     for (const variant of targetVariants) {
         const trace = saturation.perVariant[variant];
         if (!trace || trace.length === 0) {
-            console.log(`  ${variant}: 无 trace 数据`);
+            console.log(tCli('cli.saturation.variant_no_trace', lang, { variant }));
             continue;
         }
-        console.log(`  ${variant}:`);
-        console.log(`    checkpoints: ${trace.length} (N=${trace.map((p) => p.n).join(', ')})`);
-        console.log(`    最近一点 mean=${trace[trace.length - 1].mean.toFixed(3)}, CI=[${trace[trace.length - 1].ciLow.toFixed(3)}, ${trace[trace.length - 1].ciHigh.toFixed(3)}]`);
+        console.log(tCli('cli.saturation.variant_label', lang, { variant }));
+        console.log(tCli('cli.saturation.checkpoints', lang, {
+            n: trace.length, list: trace.map((p) => p.n).join(', '),
+        }));
+        const last = trace[trace.length - 1];
+        console.log(tCli('cli.saturation.last_point', lang, {
+            mean: last.mean.toFixed(3), lo: last.ciLow.toFixed(3), hi: last.ciHigh.toFixed(3),
+        }));
         if (saturation.verdicts?.[variant]) {
             const v = saturation.verdicts[variant];
-            console.log(`    持久化判定 (${v.method}): ${v.saturated ? `已饱和@N=${v.atN}` : '未饱和'} - ${v.reason}`);
+            const result = v.saturated
+                ? tCli('cli.saturation.persisted_verdict_saturated', lang, { n: v.atN ?? '?' })
+                : tCli('cli.saturation.persisted_verdict_unsaturated', lang);
+            console.log(tCli('cli.saturation.persisted_verdict', lang, {
+                method: v.method, result, reason: v.reason,
+            }));
         }
         else if (trace.length < 5) {
-            console.log(`    判定: 数据点 ${trace.length} < 5,跳过 (跑 --repeat 5 以上才输出)`);
+            console.log(tCli('cli.saturation.skipped_too_few_points', lang, { n: trace.length }));
         }
     }
     console.log('');
@@ -1551,34 +1502,16 @@ async function handleSaturation(argv) {
 // handleVerdict — one-line ship/no-ship verdict (v0.22)
 // ---------------------------------------------------------------------------
 async function handleVerdict(argv) {
+    const lang = langFromArgv(argv);
     const reportId = argv[0];
     if (!reportId || reportId === '--help' || reportId === '-h') {
-        console.log([
-            '',
-            'Usage: omk bench verdict <reportId> [options]',
-            '',
-            '聚合 bootstrap CI / 三层 ci-gate / saturation / human α 给出一行结论。',
-            '',
-            'Verdict 等级:',
-            '  PROGRESS      显著改进 + 三层全过',
-            '  CAUTIOUS      改进真实但有警告 (gate 破 / 幅度太小 / 控制组本身崩)',
-            '  REGRESS       显著回退 — 不要 ship',
-            '  NOISE         CI 跨 0,无法判定',
-            '  UNDERPOWERED  样本不足,需要扩 N 重测',
-            '  SOLO          单变体报告,无对比对象',
-            '',
-            'Options:',
-            '  --reports-dir <dir>      report store dir (default: ~/.oh-my-knowledge/reports)',
-            '  --threshold <num>        三层 gate 阈值 (default 3.5,匹配 omk bench ci)',
-            '  --trivial-diff <num>     "幅度太小"阈值 (default 0.1)',
-            '  --verbose                展开 per-pair 详情',
-            '',
-        ].join('\n'));
+        console.log(tCli('cli.help.verdict', lang));
         process.exit(reportId ? 0 : 1);
     }
     const { values } = parseArgs({
         args: argv.slice(1),
         options: {
+            ...COMMON_OPTIONS,
             'reports-dir': { type: 'string', default: DEFAULT_REPORTS_DIR },
             threshold: { type: 'string' },
             'trivial-diff': { type: 'string' },
@@ -1590,12 +1523,12 @@ async function handleVerdict(argv) {
     const store = createFileStore(resolve(values['reports-dir']));
     const report = await store.get(reportId);
     if (!report) {
-        console.error(`Report not found: ${reportId}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: reportId }));
         process.exit(1);
     }
     const { computeVerdict, formatVerdictText } = await import('./eval-core/verdict.js');
     const result = computeVerdict(report, {
-        ciThreshold: values.threshold != null ? Number(values.threshold) : undefined,
+        gateThreshold: values.threshold != null ? Number(values.threshold) : undefined,
         triviallySmallDiff: values['trivial-diff'] != null ? Number(values['trivial-diff']) : undefined,
     });
     console.log(formatVerdictText(result, { verbose: Boolean(values.verbose) }));
@@ -1614,31 +1547,16 @@ async function handleVerdict(argv) {
 // handleDiagnose — per-sample quality diagnostics (v0.23 A)
 // ---------------------------------------------------------------------------
 async function handleDiagnose(argv) {
+    const lang = langFromArgv(argv);
     const reportId = argv[0];
     if (!reportId || reportId === '--help' || reportId === '-h') {
-        console.log([
-            '',
-            'Usage: omk bench diagnose <reportId> [options]',
-            '',
-            '诊断样本集本身的质量问题:区分度低 / 重复 / 歧义 / 成本异常 / 全 fail。',
-            '回答"测评结论是否被坏样本污染"——与 omk bench verdict 互补。',
-            '',
-            'Options:',
-            '  --reports-dir <dir>      report store dir',
-            '  --samples <path>         样本文件路径 (用于 near-duplicate 检测;默认从 report.meta.request 读)',
-            '  --top <n>                每类只显示前 N 个 (默认 10,0=全部)',
-            '  --duplicate-rouge <num>  near-duplicate ROUGE-1 阈值 (默认 0.7)',
-            '  --ambiguous-stddev <num> 歧义阈值,judge stddev (默认 1.0,需要 --judge-repeat ≥ 2 数据)',
-            '  --cost-k <num>           成本异常倍数 vs median (默认 3)',
-            '  --latency-k <num>        耗时异常倍数 vs median (默认 3)',
-            '  --flat <num>             flat_scores 分差阈值 (默认 0.5)',
-            '',
-        ].join('\n'));
+        console.log(tCli('cli.help.diagnose', lang));
         process.exit(reportId ? 0 : 1);
     }
     const { values } = parseArgs({
         args: argv.slice(1),
         options: {
+            ...COMMON_OPTIONS,
             'reports-dir': { type: 'string', default: DEFAULT_REPORTS_DIR },
             samples: { type: 'string' },
             top: { type: 'string', default: '10' },
@@ -1654,7 +1572,7 @@ async function handleDiagnose(argv) {
     const store = createFileStore(resolve(values['reports-dir']));
     const report = await store.get(reportId);
     if (!report) {
-        console.error(`Report not found: ${reportId}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: reportId }));
         process.exit(1);
     }
     // Try to read the samples file for near-duplicate detection. Source order:
@@ -1669,7 +1587,9 @@ async function handleDiagnose(argv) {
             samples = loadSamples(samplesPath).samples;
         }
         catch (err) {
-            process.stderr.write(`warn: 加载 samples 文件失败 (${samplesPath}): ${err.message}\n`);
+            process.stderr.write(tCli('cli.common.warn_load_samples_failed', lang, {
+                path: samplesPath, message: err.message,
+            }));
         }
     }
     const topRaw = Number(values.top);
@@ -1694,29 +1614,16 @@ async function handleDiagnose(argv) {
 // handleFailures — LLM-driven failure clustering (v0.23 B)
 // ---------------------------------------------------------------------------
 async function handleFailures(argv) {
+    const lang = langFromArgv(argv);
     const reportId = argv[0];
     if (!reportId || reportId === '--help' || reportId === '-h') {
-        console.log([
-            '',
-            'Usage: omk bench failures <reportId> [options]',
-            '',
-            '把已有 report 的失败样本喂给一次 LLM 调用,自动聚类 + 给修复建议。',
-            '失败定义:compositeScore < threshold 或 ok=false。',
-            '',
-            'Options:',
-            '  --reports-dir <dir>      report store dir',
-            '  --judge-executor <name>  执行器 (default: claude)',
-            '  --judge-model <id>       聚类用的 model (default: 沿用 report.meta.judgeModel)',
-            '  --max-clusters <n>       最多多少 cluster (default 5)',
-            '  --threshold <num>        compositeScore < threshold 算失败 (default 3)',
-            '  --max-feed <n>           最多喂给 LLM 多少条 (default 50,超出取最差)',
-            '',
-        ].join('\n'));
+        console.log(tCli('cli.help.failures', lang));
         process.exit(reportId ? 0 : 1);
     }
     const { values } = parseArgs({
         args: argv.slice(1),
         options: {
+            ...COMMON_OPTIONS,
             'reports-dir': { type: 'string', default: DEFAULT_REPORTS_DIR },
             'judge-executor': { type: 'string', default: 'claude' },
             'judge-model': { type: 'string' },
@@ -1730,12 +1637,12 @@ async function handleFailures(argv) {
     const store = createFileStore(resolve(values['reports-dir']));
     const report = await store.get(reportId);
     if (!report) {
-        console.error(`Report not found: ${reportId}`);
+        console.error(tCli('cli.common.report_not_found', lang, { id: reportId }));
         process.exit(1);
     }
     const judgeModel = values['judge-model'] ?? report.meta?.judgeModel;
     if (!judgeModel) {
-        console.error('No judge model. Pass --judge-model <id> or ensure report has meta.judgeModel.');
+        console.error(tCli('cli.common.no_judge_model', lang));
         process.exit(1);
     }
     const { createExecutor } = await import('./executors/index.js');