oh-my-knowledge 0.33.0 → 0.35.0

package/dist/eval-core/execution-strategy.js

@@ -53,7 +53,7 @@ export function buildVariantConfig(artifact) {
     // 因此缺失时直接 throw,强迫调用者修配置,不走默认值掩盖。
     if (!artifact.experimentRole) {
         throw new Error(`artifact "${artifact.name}" 缺少 experimentRole:应该由 CLI --control/--treatment、`
-            + `--config eval.yaml 或 batch workflow 显式注入。参见 docs/specs/terminology-spec.md 三-4。`);
+            + `--config eval.yaml 或 batch workflow 显式注入。参见 docs/zh/specs/terminology-spec.md 三-4。`);
     }
     return {
         variant: artifact.name,
@@ -72,7 +72,13 @@ export function buildVariantConfig(artifact) {
     };
 }
 function extractSkillDir(artifact) {
-    if (!artifact.locator)
+    // dir-skill(任意源)优先用隔离副本执行根 execRoot:agent cwd 锚副本、skillDir 也指副本。副本经
+    // distributable filter 物化、不含 node_modules → buildExecEnv 的 .bin 检查恒不命中 → PATH 不被污染。
+    if (artifact.execRoot)
+        return artifact.execRoot;
+    // git 文件-skill 无 execRoot 且 locator 是 git spec(`review` / `skills/review`)非磁盘目录,取 dirname
+    // 会得伪 skillDir 污染 PATH/runtime fingerprint → 与 baseline 一样返回 null。本地文件-skill 取 .md 所在目录。
+    if (!artifact.locator || artifact.source === 'git')
         return null;
     return dirname(artifact.locator);
 }
@@ -162,4 +168,3 @@ export function resolveExecutionStrategy(task, model, timeoutMs, verbose, effort
             };
     }
 }
-//# sourceMappingURL=execution-strategy.js.map

package/dist/eval-core/fact-checker.d.ts

@@ -22,4 +22,3 @@ export declare function extractPathClaims(output: string): string[];
  * Check facts in agent output by verifying file paths exist in cwd.
  */
 export declare function checkFacts(output: string, cwd: string): FactCheckResult;
-//# sourceMappingURL=fact-checker.d.ts.map

package/dist/eval-core/fact-checker.js

@@ -63,4 +63,3 @@ export function checkFacts(output, cwd) {
     const verifiedRate = totalCount > 0 ? Number((verifiedCount / totalCount).toFixed(2)) : 1;
     return { claims, verifiedCount, totalCount, verifiedRate };
 }
-//# sourceMappingURL=fact-checker.js.map

package/dist/eval-core/layer-gates.d.ts

@@ -14,4 +14,3 @@ export interface LayerGateResult {
  * 抽为纯函数便于单测；commands/eval-runner.ts 负责 IO（读 config、打印、退出码）。
  */
 export declare function evaluateLayerGates(summary: Record<string, VariantSummary>, threshold: number): LayerGateResult;
-//# sourceMappingURL=layer-gates.d.ts.map

package/dist/eval-core/layer-gates.js

@@ -39,4 +39,3 @@ export function evaluateLayerGates(summary, threshold) {
     }
     return { allPass, lines };
 }
-//# sourceMappingURL=layer-gates.js.map

package/dist/eval-core/mocks-runtime.d.ts

@@ -97,4 +97,3 @@ declare function readMockHookTemplate(): string;
 export declare const _readMockHookTemplateForTest: typeof readMockHookTemplate;
 export declare function _testMakeTempConfigDir(): string;
 export {};
-//# sourceMappingURL=mocks-runtime.d.ts.map

package/dist/eval-core/mocks-runtime.js

@@ -556,4 +556,3 @@ export function _testMakeTempConfigDir() {
     mkdirSync(dir, { recursive: true });
     return dir;
 }
-//# sourceMappingURL=mocks-runtime.js.map

package/dist/eval-core/schema.d.ts

@@ -27,4 +27,3 @@ export declare function buildVariantResult(execResult: ExecResult, gradeResult:
  */
 export declare function buildVariantSummary(entries: VariantResult[]): VariantSummary;
 export {};
-//# sourceMappingURL=schema.d.ts.map

package/dist/eval-core/schema.js

@@ -306,4 +306,3 @@ function buildEnsembleAggregate(ok) {
         judgeAgreement: { ...agreement, sampleCount: alignedMatrix[0].length },
     };
 }
-//# sourceMappingURL=schema.js.map

package/dist/eval-core/statistics.d.ts

@@ -83,4 +83,3 @@ export declare function tTest(a: number[], b: number[]): {
     df: number;
     significant: boolean;
 };
-//# sourceMappingURL=statistics.d.ts.map

package/dist/eval-core/statistics.js

@@ -207,4 +207,3 @@ export function tTest(a, b) {
         significant,
     };
 }
-//# sourceMappingURL=statistics.js.map

package/dist/eval-core/task-planner.d.ts

@@ -9,4 +9,3 @@ import type { Artifact, Sample, SampleEnvironment, Task } from '../types/index.j
 export declare function renderEnvironmentSection(env: SampleEnvironment | undefined): string | null;
 export declare function buildTasks(samples: Sample[], variants: string[], skills: Record<string, string | null>): Task[];
 export declare function buildTasksFromArtifacts(samples: Sample[], artifacts: Artifact[]): Task[];
-//# sourceMappingURL=task-planner.d.ts.map

package/dist/eval-core/task-planner.js

@@ -62,11 +62,11 @@ export function buildTasksFromArtifacts(samples, artifacts) {
                 assertions: sample.assertions || null,
                 dimensions: sample.dimensions || null,
                 artifactContent: artifact.content,
-                cwd: artifact.cwd || artifact.skillRoot || sample.cwd || null,
+                // 用户显式 cwd > execRoot(隔离副本,dir-skill 执行根) > skillRoot(真源,兜底) > sample.cwd
+                cwd: artifact.cwd || artifact.execRoot || artifact.skillRoot || sample.cwd || null,
                 _sample: sample,
             });
         }
     }
     return tasks;
 }
-//# sourceMappingURL=task-planner.js.map

package/dist/eval-core/verdict.d.ts

@@ -92,4 +92,3 @@ export declare function computeVerdict(report: Report, options?: VerdictOptions)
 export declare function formatVerdictText(result: VerdictResult, options?: {
     verbose?: boolean;
 }): string;
-//# sourceMappingURL=verdict.d.ts.map

package/dist/eval-core/verdict.js

@@ -369,4 +369,3 @@ export function formatVerdictText(result, options = {}) {
     }
     return lines.join('\n');
 }
-//# sourceMappingURL=verdict.js.map

package/dist/eval-workflows/batch-evaluation-workflow.d.ts

@@ -1,8 +1,12 @@
-import type { Artifact, BatchEvaluationReport, EvaluationReport, JobStore, ProgressCallback } from '../types/index.js';
+import type { BatchEvaluationReport, EvaluationReport, JobStore, ProgressCallback, VariantSpec } from '../types/index.js';
 interface RunSingleEvaluationOptions {
     samplesPath: string;
     skillDir: string;
-    artifacts: Artifact[];
+    /** 每个 batch entry 的实验结构(baseline control vs 当前 skill treatment),由
+     *  runEvaluation 的 spec-based 解析统一绑定 role / 隔离。 */
+    variantSpecs: VariantSpec[];
+    /** strict-baseline default,透传给 per-skill runEvaluation(baseline → allowedSkills=[])。 */
+    strictBaseline?: boolean;
     model: string;
     judgeModel: string;
     outputDir: string | null;
@@ -42,6 +46,18 @@ interface CompletedBatchSkillRun {
     report: EvaluationReport;
     filePath: string | null;
 }
+/** Batch 每个 entry 的实验结构固定:baseline control vs 当前 skill treatment。
+ *  构造 VariantSpec 交给 runEvaluation 的 spec-based 解析按身份绑定 role / 隔离——与单跑
+ *  路径同一处绑定逻辑,batch 不再自管 artifact 拼装(此前两处手抄 resolveArtifacts().map()
+ *  导致 #183 角色误绑各存一份)。
+ *  两个 variant 的 allowedSkills 都从 eval.yaml variants[].allowedSkills 取:treatment 按 skill
+ *  名 entry.name 查、baseline 按保留名 `baseline` 查,挂到对应 spec 上由 prepareEvaluationRun
+ *  统一绑定。baseline 的显式声明(白名单或 `[]`)必须保留——eval.yaml variant.allowedSkills
+ *  优先于 strictBaseline 默认,漏挂会让 `--batch --config` 的 baseline 隔离配置静默失效。 */
+export declare function buildBatchVariantSpecs(entry: {
+    name: string;
+    skillPath: string;
+}, variantAllowedSkills?: Record<string, string[]>): VariantSpec[];
 export declare function buildBatchEvaluationReport({ batchRunId, skillDir, skillEntries, skillResults, model, judgeModel, noJudge, executorName, judgeExecutorName, project, owner, tags, concurrency, timeoutMs, totalCostUSD, repeat, judgeModels, noCache, bootstrap, bootstrapSamples, lengthDebias, budget, }: {
     batchRunId: string;
     skillDir: string;
@@ -127,4 +143,3 @@ export declare function executeBatchEvaluationRuns({ skillDir, skillEntries, mod
     filePath: string | null;
 }>;
 export {};
-//# sourceMappingURL=batch-evaluation-workflow.d.ts.map

package/dist/eval-workflows/batch-evaluation-workflow.js

@@ -1,9 +1,34 @@
-import { dirname, resolve } from 'node:path';
+import { dirname } from 'node:path';
 import { DEFAULT_OUTPUT_DIR, generateRunId, getCliVersion, getGitInfo, persistReport } from '../eval-core/evaluation-reporting.js';
 import { buildEvaluationRequest, createEvaluationRun, createSucceededJob, finalizeEvaluationRun } from '../eval-core/evaluation-job.js';
 import { getExecutorRuntimeFingerprint } from '../executors/runtime-fingerprint.js';
 import { createFileJobStore, DEFAULT_JOBS_DIR } from '../server/job-store.js';
-import { resolveArtifacts } from '../inputs/skill-loader.js';
+/** Batch 每个 entry 的实验结构固定:baseline control vs 当前 skill treatment。
+ *  构造 VariantSpec 交给 runEvaluation 的 spec-based 解析按身份绑定 role / 隔离——与单跑
+ *  路径同一处绑定逻辑,batch 不再自管 artifact 拼装(此前两处手抄 resolveArtifacts().map()
+ *  导致 #183 角色误绑各存一份)。
+ *  两个 variant 的 allowedSkills 都从 eval.yaml variants[].allowedSkills 取:treatment 按 skill
+ *  名 entry.name 查、baseline 按保留名 `baseline` 查,挂到对应 spec 上由 prepareEvaluationRun
+ *  统一绑定。baseline 的显式声明(白名单或 `[]`)必须保留——eval.yaml variant.allowedSkills
+ *  优先于 strictBaseline 默认,漏挂会让 `--batch --config` 的 baseline 隔离配置静默失效。 */
+export function buildBatchVariantSpecs(entry, variantAllowedSkills) {
+    const baselineAllowed = variantAllowedSkills?.baseline;
+    const treatmentAllowed = variantAllowedSkills?.[entry.name];
+    return [
+        {
+            name: 'baseline',
+            role: 'control',
+            expr: 'baseline',
+            ...(baselineAllowed !== undefined && { allowedSkills: baselineAllowed }),
+        },
+        {
+            name: entry.name,
+            role: 'treatment',
+            expr: entry.skillPath,
+            ...(treatmentAllowed !== undefined && { allowedSkills: treatmentAllowed }),
+        },
+    ];
+}
 function commonRuntime(runtimes) {
     const values = Object.values(runtimes);
     if (values.length === 0)
@@ -110,7 +135,7 @@ export function buildBatchEvaluationReport({ batchRunId, skillDir, skillEntries,
         ...(noJudge ? {} : { runtime: getExecutorRuntimeFingerprint(jc.executor, jc.model, { skillDir }) }),
     }));
     const report = {
-        kind: 'batch-evaluation',
+        reportKind: 'batch-evaluation',
         id: batchRunId,
         mode: 'skill',
         meta: {
@@ -146,21 +171,13 @@ export async function executeBatchEvaluationRuns({ skillDir, skillEntries, model
     for (let i = 0; i < skillEntries.length; i++) {
         const entry = skillEntries[i];
         onSkillProgress?.({ phase: 'start', skill: entry.name, current: i + 1, total: skillEntries.length });
-        // Batch mode 的实验结构固定为 baseline control vs 当前 skill treatment。
-        const perSkillAllowedSkills = variantAllowedSkills?.[entry.name] !== undefined
-            ? { ...variantAllowedSkills, [entry.skillPath]: variantAllowedSkills[entry.name] }
-            : variantAllowedSkills;
-        const skillArtifacts = resolveArtifacts(resolve(skillDir), ['baseline', entry.skillPath], { strictBaseline, variantAllowedSkills: perSkillAllowedSkills }).map((artifact) => {
-            if (artifact.name === entry.skillPath) {
-                return { ...artifact, name: entry.name, experimentRole: 'treatment' };
-            }
-            return { ...artifact, experimentRole: 'control' };
-        });
+        const variantSpecs = buildBatchVariantSpecs(entry, variantAllowedSkills);
         const childRunId = `${batchRunId}-${String(i + 1).padStart(2, '0')}-${safeRunIdPart(entry.name)}`;
         const { report, filePath } = await runSingleEvaluation({
             samplesPath: entry.samplesPath,
             skillDir,
-            artifacts: skillArtifacts,
+            variantSpecs,
+            strictBaseline,
             model,
             judgeModel,
             outputDir,
@@ -226,4 +243,3 @@ export async function executeBatchEvaluationRuns({ skillDir, skillEntries, model
         await resolvedJobStore.save(job.jobId, job);
     return { report: batchReport, filePath };
 }
-//# sourceMappingURL=batch-evaluation-workflow.js.map

package/dist/eval-workflows/evaluation-pipeline/preflight-warnings.d.ts

@@ -19,4 +19,3 @@ export declare function buildPowerWarnings(sampleCount: number, repeat: number,
 export declare function emitPowerWarnings(sampleCount: number, repeat: number, lang: Lang): void;
 export declare function buildIsolationWarnings(artifacts: Artifact[], strictBaseline: boolean | undefined): string[];
 export declare function emitIsolationWarnings(artifacts: Artifact[], strictBaseline: boolean | undefined): void;
-//# sourceMappingURL=preflight-warnings.d.ts.map

package/dist/eval-workflows/evaluation-pipeline/preflight-warnings.js

@@ -64,4 +64,3 @@ export function emitIsolationWarnings(artifacts, strictBaseline) {
         process.stderr.write(`${w}\n`);
     }
 }
-//# sourceMappingURL=preflight-warnings.js.map

package/dist/eval-workflows/evaluation-pipeline/report-finalize.d.ts

@@ -29,4 +29,3 @@ export declare function finalizeEvaluationReport({ report, results, artifacts, v
     samples: Sample[];
 }): Report;
 export {};
-//# sourceMappingURL=report-finalize.d.ts.map

package/dist/eval-workflows/evaluation-pipeline/report-finalize.js

@@ -54,4 +54,3 @@ export function finalizeEvaluationReport({ report, results, artifacts, variantNa
     }
     return report;
 }
-//# sourceMappingURL=report-finalize.js.map

package/dist/eval-workflows/evaluation-pipeline/run-state.d.ts

@@ -57,4 +57,3 @@ export declare function finalizeSuccessfulRun(state: EvaluationRunState): {
 };
 export declare function persistSuccessfulJob(state: EvaluationRunState, job: EvaluationJob): Promise<void>;
 export declare function persistFailedJob(state: EvaluationRunState, err: unknown): Promise<void>;
-//# sourceMappingURL=run-state.d.ts.map

package/dist/eval-workflows/evaluation-pipeline/run-state.js

@@ -85,4 +85,3 @@ export async function persistFailedJob(state, err) {
         await state.resolvedJobStore.save(state.jobId, failedJob);
     }
 }
-//# sourceMappingURL=run-state.js.map

package/dist/eval-workflows/evaluation-pipeline/test-set-hash.d.ts

@@ -18,4 +18,3 @@
  */
 export declare function computeTestSetHash(samplesPath: string, sourceFiles?: string[]): string | null;
 export declare function _computeTestSetHashForTest(samplesPath: string, sourceFiles?: string[]): string | null;
-//# sourceMappingURL=test-set-hash.d.ts.map

package/dist/eval-workflows/evaluation-pipeline/test-set-hash.js

@@ -43,4 +43,3 @@ export function computeTestSetHash(samplesPath, sourceFiles) {
 export function _computeTestSetHashForTest(samplesPath, sourceFiles) {
     return computeTestSetHash(samplesPath, sourceFiles);
 }
-//# sourceMappingURL=test-set-hash.js.map

package/dist/eval-workflows/evaluation-pipeline.d.ts

@@ -94,4 +94,3 @@ export declare function executeEvaluationPipeline({ samplesPath, samplesBaseDir,
     report: Report;
     filePath: string | null;
 }>;
-//# sourceMappingURL=evaluation-pipeline.d.ts.map

package/dist/eval-workflows/evaluation-pipeline.js

@@ -176,4 +176,3 @@ requires: _requires, layeredStats = false, repeat, batch, judgeRepeat, judgeMode
         await stopAllServers();
     }
 }
-//# sourceMappingURL=evaluation-pipeline.js.map

package/dist/eval-workflows/evaluation-preparation.d.ts

@@ -12,17 +12,14 @@ export interface PreparedEvaluationRun {
     /** Sample bundle 内参与合并的所有源文件绝对路径(已排序)。test set hash 用它遍历。 */
     samplesSourceFiles: string[];
 }
-export declare function prepareEvaluationRun({ samplesPath, skillDir, variantSpecs, artifacts, dryRun, mcpConfig, strictBaseline, variantAllowedSkills, }: {
+export declare function prepareEvaluationRun({ samplesPath, skillDir, variantSpecs, dryRun, mcpConfig, strictBaseline, }: {
     samplesPath: string;
     skillDir: string;
     variantSpecs: VariantSpec[];
-    artifacts?: Artifact[];
     dryRun: boolean;
     mcpConfig?: string;
     /**  — default true, baseline-kind 自动 allowedSkills=[]。 */
     strictBaseline?: boolean;
-    /**  — eval.yaml 显式 allowedSkills per variant (overrides strictBaseline default). */
-    variantAllowedSkills?: Record<string, string[]>;
 }): Promise<PreparedEvaluationRun>;
 export declare function buildDryRunTaskReport({ model, judgeModels, executorName, samplesPath, skillDir, tasks, variantNames, }: {
     model: string;
@@ -57,4 +54,3 @@ export declare function buildDryRunTaskReport({ model, judgeModels, executorName
         hasSystem: boolean;
     }[];
 };
-//# sourceMappingURL=evaluation-preparation.d.ts.map

package/dist/eval-workflows/evaluation-preparation.js

@@ -5,50 +5,29 @@ import { resolveArtifacts } from '../inputs/skill-loader.js';
 import { buildVariantConfig } from '../eval-core/execution-strategy.js';
 import { loadMcpConfig, resolveMcpUrls } from '../inputs/mcp-resolver.js';
 import { resolveUrls } from '../inputs/url-fetcher.js';
-export async function prepareEvaluationRun({ samplesPath, skillDir, variantSpecs, artifacts, dryRun, mcpConfig, strictBaseline, variantAllowedSkills, }) {
+export async function prepareEvaluationRun({ samplesPath, skillDir, variantSpecs, dryRun, mcpConfig, strictBaseline, }) {
     const { samples, requires, baseDir: samplesBaseDir, sourceFiles: samplesSourceFiles } = loadSamples(samplesPath);
-    // Build expressions from specs (preserving order) and resolve to artifacts.
-    const variantExpressions = variantSpecs.map((spec) => spec.expr);
-    // 不把 variantAllowedSkills 透进 resolveArtifacts:那里按解析出的 artifact 名查,
-    // 与 eval.yaml 的自定义 variant 名不一致时会漏绑。allowedSkills 统一在下面按 spec 绑定。
-    // resolveArtifacts 只保留 strictBaseline 默认(baseline → [])。
-    const resolvedArtifacts = artifacts || resolveArtifacts(resolve(skillDir), variantExpressions, { strictBaseline });
-    // Attach experimentRole to each artifact.
-    //   - 当 artifacts 由本函数从 variantSpecs.map(spec => spec.expr) 解析时,两者顺序一一对应,
-    //     按 index 绑定 role,并把 spec.name 同步成消歧后的最终唯一名。不能按 spec.name 匹配
-    //     artifact.name:同 basename 的 variant 被 ensureUniqueVariantNames 消歧后(v1/greeter /
-    //     v2/greeter)就和 spec 的派生短名(greeter)对不上,会丢 role。
-    //   - 当 artifacts 由外部传入(如 batch workflow,role 已预置)时,index 无法对齐,
-    //     退回按 name 匹配,且 if-already-set 守卫会保留预置 role。
-    if (!artifacts && variantSpecs.length === resolvedArtifacts.length) {
-        variantSpecs.forEach((spec, i) => {
-            const artifact = resolvedArtifacts[i];
-            if (!artifact.experimentRole)
-                artifact.experimentRole = spec.role;
-            // allowedSkills 按 spec 身份绑定(eval 主路径的唯一隔离绑定点,显式声明优先于
-            // strictBaseline 默认)。首选 spec.allowedSkills(eval.yaml 经 configVariantsToSpecs 带上);
-            // 旧的按变量名 map 仅作 legacy fallback,兜住「CLI roles 覆盖 config 时 spec 不带
-            // allowedSkills」等边角,行为与重构前一致。
-            // TODO: 待所有路径都按 spec 携带 allowedSkills 后,移除 variantAllowedSkills 这条 fallback,
-            //       与 resolveArtifacts 里的按名绑定一并收成单一来源。
-            const explicit = spec.allowedSkills ?? variantAllowedSkills?.[spec.name];
-            if (explicit !== undefined)
-                artifact.allowedSkills = explicit;
-            spec.name = artifact.name; // 同步唯一名,让 spec 与 report / variantNames 一致(放最后,前面还要用旧名查)
-        });
-    }
-    else {
-        const roleByName = {};
-        for (const spec of variantSpecs)
-            roleByName[spec.name] = spec.role;
-        for (const artifact of resolvedArtifacts) {
-            if (artifact.experimentRole)
-                continue; // preserve if already set (e.g. batch workflow)
-            const role = roleByName[artifact.name];
-            if (role)
-                artifact.experimentRole = role;
-        }
-    }
+    // 结构化传入 {expr, cwd}(与 variantSpecs 顺序一一对应),cwd 不再编码进 expr 字符串。
+    // resolveArtifacts 只保留 strictBaseline 默认(baseline → []);per-variant 隔离声明走
+    // spec.allowedSkills,在下面按 spec 身份绑定。
+    const variantInputs = variantSpecs.map((spec) => spec.git
+        ? { git: spec.git, cwd: spec.cwd, name: spec.name }
+        : { expr: spec.expr, cwd: spec.cwd });
+    const resolvedArtifacts = resolveArtifacts(resolve(skillDir), variantInputs, { strictBaseline });
+    // experimentRole / allowedSkills 按 spec 身份绑定。variantSpecs 与 resolvedArtifacts 顺序
+    // 一一对应(resolveArtifacts 每个 input 产出一个 artifact),按 index 绑定,并把 spec.name
+    // 同步成消歧后的最终唯一名。不能按 spec.name 匹配 artifact.name:同 basename 的 variant 被
+    // ensureUniqueVariantNames 消歧后(v1/greeter / v2/greeter)就和 spec 的派生短名(greeter)
+    // 对不上,会丢 role。
+    variantSpecs.forEach((spec, i) => {
+        const artifact = resolvedArtifacts[i];
+        artifact.experimentRole = spec.role;
+        // spec.allowedSkills 是隔离声明的单一来源(eval.yaml 经 configVariantsToSpecs、batch 经
+        // buildBatchVariantSpecs 都挂到 spec 上),显式声明优先于 strictBaseline 默认。
+        if (spec.allowedSkills !== undefined)
+            artifact.allowedSkills = spec.allowedSkills;
+        spec.name = artifact.name; // 同步唯一名,让 spec 与 report / variantNames 一致(放最后,前面还要用旧名查)
+    });
     if (!dryRun) {
         const mcpServers = loadMcpConfig(mcpConfig);
         const mcpResolved = mcpServers ? await resolveMcpUrls(samples, mcpServers) : new Set();
@@ -102,4 +81,3 @@ export function buildDryRunTaskReport({ model, judgeModels, executorName, sample
         }),
     };
 }
-//# sourceMappingURL=evaluation-preparation.js.map

package/dist/eval-workflows/messages.d.ts

@@ -1,4 +1,3 @@
 import type { Lang } from '../types/shared.js';
 export type EvalWorkflowMessageCode = 'power_warning_tiny_n' | 'power_warning_small_n' | 'power_warning_repeat_one' | 'doctor_gate_blocked';
 export declare function tEvalWorkflowMessage(code: EvalWorkflowMessageCode, lang?: Lang, params?: Record<string, string | number>): string;
-//# sourceMappingURL=messages.d.ts.map

package/dist/eval-workflows/messages.js

@@ -27,4 +27,3 @@ export function tEvalWorkflowMessage(code, lang = DEFAULT_LANG, params) {
     }
     return text;
 }
-//# sourceMappingURL=messages.js.map

package/dist/eval-workflows/run-evaluation.d.ts

@@ -63,7 +63,6 @@ export interface RunEvaluationOptions extends CommonEvaluationOptions {
     samplesPath: string;
     skillDir: string;
     variantSpecs?: VariantSpec[];
-    artifacts?: Artifact[];
     dryRun?: boolean;
     blind?: boolean;
     retry?: number;
@@ -71,12 +70,9 @@ export interface RunEvaluationOptions extends CommonEvaluationOptions {
     /** Explicit persisted run id. Used by batch workflows that need stable child ids. */
     runId?: string;
     /** strict-baseline default (CLI `--strict-baseline` default true).
-     *  When undefined, treated as true. baseline-kind variants get allowedSkills=[]
-     *  unless eval.yaml explicitly overrides per-variant. */
+     *  When undefined, treated as true. baseline-kind variants get allowedSkills=[].
+     *  per-variant 显式隔离声明走 variantSpecs[].allowedSkills(eval.yaml 经 configVariantsToSpecs)。 */
     strictBaseline?: boolean;
-    /** per-variant explicit allowedSkills override map (from eval.yaml).
-     *  Keyed by variant name. Always wins over strictBaseline default. */
-    variantAllowedSkills?: Record<string, string[]>;
 }
 export interface RunBatchEvaluationOptions extends CommonEvaluationOptions {
     skillDir: string;
@@ -122,7 +118,7 @@ export interface DryRunReport extends DryRunBase {
     samplesPath: string;
     tasks: DryRunTask[];
 }
-export declare function runEvaluation({ samplesPath, skillDir, variantSpecs, artifacts, model, outputDir, project, owner, tags, noJudge, dryRun, blind, concurrency, timeoutMs, noCache, executorName, jobStore, persistJob, onProgress, skipConnectivity, skipDoctor, lang, mcpConfig, verbose, retry, resume, runId, layeredStats, repeat, batch, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, strictBaseline, variantAllowedSkills, effort, noDiagnostic, }: RunEvaluationOptions): Promise<{
+export declare function runEvaluation({ samplesPath, skillDir, variantSpecs, model, outputDir, project, owner, tags, noJudge, dryRun, blind, concurrency, timeoutMs, noCache, executorName, jobStore, persistJob, onProgress, skipConnectivity, skipDoctor, lang, mcpConfig, verbose, retry, resume, runId, layeredStats, repeat, batch, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, strictBaseline, effort, noDiagnostic, }: RunEvaluationOptions): Promise<{
     report: Report | DryRunReport;
     filePath: string | null;
 }>;
@@ -148,4 +144,3 @@ export declare function runMultiple({ repeat, onRepeatProgress, ...config }: Run
     filePath: string | null;
 }>;
 export {};
-//# sourceMappingURL=run-evaluation.d.ts.map

package/dist/eval-workflows/run-evaluation.js

@@ -1,14 +1,14 @@
 import { resolve } from 'node:path';
 import { DEFAULT_OUTPUT_DIR, persistReport } from '../eval-core/evaluation-reporting.js';
 import { createExecutor, DEFAULT_MODEL, JUDGE_MODEL } from '../executors/index.js';
-import { discoverBatchSkills, resolveArtifacts } from '../inputs/skill-loader.js';
+import { discoverBatchSkills } from '../inputs/skill-loader.js';
 import { confidenceInterval, tTest, effectSize } from '../eval-core/statistics.js';
-import { executeBatchEvaluationRuns } from './batch-evaluation-workflow.js';
+import { executeBatchEvaluationRuns, buildBatchVariantSpecs } from './batch-evaluation-workflow.js';
 import { buildDryRunTaskReport, prepareEvaluationRun, } from './evaluation-preparation.js';
 import { executeEvaluationPipeline } from './evaluation-pipeline.js';
 import { findSaturationPoint } from '../analysis/saturation.js';
 import { bootstrapMeanCI, DEFAULT_BOOTSTRAP_ALPHA, DEFAULT_BOOTSTRAP_SAMPLES } from '../eval-core/bootstrap.js';
-export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [], artifacts, model = DEFAULT_MODEL, outputDir = DEFAULT_OUTPUT_DIR, project, owner, tags, noJudge = false, dryRun = false, blind = false, concurrency = 1, timeoutMs, noCache = false, executorName = 'claude', jobStore = null, persistJob = true, onProgress = null, skipConnectivity = false, skipDoctor = false, lang = 'zh', mcpConfig, verbose = false, retry = 0, resume, runId, layeredStats = false, repeat, batch, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, strictBaseline, variantAllowedSkills, effort, noDiagnostic, }) {
+export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [], model = DEFAULT_MODEL, outputDir = DEFAULT_OUTPUT_DIR, project, owner, tags, noJudge = false, dryRun = false, blind = false, concurrency = 1, timeoutMs, noCache = false, executorName = 'claude', jobStore = null, persistJob = true, onProgress = null, skipConnectivity = false, skipDoctor = false, lang = 'zh', mcpConfig, verbose = false, retry = 0, resume, runId, layeredStats = false, repeat, batch, judgeRepeat, judgeModels, bootstrap, bootstrapSamples, lengthDebias, budget, strictBaseline, effort, noDiagnostic, }) {
     // Unified judgeModels → derive single-judge fields for downstream pipeline / grading
     // (which still operate on string `judgeModel` + `judgeExecutorName` fields per call).
     const effectiveJudgeModels = judgeModels && judgeModels.length > 0
@@ -20,11 +20,9 @@ export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [],
         samplesPath,
         skillDir,
         variantSpecs,
-        artifacts,
         dryRun,
         mcpConfig,
         strictBaseline,
-        variantAllowedSkills,
     });
     // doctor 强制门禁: skill 静态结构 + 元数据 + 依赖 + 用例契约。
     // 在 dryRun 分支之前跑, 让 dry-run 也得到 doctor 覆盖(保护 garbage-in 的 verdict)。
@@ -99,7 +97,7 @@ export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [],
         const { createFileStore } = await import('../server/report-store.js');
         const store = createFileStore(resolve(outputDir || DEFAULT_OUTPUT_DIR));
         const existing = await store.get(resume);
-        if (existing?.kind === 'evaluation') {
+        if (existing?.reportKind === 'evaluation') {
             existingResults = {};
             for (const entry of existing.results || []) {
                 existingResults[entry.sample_id] = entry.variants;
@@ -123,7 +121,7 @@ export async function runEvaluation({ samplesPath, skillDir, variantSpecs = [],
                     + `   新 entries 不隔离 → 与现有 entries 不可比。建议恢复默认 strict-baseline。\n`);
             }
         }
-        else if (existing?.kind === 'batch-evaluation') {
+        else if (existing?.reportKind === 'batch-evaluation') {
             process.stderr.write(`\n⚠️  report ${resume} is a BatchEvaluationReport; resume needs a child EvaluationReport, starting from scratch\n`);
         }
         else {
@@ -381,23 +379,17 @@ export async function runBatchEvaluation({ skillDir, model = DEFAULT_MODEL, outp
     if (dryRun) {
         // batch dry-run 走 per-entry runEvaluation(dryRun: true) — doctor 嵌入
         // 自动复用 runEvaluation 内部的 builder + 单一逻辑,不再在 batch 这边
-        // 旁路一套 doctor / 路径推断。entry artifact 拼装方式与 batch real-run
-        // (executeBatchEvaluationRuns) 严格一致, 避免 dry-run / real-run 漂移。
-        const skillDirAbs = resolve(skillDir);
+        // 旁路一套 doctor / 路径推断。entry 的 spec 构造(buildBatchVariantSpecs)与 batch
+        // real-run(executeBatchEvaluationRuns)同一处,避免 dry-run / real-run 漂移。
         const dryArtifacts = [];
         for (const entry of skillEntries) {
-            const skillArtifacts = resolveArtifacts(skillDirAbs, ['baseline', entry.skillPath], { strictBaseline, variantAllowedSkills }).map((artifact) => {
-                if (artifact.name === entry.skillPath) {
-                    return { ...artifact, name: entry.name, experimentRole: 'treatment' };
-                }
-                return { ...artifact, experimentRole: 'control' };
-            });
+            const variantSpecs = buildBatchVariantSpecs(entry, variantAllowedSkills);
             let entryReport;
             try {
                 const result = await runEvaluation({
                     samplesPath: entry.samplesPath,
                     skillDir,
-                    artifacts: skillArtifacts,
+                    variantSpecs,
                     model,
                     executorName,
                     dryRun: true,
@@ -406,7 +398,6 @@ export async function runBatchEvaluation({ skillDir, model = DEFAULT_MODEL, outp
                     lang,
                     mcpConfig,
                     strictBaseline,
-                    variantAllowedSkills,
                     // 透传 user 显式输入的参数,让 dry-run power / isolation warning 与
                     // 真实 run 一致(否则 batch dry-run 永远按 repeat=1 报警,误导用户)。
                     repeat,
@@ -525,4 +516,3 @@ export async function runMultiple({ repeat = 1, onRepeatProgress, ...config }) {
     }
     return { report, aggregated, filePath };
 }
-//# sourceMappingURL=run-evaluation.js.map

package/dist/executors/anthropic-api.d.ts

@@ -1,3 +1,2 @@
 import type { ExecResult, ExecutorInput } from '../types/index.js';
 export declare function anthropicApiExecutor({ model, system, prompt, timeoutMs }: ExecutorInput): Promise<ExecResult>;
-//# sourceMappingURL=anthropic-api.d.ts.map

package/dist/executors/anthropic-api.js

@@ -5,7 +5,7 @@ export async function anthropicApiExecutor({ model, system, prompt, timeoutMs =
         throw new Error('ANTHROPIC_API_KEY environment variable is not set');
     const baseUrl = process.env.ANTHROPIC_BASE_URL || 'https://api.anthropic.com';
     const reqBody = {
-        model: model || 'claude-sonnet-4-5-20250514',
+        model: model || 'claude-sonnet-4-5',
         max_tokens: 8192,
         messages: [{ role: 'user', content: prompt }],
     };
@@ -40,4 +40,3 @@ export async function anthropicApiExecutor({ model, system, prompt, timeoutMs =
         return { ok: false, error, durationMs, durationApiMs: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, costUSD: 0, output: null, stopReason, numTurns: 0 };
     }
 }
-//# sourceMappingURL=anthropic-api.js.map

package/dist/executors/claude-cli.d.ts

@@ -1,3 +1,2 @@
 import type { ExecResult, ExecutorInput } from '../types/index.js';
 export declare function claudeCliExecutor({ model, system, prompt, cwd, skillDir, timeoutMs, allowedSkills, mocks, mocksBaseDir, mocksStrict, lean, effort }: ExecutorInput): Promise<ExecResult>;
-//# sourceMappingURL=claude-cli.d.ts.map

package/dist/executors/claude-cli.js

@@ -178,4 +178,3 @@ export async function claudeCliExecutor({ model, system, prompt, cwd, skillDir,
         mockHandle?.cleanup();
     }
 }
-//# sourceMappingURL=claude-cli.js.map

package/dist/executors/claude-sdk-trace.d.ts

@@ -7,4 +7,3 @@ export declare function extractAgentTrace(messages: ClaudeSdkBaseMessage[], time
     fullNumTurns: number;
     numSubAgents: number;
 };
-//# sourceMappingURL=claude-sdk-trace.d.ts.map

package/dist/executors/claude-sdk-trace.js

@@ -103,4 +103,3 @@ export function extractAgentTrace(messages, timestamps) {
     }
     return { turns, toolCalls, fullNumTurns, numSubAgents };
 }
-//# sourceMappingURL=claude-sdk-trace.js.map

package/dist/executors/claude-sdk.d.ts

@@ -13,4 +13,3 @@ export declare function buildSdkIsolationOptions(allowedSkills: string[] | undef
     disallowedTools?: string[];
 };
 export declare function claudeSdkExecutor({ model, system, prompt, cwd, skillDir, timeoutMs, verbose, allowedSkills, mocks, mocksBaseDir, mocksStrict, lean, effort }: ExecutorInput): Promise<ExecResult>;
-//# sourceMappingURL=claude-sdk.d.ts.map

package/dist/executors/claude-sdk.js

@@ -198,4 +198,3 @@ export async function claudeSdkExecutor({ model, system, prompt, cwd, skillDir,
         };
     }
 }
-//# sourceMappingURL=claude-sdk.js.map

package/dist/executors/codex-cli-trace.d.ts

@@ -7,4 +7,3 @@ export declare function extractCodexTrace(events: CodexEvent[]): {
     fullNumTurns: number;
     numSubAgents: number;
 };
-//# sourceMappingURL=codex-cli-trace.d.ts.map

package/dist/executors/codex-cli-trace.js

@@ -120,4 +120,3 @@ export function extractCodexTrace(events) {
     flushTurn();
     return { turns, toolCalls, fullNumTurns, numSubAgents };
 }
-//# sourceMappingURL=codex-cli-trace.js.map