oh-my-knowledge 0.33.0 → 0.34.0

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 +5,27 @@ 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) => ({ 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 +79,3 @@ 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

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

@@ -15,4 +15,3 @@ export declare function buildCodexArgs({ model, cwd, prompt }: {
     prompt: string;
 }): string[];
 export declare function codexCliExecutor({ model, system, prompt, cwd, skillDir, timeoutMs, allowedSkills, verbose }: ExecutorInput): Promise<ExecResult>;
-//# sourceMappingURL=codex-cli.d.ts.map

package/dist/executors/codex-cli.js

@@ -251,4 +251,3 @@ export async function codexCliExecutor({ model, system, prompt, cwd, skillDir, t
         };
     }
 }
-//# sourceMappingURL=codex-cli.js.map

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

@@ -15,4 +15,3 @@ export declare function buildCodexSdkThreadOptions({ model, cwd }: {
 export declare function buildCodexSdkClientOptions(env: NodeJS.ProcessEnv): CodexOptions;
 export declare function createCodexSdkClient(env: NodeJS.ProcessEnv): Promise<Codex>;
 export declare function codexSdkExecutor({ model, system, prompt, cwd, skillDir, timeoutMs, allowedSkills, verbose }: ExecutorInput): Promise<ExecResult>;
-//# sourceMappingURL=codex-sdk.d.ts.map

package/dist/executors/codex-sdk.js

@@ -211,4 +211,3 @@ export async function codexSdkExecutor({ model, system, prompt, cwd, skillDir, t
         process.removeListener('SIGINT', sigintListener);
     }
 }
-//# sourceMappingURL=codex-sdk.js.map

package/dist/executors/gemini.d.ts

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

package/dist/executors/gemini.js

@@ -79,4 +79,3 @@ export async function geminiExecutor({ model, system, prompt, timeoutMs = DEFAUL
         };
     }
 }
-//# sourceMappingURL=gemini.js.map

package/dist/executors/index.d.ts

@@ -4,4 +4,3 @@ import { createScriptExecutor } from './script.js';
 import { DEFAULT_MODEL, JUDGE_MODEL } from './shared.js';
 export { DEFAULT_MODEL, JUDGE_MODEL, extractAgentTrace, createScriptExecutor };
 export declare function createExecutor(name?: string): ExecutorFn;
-//# sourceMappingURL=index.d.ts.map

package/dist/executors/index.js

@@ -24,4 +24,3 @@ export { DEFAULT_MODEL, JUDGE_MODEL, extractAgentTrace, createScriptExecutor };
 export function createExecutor(name = 'claude') {
     return EXECUTOR_REGISTRY[name] || createScriptExecutor(name);
 }
-//# sourceMappingURL=index.js.map

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

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

package/dist/executors/openai-api.js

@@ -37,4 +37,3 @@ export async function openAiApiExecutor({ model, system, prompt, timeoutMs = DEF
         return { ok: false, error, durationMs, durationApiMs: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, costUSD: 0, output: null, stopReason, numTurns: 0 };
     }
 }
-//# sourceMappingURL=openai-api.js.map

package/dist/executors/runtime-fingerprint.d.ts

@@ -4,4 +4,3 @@ export interface ExecutorRuntimeFingerprintOptions {
     env?: NodeJS.ProcessEnv;
 }
 export declare function getExecutorRuntimeFingerprint(executorName: string, model: string, options?: ExecutorRuntimeFingerprintOptions): ExecutorRuntimeFingerprint;
-//# sourceMappingURL=runtime-fingerprint.d.ts.map

package/dist/executors/runtime-fingerprint.js

@@ -140,7 +140,7 @@ function withFingerprint(input) {
     const stablePayload = {
         executor: input.executor,
         model: input.model,
-        kind: input.kind,
+        kind: input.runtimeKind,
         binary: input.binary
             ? {
                 name: input.binary.name,
@@ -171,7 +171,7 @@ function runtime(executor, model, kind, capabilities, extra = {}) {
     return withFingerprint({
         executor,
         model,
-        kind,
+        runtimeKind: kind,
         capabilities,
         ...extra,
     });
@@ -274,4 +274,3 @@ export function getExecutorRuntimeFingerprint(executorName, model, options = {})
     RUNTIME_CACHE.set(cacheKey, fp);
     return fp;
 }
-//# sourceMappingURL=runtime-fingerprint.js.map

package/dist/executors/script.d.ts

@@ -1,3 +1,2 @@
 import type { ExecutorFn } from '../types/index.js';
 export declare function createScriptExecutor(command: string): ExecutorFn;
-//# sourceMappingURL=script.d.ts.map

package/dist/executors/script.js

@@ -74,4 +74,3 @@ export function createScriptExecutor(command) {
         }
     };
 }
-//# sourceMappingURL=script.js.map

package/dist/executors/shared.d.ts

@@ -128,7 +128,6 @@ export interface CodexEvent {
         results?: unknown[];
         changes?: Array<{
             path?: string;
-            kind?: string;
         }>;
         server?: string;
         tool?: string;
@@ -191,4 +190,3 @@ export declare function spawnWithSigintPropagation(command: string, args: string
     child: ChildProcess;
     done: Promise<SpawnHelperResult>;
 };
-//# sourceMappingURL=shared.d.ts.map

package/dist/executors/shared.js

@@ -250,4 +250,3 @@ export function spawnWithSigintPropagation(command, args, options = {}) {
     });
     return { child, done };
 }
-//# sourceMappingURL=shared.js.map

package/dist/grading/assertions.d.ts

@@ -36,4 +36,3 @@ export declare function runAssertions(output: string, assertions: Assertion[], c
     };
 }): AssertionResults;
 export declare function runAsyncAssertions(output: string, assertions: Assertion[], { executor, judgeModel, sample, samplesDir }: AsyncAssertionContext): Promise<AssertionResults>;
-//# sourceMappingURL=assertions.d.ts.map

package/dist/grading/assertions.js

@@ -551,4 +551,3 @@ async function runRagJudge(assertion, output, sample, executor, judgeModel) {
         return { passed: false, message: 'failed to parse judge response', costUSD: result.costUSD || 0, ...reported };
     }
 }
-//# sourceMappingURL=assertions.js.map

package/dist/grading/debias-validate.d.ts

@@ -81,4 +81,3 @@ export interface DebiasValidateResult {
  */
 export declare function validateLengthDebias(input: DebiasValidateInput): Promise<DebiasValidateResult>;
 export declare function formatDebiasValidate(result: DebiasValidateResult): string;
-//# sourceMappingURL=debias-validate.d.ts.map

package/dist/grading/debias-validate.js

@@ -170,4 +170,3 @@ export function formatDebiasValidate(result) {
     }
     return lines.join('\n');
 }
-//# sourceMappingURL=debias-validate.js.map

package/dist/grading/diagnostic.d.ts

@@ -37,4 +37,3 @@ export declare function salvagePartialJson(blob: string): {
 };
 export declare function runDiagnostic(opts: RunDiagnosticOptions): Promise<DiagnosticResult>;
 export {};
-//# sourceMappingURL=diagnostic.d.ts.map

package/dist/grading/diagnostic.js

@@ -447,4 +447,3 @@ function parseFailureModes(raw) {
     }
     return out;
 }
-//# sourceMappingURL=diagnostic.js.map

package/dist/grading/gold-cli.d.ts

@@ -82,4 +82,3 @@ export declare function validateGoldDataset(dir: string): {
     issues: string[];
     sampleCount: number;
 };
-//# sourceMappingURL=gold-cli.d.ts.map

package/dist/grading/gold-cli.js

@@ -244,4 +244,3 @@ export function validateGoldDataset(dir) {
         sampleCount: dataset?.annotations.length ?? 0,
     };
 }
-//# sourceMappingURL=gold-cli.js.map

package/dist/grading/gold-dataset.d.ts

@@ -70,4 +70,3 @@ export interface LoadResult {
 export declare function loadGoldDataset(dir: string): LoadResult;
 /** Re-export YAML serializer so the `gold init` CLI can write a template. */
 export declare function dumpYaml(value: unknown): string;
-//# sourceMappingURL=gold-dataset.d.ts.map

package/dist/grading/gold-dataset.js

@@ -158,4 +158,3 @@ function validateAnnotation(raw, path, index) {
 export function dumpYaml(value) {
     return yaml.dump(value, { lineWidth: 100, noRefs: true });
 }
-//# sourceMappingURL=gold-dataset.js.map

package/dist/grading/human-gold.d.ts

@@ -99,4 +99,3 @@ export declare function computeAgreementWithCI(pairs: RatingPair[], options?: {
         max: number;
     };
 }): AgreementResult;
-//# sourceMappingURL=human-gold.d.ts.map

package/dist/grading/human-gold.js

@@ -185,4 +185,3 @@ function roundOrNaN(x) {
         return NaN;
     return Math.round(x * 10000) / 10000;
 }
-//# sourceMappingURL=human-gold.js.map

package/dist/grading/index.d.ts

@@ -58,4 +58,3 @@ interface GradeOptions {
 export declare function grade({ output, sample, judgeModels, judgeExecutors, allowLlmJudge, execMetrics, samplesDir, judgeRepeat, lengthDebias }: GradeOptions): Promise<GradeResult>;
 export { runAssertions, validateJsonSchema } from './assertions.js';
 export { buildTraceSummary } from './judge.js';
-//# sourceMappingURL=index.d.ts.map

package/dist/grading/index.js

@@ -137,4 +137,3 @@ export async function grade({ output, sample, judgeModels, judgeExecutors, allow
 }
 export { runAssertions, validateJsonSchema } from './assertions.js';
 export { buildTraceSummary } from './judge.js';
-//# sourceMappingURL=index.js.map

package/dist/grading/judge.d.ts

@@ -73,4 +73,3 @@ export declare function computeJudgeAgreement(judgeScores: number[][]): JudgeAgr
  */
 export declare function llmJudgeEnsemble(options: LlmJudgeOptions, judges: JudgeConfig[], executorByName: (name: string) => ExecutorFn, judgeRepeat?: number): Promise<DimensionResult>;
 export {};
-//# sourceMappingURL=judge.d.ts.map

package/dist/grading/judge.js

@@ -473,4 +473,3 @@ export async function llmJudgeEnsemble(options, judges, executorByName, judgeRep
         agreement,
     };
 }
-//# sourceMappingURL=judge.js.map

package/dist/grading/layered-scores.d.ts

@@ -10,4 +10,3 @@ export declare function computeLayeredScores(results: CompositeInput): {
     layeredScores: LayeredScores;
 };
 export {};
-//# sourceMappingURL=layered-scores.d.ts.map

package/dist/grading/layered-scores.js

@@ -60,4 +60,3 @@ export function computeLayeredScores(results) {
     const compositeScore = scores.length > 0 ? Number((scores.reduce((a, b) => a + b, 0) / scores.length).toFixed(2)) : 0;
     return { compositeScore, layeredScores: layered };
 }
-//# sourceMappingURL=layered-scores.js.map

package/dist/inputs/eval-config.d.ts

@@ -10,4 +10,3 @@ export declare function loadEvalConfig(configPath: string): EvalConfig;
  * so the downstream variant resolver can treat CLI and config uniformly.
  */
 export declare function configVariantsToSpecs(variants: EvalConfigVariant[]): VariantSpec[];
-//# sourceMappingURL=eval-config.d.ts.map

package/dist/inputs/eval-config.js

@@ -26,7 +26,8 @@ export function configVariantsToSpecs(variants) {
     return variants.map((v) => ({
         name: v.name,
         role: v.role,
-        expr: v.cwd ? `${v.artifact}@${v.cwd}` : v.artifact,
+        expr: v.artifact, // 纯 artifact 身份;cwd 结构化携带,不再编码进 expr
+        ...(v.cwd !== undefined && { cwd: v.cwd }),
         ...(v.allowedSkills !== undefined && { allowedSkills: v.allowedSkills }),
     }));
 }
@@ -249,4 +250,3 @@ function resolveConfigPaths(config, configDir) {
         })),
     };
 }
-//# sourceMappingURL=eval-config.js.map

package/dist/inputs/load-samples.d.ts

@@ -33,4 +33,3 @@ export interface LoadSamplesResult {
  */
 export declare function loadSamples(samplesPath: string): LoadSamplesResult;
 export declare function validateSamples(samples: Sample[]): void;
-//# sourceMappingURL=load-samples.d.ts.map

package/dist/inputs/load-samples.js

@@ -245,4 +245,3 @@ function loadSampleFile(samplesPath) {
     validateSamples(samples);
     return { samples, requires };
 }
-//# sourceMappingURL=load-samples.js.map

package/dist/inputs/mcp-resolver.d.ts

@@ -47,4 +47,3 @@ export declare function resolveMcpUrls(samples: Sample[], mcpServers: McpServers
  * Should be called when the evaluation run finishes.
  */
 export declare function stopAllServers(): Promise<void>;
-//# sourceMappingURL=mcp-resolver.d.ts.map

package/dist/inputs/mcp-resolver.js

@@ -304,4 +304,3 @@ async function fetchFromServer(serverName, serverDef, urls, resultMap) {
     });
     await Promise.allSettled(promises);
 }
-//# sourceMappingURL=mcp-resolver.js.map