oh-my-knowledge 0.32.0 → 0.34.0

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 +26,9 @@ 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 }),
     }));
 }
 function validateEvalConfig(parsed, configPath) {
@@ -248,4 +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

package/dist/inputs/skill-loader.d.ts

@@ -1,4 +1,71 @@
 import type { Artifact } from '../types/index.js';
+export declare function gitShowFile(ref: string, filePath: string, cwd?: string): string | null;
+/**
+ * 取 `<ref>:<filePath>` 的原始字节(二进制资产用,不做 utf-8 解码);不存在/非 blob/出错返回 null。
+ * 同 gitShowFile 用 `cat-file blob`:对目录会非零退出,不会把树清单字节当文件内容物化。
+ */
+export declare function gitShowBytes(ref: string, filePath: string, cwd?: string): Buffer | null;
+export interface GitTreeEntry {
+    /** git 文件模式:100644/100755=普通文件,120000=软链,160000=submodule。 */
+    mode: string;
+    /** 相对所列 tree 的路径。 */
+    path: string;
+}
+/**
+ * 递归列出 `<ref>:<treePath>` 子树下的叶子条目(blob / 软链 / submodule)。tree 不存在返回 []。
+ * 用 `-z`(NUL 分隔):git 不会对含换行 / 非 ASCII 的路径做 C-quote,路径原样可回喂 git show。
+ * 用 `--full-tree`:treePath 是仓库根相对路径,不受当前 cwd 前缀限制 —— 否则在仓库子目录执行时
+ * `ls-tree HEAD:skills/review` 会返回空(而 gitShowFile 探测不受限,导致"探测命中→物化为空"的发散)。
+ * `treePath` 为空时列整棵根 tree(`<ref>:`)。
+ */
+export declare function gitLsTreeBlobs(ref: string, treePath: string, cwd?: string): GitTreeEntry[];
+export interface GitRepoContext {
+    /** 仓库根(realpath 归一,避免 macOS /var ↔ /private/var 等价路径算错)。所有 git helper 以此为 cwd。 */
+    repoRoot: string;
+    /** fromPath 相对仓库根的路径(realpath 归一后计算)。 */
+    relDir: string;
+}
+/**
+ * 解出 `fromPath` 所属 git 仓库的上下文。**关键**:`git rev-parse` 在 `fromPath` 处执行(-C),
+ * 而非进程 cwd —— 否则 eval 从别处调用、skillDir 指向另一个 repo 时,会拿进程 cwd 的 repo root 和
+ * HEAD 去解析,轻则 not_found、重则评测错 repo 的同名内容。repoRoot / relDir 都 realpath 归一,
+ * 消除 /var ↔ /private/var 这类等价路径导致 relative 算错。不在 git 仓库时 rev-parse 抛错,由上层转
+ * not_a_git_repo。stderr 吞掉,不双重打印 `fatal:`。
+ */
+export declare function resolveGitRepoContext(fromPath: string): GitRepoContext;
+/** 拼 git 仓库相对路径:始终用 `/`,空 base 直接返回 b(避免 join 产生 `.` 这类退化 tree-ish)。 */
+export declare function gitJoin(base: string, sub: string): string;
+/**
+ * 解析 `git:<ref>:<spec>` —— **install 与 eval 共用此一处**,保证两侧对 ref/spec 的切分与合法性
+ * 判定一致。`git:<spec>` 缺省 ref=HEAD;spec 可含 `/`(如 skills/review)。空 ref / 空 spec 都非法
+ * 并返回 null:空 ref(`git::x`)会被 git 当 index/stage-0 解析,量到不可复取、依赖本地暂存区的内容,
+ * 破坏 git variant 的可复现语义;空 spec(`git:HEAD:`)会退化去探 `.md`/`SKILL.md` 误命中。
+ */
+export declare function parseGitInput(input: string): {
+    ref: string;
+    spec: string;
+} | null;
+export interface GitSkillRef {
+    isDir: boolean;
+    /** dir-skill 的子树根(仓库相对);file-skill 为空。 */
+    treePath: string;
+    /** file-skill 的 .md 路径(仓库相对);dir-skill 为空。 */
+    fileSkillPath: string;
+    name: string;
+}
+/**
+ * 把 git spec 解析成 dir / file skill —— **install 与 eval 共用此一处**,保证两条路径对同一
+ * `git:<ref>:<spec>` 的 file-vs-dir 归类绝不发散(发散会让 eval 量文件、install 注册目录,
+ * 两边对不上)。接受三种写法,与本地路径安装对称:
+ *   - 显式 SKILL.md:`skills/dir/SKILL.md` → 目录-skill,name 取父目录名;
+ *   - 显式 .md:`skills/foo.md` → 文件-skill;
+ *   - 裸 spec:`skills/review` → **文件优先**(先试 `<spec>.md`,再试 `<spec>/SKILL.md`)。
+ * 裸 spec 的文件优先必须与 eval 历史顺序一致 —— 否则同名同时存在 .md 与 dir/SKILL.md 时,
+ * eval 量文件、install 注册目录,evidence 读时按 hash 门控被静默剥离、记录永久 stale。
+ * `gitRelDir` 是各调用方的解析基准(install=cwd 相对仓库根、eval=skillDir 相对仓库根),作为显式
+ * 参数传入,基准差异留给调用方、归类逻辑单一来源。任一探测命中即返回;都不中返回 null。
+ */
+export declare function classifyGitSkillRef(ref: string, gitRelDir: string, spec: string, cwd?: string): GitSkillRef | null;
 export declare function discoverVariants(skillDir: string): string[];
 export declare function discoverBatchSkills(skillDir: string): Array<{
     name: string;
@@ -8,12 +75,10 @@ export declare function discoverBatchSkills(skillDir: string): Array<{
 export declare function loadSkills(skillDir: string, variants: string[]): Record<string, string | null>;
 /** opts for resolveArtifacts skill-isolation wiring. */
 export interface ResolveArtifactsOptions {
-    /** Default true. When true, baseline-kind artifacts get allowedSkills=[] auto-injected
-     *  unless overridden by variantAllowedSkills. */
+    /** Default true. When true, baseline-kind artifacts get allowedSkills=[] auto-injected.
+     *  显式 per-variant 隔离声明走 spec.allowedSkills(prepareEvaluationRun 按 spec 身份绑定),
+     *  不经此处——resolveArtifacts 只认 strictBaseline 默认,隔离绑定收成单一来源。 */
     strictBaseline?: boolean;
-    /** Per-variant explicit allowedSkills (from eval.yaml). Keyed by variant name (post
-     *  parseVariantCwd, matches Artifact.name). Always overrides strictBaseline default. */
-    variantAllowedSkills?: Record<string, string[]>;
 }
 /**
  * Parse variant expression, extracting optional cwd suffix.
@@ -23,5 +88,32 @@ export declare function parseVariantCwd(variant: string): {
     name: string;
     cwd?: string;
 };
-export declare function resolveArtifacts(skillDir: string, variants: string[], opts?: ResolveArtifactsOptions): Artifact[];
-//# sourceMappingURL=skill-loader.d.ts.map
+/** 把一个 variant 表达式规范化成稳定的物理身份,用于「同一 variant 不能既是 control 又是
+ *  treatment」的判重。同一份 skill 的不同写法必须折叠成同一个 key:
+ *    - `./x.md` 与 `x.md`、符号链接、大小写不敏感卷上的等价写法:resolve 后取 `dev:ino` 物理身份。
+ *    - 目录 `dir` 与 `dir/SKILL.md`：先折叠到同一锚点再取物理身份。
+ *    - 裸短名(传了 `skillDir` 时):按 resolveArtifacts 的解析基准(`skillDir/name.md` 或
+ *      `skillDir/name/SKILL.md`)取物理身份,这样裸名 `greeter` 与指向同一文件的 `./greeter.md`
+ *      能判为重复;没传 skillDir(如纯单测)则退回字面名。
+ *    - `git:` / `baseline`：本身就是稳定标识,原样返回。
+ *  结构化的 `cwd`(第三参数)按物理身份纳入 key —— 同一份 skill 绑不同 cwd 是不同 runtime
+ *  context,不算重复。不要用派生短名判重:`v1/greeter.md` 与 `v2/greeter.md` 短名都是 greeter
+ *  却是两个 variant。 */
+export declare function variantIdentity(expr: string, skillDir?: string, cwd?: string): string;
+/** 从已解析的 skill 路径取短名:`SKILL.md` 取其父目录名,否则取去掉 `.md` 后缀的 basename。
+ *  `variantExprToSkillName`(expr → 短名)与 `resolveArtifacts` 的 file-path 命名共用这一处,
+ *  避免两份各写一遍后悄悄发散——report 键就来自 resolveArtifacts 这一支。 */
+export declare function skillNameFromPath(filePath: string): string;
+/** 从 variant 表达式(纯 artifact 身份,可能是路径)取短名。 */
+export declare function variantExprToSkillName(expr: string): string;
+/** variant 输入:纯 artifact 表达式字符串,或结构化的 `{expr, cwd?}`。cwd 不再编码进 expr。 */
+export type VariantInput = string | {
+    expr: string;
+    cwd?: string;
+};
+export declare function resolveArtifacts(skillDir: string, variants: VariantInput[], opts?: ResolveArtifactsOptions): Artifact[];
+/** 消歧:多个 variant resolve 出同名时(如 `v1/greeter.md` vs `v2/greeter.md` 都叫 `greeter`),
+ *  用父目录逐段限定恢复唯一性。否则 `results[sample][variant]` / `summary[variant]` 按 name 键时
+ *  后写覆盖前写,把对照 / 实验组的结果搅在一起、verdict 拿同名跟自己比 —— 静默破坏对比可信度。
+ *  没有可用路径(baseline / git)时退化加 `#n` 序号兜底,保证返回的 name 一定两两不同。 */
+export declare function ensureUniqueVariantNames(artifacts: Artifact[]): void;