oh-my-knowledge 0.47.0 → 0.49.0

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

@@ -11,19 +11,23 @@ export const initDict = {
     // 跑出第一份报告是冷启动最该先发生的事;「换成你自己的」放到跑通之后。这也消除了
     // 主 README「不用改任何文件」与旧 init「先编辑」的矛盾。
     'cli.init.next_step_run': {
-        zh: '  1. 直接跑通（无需先改任何文件）：omk eval --control code-review-v1 --treatment code-review-v2',
-        en: '  1. Run it as-is (no edits needed): omk eval --control code-review-v1 --treatment code-review-v2',
+        zh: '  1. 直接跑通（无需先改任何文件）：{command}',
+        en: '  1. Run it as-is (no edits needed): {command}',
+    },
+    'cli.init.next_step_report': {
+        zh: '  2. 看报告里的 verdict 和“下一步”：PROGRESS 才能发布；UNDERPOWERED / NOISE 先扩样到约 20 条以上后重跑。',
+        en: '  2. Read the report verdict and Next line: PROGRESS can ship; UNDERPOWERED / NOISE means grow to roughly 20+ samples and re-run.',
     },
     'cli.init.next_step_executor': {
-        zh: '     默认执行器与评委用 claude CLI，需先装好并登录；想换别的模型或离线跑（无需 API key）见文档「执行器」。',
-        en: '     The default executor and judge use the claude CLI (install and log in first); to use another model or run offline (no API key) see the Executors docs.',
+        zh: '     executor / judge 会按运行环境选择；Codex 任务自动使用本机 Codex 配置。也可用 OMK_EXECUTOR / OMK_MODEL 固定环境偏好，详见 https://oh-my-knowledge.pages.dev/zh/reference/executors。',
+        en: '     The executor / judge follow the runtime environment; Codex tasks use the local Codex configuration automatically. OMK_EXECUTOR / OMK_MODEL pin environment preferences. See https://oh-my-knowledge.pages.dev/reference/executors.',
     },
     'cli.init.next_step_customize': {
-        zh: '  2. 跑通后，把 skills/code-review-v1/SKILL.md 和 skills/code-review-v2/SKILL.md 与 eval-samples.json 换成你自己的 skill 和用例',
-        en: '  2. Once it runs, replace skills/code-review-v1/SKILL.md and skills/code-review-v2/SKILL.md and eval-samples.json with your own skills and cases',
+        zh: '  3. 跑通后，替换为你自己的 skill 和 eval-samples.json；还没有用例时先运行 omk sample <skill-path>。',
+        en: '  3. Once it runs, replace the starter skills and eval-samples.json with your own; if you have no samples yet, run omk sample <skill-path>.',
     },
-    'cli.init.note_codex_executor': {
-        zh: '\n注: omk 评测时把 SKILL.md 整文(含 frontmatter)作为 system prompt 注入——跨 executor 一致(claude / codex / openai-api / gemini 都走同一条路径,不依赖任何 executor 的 native skill auto-discovery 或 Skill 工具机制)。frontmatter 在 prompt 头部对 model 行为无显著影响。\n模板带 Claude Code 兼容的 frontmatter(name + description)是为了让同一份 directory-skill 也能 deploy 到 Claude Code:把整个目录复制到 ~/.claude/skills/code-review-v1/(整目录,不是单个 SKILL.md),Claude SDK 才能识别。这是 omk 评测之外的 bonus,一份文件双向 dogfood。',
-        en: '\nNote: during omk evaluation the full SKILL.md (frontmatter included) is injected as the system prompt — uniformly across executors (claude / codex / openai-api / gemini all take the same path; omk does not rely on any executor\'s native skill auto-discovery or Skill tool). Frontmatter has no measurable impact on model behavior in this position.\nThe template ships with Claude Code-compatible frontmatter (name + description) so the same directory-skill can also be deployed to Claude Code: copy the whole directory to ~/.claude/skills/code-review-v1/ (the directory, not just SKILL.md) so Claude SDK can recognize it. That is a bonus beyond omk evaluation — one source, two-way dogfood.',
+    'cli.init.note_skill_injection': {
+        zh: '     注：omk eval 会把 SKILL.md 作为 system prompt 注入；模板 frontmatter 只是方便同一目录复用为 agent skill。',
+        en: '     Note: omk eval injects SKILL.md as the system prompt; the starter frontmatter only helps reuse the same directory as an agent skill.',
     },
 };

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

@@ -1,3 +1,3 @@
 import type { CliMessage } from './types.js';
-export type RunMessageKey = 'cli.progress.preflight_starting' | 'cli.progress.sample_retry' | 'cli.progress.sample_error' | 'cli.progress.sample_executing' | 'cli.progress.sample_exec_done' | 'cli.progress.output_preview' | 'cli.progress.judging' | 'cli.progress.judged' | 'cli.progress.skipped' | 'cli.progress.sample_done' | 'cli.progress.sample_failed_done' | 'cli.run.invalid_repeat' | 'cli.run.invalid_holdout_ratio' | 'cli.run.invalid_judge_repeat' | 'cli.run.no_debias_length_active' | 'cli.run.invalid_bootstrap_samples' | 'cli.run.bootstrap_samples_too_large' | 'cli.run.dry_run_no_scores' | 'cli.run.skill_section' | 'cli.run.run_section' | 'cli.run.batch_complete' | 'cli.run.batch_verdict_header' | 'cli.run.batch_verdict_next_step' | 'cli.run.batch_child_report_missing' | 'cli.run.eval_complete' | 'cli.run.tally' | 'cli.run.report_saved' | 'cli.run.evidence_recorded' | 'cli.run.evidence_recorded_unbound' | 'cli.run.report_only_gate_skipped' | 'cli.run.report_server_running' | 'cli.run.report_server_view' | 'cli.run.report_server_stop' | 'cli.run.no_serve_in_non_tty' | 'cli.run.no_serve_view_hint' | 'cli.run.gold_load_failed' | 'cli.run.gold_load_issue' | 'cli.run.contamination_warning' | 'cli.run.skip_connectivity_warning';
+export type RunMessageKey = 'cli.progress.preflight_starting' | 'cli.progress.sample_retry' | 'cli.progress.sample_error' | 'cli.progress.sample_executing' | 'cli.progress.sample_exec_done' | 'cli.progress.output_preview' | 'cli.progress.judging' | 'cli.progress.judged' | 'cli.progress.skipped' | 'cli.progress.sample_done' | 'cli.progress.sample_failed_done' | 'cli.run.invalid_repeat' | 'cli.run.invalid_holdout_ratio' | 'cli.run.invalid_judge_repeat' | 'cli.run.no_debias_length_active' | 'cli.run.invalid_bootstrap_samples' | 'cli.run.bootstrap_samples_too_large' | 'cli.run.dry_run_no_scores' | 'cli.run.skill_section' | 'cli.run.run_section' | 'cli.run.batch_complete' | 'cli.run.batch_verdict_header' | 'cli.run.batch_verdict_next_step' | 'cli.run.batch_child_report_missing' | 'cli.run.eval_complete' | 'cli.run.tally' | 'cli.run.report_saved' | 'cli.run.evidence_recorded' | 'cli.run.evidence_recorded_promotable' | 'cli.run.evidence_recorded_unbound' | 'cli.run.report_only_gate_skipped' | 'cli.run.report_server_running' | 'cli.run.report_server_view' | 'cli.run.report_server_stop' | 'cli.run.no_serve_in_non_tty' | 'cli.run.no_serve_view_hint' | 'cli.run.gold_load_failed' | 'cli.run.gold_load_issue' | 'cli.run.contamination_warning' | 'cli.run.codex_fallback_hint' | 'cli.run.codex_auth_hint' | 'cli.run.codex_model_hint' | 'cli.run.openai_api_auth_hint' | 'cli.run.openai_api_model_hint' | 'cli.run.anthropic_api_auth_hint' | 'cli.run.anthropic_api_model_hint' | 'cli.run.skip_connectivity_warning';
 export declare const runDict: Record<RunMessageKey, CliMessage>;

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

@@ -68,8 +68,8 @@ export const runDict = {
         en: '⚠ --bootstrap-samples {n} is large and may take several seconds. 1000 is the industry standard and usually sufficient.\n',
     },
     'cli.run.dry_run_no_scores': {
-        zh: 'eval dry-run：仅预览任务，不检查分数',
-        en: 'Eval dry-run: no scores to check',
+        zh: 'eval dry-run：仅预览任务，不检查分数。下一步：确认任务无误后，去掉 --dry-run 运行正式评测。',
+        en: 'Eval dry-run: no scores checked. Next: remove --dry-run to run the eval.',
     },
     'cli.run.skill_section': {
         zh: '\n=== [{i}/{n}] Skill: {skill} ===\n',
@@ -111,6 +111,10 @@ export const runDict = {
         zh: '🔖 已为受管 skill「{name}」记录评测证据 → measurable\n',
         en: '🔖 Recorded eval evidence for managed skill "{name}" → measurable\n',
     },
+    'cli.run.evidence_recorded_promotable': {
+        zh: '🔖 已为受管 skill「{name}」记录评测证据 → measurable。运行 {command} 接受当前版本。\n',
+        en: '🔖 Recorded eval evidence for managed skill "{name}" → measurable. Run {command} to accept this version.\n',
+    },
     'cli.run.evidence_recorded_unbound': {
         zh: '🔖 受管 skill「{name}」：评测内容与当前安装版本指纹不一致，证据已留存但不绑当前版本\n',
         en: '🔖 Managed skill "{name}": eval content differs from the installed version; evidence kept but not bound to current\n',
@@ -151,6 +155,34 @@ export const runDict = {
         zh: '\n⚠ {warning}\n',
         en: '\n⚠ {warning}\n',
     },
+    'cli.run.codex_fallback_hint': {
+        zh: '\n提示：当前失败的是 Claude 系列执行器。先确认 Claude Code 已登录；如果你在 Codex 环境里，也可以把模型运行参数改为：{flags}。{codexModelHint}codex 执行器目前不会报告 costUSD。',
+        en: '\nHint: the failing runtime is Claude-based. First confirm Claude Code is authenticated; in a Codex environment, you can also switch the model runtime flags to: {flags}. {codexModelHint} The codex executor does not report costUSD yet.',
+    },
+    'cli.run.codex_auth_hint': {
+        zh: '\n提示：当前失败的是 Codex 系列执行器。先确认 Codex CLI / SDK 已安装并完成登录；如果你有 Claude Code 可用，可以改走 Claude：{claudeFlags}；如果要继续走 OpenAI API，可以改为：{openaiFlags}，并设置 OPENAI_API_KEY。openai-api 会按 API 响应记录 token / cost。',
+        en: '\nHint: the failing runtime is Codex-based. First confirm the Codex CLI / SDK is installed and authenticated; if Claude Code is available, switch to Claude: {claudeFlags}; to stay on the OpenAI API path, switch to: {openaiFlags}, and set OPENAI_API_KEY. openai-api records token / cost from API responses.',
+    },
+    'cli.run.codex_model_hint': {
+        zh: '\n提示：当前失败的是 Codex 系列执行器，但模型名看起来不可用。可以先按本机 Codex 配置重试：{codexFlags}（{codexModelHint}）；也可以先运行 `{codexExec}` 验证模型是否可用。若只是想先跑通，可以改走 Claude：{claudeFlags}；或继续走 OpenAI API：{openaiFlags}，并设置 OPENAI_API_KEY。',
+        en: '\nHint: the failing runtime is Codex-based, but the model name appears unavailable. Retry with the local Codex config model: {codexFlags} ({codexModelHint}); you can also run `{codexExec}` to verify the model. To just get a first run through, switch to Claude: {claudeFlags}; or stay on the OpenAI API path: {openaiFlags}, and set OPENAI_API_KEY.',
+    },
+    'cli.run.openai_api_auth_hint': {
+        zh: '\n提示：当前失败的是 OpenAI API 执行器。请检查 OPENAI_API_KEY / OPENAI_BASE_URL 是否可用，并确认模型名对当前端点可用。',
+        en: '\nHint: the failing runtime is the OpenAI API executor. Check OPENAI_API_KEY / OPENAI_BASE_URL and confirm the model is available on that endpoint.',
+    },
+    'cli.run.openai_api_model_hint': {
+        zh: '\n提示：当前失败的是 OpenAI API 执行器，但模型名看起来对当前端点不可用。请检查 --model / --judge-models、OPENAI_BASE_URL 与账号权限是否匹配。',
+        en: '\nHint: the failing runtime is the OpenAI API executor, but the model name appears unavailable on the current endpoint. Check --model / --judge-models, OPENAI_BASE_URL, and account access.',
+    },
+    'cli.run.anthropic_api_auth_hint': {
+        zh: '\n提示：当前失败的是 Anthropic API 执行器。请检查 ANTHROPIC_API_KEY / ANTHROPIC_BASE_URL 是否可用，并确认模型名对当前端点可用。',
+        en: '\nHint: the failing runtime is the Anthropic API executor. Check ANTHROPIC_API_KEY / ANTHROPIC_BASE_URL and confirm the model is available on that endpoint.',
+    },
+    'cli.run.anthropic_api_model_hint': {
+        zh: '\n提示：当前失败的是 Anthropic API 执行器，但模型名看起来对当前端点不可用。请检查 --model / --judge-models、ANTHROPIC_BASE_URL 与账号权限是否匹配。',
+        en: '\nHint: the failing runtime is the Anthropic API executor, but the model name appears unavailable on the current endpoint. Check --model / --judge-models, ANTHROPIC_BASE_URL, and account access.',
+    },
     'cli.run.skip_connectivity_warning': {
         zh: '⚠️  --skip-connectivity 已启用: 跳过 LLM 模型连通性检测。请确保 executor / judge 已通过其他方式验证可达。',
         en: '⚠️  --skip-connectivity enabled: LLM connectivity check skipped. Verify executor / judge are reachable by other means.',

package/dist/cli/lib/llm-failure-classifier.d.ts

@@ -0,0 +1,2 @@
+export declare function looksLikeModelUnavailableFailure(message: string): boolean;
+export declare function looksLikeLlmSetupFailure(message: string): boolean;

package/dist/cli/lib/llm-failure-classifier.js

@@ -0,0 +1,8 @@
+const SETUP_FAILURE_RE = /auth|login|credential|API[_ ]?KEY|BASE_URL|API error|ENOENT|not found|ECONN|ENOTFOUND|ETIMEDOUT|timeout|timed out|401|403|404|invalid_request_error/i;
+const MODEL_UNAVAILABLE_RE = /model .*not supported|model is not supported|unsupported model|invalid_request_error.*model|model.*invalid_request_error|model_not_found|model .*not found|model .*does not exist|no such model|not have access to (the )?model|model .*not available|model .*is not available/i;
+export function looksLikeModelUnavailableFailure(message) {
+    return MODEL_UNAVAILABLE_RE.test(message);
+}
+export function looksLikeLlmSetupFailure(message) {
+    return SETUP_FAILURE_RE.test(message) || looksLikeModelUnavailableFailure(message);
+}

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

@@ -18,20 +18,21 @@
  * 反而失去「精度链一目了然」的可读性。
  */
 import type { EvalConfig, VariantSpec, JudgeConfig, EvalBudget, ProgressCallback } from '../../types/index.js';
+import { type RuntimeResolutionOptions } from './runtime-defaults.js';
 export { parseJudgeModelsArg, parseJudgeModelsArgOrExit } from './parse-run-config/judge-models.js';
 export interface RunConfig {
     samplesPath: string;
     skillDir: string;
     variantSpecs: VariantSpec[];
-    model: string | undefined;
+    model: string;
     outputDir: string;
     noJudge: boolean | undefined;
     noCache: boolean | undefined;
     dryRun: boolean | undefined;
     concurrency: number;
     timeoutMs: number;
-    executorName: string | undefined;
-    /** 跳过 LLM 模型连通性检测。--resume 时自动 true(已经验过)。 */
+    executorName: string;
+    /** 跳过 LLM 模型连通性检测。仅当 --resume 报告通过完整契约校验时自动 true。 */
     skipConnectivity: boolean | undefined;
     /** 跳过 doctor 健康检查门禁(--skip-doctor)。escape hatch — 默认 false。
      *  开启后 doctor 整段不跑(节省静态检查时间);doctor 失败也不再阻断 eval。
@@ -51,7 +52,7 @@ export interface RunConfig {
     /** --judge-repeat N. Calls LLM judge N times per (sample × dimension). Default 1. */
     judgeRepeat?: number;
     /** Unified judge config. Always non-empty; 1 entry = single judge, ≥ 2 = ensemble.
-     *  parseRunConfig guarantees at least `[{executor: <executor>, model: 'haiku'}]`. */
+     *  Defaults to the selected runtime's judge model. */
     judgeModels: JudgeConfig[];
     /** --bootstrap. Adds bootstrap CI to summary (per-variant mean + pairwise diff). */
     bootstrap?: boolean;
@@ -90,4 +91,4 @@ export interface ParseRunConfigResult {
  * 上游对未知 flag 拦截 exit 2,这里不再 parseArgs。eval-runner 等业务 caller
  * 把 oclif flags 当 values 喂进来。
  */
-export declare function parseRunConfig(values: Record<string, unknown>): ParseRunConfigResult;
+export declare function parseRunConfig(values: Record<string, unknown>, runtimeOptions?: RuntimeResolutionOptions): ParseRunConfigResult;

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

@@ -20,17 +20,18 @@
 import { resolve } from 'node:path';
 import { projectReportsDir, globalReportsDir } from '../../eval-core/measurement-dirs.js';
 import { loadEvalConfig } from '../../inputs/eval-config.js';
-import { DEFAULT_MODEL } from '../../executors/shared.js';
+import { setOwnRecordValue } from '../../shared/record-count.js';
 import { parseJudgeModelsArgOrExit } from './parse-run-config/judge-models.js';
 import { discoverSamplesPath } from './parse-run-config/samples-discovery.js';
 import { resolveVariantSpecs } from './parse-run-config/variant-resolution.js';
+import { envJudgeModels, resolveRuntimeSelection, } from './runtime-defaults.js';
 export { parseJudgeModelsArg, parseJudgeModelsArgOrExit } from './parse-run-config/judge-models.js';
 /**
  * 接 typed flags(来自 oclif Command.parse() 输出)。oclif strict 模式已经在
  * 上游对未知 flag 拦截 exit 2,这里不再 parseArgs。eval-runner 等业务 caller
  * 把 oclif flags 当 values 喂进来。
  */
-export function parseRunConfig(values) {
+export function parseRunConfig(values, runtimeOptions = {}) {
     if (values.variants !== undefined) {
         throw new Error(`--variants 已在 v0.16 废除，请改用 --control <expr> 与 --treatment <v1,v2,...>\n`
             + `  迁移示例：--variants baseline,my-skill  →  --control baseline --treatment my-skill\n`
@@ -59,18 +60,24 @@ export function parseRunConfig(values) {
     // 3) Resolve variantSpecs: CLI > config > batch > error。dedup 在 helper 里。
     const variantSpecs = resolveVariantSpecs(values, evalConfig, skillDir);
     // 4) Apply CLI > config > hard-coded default for all other fields.
-    const executorName = values.executor ?? evalConfig?.executor ?? 'claude';
-    // model fallback 链:CLI > eval.yaml > DEFAULT_MODEL(opus 4.7)。
-    // 改 default 时同步 src/cli/commands/eval.ts 里 --model flag description 的默认值文案。
-    const model = values.model ?? evalConfig?.model ?? DEFAULT_MODEL;
+    const runtime = resolveRuntimeSelection({
+        executor: values.executor ?? evalConfig?.executor,
+        model: values.model ?? evalConfig?.model,
+    }, runtimeOptions);
+    const executorName = runtime.executor;
+    const model = runtime.model;
     // judgeModels: unified judge config. Parse --judge-models (CLI) or evalConfig.judgeModels (yaml).
     // 1 entry = single judge, ≥ 2 entries = ensemble. Format `executor:model[,executor:model]`.
-    // 出口 RunConfig.judgeModels 保证非空 (default `[{executor, model: 'haiku'}]`)。
+    // 出口 RunConfig.judgeModels 保证非空。Codex 默认沿用任务模型，避免把 Claude 的
+    // `haiku` alias 误传给 Codex。
     const cliJudgesRaw = values['judge-models'];
     const parsedJudges = cliJudgesRaw !== undefined ? parseJudgeModelsArgOrExit(cliJudgesRaw) : undefined;
+    const envJudgesRaw = envJudgeModels(runtimeOptions.env);
+    const parsedEnvJudges = envJudgesRaw !== undefined ? parseJudgeModelsArgOrExit(envJudgesRaw) : undefined;
     const judgeModels = parsedJudges
         ?? evalConfig?.judgeModels
-        ?? [{ executor: executorName, model: 'haiku' }];
+        ?? parsedEnvJudges
+        ?? [{ executor: executorName, model: runtime.judgeModel }];
     // 报告默认落项目 `.omk/reports`(绑用例集,construct validity);--global 写全局;--output-dir 最高优先。
     // 同 observe / doctor 写入侧口径。读取侧(studio / resume / gold-compare / 复用)走 overlay 项目→全局兜底。
     const outputDir = resolve(values['output-dir']
@@ -110,7 +117,7 @@ export function parseRunConfig(values) {
     if (evalConfig?.variants) {
         for (const v of evalConfig.variants) {
             if (v.allowedSkills !== undefined) {
-                variantAllowedSkills[v.name] = v.allowedSkills;
+                setOwnRecordValue(variantAllowedSkills, v.name, v.allowedSkills);
             }
         }
     }

package/dist/cli/lib/runtime-defaults.d.ts

@@ -0,0 +1,21 @@
+export interface RuntimeResolutionOptions {
+    env?: NodeJS.ProcessEnv;
+    commandExists?: (command: string, env: NodeJS.ProcessEnv) => boolean;
+    lang?: 'zh' | 'en';
+}
+export interface RuntimeSelection {
+    executor: string;
+    model: string;
+    judgeModel: string;
+}
+export declare function commandExistsOnPath(command: string, env?: NodeJS.ProcessEnv): boolean;
+export declare function isCodexHost(env?: NodeJS.ProcessEnv): boolean;
+export declare function isCodexExecutor(executor: string): boolean;
+export declare function resolveCliExecutor(explicitExecutor?: string, options?: RuntimeResolutionOptions): string;
+export declare function resolveCliModel(executor: string, explicitModel?: string, options?: RuntimeResolutionOptions): string;
+export declare function defaultJudgeModel(executor: string, taskModel: string): string;
+export declare function resolveRuntimeSelection(input: {
+    executor?: string;
+    model?: string;
+}, options?: RuntimeResolutionOptions): RuntimeSelection;
+export declare function envJudgeModels(env?: NodeJS.ProcessEnv): string | undefined;

package/dist/cli/lib/runtime-defaults.js

@@ -0,0 +1,79 @@
+import { accessSync, constants } from 'node:fs';
+import { delimiter, join } from 'node:path';
+import { getCodexModelSuggestion } from './codex-model-hint.js';
+import { DEFAULT_MODEL, JUDGE_MODEL } from '../../executors/shared.js';
+function nonEmpty(value) {
+    const trimmed = value?.trim();
+    return trimmed ? trimmed : undefined;
+}
+export function commandExistsOnPath(command, env = process.env) {
+    const pathEntries = (env.PATH ?? '').split(delimiter).filter(Boolean);
+    const extensions = process.platform === 'win32'
+        ? (env.PATHEXT ?? '.EXE;.CMD;.BAT;.COM').split(';')
+        : [''];
+    return pathEntries.some((entry) => extensions.some((extension) => {
+        try {
+            accessSync(join(entry, `${command}${extension}`), constants.X_OK);
+            return true;
+        }
+        catch {
+            return false;
+        }
+    }));
+}
+export function isCodexHost(env = process.env) {
+    return Boolean(nonEmpty(env.CODEX_THREAD_ID)
+        || nonEmpty(env.CODEX_INTERNAL_ORIGINATOR_OVERRIDE)
+        || nonEmpty(env.CODEX_CI));
+}
+export function isCodexExecutor(executor) {
+    return executor === 'codex' || executor === 'codex-sdk';
+}
+export function resolveCliExecutor(explicitExecutor, options = {}) {
+    const env = options.env ?? process.env;
+    const explicit = nonEmpty(explicitExecutor);
+    if (explicit)
+        return explicit;
+    const envExecutor = nonEmpty(env.OMK_EXECUTOR);
+    if (envExecutor)
+        return envExecutor;
+    if (isCodexHost(env))
+        return 'codex';
+    const commandExists = options.commandExists ?? commandExistsOnPath;
+    if (commandExists('codex', env) && !commandExists('claude', env))
+        return 'codex';
+    return 'claude';
+}
+export function resolveCliModel(executor, explicitModel, options = {}) {
+    const env = options.env ?? process.env;
+    const explicit = nonEmpty(explicitModel);
+    if (explicit)
+        return explicit;
+    const envModel = nonEmpty(env.OMK_MODEL);
+    if (envModel)
+        return envModel;
+    if (!isCodexExecutor(executor))
+        return DEFAULT_MODEL;
+    const suggestion = getCodexModelSuggestion(env);
+    if (suggestion.fromConfig)
+        return suggestion.model;
+    const lang = options.lang ?? 'zh';
+    throw new Error(lang === 'zh'
+        ? `Codex 执行器需要明确模型。请用 --model <model>、设置 OMK_MODEL，或在 ${suggestion.configPath} 配置顶层 model。`
+        : `The Codex executor needs an explicit model. Pass --model <model>, set OMK_MODEL, or configure a top-level model in ${suggestion.configPath}.`);
+}
+export function defaultJudgeModel(executor, taskModel) {
+    return isCodexExecutor(executor) ? taskModel : JUDGE_MODEL;
+}
+export function resolveRuntimeSelection(input, options = {}) {
+    const executor = resolveCliExecutor(input.executor, options);
+    const model = resolveCliModel(executor, input.model, options);
+    return {
+        executor,
+        model,
+        judgeModel: defaultJudgeModel(executor, model),
+    };
+}
+export function envJudgeModels(env = process.env) {
+    return nonEmpty(env.OMK_JUDGE_MODELS);
+}

package/dist/diagnosis/observe-mapper.js

@@ -1,4 +1,6 @@
 import { createHash } from 'node:crypto';
+import { maxDiagnosisLifecycle } from './types.js';
+import { ownRecordValue, setOwnRecordValue, sumRecordCounts, } from '../shared/record-count.js';
 const OBSERVE_ONLY_COVERAGE = { observe: true, doctor: false, eval: false };
 export function buildObserveDiagnostics(input) {
     const byStableKey = new Map();
@@ -23,8 +25,9 @@ export function buildObserveDiagnostics(input) {
     }
     const bySkill = {};
     for (const diagnosis of byStableKey.values()) {
-        bySkill[diagnosis.skillName] ??= [];
-        bySkill[diagnosis.skillName].push(diagnosis);
+        const group = ownRecordValue(bySkill, diagnosis.skillName) ?? [];
+        group.push(diagnosis);
+        setOwnRecordValue(bySkill, diagnosis.skillName, group);
     }
     for (const diagnoses of Object.values(bySkill)) {
         diagnoses.sort((a, b) => severityRank(b.severity) - severityRank(a.severity) || a.signal.localeCompare(b.signal));
@@ -158,7 +161,7 @@ function makeDiagnosis(generatedAt, input) {
         `target:${input.target}`,
     ].join('|');
     const occurrence = {
-        id: hashParts('occurrence', stableKey, input.sourceKind, input.sourceId),
+        id: hashParts('occurrence', stableKey, input.sourceKind, input.sourceId, input.timestamp ?? generatedAt),
         diagnosisStableKey: stableKey,
         source: 'observe',
         sourceId: input.sourceId,
@@ -198,10 +201,14 @@ function upsertDiagnosis(byStableKey, next) {
         byStableKey.set(next.stableKey, next);
         return;
     }
-    existing.occurrences.push(...next.occurrences);
-    existing.occurrenceCount = existing.occurrences.length;
+    const existingOccurrenceIds = new Set(existing.occurrences.map((occurrence) => occurrence.id));
+    const hasOverlap = next.occurrences.some((occurrence) => existingOccurrenceIds.has(occurrence.id));
+    existing.occurrences.push(...next.occurrences.filter((occurrence) => !existingOccurrenceIds.has(occurrence.id)));
+    existing.occurrenceCount = hasOverlap
+        ? Math.max(existing.occurrenceCount, next.occurrenceCount)
+        : sumRecordCounts(existing.occurrenceCount, next.occurrenceCount);
     existing.severity = maxSeverity(existing.severity, next.severity);
-    existing.lifecycle = maxLifecycle(existing.lifecycle, next.lifecycle);
+    existing.lifecycle = maxDiagnosisLifecycle(existing.lifecycle, next.lifecycle);
 }
 function typeForPattern(bucket, signal) {
     if (bucket === 'definition_gap')
@@ -251,15 +258,7 @@ export function maxLifecycle(a, b) {
     // 自动 detected 信号,应保留 resolved 而不是把 confirmation 擦掉。
     // 「resolved 后新 occurrence 进来 reopen 到 detected」属于跨时间的 lifecycle 状态机,
     // 由上层 review-state store 决定,不在此处处理。
-    const rank = {
-        resolved: 6,
-        rejected: 5,
-        detected: 4,
-        candidate: 3,
-        confirmed: 2,
-        stale: 1,
-    };
-    return rank[a] >= rank[b] ? a : b;
+    return maxDiagnosisLifecycle(a, b);
 }
 function severityRank(severity) {
     if (severity === 'high')

package/dist/diagnosis/observe-producer.js

@@ -1,6 +1,7 @@
 import { buildObservationSkillChain, buildObservationSkillChains } from '../observability/skill-chain.js';
 import { getSkillChainAdvisory, resolveAdvisoryCommand } from '../observability/skill-chain-advisories.js';
 import { buildObserveDiagnostics } from './observe-mapper.js';
+import { setOwnRecordValue } from '../shared/record-count.js';
 export function buildObserveDiagnosticsFromReport(report, options = {}) {
     const experienceReports = report.experience ? [report.experience] : [];
     const skillNames = skillNamesFromReport(report);
@@ -33,7 +34,7 @@ function resolveSkillChains(report, skillNames, experienceReports, options) {
     for (const skillName of skillNames) {
         const cwd = cwdBySkill.get(skillName);
         if (cwd) {
-            chains[skillName] = buildObservationSkillChain(skillName, cwd, experienceReports);
+            setOwnRecordValue(chains, skillName, buildObservationSkillChain(skillName, cwd, experienceReports));
         }
     }
     return chains;
@@ -184,6 +185,7 @@ function evidenceRef(ref) {
     return {
         id: ref.id,
         kind: ref.kind,
+        traceId: ref.traceId,
         sourceTrace: ref.sourceTrace,
         sessionId: ref.sessionId,
         messageIndex: ref.messageIndex,

package/dist/diagnosis/studio-projection.js

@@ -1,4 +1,5 @@
-import { isActiveDiagnosisLifecycle } from './types.js';
+import { isActiveDiagnosisLifecycle, maxDiagnosisLifecycle, } from './types.js';
+import { incrementRecordCount, ownRecordValue, setOwnRecordValue, sumRecordCounts, } from '../shared/record-count.js';
 export function buildStudioDiagnosisSummary(bundle) {
     const sourceCoverage = bundle?.sourceCoverage ?? { observe: false, doctor: false, eval: false };
     const diagnostics = Object.values(bundle?.bySkill ?? {}).flat();
@@ -7,8 +8,8 @@ export function buildStudioDiagnosisSummary(bundle) {
     const byAudience = {};
     for (const diagnosis of diagnostics) {
         bySeverity[diagnosis.severity] += 1;
-        bySkill[diagnosis.skillName] = (bySkill[diagnosis.skillName] ?? 0) + 1;
-        byAudience[diagnosis.audience] = (byAudience[diagnosis.audience] ?? 0) + 1;
+        incrementRecordCount(bySkill, diagnosis.skillName);
+        incrementRecordCount(byAudience, diagnosis.audience);
     }
     return {
         sourceCoverage,
@@ -40,21 +41,27 @@ export function mergeDiagnosisBundles(bundles, generatedAt = new Date().toISOStr
                 });
                 continue;
             }
-            existing.occurrences.push(...diagnosis.occurrences);
+            const existingOccurrenceIds = new Set(existing.occurrences.map((occurrence) => occurrence.id));
+            const hasOverlap = diagnosis.occurrences.some((occurrence) => existingOccurrenceIds.has(occurrence.id));
+            existing.occurrences.push(...diagnosis.occurrences.filter((occurrence) => !existingOccurrenceIds.has(occurrence.id)));
             // 累加而不是用 occurrences.length 覆盖:聚合型 Diagnosis(如 problemPattern)的
             // occurrenceCount 是「N 次真实发生」的语义,不是「N 条 source occurrence 条数」。
             // 用 length 会把 `3 次 + 4 次` 真实发生压成 `2`(两条 source 条目),Studio 排序和
             // affectedCount 显示偏小。
-            existing.occurrenceCount = existing.occurrenceCount + diagnosis.occurrenceCount;
+            existing.occurrenceCount = hasOverlap
+                ? Math.max(existing.occurrenceCount, diagnosis.occurrenceCount)
+                : sumRecordCounts(existing.occurrenceCount, diagnosis.occurrenceCount);
             existing.severity = severityRank(existing.severity) >= severityRank(diagnosis.severity)
                 ? existing.severity
                 : diagnosis.severity;
+            existing.lifecycle = maxDiagnosisLifecycle(existing.lifecycle, diagnosis.lifecycle);
         }
     }
     const bySkill = {};
     for (const diagnosis of byStableKey.values()) {
-        bySkill[diagnosis.skillName] ??= [];
-        bySkill[diagnosis.skillName].push(diagnosis);
+        const group = ownRecordValue(bySkill, diagnosis.skillName) ?? [];
+        group.push(diagnosis);
+        setOwnRecordValue(bySkill, diagnosis.skillName, group);
     }
     for (const values of Object.values(bySkill)) {
         values.sort((a, b) => severityRank(b.severity) - severityRank(a.severity) || a.signal.localeCompare(b.signal));

package/dist/diagnosis/types.d.ts

@@ -1,3 +1,5 @@
 export type { Diagnosis, DiagnosisAudience, DiagnosisBundle, DiagnosisEvidenceRef, DiagnosisLifecycle, DiagnosisOccurrence, DiagnosisPatch, DiagnosisScope, DiagnosisSeverity, DiagnosisSource, DiagnosisSourceCoverage, DiagnosisType, StudioDiagnosisSummary, } from '../types/diagnosis.js';
 import type { DiagnosisLifecycle } from '../types/diagnosis.js';
 export declare function isActiveDiagnosisLifecycle(lifecycle: DiagnosisLifecycle): boolean;
+/** Merge lifecycle states without making the result depend on bundle order. */
+export declare function maxDiagnosisLifecycle(a: DiagnosisLifecycle, b: DiagnosisLifecycle): DiagnosisLifecycle;

package/dist/diagnosis/types.js

@@ -17,3 +17,15 @@ const ACTIVE_DIAGNOSIS_LIFECYCLES = new Set([
 export function isActiveDiagnosisLifecycle(lifecycle) {
     return ACTIVE_DIAGNOSIS_LIFECYCLES.has(lifecycle);
 }
+/** Merge lifecycle states without making the result depend on bundle order. */
+export function maxDiagnosisLifecycle(a, b) {
+    const rank = {
+        resolved: 6,
+        rejected: 5,
+        detected: 4,
+        candidate: 3,
+        confirmed: 2,
+        stale: 1,
+    };
+    return rank[a] >= rank[b] ? a : b;
+}

package/dist/doctor/endpoint-rule.js

@@ -25,6 +25,7 @@
  */
 import { existsSync, readFileSync, readdirSync, lstatSync, openSync, readSync, closeSync } from 'node:fs';
 import { join } from 'node:path';
+import { setOwnRecordValue } from '../shared/record-count.js';
 const DEFAULT_MAX_FILE_BYTES = 200 * 1024;
 const DEFAULT_MAX_TOTAL_BYTES = 2 * 1024 * 1024;
 /** 二进制 / 大文件不可读时跳过;只收文本。简单按扩展名 + 内容嗅探。 */
@@ -125,7 +126,7 @@ function collectFiles(skillRoot, maxFileBytes, maxTotalBytes) {
                 const over = buf.length > cap;
                 const slice = over ? buf.subarray(0, cap) : buf;
                 const text = slice.toString('utf-8') + (over ? '\n…[truncated]' : '');
-                files[relPath] = text;
+                setOwnRecordValue(files, relPath, text);
                 total += Buffer.byteLength(text, 'utf-8');
             }
         }

package/dist/eval-core/artifact-file-names.js

@@ -1,3 +1,4 @@
+import { createHash } from 'node:crypto';
 import { join } from 'node:path';
 export const REPORT_FILE_SUFFIX = '.report.json';
 export const GRAPH_FILE_SUFFIX = '.graph.json';
@@ -41,5 +42,21 @@ export function stripDomainPrefix(id, domain) {
     return safeId.startsWith(prefix) ? safeId.slice(prefix.length) : safeId;
 }
 export function doctorReportFileStem(skillName, reportId) {
-    return `${safeArtifactFileStem(skillName)}-${stripDomainPrefix(reportId, 'doctor')}`;
+    const reportSuffix = reportId.startsWith('doctor-')
+        ? reportId.slice('doctor-'.length)
+        : reportId;
+    return [
+        collisionResistantFileComponent(skillName),
+        collisionResistantFileComponent(reportSuffix),
+    ].join('-');
+}
+function collisionResistantFileComponent(value) {
+    const safe = safeArtifactFileStem(value);
+    if (safe === value)
+        return safe;
+    const fingerprint = createHash('sha256')
+        .update(value)
+        .digest('hex')
+        .slice(0, 12);
+    return `${safe}-${fingerprint}`;
 }

package/dist/eval-core/artifact-index.d.ts

@@ -1,5 +1,4 @@
 import type { ReportDocument, ReportIndexCard } from '../types/index.js';
-import type { SkillDoctorSnapshot } from '../types/skill-index.js';
 import type { DoctorSkillStatus } from '../types/doctor.js';
 export type ArtifactDomain = 'report' | 'doctor' | 'observe-health';
 /** 某域的索引目录 `<root>/<domain>/`。 */
@@ -17,15 +16,11 @@ export declare function cardTargetSentinel(domain: ArtifactDomain): string;
  * persistReport 的索引钩子:报告落盘后 best-effort 追加卡片。永不抛、永不阻断报告落盘。
  * 入参故意收 `{id}` 宽松(同 persistReport 的 PersistableReport):非完整报告(无 canonical kind)防御式跳过。
  */
-export declare function indexReportWrite(report: {
-    id: string;
-}, sourcePath: string, outputDir: string): void;
+export declare function indexReportWrite(report: ReportDocument, sourcePath: string, outputDir: string): void;
 /** 读 report 域全部卡片(跳过坏文件 / 缺字段 / 坏 kind)。 */
 export declare function listReportCards(): ReportIndexCard[];
 /** report 卡片(过滤悬空真身):供 studio 机器级 list / findBy 展示。 */
 export declare function listLiveReportCards(): ReportIndexCard[];
-/** 卡片 → ReportDocument(results:[]):供 studio list / buildSkillIndex / trend 消费(它们不读 results)。 */
-export declare function cardToReportDocument(card: ReportIndexCard): ReportDocument;
 /** 删 report 域某 id 的卡片(DELETE 报告时连卡片一起删,使其从机器级 list 消失)。幂等、best-effort。 */
 export declare function removeReportCard(id: string): boolean;
 /** doctor 卡片:per-skill 文件(`{skillName}-{reportId}.json`)一张。id = 文件名 stem(唯一),
@@ -49,22 +44,23 @@ export declare function indexDoctorWrite(card: Omit<DoctorIndexCard, 'domain'>,
 export declare function listDoctorCards(): DoctorIndexCard[];
 /** doctor 卡片(过滤悬空真身):供 buildSkillIndex 机器级合并。 */
 export declare function listLiveDoctorCards(): DoctorIndexCard[];
-/** doctor 卡片 → (skillName, SkillDoctorSnapshot)。results:[](逐规则详情需回源项目看)。 */
-export declare function cardToDoctorSnapshot(card: DoctorIndexCard): {
-    skillName: string;
-    snap: SkillDoctorSnapshot;
-};
 /** 删 doctor 域某 id(文件 stem)的卡片。doctor 历史按 50/skill prune,删正文时必须连卡片一起删,
  *  否则被 prune 掉的报告会经卡片合并在本项目 studio「复活」。 */
 export declare function removeDoctorCard(id: string): boolean;
 /** observe 报告投影到卡片所需的结构子集(不 import observability/SkillHealthReport,保持 eval-core 解耦)。 */
 interface ObserveCardSkill {
     toolFailureRate: number;
+    toolFailureCount?: number;
+    toolCallCount?: number;
+    toolResolvedCount?: number;
+    toolCancelledCount?: number;
+    toolUnknownCount?: number;
     segmentCount: number;
     gap?: {
         weightedGapRate?: number;
     };
     confidence?: 'high' | 'low' | 'underpowered';
+    stability?: 'stable' | 'unstable' | 'very-unstable' | 'unknown';
 }
 interface ObserveSource {
     meta: {