oh-my-knowledge 0.49.0 → 0.51.0

package/dist/assets/agent-skills/omk/references/commands.md

@@ -599,7 +599,7 @@ omk sample [skillPath] [flags]
 - `--executor` `option`:执行器名。Codex 任务内自动用 codex；也可用 OMK_EXECUTOR 设置环境偏好。
 - `--fix` `boolean`:fix 模式：基于最近评测报告自动修复 sample_design 类型失败。
 - `--focus` `option`:生成焦点（自然语言提示）。控制 LLM 偏向哪类用例。
-- `--from-traces` `boolean`:from-traces 模式：从 observe inbox 的失败信号回流生成回归用例草稿（provenance: production-trace），落草稿待人工 review。
+- `--from-traces` `boolean`:from-traces 模式：从 observe inbox 的失败信号回流生成评测用例草稿（provenance: production-trace），落草稿待人工 review。
 - `--lang` `option` (默认 `zh`):输出语言 zh|en，优先级 CLI > OMK_LANG env > zh。
 - `--model` `option`:生成 LLM model 名。Codex 自动读取本机配置；也可用 OMK_MODEL 设置环境偏好。
 - `--no-mock` `boolean`:不生成 mocks，eval 时所有工具调用真实执行。
@@ -629,7 +629,7 @@ omk sample --batch --skill-dir skills
 omk sample skills/my-skill/SKILL.md --fix
 ```
 
-> 从 observe inbox 的失败信号回流生成回归用例草稿
+> 从 observe inbox 的失败信号回流生成评测用例草稿
 
 ```bash
 omk sample --from-traces

package/dist/cli/commands/observe/inbox.js

@@ -100,7 +100,7 @@ export async function runObserveInbox(_args, flags, lang) {
         }
         if (recyclableCount > 0) {
             console.log(lang === 'zh'
-                ? `提示：确认信号后生成回归用例草稿：${sampleCommand}`
+                ? `提示：确认信号后生成评测用例草稿：${sampleCommand}`
                 : `Tip: after confirming signals, draft regression samples: ${sampleCommand}`);
         }
         return;
@@ -142,7 +142,7 @@ export async function runObserveInbox(_args, flags, lang) {
         : 'Tip: omk observe inbox --explore 10 --include-noise  # explicitly include the noise bucket');
     if (recyclableCount > 0) {
         console.log(lang === 'zh'
-            ? `提示：确认高风险或抽样信号后生成回归用例草稿：${sampleCommand}`
+            ? `提示：确认高风险或抽样信号后生成评测用例草稿：${sampleCommand}`
             : `Tip: after confirming high-risk / sampled signals, draft regression samples: ${sampleCommand}`);
     }
 }

package/dist/cli/commands/observe/ingest.js

@@ -103,7 +103,7 @@ export default class ObserveIngest extends BaseCommand {
             const inboxCommand = `omk observe inbox --input-dir ${shellQuoteArg(outDir)}`;
             const sampleCommand = `omk sample --from-traces --observations-dir ${shellQuoteArg(outDir)}`;
             process.stderr.write(lang === 'zh'
-                ? `observe inbox 已写入：${path}\n下一步：${inboxCommand}\n确认高风险或抽样信号后，可生成回归用例草稿：${sampleCommand}\n`
+                ? `observe inbox 已写入：${path}\n下一步：${inboxCommand}\n确认高风险或抽样信号后，可生成评测用例草稿：${sampleCommand}\n`
                 : `observe inbox written to: ${path}\nNext: ${inboxCommand}\nAfter confirming high-risk / sampled signals, draft regression samples: ${sampleCommand}\n`);
         });
     }

package/dist/cli/commands/sample.js

@@ -360,7 +360,7 @@ export async function runSampleFromTraces(flags, lang) {
     }
     const count = flags.count !== undefined ? Math.max(1, Number(flags.count) || 5) : undefined;
     process.stderr.write(lang === 'zh'
-        ? `🔭 发现 ${items.length} 个${flags.skill ? ` ${flags.skill} 的` : ''}失败信号，正在生成回归用例草稿...\n`
+        ? `🔭 发现 ${items.length} 个${flags.skill ? ` ${flags.skill} 的` : ''}失败信号，正在生成评测用例草稿...\n`
         : `🔭 Found ${items.length}${flags.skill ? ` ${flags.skill}` : ''} failure signal(s); generating regression-sample drafts...\n`);
     try {
         const { samples, costUSD } = await generateSamplesFromTraces({ items, count, model, executorName });
@@ -598,7 +598,7 @@ export default class Sample extends BaseCommand {
         },
         {
             description: bilingual({
-                zh: '从 observe inbox 的失败信号回流生成回归用例草稿',
+                zh: '从 observe inbox 的失败信号回流生成评测用例草稿',
                 en: 'Recycle observe-inbox failure signals into draft regression samples',
             }),
             command: '<%= config.bin %> sample --from-traces',
@@ -689,7 +689,7 @@ export default class Sample extends BaseCommand {
         }),
         'from-traces': Flags.boolean({
             description: bilingual({
-                zh: 'from-traces 模式：从 observe inbox 的失败信号回流生成回归用例草稿（provenance: production-trace），落草稿待人工 review。',
+                zh: 'from-traces 模式：从 observe inbox 的失败信号回流生成评测用例草稿（provenance: production-trace），落草稿待人工 review。',
                 en: 'from-traces mode: recycle observe-inbox failure signals into draft regression samples (provenance: production-trace) for review.',
             }),
             default: false,

package/dist/observability/experience.js

@@ -3262,21 +3262,23 @@ function skillSegmentsWithOrchestrationRoles(skillSegments, orchestrationEdges)
     const parentIds = new Set(orchestrationEdges.map((edge) => edge.parentSkillSegmentId).filter((value) => Boolean(value)));
     const executorIds = new Set(orchestrationEdges.map((edge) => edge.executorSkillSegmentId).filter((value) => Boolean(value)));
     return skillSegments.map((segment) => {
-        if (parentIds.has(segment.id) && segment.skillType === 'unknown') {
+        if (parentIds.has(segment.id)) {
+            const inferredFromTrace = !segment.declaredSkillType;
             return {
                 ...segment,
-                skillType: 'router',
-                skillTypeSource: 'trace',
-                traceInferredSkillType: 'router',
+                skillType: inferredFromTrace ? 'router' : segment.skillType,
+                skillTypeSource: inferredFromTrace ? 'trace' : segment.skillTypeSource,
+                traceInferredSkillType: inferredFromTrace ? 'router' : segment.traceInferredSkillType,
                 episodeRole: 'router',
             };
         }
-        if (executorIds.has(segment.id) && segment.episodeRole === 'supporting') {
+        if (executorIds.has(segment.id)) {
+            const inferredFromTrace = !segment.declaredSkillType;
             return {
                 ...segment,
-                skillType: segment.skillType === 'unknown' ? 'executor' : segment.skillType,
-                skillTypeSource: segment.skillType === 'unknown' ? 'trace' : segment.skillTypeSource,
-                traceInferredSkillType: segment.traceInferredSkillType ?? 'executor',
+                skillType: inferredFromTrace ? 'executor' : segment.skillType,
+                skillTypeSource: inferredFromTrace ? 'trace' : segment.skillTypeSource,
+                traceInferredSkillType: inferredFromTrace ? 'executor' : segment.traceInferredSkillType,
                 episodeRole: 'main_executor',
             };
         }
@@ -3496,40 +3498,109 @@ function sessionStoryOrchestrationEdges(episodeId, skillSegments, subagentDispat
             });
         }
     }
-    if (edges.length === 0) {
-        const fallbackEdge = fallbackOrchestrationEdgeFromRuntime(episodeId, skillSegments, invocations);
-        if (fallbackEdge)
-            edges.push(fallbackEdge);
-    }
     for (const dispatch of subagentDispatches) {
-        const parentSegment = delegator ?? router ?? runnerOwner;
-        const executor = skillSegments.find((segment) => segment.id !== parentSegment?.id
-            && segment.evidenceRefs.some((ref) => ref.traceId === dispatch.traceId)) ?? skillSegments.find((segment) => segment.id !== parentSegment?.id && segment.episodeRole === 'main_executor');
-        const dispatchRunnerRef = runnerRef && (!dispatch.attachTo?.callInstanceId
-            ? !dispatch.attachTo?.toolUseId
-                || dispatch.attachTo.toolUseId === runnerRef.toolUseId
-            : dispatch.attachTo.callInstanceId === runnerRef.callInstanceId)
-            ? runnerRef
-            : undefined;
+        const executor = skillSegmentForTrace(skillSegments, dispatch.traceId, dispatch.sourceTrace);
+        const parentSegment = dispatchParentSkillSegment(skillSegments, executor, session, dispatch, router, delegator, runnerOwner);
+        const distinctExecutor = executor?.id === parentSegment?.id ? undefined : executor;
+        const dispatchRunnerRef = dispatchAttachmentEvidenceRef(session, dispatch);
+        const terminal = dispatchTerminalLifecycle(session, dispatch);
         edges.push({
             id: hashParts('session-story-edge', episodeId, dispatch.id),
             episodeId,
             edgeKind: 'external_child_session',
             parentSkillSegmentId: parentSegment?.id,
-            executorSkillSegmentId: executor?.id,
+            executorSkillSegmentId: distinctExecutor?.id,
             childSessionId: dispatch.childSessionId,
             runnerStartedRef: dispatchRunnerRef,
-            status: 'started',
+            runnerCompletedRef: terminal?.status === 'completed' ? terminal.evidenceRef : undefined,
+            status: terminal?.status ?? 'started',
             evidenceRefs: uniqueEvidenceRefs([
                 ...(parentSegment?.evidenceRefs.slice(0, 2) ?? []),
-                ...(executor?.evidenceRefs.slice(0, 2) ?? []),
+                ...(distinctExecutor?.evidenceRefs.slice(0, 2) ?? []),
                 ...(dispatchRunnerRef ? [dispatchRunnerRef] : []),
+                ...(terminal ? [terminal.evidenceRef] : []),
                 ...dispatch.evidenceRefs,
             ]).slice(0, 6),
         });
     }
+    if (edges.length === 0 && subagentDispatches.length === 0) {
+        const fallbackEdge = fallbackOrchestrationEdgeFromRuntime(episodeId, skillSegments, invocations);
+        if (fallbackEdge)
+            edges.push(fallbackEdge);
+    }
     return edges;
 }
+function skillSegmentForTrace(skillSegments, traceId, sourceTrace) {
+    return skillSegments
+        .filter((segment) => (segment.messageRanges ?? []).some((range) => range.traceId === traceId || range.sourceTrace === sourceTrace)
+        || segment.evidenceRefs.some((ref) => ref.traceId === traceId || ref.sourceTrace === sourceTrace))
+        .sort((a, b) => a.order - b.order)[0];
+}
+function dispatchParentSkillSegment(skillSegments, executor, session, dispatch, router, delegator, runnerOwner) {
+    const mainTraceIds = new Set((session.timelineTree?.main ?? [])
+        .map((event) => event.traceId)
+        .filter((value) => Boolean(value)));
+    const mainSourceTraces = new Set((session.timelineTree?.main ?? [])
+        .map((event) => event.sourceTrace)
+        .filter(Boolean));
+    const isMainline = (segment) => (segment.messageRanges ?? []).some((range) => Boolean(range.traceId && mainTraceIds.has(range.traceId))
+        || Boolean(range.sourceTrace && mainSourceTraces.has(range.sourceTrace)))
+        || segment.evidenceRefs.some((ref) => ref.traceRole === 'main'
+            || ref.traceRole === 'standalone'
+            || Boolean(ref.traceId && mainTraceIds.has(ref.traceId))
+            || mainSourceTraces.has(ref.sourceTrace));
+    const preferred = [router, delegator]
+        .filter((segment) => Boolean(segment))
+        .find((segment) => segment.id !== executor?.id && isMainline(segment));
+    if (preferred)
+        return preferred;
+    const dispatchStart = minString(dispatch.evidenceRefs.map((ref) => ref.timestamp));
+    const mainline = skillSegments
+        .filter((segment) => segment.id !== executor?.id && isMainline(segment))
+        .filter((segment) => !dispatchStart || segment.startTimestamp <= dispatchStart)
+        .sort((a, b) => b.startTimestamp.localeCompare(a.startTimestamp)
+        || b.order - a.order)[0];
+    if (mainline)
+        return mainline;
+    if (runnerOwner?.id !== executor?.id && runnerOwner && isMainline(runnerOwner))
+        return runnerOwner;
+    return executor && isMainline(executor) ? executor : undefined;
+}
+function dispatchAttachmentEvidenceRef(session, dispatch) {
+    const attachTo = dispatch.attachTo;
+    if (!attachTo)
+        return undefined;
+    if (!attachTo.callInstanceId && !attachTo.toolUseId && attachTo.messageIndex === undefined) {
+        return undefined;
+    }
+    const event = (session.timelineTree?.main ?? []).find((candidate) => {
+        if (candidate.kind !== 'tool_use')
+            return false;
+        if (attachTo.callInstanceId) {
+            return candidate.callInstanceId === attachTo.callInstanceId;
+        }
+        if (attachTo.toolUseId && candidate.toolUseId !== attachTo.toolUseId)
+            return false;
+        return attachTo.messageIndex === undefined
+            || candidate.messageIndex === attachTo.messageIndex;
+    });
+    return event ? evidenceRefFromTimeline(event) : undefined;
+}
+function dispatchTerminalLifecycle(session, dispatch) {
+    const branch = session.timelineTree?.branches.find((candidate) => candidate.id === dispatch.branchId
+        || candidate.traceId === dispatch.traceId
+        || candidate.sourceTrace === dispatch.sourceTrace);
+    const event = branch?.events.at(-1);
+    if (event?.kind !== 'runtime_context')
+        return undefined;
+    if (event.label === 'turn_completed' || event.label === 'session_ended') {
+        return { status: 'completed', evidenceRef: evidenceRefFromTimeline(event) };
+    }
+    if (event.label === 'turn_aborted' || event.label === 'turn_interrupted') {
+        return { status: 'failed', evidenceRef: evidenceRefFromTimeline(event) };
+    }
+    return undefined;
+}
 function bestPriorUpstreamSkillSegmentForRuntime(runnerOwner, runnerEvent, timeline, skillSegments) {
     const runnerMessageIndex = runnerEvent?.messageIndex;
     const contextText = timeline
@@ -3641,16 +3712,8 @@ function isOrchestrationRuntimeEvent(event) {
         return /^(?:Task|Agent)$/i.test(event.toolName ?? '')
             || /runner\.js|send-input\.js|check-session\.js/i.test(text);
     }
-    if (event.kind === 'tool_result') {
-        const trimmed = text.trim();
-        if (/^---\s*\n?\s*name:/i.test(trimmed) || /^---\s+name:/i.test(trimmed))
-            return false;
-        return /"event"\s*:\s*"started"|Command still running \(session|Process exited with code|Process exited with signal/i.test(text)
-            || /["']?(?:child_?session_?id|agent_?id|thread_?id|session_?id)["']?\s*[:=]/i.test(text)
-            || /\b(?:session|thread|agent)(?:\s+id)?\s*[:=]\s*[a-z0-9][a-z0-9._-]*/i.test(text);
-    }
     if (event.kind === 'assistant_message') {
-        return /ttyd|(?:已启动|启动了|spawned|started).*(?:subagent|sub-agent|child agent|子\s*(?:agent|代理|智能体|Claude|Codex))|(?:subagent|sub-agent|child agent|子\s*(?:agent|代理|智能体|Claude|Codex)).*(?:执行|运行|分析|started|running)/i.test(text)
+        return /ttyd|(?:已启动|启动了|spawned|started).{0,80}(?:subagent|sub-agent|child agent|子\s*(?:agent|代理|智能体|Claude|Codex))|(?:subagent|sub-agent|child agent|子\s*(?:agent|代理|智能体|Claude|Codex)).{0,48}(?:已启动|启动中|正在(?:执行|运行|分析)|spawned|started|running)/i.test(text)
             || /\b(?:session|thread|agent)(?:\s+id)?\s*[:=]\s*[a-z0-9][a-z0-9._-]*/i.test(text);
     }
     return false;
@@ -4209,16 +4272,10 @@ function inferSkillRole(group, allInvocations, session) {
     return 'unknown';
 }
 function routingEvidenceEvents(invocation) {
-    return invocation.timeline.filter((event) => {
-        if (event.kind !== 'assistant_message' && event.kind !== 'tool_use')
-            return false;
-        if (isOrchestrationRuntimeEvent(event))
-            return true;
-        if (isDifferentSkillInvocationEvent(event, invocation.skillName))
-            return true;
-        const text = event.fullText ?? event.snippet ?? '';
-        return /subagent|sub-agent|child agent|子\s*(?:agent|代理|智能体|Claude|Codex)|子任务|分发|委派|路由|delegate|dispatch|route|调用.+skill|走\s*`?[\w-]+`?\s*skill/i.test(text);
-    }).slice(0, 3);
+    return invocation.timeline
+        .filter((event) => isOrchestrationRuntimeEvent(event)
+        || isDifferentSkillInvocationEvent(event, invocation.skillName))
+        .slice(0, 3);
 }
 function isDifferentSkillInvocationEvent(event, currentSkillName) {
     if (event.kind !== 'tool_use' || !/^Skill$/i.test(event.toolName ?? ''))

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-knowledge",
-  "version": "0.49.0",
+  "version": "0.51.0",
   "packageManager": "yarn@4.16.0",
   "description": "Evaluation framework for LLM knowledge inputs — prompts, RAG corpora, skills, agent workflows. Fix the model, vary the artifact. Built-in statistical rigor: bootstrap CI, Krippendorff α, length-debias, saturation curves.",
   "type": "module",