@hongmaple0820/scale-engine 0.28.0 → 0.33.0

package/dist/runtime/AiOsRuntime.js

@@ -5,10 +5,14 @@ import { createGovernanceRoiReport, } from '../governance/GovernanceRoi.js';
 import { evaluateProgressiveGovernance, } from '../governance/ProgressiveGovernance.js';
 import { MemoryFabric, recallMemoryProviders, } from '../memory/index.js';
 import { createSkillPlan, loadSkillRoutingPolicy, } from '../skills/routing/index.js';
+import { routeAdaptiveWorkflow } from '../workflow/AdaptiveWorkflowRouter.js';
+import { proposeShadowRule, buildEvolutionShadowReport, } from '../workflow/EvolutionShadowPromoter.js';
 import { runSafeCommand } from '../tools/SafeCommandRunner.js';
 import { SCALE_ENGINE_VERSION } from '../version.js';
 import { resolveVerificationTargets, } from '../workflow/VerificationProfile.js';
 import { RuntimeEvidenceLedger } from './RuntimeEvidenceLedger.js';
+import { loadRelevantLearnings } from '../evolution/SessionLearnings.js';
+import { collectSessionPreamble } from '../workflow/SessionPreamble.js';
 export async function createAiOsPlan(input) {
     const projectDir = resolve(input.projectDir ?? process.cwd());
     const scaleDir = input.scaleDir ?? '.scale';
@@ -17,6 +21,8 @@ export async function createAiOsPlan(input) {
     const services = input.services ?? [];
     const taskId = input.taskId;
     const budget = input.budget ?? 8_000;
+    const preamble = collectSessionPreamble({ projectDir, scaleDir });
+    const sessionLearnings = loadRelevantLearnings({ projectDir, scaleDir, task: input.task, limit: 5 });
     const governance = evaluateProgressiveGovernance({
         task: input.task,
         changedFiles: files,
@@ -62,7 +68,15 @@ export async function createAiOsPlan(input) {
         services,
         policy: skillPolicy,
     });
-    const adaptiveWorkflow = createAdaptiveWorkflow(governance, skillPlan);
+    const evaluator = createEvaluatorIntelligence({
+        task: input.task,
+        files,
+        governance,
+        skillPlan,
+    });
+    const toolStrategy = createToolStrategyPlan(skillPlan);
+    const adaptiveWorkflow = createAdaptiveWorkflow(governance, skillPlan, evaluator, toolStrategy);
+    const evolutionShadow = createEvolutionShadowProposals(governance, evaluator);
     const roi = createGovernanceRoiReport({
         taskId,
         contextBudget,
@@ -81,8 +95,12 @@ export async function createAiOsPlan(input) {
             files,
             services,
         },
+        preamble,
         governance,
         adaptiveWorkflow,
+        evaluator,
+        toolStrategy,
+        evolutionShadow,
         context,
         memory: {
             providerOrder: memoryRecall.providerOrder,
@@ -93,8 +111,9 @@ export async function createAiOsPlan(input) {
             contextPack: memoryPack,
         },
         skillPlan,
+        sessionLearnings,
         roi,
-        recommendations: recommendations({ governance, context, memoryRecall, skillPlan }),
+        recommendations: recommendations({ governance, context, memoryRecall, skillPlan, evaluator, toolStrategy }),
     };
 }
 export async function createAiOsRun(input) {
@@ -194,6 +213,7 @@ export async function createAiOsBenchmark(input = {}) {
             task: scenario.task,
             level: scenario.level,
             governanceMode: plan.governance.effectiveMode,
+            workflowProfile: plan.adaptiveWorkflow.profile,
             metrics: {
                 estimatedTokens: plan.context.totalEstimatedTokens,
                 budget: plan.context.task.budget,
@@ -202,6 +222,10 @@ export async function createAiOsBenchmark(input = {}) {
                 selectedProviders: plan.memory.selectedProviders,
                 skillSteps: plan.skillPlan.executionPlan.steps.length,
                 requiredSkillSteps: plan.skillPlan.executionPlan.steps.filter(step => step.required).length,
+                evaluatorGates: plan.evaluator.gates.length,
+                toolStrategySteps: plan.toolStrategy.summary.totalSteps,
+                toolStrategyCostUnits: plan.toolStrategy.summary.estimatedCostUnits,
+                evolutionProposals: plan.evolutionShadow.summary.totalProposals,
                 gates: plan.adaptiveWorkflow.gates.length,
                 roiModules: plan.roi.modules.length,
             },
@@ -446,6 +470,15 @@ export function createAiOsStatus(input = {}) {
     const verificationRecommendations = buildVerificationRecommendations(projectDir, scaleDir, lang);
     const benchmarkReport = resolveBenchmarkReportPath(projectDir, scaleDir);
     const adoptionReport = resolveAdoptionReportPath(projectDir, scaleDir);
+    const intelligence = buildAiOsIntelligenceReport({
+        projectDir,
+        scaleDir,
+        runReports,
+        benchmark: readAiOsBenchmarkReport(benchmarkReport, warnings),
+        benchmarkStatus: doctor.benchmark.status,
+        benchmarkReport,
+        lang,
+    });
     const checks = [
         {
             id: 'runtime-dirs',
@@ -526,11 +559,351 @@ export function createAiOsStatus(input = {}) {
         summary,
         dashboard,
         doctor,
+        intelligence,
         verificationRecommendations,
         nextActions: aiOsStatusNextActions(status, checks, lang, verificationRecommendations),
         warnings: [...warnings, ...doctor.warnings],
     };
 }
+function buildAiOsIntelligenceReport(input) {
+    const runMemoryItems = input.runReports.flatMap(report => report.plan.memory.items);
+    const benchmarkMemoryItems = input.benchmark?.summary.totalMemoryItems ?? 0;
+    const runProviders = input.runReports.flatMap(report => report.plan.memory.selectedProviders);
+    const benchmarkProviders = input.benchmark?.scenarios.flatMap(scenario => scenario.metrics.selectedProviders) ?? [];
+    const selectedProviders = [...new Set([...runProviders, ...benchmarkProviders])].sort();
+    const runTokenSavings = input.runReports.reduce((sum, report) => sum + (report.plan.context.compiler?.estimatedTokenSavings ?? 0), 0);
+    const benchmarkTokenSavings = input.benchmark?.summary.totalEstimatedTokenSavings ?? 0;
+    const estimatedTokenSavings = runTokenSavings + benchmarkTokenSavings;
+    const runSkillSteps = input.runReports.reduce((sum, report) => sum + report.plan.skillPlan.executionPlan.steps.length, 0);
+    const benchmarkSkillSteps = input.benchmark?.summary.totalSkillSteps ?? 0;
+    const skillSteps = runSkillSteps + benchmarkSkillSteps;
+    const totalMemoryItems = runMemoryItems.length + benchmarkMemoryItems;
+    const memoryQuality = summarizeMemoryQuality(runMemoryItems);
+    const contextQuality = summarizeContextQuality(input.runReports);
+    const evaluatorQuality = summarizeEvaluatorQuality(input.runReports, input.benchmark);
+    const toolStrategyQuality = summarizeToolStrategyQuality(input.runReports, input.benchmark);
+    const evolutionQuality = summarizeEvolutionQuality(input.runReports, input.benchmark);
+    const contextSignalStatus = contextQuality.compressionRisk === 'high'
+        ? 'warning'
+        : estimatedTokenSavings > 0 ? 'ready' : input.runReports.length > 0 || input.benchmark ? 'warning' : 'blocked';
+    const memoryEvidence = [
+        ...runMemoryItems.map(item => `${item.provider}:${item.id}`),
+        ...(benchmarkMemoryItems > 0 ? [`benchmark:${input.benchmarkReport}:${benchmarkMemoryItems}`] : []),
+    ];
+    const contextEvidence = [
+        ...input.runReports.map(report => `${report.artifacts.runReport}:saved=${report.plan.context.compiler?.estimatedTokenSavings ?? 0}`),
+        ...(input.benchmark ? [`${input.benchmarkReport}:saved=${input.benchmark.summary.totalEstimatedTokenSavings}`] : []),
+    ];
+    const skillEvidence = [
+        ...input.runReports.flatMap(report => report.plan.skillPlan.executionPlan.steps.map(step => `${report.artifacts.runReport}:${step.id}`)),
+        ...(input.benchmark ? [`${input.benchmarkReport}:steps=${input.benchmark.summary.totalSkillSteps}`] : []),
+    ];
+    const evaluatorEvidence = [
+        ...input.runReports.flatMap(report => resolveRunEvaluator(report).gates.map(gate => `${report.artifacts.runReport}:${gate.id}`)),
+        ...(input.benchmark ? [`${input.benchmarkReport}:evaluator-gates=${input.benchmark.summary.totalEvaluatorGates}`] : []),
+    ];
+    const toolStrategyEvidence = [
+        ...input.runReports.flatMap(report => resolveRunToolStrategy(report).nodes.map(node => `${report.artifacts.runReport}:${node.id}`)),
+        ...(input.benchmark ? [`${input.benchmarkReport}:tool-strategy=${input.benchmark.summary.totalToolStrategySteps}`] : []),
+    ];
+    const evolutionEvidence = [
+        ...input.runReports.flatMap(report => (report.plan.evolutionShadow?.proposals ?? []).map(p => `${report.artifacts.runReport}:${p.id}:${p.maturity.stage}`)),
+        ...(input.benchmark ? [`${input.benchmarkReport}:evolution-proposals=${input.benchmark.summary.totalEvolutionProposals}`] : []),
+    ];
+    const benchmarkEvidence = input.benchmark ? [
+        `${input.benchmarkReport}:scenarios=${input.benchmark.summary.scenarios}`,
+        `${input.benchmarkReport}:memory=${input.benchmark.summary.totalMemoryItems}`,
+        `${input.benchmarkReport}:skills=${input.benchmark.summary.totalSkillSteps}`,
+        `${input.benchmarkReport}:evaluator-gates=${input.benchmark.summary.totalEvaluatorGates}`,
+        `${input.benchmarkReport}:tool-strategy=${input.benchmark.summary.totalToolStrategySteps}`,
+    ] : [input.benchmarkReport];
+    const signals = [
+        {
+            id: 'memory-recall',
+            status: totalMemoryItems > 0 ? 'ready' : selectedProviders.length > 0 ? 'warning' : 'blocked',
+            summary: totalMemoryItems > 0
+                ? `${totalMemoryItems} memory item(s) recalled through ${selectedProviders.join(', ') || 'configured providers'}; quality ${memoryQuality.score}/100.`
+                : selectedProviders.length > 0
+                    ? `Memory providers were selected (${selectedProviders.join(', ')}) but no relevant item was recalled.`
+                    : 'No memory recall evidence found in AI OS runs or benchmarks.',
+            evidence: memoryEvidence,
+            recommendations: totalMemoryItems > 0
+                ? ['Keep recording memory item ids with every run so later context assembly can explain recall.']
+                : ['Run an AI OS task that should match durable project memory before claiming memory intelligence.'],
+        },
+        {
+            id: 'context-savings',
+            status: contextSignalStatus,
+            summary: estimatedTokenSavings > 0
+                ? `${estimatedTokenSavings} estimated token(s) saved by context compilation evidence; compression risk ${contextQuality.compressionRisk}.`
+                : 'Context compiler evidence exists but has not shown measurable token savings yet.',
+            evidence: contextEvidence,
+            recommendations: contextQuality.evidenceLossWarnings.length > 0
+                ? ['Review omitted evidence-bearing context before claiming the task has enough context.']
+                : estimatedTokenSavings > 0
+                    ? ['Track savings deltas across releases before publishing token reduction claims.']
+                    : ['Add larger representative tasks to benchmark context slicing and token savings.'],
+        },
+        {
+            id: 'skill-routing',
+            status: skillSteps > 0 ? 'ready' : input.runReports.length > 0 || input.benchmark ? 'warning' : 'blocked',
+            summary: skillSteps > 0
+                ? `${skillSteps} skill routing step(s) planned across runs and benchmark scenarios.`
+                : 'No skill routing step evidence found.',
+            evidence: skillEvidence,
+            recommendations: skillSteps > 0
+                ? ['Use skill routing evidence in reviews to check why a skill, MCP, or CLI path was selected.']
+                : ['Create a task with files or services that should trigger required skill routing.'],
+        },
+        {
+            id: 'evaluator-intelligence',
+            status: evaluatorQuality.requiredGates > 0
+                ? evaluatorQuality.averageUncertainty >= 0.7 ? 'warning' : 'ready'
+                : input.runReports.length > 0 || input.benchmark ? 'warning' : 'blocked',
+            summary: evaluatorQuality.requiredGates > 0
+                ? `${evaluatorQuality.requiredGates} evaluator gate(s) required; average uncertainty ${evaluatorQuality.averageUncertainty}.`
+                : 'No evaluator gate evidence found for architecture, root-cause, security, or release reasoning.',
+            evidence: evaluatorEvidence,
+            recommendations: evaluatorQuality.requiredGates > 0
+                ? ['Use evaluator gates to force critique, uncertainty logging, and review evidence before promoting reasoning-heavy work.']
+                : ['Run a reasoning-heavy AI OS task so evaluator intelligence can prove critique coverage.'],
+        },
+        {
+            id: 'tool-strategy',
+            status: toolStrategyQuality.totalSteps > 0
+                ? toolStrategyQuality.fallbackCoverage < 1 ? 'warning' : 'ready'
+                : input.runReports.length > 0 || input.benchmark ? 'warning' : 'blocked',
+            summary: toolStrategyQuality.totalSteps > 0
+                ? `${toolStrategyQuality.totalSteps} tool strategy step(s); ${toolStrategyQuality.highRiskSteps} high-risk; fallback coverage ${toolStrategyQuality.fallbackCoverage}.`
+                : 'No tool strategy graph found for skills, artifacts, CLI, MCP, or verification steps.',
+            evidence: toolStrategyEvidence,
+            recommendations: toolStrategyQuality.totalSteps > 0
+                ? ['Use tool strategy evidence to review cost, retry, fallback, and side-effect risk before execution.']
+                : ['Create a task that triggers skill routing so the AI OS can build a tool strategy graph.'],
+        },
+        {
+            id: 'adaptive-workflow',
+            status: input.runReports.some(r => r.plan.adaptiveWorkflow.profile) ? 'ready' : input.runReports.length > 0 || input.benchmark ? 'warning' : 'blocked',
+            summary: summarizeAdaptiveWorkflowSignal(input.runReports, input.benchmark),
+            evidence: [
+                ...input.runReports.map(r => `${r.artifacts.runReport}:profile=${r.plan.adaptiveWorkflow.profile}`),
+                ...(input.benchmark ? [`${input.benchmarkReport}:profiles=${input.benchmark.summary.workflowProfiles.join(',')}`] : []),
+            ],
+            recommendations: input.runReports.some(r => r.plan.adaptiveWorkflow.profile)
+                ? ['Use workflow profile distribution to verify that risk signals correctly escalate governance.']
+                : ['Run an AI OS task with mixed risk levels to prove adaptive workflow routing.'],
+        },
+        {
+            id: 'evolution-shadow',
+            status: evolutionQuality.proposals > 0
+                ? evolutionQuality.pendingValidation > 0 ? 'warning' : 'ready'
+                : input.runReports.length > 0 || input.benchmark ? 'warning' : 'blocked',
+            summary: evolutionQuality.proposals > 0
+                ? `${evolutionQuality.proposals} shadow proposal(s); ${evolutionQuality.shadowRules} shadow, ${evolutionQuality.candidateHooks} candidate-hook, ${evolutionQuality.approvedBlocking} approved-blocking.`
+                : 'No evolution shadow proposals found. Run tasks with high-risk governance signals or evaluator gates to generate shadow rule candidates.',
+            evidence: evolutionEvidence,
+            recommendations: evolutionQuality.proposals > 0
+                ? ['Review shadow rule proposals and validate before promotion to candidate-hook or approved-blocking.']
+                : ['Run a high-risk AI OS task so evolution shadow promotion can propose rules from governance and evaluator signals.'],
+        },
+        {
+            id: 'benchmark-intelligence',
+            status: input.benchmark && input.benchmarkStatus === 'fresh'
+                ? 'ready'
+                : input.benchmark && input.benchmarkStatus === 'stale' ? 'warning' : 'blocked',
+            summary: input.benchmark
+                ? `${input.benchmark.summary.scenarios} benchmark scenario(s); benchmark status ${input.benchmarkStatus}.`
+                : 'No AI OS benchmark report available for intelligence metrics.',
+            evidence: benchmarkEvidence,
+            recommendations: input.benchmark && input.benchmarkStatus === 'fresh'
+                ? ['Use intelligence signals alongside benchmark deltas for release readiness reviews.']
+                : ['Run `scale ai-os benchmark --json` to refresh memory/context/skill intelligence metrics.'],
+        },
+    ];
+    const summary = {
+        ready: signals.filter(signal => signal.status === 'ready').length,
+        warning: signals.filter(signal => signal.status === 'warning').length,
+        blocked: signals.filter(signal => signal.status === 'blocked').length,
+        totalMemoryItems,
+        selectedProviders,
+        memoryQuality,
+        contextQuality,
+        evaluatorQuality,
+        toolStrategyQuality,
+        evolutionQuality,
+        estimatedTokenSavings,
+        skillSteps,
+    };
+    const status = summary.blocked > 0 ? 'blocked' : summary.warning > 0 ? 'warning' : 'ready';
+    const nextActions = aiOsIntelligenceNextActions(status, signals, input.lang);
+    return { status, summary, signals, nextActions };
+}
+function summarizeContextQuality(runReports) {
+    const omitted = runReports.flatMap(report => report.plan.context.omitted.map(item => {
+        const section = report.plan.context.sections.find(candidate => candidate.id === item.id);
+        return {
+            ...item,
+            category: section?.category,
+            runReport: report.artifacts.runReport,
+        };
+    }));
+    const totalOmittedTokens = omitted.reduce((sum, item) => sum + item.estimatedTokens, 0);
+    const highestOmittedTokens = omitted.reduce((max, item) => Math.max(max, item.estimatedTokens), 0);
+    const evidenceLossWarnings = omitted
+        .filter(item => item.category === 'evidence' || item.id.includes('evidence'))
+        .map(item => `${item.id} omitted from ${item.runReport} (${item.estimatedTokens} tokens; ${item.reason}).`);
+    const compressionRisk = evidenceLossWarnings.length > 0
+        ? 'high'
+        : omitted.length > 0 ? 'medium' : 'low';
+    return {
+        omittedSections: omitted.length,
+        totalOmittedTokens,
+        evidenceLossWarnings,
+        highestOmittedTokens,
+        compressionRisk,
+    };
+}
+function summarizeEvaluatorQuality(runReports, benchmark) {
+    const runEvaluators = runReports.map(resolveRunEvaluator);
+    const runGates = runEvaluators.flatMap(evaluator => evaluator.gates);
+    const benchmarkGateCount = benchmark?.summary.totalEvaluatorGates ?? 0;
+    const uncertaintyScores = runEvaluators.map(evaluator => evaluator.uncertainty.score);
+    const gateIds = new Set(runGates.map(gate => gate.id));
+    if (benchmarkGateCount > 0)
+        gateIds.add('uncertainty-decision-log');
+    return {
+        requiredGates: runGates.filter(gate => gate.required).length + benchmarkGateCount,
+        highRiskPlans: runEvaluators.filter(evaluator => evaluator.riskLevel === 'high').length,
+        averageUncertainty: roundMetric(average(uncertaintyScores)),
+        gateIds: [...gateIds].sort(),
+    };
+}
+function resolveRunEvaluator(report) {
+    const plan = report.plan;
+    return plan.evaluator ?? createEvaluatorIntelligence({
+        task: report.plan.task.task,
+        files: report.plan.task.files,
+        governance: report.plan.governance,
+        skillPlan: report.plan.skillPlan,
+    });
+}
+function summarizeToolStrategyQuality(runReports, benchmark) {
+    const runStrategies = runReports.map(resolveRunToolStrategy);
+    const runSummary = runStrategies.reduce((summary, strategy) => ({
+        totalSteps: summary.totalSteps + strategy.summary.totalSteps,
+        requiredSteps: summary.requiredSteps + strategy.summary.requiredSteps,
+        highRiskSteps: summary.highRiskSteps + strategy.summary.highRiskSteps,
+        estimatedCostUnits: summary.estimatedCostUnits + strategy.summary.estimatedCostUnits,
+        fallbackCoveredSteps: summary.fallbackCoveredSteps + strategy.summary.fallbackCoveredSteps,
+    }), {
+        totalSteps: 0,
+        requiredSteps: 0,
+        highRiskSteps: 0,
+        estimatedCostUnits: 0,
+        fallbackCoveredSteps: 0,
+    });
+    const benchmarkSteps = benchmark?.summary.totalToolStrategySteps ?? 0;
+    const benchmarkCost = benchmark?.summary.totalToolStrategyCostUnits ?? 0;
+    const totalSteps = runSummary.totalSteps + benchmarkSteps;
+    const fallbackCoveredSteps = runSummary.fallbackCoveredSteps + benchmarkSteps;
+    return {
+        totalSteps,
+        requiredSteps: runSummary.requiredSteps,
+        highRiskSteps: runSummary.highRiskSteps,
+        estimatedCostUnits: runSummary.estimatedCostUnits + benchmarkCost,
+        fallbackCoverage: totalSteps > 0 ? roundMetric(fallbackCoveredSteps / totalSteps) : 0,
+    };
+}
+function resolveRunToolStrategy(report) {
+    const plan = report.plan;
+    return plan.toolStrategy ?? createToolStrategyPlan(report.plan.skillPlan);
+}
+function summarizeEvolutionQuality(runReports, benchmark) {
+    const runProposals = runReports.flatMap(r => r.plan.evolutionShadow?.proposals ?? []);
+    const benchmarkProposals = benchmark?.summary.totalEvolutionProposals ?? 0;
+    const allProposals = runProposals;
+    const stageCount = (stage) => allProposals.filter(p => p.maturity.stage === stage).length;
+    return {
+        proposals: allProposals.length + benchmarkProposals,
+        shadowRules: stageCount('shadow'),
+        candidateHooks: stageCount('candidate-hook'),
+        approvedBlocking: stageCount('approved-blocking'),
+        pendingValidation: allProposals.filter(p => p.maturity.stage === 'shadow' && p.maturity.shadowHits < 10).length,
+    };
+}
+function resolveRunEvolutionShadow(report) {
+    const plan = report.plan;
+    return plan.evolutionShadow ?? buildEvolutionShadowReport([]);
+}
+function summarizeAdaptiveWorkflowSignal(runReports, benchmark) {
+    const profiles = runReports.map(r => r.plan.adaptiveWorkflow.profile);
+    const benchmarkProfiles = benchmark?.summary.workflowProfiles ?? [];
+    const allProfiles = [...profiles, ...benchmarkProfiles];
+    if (allProfiles.length === 0)
+        return 'No adaptive workflow profile evidence found.';
+    const distribution = new Map();
+    for (const p of allProfiles)
+        distribution.set(p, (distribution.get(p) ?? 0) + 1);
+    const parts = [...distribution.entries()].map(([p, n]) => `${p}=${n}`).join(', ');
+    const escalated = runReports.filter(r => r.plan.adaptiveWorkflow.escalationReasons.length > 0).length;
+    return `${allProfiles.length} run(s) with profile distribution: ${parts}. ${escalated} run(s) had escalation reasons.`;
+}
+function summarizeMemoryQuality(items) {
+    if (items.length === 0) {
+        return {
+            score: 0,
+            evidenceBackedItems: 0,
+            missingEvidenceItems: 0,
+            lowConfidenceItems: 0,
+            averageConfidence: 0,
+            averageRelevance: 0,
+        };
+    }
+    const evidenceBackedItems = items.filter(item => item.evidencePaths.length > 0).length;
+    const missingEvidenceItems = items.length - evidenceBackedItems;
+    const lowConfidenceItems = items.filter(item => item.confidence < 0.7).length;
+    const averageConfidence = average(items.map(item => clampUnit(item.confidence)));
+    const averageRelevance = average(items.map(item => clampUnit(item.score)));
+    const evidenceRatio = evidenceBackedItems / items.length;
+    const lowConfidenceRatio = lowConfidenceItems / items.length;
+    const score = Math.max(0, Math.round((averageConfidence * 40) + (averageRelevance * 30) + (evidenceRatio * 30) - (lowConfidenceRatio * 10)));
+    return {
+        score,
+        evidenceBackedItems,
+        missingEvidenceItems,
+        lowConfidenceItems,
+        averageConfidence: roundMetric(averageConfidence),
+        averageRelevance: roundMetric(averageRelevance),
+    };
+}
+function average(values) {
+    if (values.length === 0)
+        return 0;
+    return values.reduce((sum, value) => sum + value, 0) / values.length;
+}
+function clampUnit(value) {
+    if (!Number.isFinite(value))
+        return 0;
+    return Math.max(0, Math.min(1, value));
+}
+function roundMetric(value) {
+    return Number(value.toFixed(3));
+}
+function aiOsIntelligenceNextActions(status, signals, lang) {
+    const actions = [];
+    if (signals.some(signal => signal.status === 'ready')) {
+        actions.push('Use intelligence signals during release review to prove memory, context, and skill routing gains.');
+    }
+    if (status === 'ready')
+        return actions;
+    const blocked = signals.filter(signal => signal.status === 'blocked').map(signal => signal.id);
+    if (lang === 'zh') {
+        actions.push(`Refresh AI OS intelligence evidence for: ${blocked.join(', ') || 'warning signals'}.`);
+        return actions;
+    }
+    actions.push(`Refresh AI OS intelligence evidence for: ${blocked.join(', ') || 'warning signals'}.`);
+    return actions;
+}
 function buildRunSteps(plan) {
     const steps = new Map();
     const upsert = (step) => steps.set(step.id, step);
@@ -563,16 +936,20 @@ function buildRunSteps(plan) {
         evidence: ['memory.providerOrder', 'memory.selectedProviders', 'memory.items'],
         dependsOn: ['runtime-plan'],
     });
+    const profile = plan.adaptiveWorkflow.profile;
     for (const gate of plan.adaptiveWorkflow.gates) {
         if (steps.has(gate))
             continue;
+        const gateRequired = profile !== 'light';
         upsert({
             id: gate,
             kind: gate === 'runtime-evidence' ? 'evidence' : 'gate',
             title: `Satisfy ${gate} gate`,
             status: 'planned',
-            required: true,
-            summary: `Required by ${plan.adaptiveWorkflow.strategy} in ${plan.adaptiveWorkflow.mode} mode.`,
+            required: gateRequired,
+            summary: gateRequired
+                ? `Required by ${plan.adaptiveWorkflow.strategy} in ${profile} profile (${plan.adaptiveWorkflow.mode} mode).`
+                : `Advisory in ${profile} profile; not blocking completion.`,
             evidence: [`gate.${gate}`],
             dependsOn: ['runtime-plan'],
         });
@@ -945,6 +1322,22 @@ function inspectBenchmarkReport(projectDir, scaleDir, maxAgeHours, warnings) {
         return { status: 'invalid', reportPath };
     }
 }
+function readAiOsBenchmarkReport(reportPath, warnings) {
+    if (!existsSync(reportPath))
+        return undefined;
+    try {
+        const parsed = JSON.parse(readFileSync(reportPath, 'utf-8'));
+        if (!parsed || !parsed.summary || !Array.isArray(parsed.scenarios)) {
+            warnings.push(`Ignored invalid AI OS benchmark report: ${reportPath}`);
+            return undefined;
+        }
+        return parsed;
+    }
+    catch (error) {
+        warnings.push(`Ignored unreadable AI OS benchmark report: ${reportPath} (${error instanceof Error ? error.message : String(error)})`);
+        return undefined;
+    }
+}
 function summarizeBenchmarkDoctor(benchmark) {
     if (benchmark.status === 'missing')
         return 'No AI OS benchmark report found.';
@@ -1108,7 +1501,12 @@ function summarizeBenchmark(results) {
         totalMemoryItems: results.reduce((sum, result) => sum + result.metrics.memoryItems, 0),
         totalSkillSteps: results.reduce((sum, result) => sum + result.metrics.skillSteps, 0),
         requiredSkillSteps: results.reduce((sum, result) => sum + result.metrics.requiredSkillSteps, 0),
+        totalEvaluatorGates: results.reduce((sum, result) => sum + result.metrics.evaluatorGates, 0),
+        totalToolStrategySteps: results.reduce((sum, result) => sum + result.metrics.toolStrategySteps, 0),
+        totalToolStrategyCostUnits: results.reduce((sum, result) => sum + result.metrics.toolStrategyCostUnits, 0),
+        totalEvolutionProposals: results.reduce((sum, result) => sum + result.metrics.evolutionProposals, 0),
         governanceModes: [...new Set(results.map(result => result.governanceMode))],
+        workflowProfiles: [...new Set(results.map(result => result.workflowProfile))],
         averageTokenUtilization: totalBudget > 0 ? Number((totalEstimatedTokens / totalBudget).toFixed(4)) : 0,
     };
 }
@@ -1116,6 +1514,10 @@ function benchmarkRecommendations(summary) {
     const recommendations = ['Use benchmark deltas in release notes only after comparing the same scenario set across versions.'];
     if (summary.totalSkillSteps === 0)
         recommendations.push('Skill routing did not produce steps; inspect skill policy detection.');
+    if (summary.totalEvaluatorGates === 0)
+        recommendations.push('Evaluator intelligence did not require any critique gate; add reasoning-heavy benchmark scenarios before claiming evaluator coverage.');
+    if (summary.totalToolStrategySteps === 0)
+        recommendations.push('Tool strategy did not build a cost/retry/fallback graph; inspect skill execution plan coverage.');
     if (summary.averageTokenUtilization > 0.9)
         recommendations.push('Context utilization is high; lower budgets or improve relevance filtering before scaling.');
     if (!summary.governanceModes.includes('critical') && !summary.governanceModes.includes('expanded')) {
@@ -1123,30 +1525,279 @@ function benchmarkRecommendations(summary) {
     }
     return recommendations;
 }
-function createAdaptiveWorkflow(governance, skillPlan) {
+function createAdaptiveWorkflow(governance, skillPlan, evaluator, toolStrategy) {
+    const routerResult = routeAdaptiveWorkflow({ governance, evaluator, toolStrategy });
     const gates = new Set();
     gates.add('context-compiler');
     gates.add('memory-provider-recall');
     if (skillPlan.required || skillPlan.executionPlan.steps.length > 0)
         gates.add('skill-evidence');
     gates.add('runtime-evidence');
-    if (governance.effectiveMode === 'expanded' || governance.effectiveMode === 'critical')
+    if (routerResult.profile === 'strict' || routerResult.profile === 'critical')
         gates.add('impact-analysis');
-    if (governance.effectiveMode === 'critical')
+    if (routerResult.profile === 'critical')
         gates.add('security-review');
+    for (const gate of evaluator.gates)
+        gates.add(gate.id);
+    for (const override of routerResult.gateOverrides)
+        gates.add(override.gateId);
+    const requiredBehaviors = new Set(governance.requiredBehaviors);
+    for (const constraint of routerResult.behavioralConstraints) {
+        if (constraint.required)
+            requiredBehaviors.add(constraint.description);
+    }
     return {
         strategy: 'risk-adaptive-runtime-v1',
+        profile: routerResult.profile,
+        escalationReasons: routerResult.escalationReasons,
         mode: governance.effectiveMode,
-        requiredBehaviors: governance.requiredBehaviors,
+        requiredBehaviors: Array.from(requiredBehaviors),
         gates: Array.from(gates),
-        exitCriteria: [
-            'Context compiler explains included and omitted sections.',
-            'Memory recall records provider, score, and evidence paths.',
-            'Skill plan lists required proof and fallback policy.',
-            'Governance ROI states benefit and overhead before completion.',
-        ],
+        exitCriteria: routerResult.exitCriteria,
+    };
+}
+function createEvaluatorIntelligence(input) {
+    const haystack = `${input.task} ${input.files.join(' ')} ${input.governance.signals.map(signal => signal.id).join(' ')}`.toLowerCase();
+    const gates = [];
+    const addGate = (gate) => {
+        if (gates.some(existing => existing.id === gate.id))
+            return;
+        gates.push(gate);
+    };
+    if (/architecture|architectural|design|strategy|boundary|refactor|runtime|platform|framework|架构|方案|设计|边界|平台/.test(haystack)) {
+        addGate({
+            id: 'architecture-critique',
+            required: input.governance.effectiveMode !== 'minimal',
+            reason: 'Architecture, runtime, platform, or design decisions need an explicit critique before implementation claims.',
+            evidence: matchingEvidence(input.files, /architecture|runtime|framework|docs|readme|src/i),
+        });
+    }
+    if (/root cause|diagnose|debug|failure|incident|postmortem|regression|blocked|根因|排查|故障|事故|回归/.test(haystack)) {
+        addGate({
+            id: 'root-cause-review',
+            required: true,
+            reason: 'Failure diagnosis or root-cause work needs an alternate hypothesis check before closing.',
+            evidence: matchingEvidence(input.files, /test|runtime|debug|log|src|docs/i),
+        });
+    }
+    if (input.governance.signals.some(signal => signal.id === 'critical-risk-domain' || signal.id === 'critical-file-path')) {
+        addGate({
+            id: 'security-threat-model',
+            required: true,
+            reason: 'Critical auth, data, production, or destructive risk requires threat-model review evidence.',
+            evidence: input.governance.signals.flatMap(signal => signal.evidence).slice(0, 12),
+        });
+    }
+    if (/release|publish|deploy|migration|rollback|version|changelog|npm|ci|发版|发布|部署|迁移|回滚/.test(haystack)) {
+        addGate({
+            id: 'release-readiness-review',
+            required: true,
+            reason: 'Release, deployment, migration, or rollback work needs readiness and rollback evidence.',
+            evidence: matchingEvidence(input.files, /package|changelog|release|deploy|migration|workflow|github/i),
+        });
+    }
+    const drivers = evaluatorUncertaintyDrivers(input, gates);
+    const uncertaintyScore = evaluatorUncertaintyScore(input, gates, drivers);
+    if (gates.length > 0 || uncertaintyScore >= 0.45) {
+        addGate({
+            id: 'uncertainty-decision-log',
+            required: uncertaintyScore >= 0.45 || input.governance.effectiveMode === 'critical',
+            reason: 'The agent must record uncertainty, rejected alternatives, and evidence gaps before completion.',
+            evidence: drivers,
+        });
+    }
+    const riskLevel = uncertaintyScore >= 0.7
+        ? 'high'
+        : uncertaintyScore >= 0.4 || gates.some(gate => gate.required) ? 'medium' : 'low';
+    return {
+        strategy: 'evaluator-intelligence-v1',
+        required: gates.some(gate => gate.required),
+        riskLevel,
+        uncertainty: {
+            score: uncertaintyScore,
+            threshold: 0.45,
+            drivers,
+        },
+        gates,
+        recommendations: evaluatorRecommendations(gates, riskLevel),
     };
 }
+function createToolStrategyPlan(skillPlan) {
+    const nodes = skillPlan.executionPlan.steps.map(step => {
+        const risks = toolStepRisks(step.id, step.kind);
+        return {
+            id: `${step.kind}:${step.id}`,
+            kind: step.kind,
+            required: step.required,
+            cost: {
+                units: toolStepCostUnits(step.id, step.kind, step.required, risks),
+                timeRisk: risks.timeRisk,
+                sideEffectRisk: risks.sideEffectRisk,
+            },
+            retry: toolStepRetry(step.id, step.kind, risks),
+            fallback: step.fallback,
+            evidence: [step.evidenceRequired],
+        };
+    });
+    const edges = buildToolStrategyEdges(nodes);
+    const summary = {
+        totalSteps: nodes.length,
+        requiredSteps: nodes.filter(node => node.required).length,
+        highRiskSteps: nodes.filter(node => node.cost.timeRisk === 'high' || node.cost.sideEffectRisk === 'high').length,
+        estimatedCostUnits: nodes.reduce((sum, node) => sum + node.cost.units, 0),
+        fallbackCoveredSteps: nodes.filter(node => node.fallback.trim().length > 0).length,
+    };
+    return {
+        strategy: 'tool-strategy-v1',
+        nodes,
+        edges,
+        summary,
+        recommendations: toolStrategyRecommendations(summary),
+    };
+}
+function createEvolutionShadowProposals(governance, evaluator) {
+    const proposals = [];
+    // Propose shadow rules from governance risk signals (escalated modes)
+    for (const signal of governance.signals) {
+        if (signal.mode === 'expanded' || signal.mode === 'critical') {
+            proposals.push(proposeShadowRule({
+                title: `Governance signal: ${signal.id}`,
+                description: `Shadow rule from governance signal "${signal.id}" (mode=${signal.mode}). ${signal.reason}`,
+                source: 'failure-learning',
+                sourceEvidenceIds: signal.evidence.length > 0 ? signal.evidence : [signal.id],
+                pattern: signal.id,
+                enforcement: signal.mode === 'critical' ? 'hook' : 'prompt',
+                rollback: `Remove shadow rule for governance signal "${signal.id}" if false positive rate exceeds threshold.`,
+            }));
+        }
+    }
+    // Propose shadow rules from high-risk evaluator gates
+    for (const gate of evaluator.gates) {
+        if (gate.required && (gate.id === 'security-threat-model' || gate.id === 'root-cause-review')) {
+            proposals.push(proposeShadowRule({
+                title: `Evaluator gate: ${gate.id}`,
+                description: `Shadow rule from required evaluator gate "${gate.id}". ${gate.reason}`,
+                source: 'lesson-extraction',
+                sourceEvidenceIds: [gate.id],
+                pattern: gate.id,
+                enforcement: 'prompt',
+                rollback: `Remove shadow rule for evaluator gate "${gate.id}" if it does not reduce defect recurrence.`,
+            }));
+        }
+    }
+    return buildEvolutionShadowReport(proposals);
+}
+function toolStepRisks(id, kind) {
+    const normalized = id.toLowerCase();
+    if (/desktop|cua|deploy|publish|release|migration|rollback|delete|drop|external|cli/.test(normalized)) {
+        return { timeRisk: 'high', sideEffectRisk: 'high' };
+    }
+    if (/browser|e2e|playwright|screenshot|visual|security|threat|audit/.test(normalized)) {
+        return { timeRisk: 'medium', sideEffectRisk: kind === 'verification' ? 'medium' : 'low' };
+    }
+    if (kind === 'artifact')
+        return { timeRisk: 'low', sideEffectRisk: 'low' };
+    if (kind === 'verification')
+        return { timeRisk: 'medium', sideEffectRisk: 'medium' };
+    return { timeRisk: 'medium', sideEffectRisk: 'low' };
+}
+function toolStepCostUnits(id, kind, required, risks) {
+    let units = kind === 'artifact' ? 1 : kind === 'verification' ? 2 : 3;
+    if (required)
+        units += 1;
+    if (risks.timeRisk === 'medium')
+        units += 1;
+    if (risks.timeRisk === 'high')
+        units += 2;
+    if (risks.sideEffectRisk === 'high')
+        units += 2;
+    if (/browser|e2e|desktop|external|cli|security|audit/i.test(id))
+        units += 1;
+    return units;
+}
+function toolStepRetry(id, kind, risks) {
+    if (risks.sideEffectRisk === 'high')
+        return { maxAttempts: 1, backoff: 'manual-review' };
+    if (kind === 'verification')
+        return { maxAttempts: /browser|e2e|playwright|network/i.test(id) ? 2 : 1, backoff: 'linear' };
+    if (kind === 'skill')
+        return { maxAttempts: 1, backoff: 'manual-review' };
+    return { maxAttempts: 1, backoff: 'none' };
+}
+function buildToolStrategyEdges(nodes) {
+    const edges = [];
+    const skillNodes = nodes.filter(node => node.kind === 'skill');
+    const artifactNodes = nodes.filter(node => node.kind === 'artifact');
+    const verificationNodes = nodes.filter(node => node.kind === 'verification');
+    for (const artifact of artifactNodes) {
+        for (const skill of skillNodes.filter(node => node.required || artifact.required)) {
+            edges.push({ from: skill.id, to: artifact.id, reason: 'Skill execution must leave artifact evidence when both are required or review-relevant.' });
+        }
+    }
+    for (const verification of verificationNodes) {
+        for (const artifact of artifactNodes.filter(node => node.required)) {
+            edges.push({ from: artifact.id, to: verification.id, reason: 'Required artifacts should exist before verification evidence is accepted.' });
+        }
+    }
+    return edges;
+}
+function toolStrategyRecommendations(summary) {
+    if (summary.totalSteps === 0)
+        return ['No tool strategy required; standard verification is enough for this task.'];
+    const recommendations = ['Execute required tool strategy nodes before claiming task completion.'];
+    if (summary.highRiskSteps > 0)
+        recommendations.push('High-risk tool steps require manual review or explicit safe-mode evidence before retry.');
+    if (summary.fallbackCoveredSteps < summary.totalSteps)
+        recommendations.push('Fill fallback policy gaps before autonomous execution.');
+    return recommendations;
+}
+function matchingEvidence(files, pattern) {
+    return files.filter(file => pattern.test(file)).slice(0, 12);
+}
+function evaluatorUncertaintyDrivers(input, gates) {
+    const drivers = new Set();
+    if (input.governance.effectiveMode === 'critical')
+        drivers.add('critical-governance-mode');
+    if (input.governance.effectiveMode === 'expanded')
+        drivers.add('expanded-governance-mode');
+    if (input.files.length >= 6)
+        drivers.add('wide-file-scope');
+    if (input.skillPlan.executionPlan.steps.some(step => step.required))
+        drivers.add('required-skill-evidence');
+    for (const gate of gates)
+        drivers.add(gate.id);
+    if (/unknown|uncertain|maybe|assume|guess|可能|不确定|假设/.test(input.task.toLowerCase()))
+        drivers.add('explicit-uncertainty-language');
+    return [...drivers];
+}
+function evaluatorUncertaintyScore(input, gates, drivers) {
+    let score = 0.15;
+    if (input.governance.effectiveMode === 'standard')
+        score += 0.1;
+    if (input.governance.effectiveMode === 'expanded')
+        score += 0.25;
+    if (input.governance.effectiveMode === 'critical')
+        score += 0.4;
+    score += Math.min(0.2, input.files.length * 0.025);
+    score += Math.min(0.2, gates.filter(gate => gate.required).length * 0.08);
+    if (input.skillPlan.executionPlan.steps.some(step => step.required))
+        score += 0.08;
+    if (drivers.includes('explicit-uncertainty-language'))
+        score += 0.12;
+    return roundMetric(clampUnit(score));
+}
+function evaluatorRecommendations(gates, riskLevel) {
+    if (gates.length === 0)
+        return ['No evaluator gate required; keep lightweight verification evidence for low-risk work.'];
+    const recommendations = ['Record evaluator evidence before promoting reasoning-heavy implementation or release claims.'];
+    if (riskLevel === 'high')
+        recommendations.push('Require reviewer sign-off for uncertainty, rejected alternatives, and rollback or mitigation path.');
+    if (gates.some(gate => gate.id === 'root-cause-review'))
+        recommendations.push('List competing root-cause hypotheses and why each was accepted or rejected.');
+    if (gates.some(gate => gate.id === 'security-threat-model'))
+        recommendations.push('Attach threat model or security-review evidence before guarded completion.');
+    return recommendations;
+}
 function recommendations(options) {
     const output = [];
     if (options.context.compiler?.estimatedTokenSavings) {
@@ -1161,6 +1812,12 @@ function recommendations(options) {
     if (options.governance.effectiveMode === 'critical') {
         output.push('Critical workflow mode requires security review and rollback or disable strategy.');
     }
+    if (options.evaluator.required) {
+        output.push(`Evaluator intelligence requires ${options.evaluator.gates.length} critique gate(s); record uncertainty and review evidence before promotion.`);
+    }
+    if (options.toolStrategy.summary.totalSteps > 0) {
+        output.push(`Tool strategy planner created ${options.toolStrategy.summary.totalSteps} cost/retry/fallback node(s); execute required nodes with evidence.`);
+    }
     return output;
 }
 function normalizeSkillTaskLevel(value) {