newmark-agent 0.6.4 → 0.6.8

package/dist/core/autoRouter.js

@@ -1,18 +1,40 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.AutoRouter = void 0;
+exports.RouteExecutionController = exports.ROUTE_POLICY_VERSION = void 0;
 exports.normalizeAutoPreference = normalizeAutoPreference;
 exports.defaultRoutePolicy = defaultRoutePolicy;
 exports.classifyTaskClasses = classifyTaskClasses;
 exports.classifyRouteFailure = classifyRouteFailure;
+exports.createRouteId = createRouteId;
+exports.catalogSnapshotHash = catalogSnapshotHash;
+exports.validationEvidenceIsStale = validationEvidenceIsStale;
+/**
+ * Auto Router v2 — shared route contracts, failure classification, operational
+ * endpoint health and provider-local recovery planning.
+ *
+ * This file no longer selects models. Model selection is the exclusive
+ * responsibility of the JEV decision engine (`jevDecision.ts`) inside the
+ * candidate space produced by the deterministic eligibility guard
+ * (`routeEligibility.ts`). What remains here is execution plumbing:
+ *
+ *   - shared route types (`RouteDecision`, `RouteAttempt`, `RoutePolicy`, …)
+ *   - provider failure classification
+ *   - endpoint health / circuit breaker (operational protection only)
+ *   - build-scoped transaction pinning (execution consistency inside one Build)
+ *   - provider-local recovery planning (retry / equivalent / fallback / JEV alternate)
+ *
+ * Removed on purpose: the feedback map, EWMA preference learning,
+ * `quality_by_task` scoring, cache affinity, switch thresholds, and every
+ * weighted cost/reliability/speed utility pass. Nothing in this file may
+ * reconstruct a second model selector.
+ */
 const crypto_1 = require("crypto");
+const routeEligibility_1 = require("./routeEligibility");
 const DAY_MS = 24 * 60 * 60 * 1_000;
 const VALIDATION_TTL_MS = 7 * DAY_MS;
-const AFFINITY_TTL_MS = 5 * 60_000;
 const HEALTH_WINDOW_MS = 30 * 60_000;
 const CIRCUIT_COOLDOWN_MS = 60_000;
-const FEEDBACK_WINDOW_MS = 30 * DAY_MS;
-const FEEDBACK_HALF_LIFE_MS = 30 * DAY_MS;
+exports.ROUTE_POLICY_VERSION = 'newmark-auto-v2';
 function finite(value) {
     const number = Number(value);
     return Number.isFinite(number) ? number : undefined;
@@ -23,32 +45,17 @@ function clamp(value, min = 0, max = 1) {
 function deploymentKey(deployment) {
     return `${deployment.providerId}\u0000${deployment.modelId}\u0000${deployment.logicalModelGroupId || ''}`;
 }
-function sameDeployment(left, right) {
-    return left.providerId === right.providerId && left.modelId === right.modelId;
-}
-function inSubset(deployment, subset) {
-    if (!subset?.length)
-        return true;
-    return subset.some(item => sameDeployment(item, deployment));
-}
-function inScope(deployment, scope) {
-    return scope.kind === 'global' || deployment.providerId === scope.providerId;
-}
 function normalizeAutoPreference(value) {
-    switch (String(value || '').toLowerCase()) {
-        case 'performance':
-        case 'quality': return 'quality';
-        case 'cheap_save':
-        case 'cost': return 'cost';
-        case 'speed': return 'speed';
-        default: return 'balanced';
-    }
+    const normalized = String(value || '').trim().toLowerCase();
+    if (normalized === 'conservative' || normalized === 'balanced' || normalized === 'aggressive')
+        return normalized;
+    // Previous values selected fixed quality/cost/speed objectives. They do not
+    // have a reliable one-to-one mapping to a model decision tendency.
+    return 'balanced';
 }
 function defaultRoutePolicy(mode = 'balanced') {
-    const maxQualityLoss = mode === 'quality' ? 0 : mode === 'balanced' ? 0.02 : mode === 'cost' ? 0.06 : 0.04;
     return {
         mode,
-        maxQualityLoss,
         allowPreview: false,
         privacy: 'default',
         requiredCapabilities: [],
@@ -134,132 +141,76 @@ function classifyRouteFailure(error) {
     }
     return { type: 'execution_error', retryable: false, switchAllowed: false, statusCode };
 }
-class AutoRouter {
+function createRouteId() {
+    return (0, crypto_1.randomUUID)();
+}
+function catalogSnapshotHash(candidates) {
+    const snapshot = candidates.map(candidate => ({
+        deployment: candidate.deployment,
+        enabled: candidate.enabled,
+        unavailableReason: candidate.unavailableReason,
+        validation: candidate.validation,
+        capabilities: [...candidate.capabilities].sort(),
+        maxContextTokens: candidate.maxContextTokens,
+        preview: candidate.preview,
+        privacy: [...candidate.privacy].sort(),
+        dataRegions: [...(candidate.dataRegions || [])].sort(),
+        supportedProtocolParameters: [...(candidate.supportedProtocolParameters || [])].sort(),
+        expectedInputCostUsdPerM: candidate.expectedInputCostUsdPerM,
+        expectedOutputCostUsdPerM: candidate.expectedOutputCostUsdPerM,
+        latencyMs: candidate.latencyMs,
+        reliability: candidate.reliability,
+        toolValidity: candidate.toolValidity,
+        throughput: candidate.throughput,
+        circuit: candidate.circuit,
+        preference: candidate.preference,
+        fallbackOnly: !!candidate.fallbackOnly,
+    })).sort((left, right) => deploymentKey(left.deployment).localeCompare(deploymentKey(right.deployment)));
+    return (0, crypto_1.createHash)('sha256').update(JSON.stringify(snapshot)).digest('hex');
+}
+/**
+ * Operational execution controller.
+ *
+ * It owns endpoint health, the circuit breaker, build-scoped transaction pins
+ * and the provider-local recovery ladder. It never ranks models and never
+ * learns a preference; health observations only decide whether an endpoint is
+ * attempted right now.
+ */
+class RouteExecutionController {
     now;
     policyVersion;
-    affinityTtlMs;
-    switchThreshold;
     transactionPins = new Map();
-    affinities = new Map();
     endpointHealth = new Map();
-    feedback = new Map();
     constructor(options = {}) {
         this.now = options.now || Date.now;
-        this.policyVersion = options.policyVersion || 'newmark-auto-v1';
-        this.affinityTtlMs = Math.max(0, options.affinityTtlMs ?? AFFINITY_TTL_MS);
-        this.switchThreshold = Math.max(0, options.switchThreshold ?? 0.15);
+        this.policyVersion = options.policyVersion || exports.ROUTE_POLICY_VERSION;
     }
-    route(selection, policy, candidates, request) {
-        const now = this.now();
-        const taskClasses = classifyTaskClasses(request.taskText, [...policy.requiredCapabilities, ...request.requiredCapabilities], request.estimatedInputTokens);
-        const decision = {
-            routeId: (0, crypto_1.randomUUID)(),
-            requestedSelection: cloneSelection(selection),
-            policyVersion: this.policyVersion,
-            catalogSnapshotHash: catalogHash(candidates),
-            taskClasses,
-            excludedCandidates: [],
-            rankedCandidates: [],
-            attempts: [],
-            finalStatus: 'no_candidate',
-            retryBudgetMs: request.batch ? 15_000 : 5_000,
-        };
-        if (selection.kind === 'fixed') {
-            const fixed = candidates.find(item => sameDeployment(item.deployment, selection.deployment));
-            if (!fixed || fixed.enabled === false) {
-                decision.finalStatus = 'fixed_unavailable';
-                return decision;
-            }
-            const ranked = this.rankCandidate(fixed, taskClasses, policy, request, false, now);
-            decision.rankedCandidates = [ranked];
-            decision.resolvedDeployment = { ...fixed.deployment };
-            decision.attempts = [{ deployment: { ...fixed.deployment }, kind: 'initial', status: 'planned' }];
-            decision.finalStatus = 'resolved';
-            this.claimEndpointAttempt(fixed.deployment);
-            return decision;
-        }
-        const eligible = [];
-        for (const candidate of candidates) {
-            const reasons = [];
-            if (!candidate.enabled)
-                reasons.push('disabled');
-            if (!inScope(candidate.deployment, selection.scope))
-                reasons.push('outside_scope');
-            if (!inSubset(candidate.deployment, selection.subset))
-                reasons.push('outside_subset');
-            if (candidate.fallbackOnly)
-                reasons.push('fallback_only');
-            if (candidate.preview && !policy.allowPreview)
-                reasons.push('preview_disallowed');
-            if (request.estimatedInputTokens + request.expectedOutputTokens > Math.max(0, candidate.maxContextTokens || 0))
-                reasons.push('context_too_small');
-            if (policy.privacy !== 'default' && !candidate.privacy.includes(policy.privacy))
-                reasons.push(`privacy:${policy.privacy}`);
-            if (policy.dataRegion) {
-                const requiredRegion = policy.dataRegion.toLowerCase();
-                const regions = new Set((candidate.dataRegions || []).map(item => String(item).toLowerCase()));
-                if (!regions.has(requiredRegion))
-                    reasons.push(`data_region:${policy.dataRegion}`);
-            }
-            const supportedParameters = new Set((candidate.supportedProtocolParameters || []).map(item => String(item).toLowerCase()));
-            for (const parameter of policy.requiredProtocolParameters || []) {
-                const requiredParameter = String(parameter).toLowerCase();
-                if (requiredParameter && !supportedParameters.has(requiredParameter))
-                    reasons.push(`protocol_parameter:${requiredParameter}`);
-            }
-            const expectedCost = expectedRequestCost(candidate, request);
-            if (policy.maxExpectedCostUsd !== undefined && (expectedCost === undefined || expectedCost > policy.maxExpectedCostUsd)) {
-                reasons.push(expectedCost === undefined ? 'unknown_cost' : 'budget_exceeded');
-            }
-            if (reasons.length)
-                decision.excludedCandidates.push({ deployment: { ...candidate.deployment }, reasons: [...new Set(reasons)] });
-            else
-                eligible.push(candidate);
-        }
-        if (!eligible.length)
-            return decision;
-        const affinityKey = this.affinityKey(request.affinityKey, selection.scope, policy.mode);
-        let affinity = this.affinities.get(affinityKey);
-        if (affinity && affinity.expiresAt <= now) {
-            this.affinities.delete(affinityKey);
-            affinity = undefined;
-        }
-        const affinityDeployment = affinity?.deployment;
-        const transactionPin = this.transactionPins.get(request.transactionId);
-        if (transactionPin) {
-            const pinned = eligible.find(item => sameDeployment(item.deployment, transactionPin));
-            if (pinned) {
-                decision.rankedCandidates = this.rankEligible(eligible, taskClasses, policy, request, now, affinityDeployment);
-                decision.resolvedDeployment = { ...pinned.deployment };
-                decision.pinReason = 'transaction';
-                decision.attempts = [{ deployment: { ...pinned.deployment }, kind: 'initial', status: 'planned' }];
-                decision.finalStatus = 'resolved';
-                this.claimEndpointAttempt(pinned.deployment);
-                return decision;
-            }
-            this.transactionPins.delete(request.transactionId);
-        }
-        const ranked = this.rankEligible(eligible, taskClasses, policy, request, now, affinityDeployment);
-        decision.rankedCandidates = ranked;
-        let selected = ranked[0];
-        if (affinity) {
-            const incumbent = ranked.find(item => sameDeployment(item.deployment, affinity.deployment));
-            if (incumbent && selected.utility - incumbent.utility < this.switchThreshold) {
-                selected = incumbent;
-                decision.pinReason = 'cache_affinity';
-            }
-        }
-        decision.resolvedDeployment = { ...selected.deployment };
-        decision.attempts = [{ deployment: { ...selected.deployment }, kind: 'initial', status: 'planned' }];
-        decision.finalStatus = 'resolved';
-        this.claimEndpointAttempt(selected.deployment);
-        this.transactionPins.set(request.transactionId, { ...selected.deployment });
-        this.affinities.set(affinityKey, { deployment: { ...selected.deployment }, expiresAt: now + this.affinityTtlMs });
-        return decision;
+    version() {
+        return this.policyVersion;
+    }
+    /** Execution-consistency pin for one Build / route transaction. */
+    pinTransaction(transactionId, deployment) {
+        if (!transactionId || !deployment.providerId || !deployment.modelId)
+            return;
+        this.transactionPins.set(transactionId, { ...deployment });
+    }
+    transactionPin(transactionId) {
+        const pinned = this.transactionPins.get(transactionId);
+        return pinned ? { ...pinned } : undefined;
     }
     endTransaction(transactionId) {
         this.transactionPins.delete(transactionId);
     }
+    /**
+     * Provider-local recovery ladder.
+     *
+     * Order of preference: retry the same deployment (when the failure is
+     * retryable inside the retry budget), then an equivalent deployment of the
+     * same logical model group, then the alternates the decision engine already
+     * validated (in the order it produced them), then explicit fallback-only
+     * models. Recovery never crosses the failed provider boundary and never
+     * re-runs model selection.
+     */
     planAttempts(decision, candidates, failure) {
         const current = decision.resolvedDeployment;
         if (!current || !failure.error.switchAllowed || failure.streamCommitted || failure.sideEffectCommitted)
@@ -270,7 +221,7 @@ class AutoRouter {
         const attempts = [];
         const retryDelayMs = failure.error.retryAfterMs ?? 250;
         const alreadyRetriedCurrent = decision.attempts.some(attempt => attempt.kind === 'retry_same_deployment'
-            && sameDeployment(attempt.deployment, current));
+            && (0, routeEligibility_1.sameDeploymentRef)(attempt.deployment, current));
         if (failure.error.retryable && !alreadyRetriedCurrent && retryDelayMs <= (decision.retryBudgetMs ?? 5_000)) {
             attempts.push({
                 deployment: { ...current },
@@ -287,11 +238,9 @@ class AutoRouter {
         const selection = decision.requestedSelection;
         if (selection.kind === 'fixed')
             return attempts.slice(0, remainingAttempts);
-        const scope = selection.kind === 'auto' ? selection.scope : { kind: 'provider', providerId: current.providerId };
         const subset = selection.kind === 'auto' ? selection.subset : undefined;
         const currentGroup = current.logicalModelGroupId;
         const currentProviderId = current.providerId;
-        const now = this.now();
         const attemptedDeployments = decision.attempts.map(attempt => attempt.deployment);
         // Fallback is a recovery operation, not a new global routing decision.
         // It must never cross the provider boundary, even when the original Auto
@@ -299,26 +248,33 @@ class AutoRouter {
         // switching credentials/endpoints (for example provider A/model X to
         // provider B/model X).
         const eligible = candidates.filter(candidate => candidate.enabled
+            && !candidate.unavailableReason
             && candidate.deployment.providerId === currentProviderId
-            && inScope(candidate.deployment, scope)
-            && inSubset(candidate.deployment, subset)
-            && !sameDeployment(candidate.deployment, current)
-            && !attemptedDeployments.some(attempted => sameDeployment(candidate.deployment, attempted))
+            && (0, routeEligibility_1.inAutoScope)(candidate.deployment, { kind: 'provider', providerId: currentProviderId })
+            && (0, routeEligibility_1.inAutoSubset)(candidate.deployment, subset)
+            && !(0, routeEligibility_1.sameDeploymentRef)(candidate.deployment, current)
+            && !attemptedDeployments.some(attempted => (0, routeEligibility_1.sameDeploymentRef)(candidate.deployment, attempted))
             && this.passedInitialHardFilters(decision, candidate));
         const equivalent = currentGroup
             ? eligible.find(candidate => candidate.deployment.logicalModelGroupId === currentGroup && !candidate.fallbackOnly)
             : undefined;
         const fallback = eligible.find(candidate => candidate.fallbackOnly);
-        const rankedAlternates = decision.rankedCandidates
-            .map(ranked => eligible.find(candidate => sameDeployment(candidate.deployment, ranked.deployment)))
-            .filter((candidate) => !!candidate && candidate !== equivalent && candidate !== fallback && !candidate.fallbackOnly);
+        const orderedAlternates = [];
+        for (const alternate of decision.alternatives) {
+            const match = eligible.find(candidate => (0, routeEligibility_1.sameDeploymentRef)(candidate.deployment, alternate));
+            if (!match || match === equivalent || match === fallback || match.fallbackOnly)
+                continue;
+            if (orderedAlternates.some(existing => (0, routeEligibility_1.sameDeploymentRef)(existing.deployment, match.deployment)))
+                continue;
+            orderedAlternates.push(match);
+        }
         for (const candidate of eligible) {
             if (candidate === equivalent || candidate === fallback || candidate.fallbackOnly
-                || rankedAlternates.some(existing => sameDeployment(existing.deployment, candidate.deployment)))
+                || orderedAlternates.some(existing => (0, routeEligibility_1.sameDeploymentRef)(existing.deployment, candidate.deployment)))
                 continue;
-            rankedAlternates.push(candidate);
+            orderedAlternates.push(candidate);
         }
-        for (const next of [equivalent, fallback, ...rankedAlternates]) {
+        for (const next of [equivalent, ...orderedAlternates, fallback]) {
             if (!next || attempts.length >= 2)
                 continue;
             attempts.push({
@@ -396,96 +352,6 @@ class AutoRouter {
             circuit: this.circuitState(deployment, now, false),
         };
     }
-    recordFeedback(event) {
-        const at = event.at ?? this.now();
-        const safe = {
-            deployment: { ...event.deployment },
-            taskClass: event.taskClass,
-            score: clamp(Number(event.score) || 0, -1, 1),
-            source: event.source,
-            at,
-        };
-        const key = feedbackKey(safe.deployment, safe.taskClass);
-        const events = (this.feedback.get(key) || []).filter(item => at - Number(item.at || 0) <= FEEDBACK_WINDOW_MS);
-        events.push(safe);
-        this.feedback.set(key, events);
-        return events.length >= 3;
-    }
-    clearLearnedPreferences() {
-        this.feedback.clear();
-    }
-    learnedPreference(deployment, taskClasses) {
-        const now = this.now();
-        const values = [];
-        for (const taskClass of taskClasses) {
-            const events = (this.feedback.get(feedbackKey(deployment, taskClass)) || [])
-                .filter(event => now - Number(event.at || 0) <= FEEDBACK_WINDOW_MS)
-                .sort((left, right) => Number(left.at || 0) - Number(right.at || 0));
-            if (events.length < 3)
-                continue;
-            let ewma = events[0].score;
-            let previousAt = Number(events[0].at || now);
-            for (const event of events.slice(1)) {
-                const eventAt = Number(event.at || previousAt);
-                const betweenDecay = Math.pow(0.5, Math.max(0, eventAt - previousAt) / FEEDBACK_HALF_LIFE_MS);
-                ewma = 0.2 * event.score + 0.8 * ewma * betweenDecay;
-                previousAt = eventAt;
-            }
-            const newestAt = Number(events[events.length - 1].at || now);
-            const decay = Math.pow(0.5, Math.max(0, now - newestAt) / FEEDBACK_HALF_LIFE_MS);
-            values.push(clamp(ewma * decay, -1, 1));
-        }
-        return values.length ? values.reduce((sum, value) => sum + value, 0) / values.length : 0;
-    }
-    rankEligible(candidates, taskClasses, policy, request, now, affinityDeployment) {
-        const qualities = candidates.map(candidate => ({ candidate, quality: qualityScore(candidate, taskClasses) }));
-        const bestQuality = Math.max(...qualities.map(item => item.quality));
-        const band = qualities.filter(item => bestQuality - item.quality <= policy.maxQualityLoss + Number.EPSILON);
-        return band.map(item => this.rankCandidate(item.candidate, taskClasses, policy, request, !!affinityDeployment && sameDeployment(item.candidate.deployment, affinityDeployment), now, item.quality))
-            .sort((left, right) => right.utility - left.utility || right.quality - left.quality || deploymentKey(left.deployment).localeCompare(deploymentKey(right.deployment)));
-    }
-    rankCandidate(candidate, taskClasses, policy, request, cache, _now, quality = qualityScore(candidate, taskClasses)) {
-        const expectedCostUsd = expectedRequestCost(candidate, request);
-        const cost = expectedCostUsd === undefined ? 0 : 1 / (1 + Math.max(0, expectedCostUsd) * 1_000);
-        const endpointReliability = candidate.reliability === undefined ? 0.5 : clamp(candidate.reliability);
-        const toolRelevant = taskClasses.some(taskClass => taskClass === 'tool_use' || taskClass === 'computer_use');
-        const reliability = toolRelevant && candidate.toolValidity !== undefined
-            ? (endpointReliability + clamp(candidate.toolValidity)) / 2
-            : endpointReliability;
-        const latency = candidate.latencyMs === undefined ? 0.5 : 1 / (1 + Math.max(0, candidate.latencyMs) / 1_000);
-        const throughput = candidate.throughput === undefined ? 0.5 : Math.max(0, candidate.throughput) / (Math.max(0, candidate.throughput) + 40);
-        const speed = (latency + throughput) / 2;
-        const learned = this.learnedPreference(candidate.deployment, taskClasses);
-        const configuredPreference = clamp(Number(candidate.preference) || 0, -1, 1);
-        const effectivePreference = configuredPreference === 0 ? learned : configuredPreference;
-        const preference = clamp((effectivePreference + 1) / 2);
-        const components = { cost, reliability, speed, cache: cache ? 1 : 0, preference };
-        let utility;
-        switch (policy.mode) {
-            case 'quality':
-                utility = quality * 0.55 + reliability * 0.25 + speed * 0.15 + cost * 0.05;
-                break;
-            case 'cost':
-                utility = cost * 0.65 + reliability * 0.15 + speed * 0.10 + components.cache * 0.05 + preference * 0.05;
-                break;
-            case 'speed':
-                utility = speed * 0.60 + reliability * 0.20 + cost * 0.10 + components.cache * 0.05 + preference * 0.05;
-                break;
-            default:
-                utility = cost * 0.40 + reliability * 0.25 + speed * 0.20 + components.cache * 0.10 + preference * 0.05;
-                break;
-        }
-        return {
-            deployment: { ...candidate.deployment },
-            quality,
-            utility: clamp(utility),
-            expectedCostUsd,
-            components,
-        };
-    }
-    affinityKey(key, scope, mode) {
-        return `${key}\u0000${scope.kind === 'global' ? 'global' : `provider:${scope.providerId}`}\u0000${mode}`;
-    }
     healthFor(deployment, now) {
         const key = deploymentKey(deployment);
         let health = this.endpointHealth.get(key);
@@ -497,12 +363,10 @@ class AutoRouter {
         return health;
     }
     passedInitialHardFilters(decision, candidate) {
-        const exclusion = decision.excludedCandidates.find(entry => sameDeployment(entry.deployment, candidate.deployment));
+        const exclusion = decision.excludedCandidates.find(entry => (0, routeEligibility_1.sameDeploymentRef)(entry.deployment, candidate.deployment));
         if (!exclusion)
             return true;
-        return !!candidate.fallbackOnly
-            && exclusion.reasons.length > 0
-            && exclusion.reasons.every(reason => reason === 'fallback_only');
+        return !!candidate.fallbackOnly && (0, routeEligibility_1.excludedOnlyForFallbackPool)(exclusion);
     }
     circuitState(deployment, now, claimHalfOpen = true) {
         const health = this.healthFor(deployment, now);
@@ -519,61 +383,12 @@ class AutoRouter {
         return 'closed';
     }
 }
-exports.AutoRouter = AutoRouter;
-function cloneSelection(selection) {
-    if (selection.kind === 'fixed')
-        return { kind: 'fixed', deployment: { ...selection.deployment } };
-    return {
-        kind: 'auto',
-        scope: selection.scope.kind === 'global' ? { kind: 'global' } : { kind: 'provider', providerId: selection.scope.providerId },
-        policyId: selection.policyId,
-        subset: selection.subset?.map(item => ({ ...item })),
-    };
-}
-function qualityScore(candidate, taskClasses) {
-    const scores = taskClasses.map(taskClass => {
-        const metric = candidate.qualityByTask?.[taskClass];
-        if (!metric)
-            return 0.5;
-        const attempts = Math.max(0, Math.floor(Number(metric.attempts) || 0));
-        const successes = clamp(Number(metric.successes) || 0, 0, attempts);
-        return (successes + 2) / (attempts + 4);
-    });
-    return scores.reduce((sum, value) => sum + value, 0) / Math.max(1, scores.length);
-}
-function expectedRequestCost(candidate, request) {
-    const input = finite(candidate.expectedInputCostUsdPerM);
-    const output = finite(candidate.expectedOutputCostUsdPerM);
-    if (input === undefined || output === undefined)
-        return undefined;
-    return input * Math.max(0, request.estimatedInputTokens) / 1_000_000
-        + output * Math.max(0, request.expectedOutputTokens) / 1_000_000;
-}
-function catalogHash(candidates) {
-    const snapshot = candidates.map(candidate => ({
-        deployment: candidate.deployment,
-        enabled: candidate.enabled,
-        validation: candidate.validation,
-        capabilities: [...candidate.capabilities].sort(),
-        maxContextTokens: candidate.maxContextTokens,
-        preview: candidate.preview,
-        privacy: [...candidate.privacy].sort(),
-        dataRegions: [...(candidate.dataRegions || [])].sort(),
-        supportedProtocolParameters: [...(candidate.supportedProtocolParameters || [])].sort(),
-        expectedInputCostUsdPerM: candidate.expectedInputCostUsdPerM,
-        expectedOutputCostUsdPerM: candidate.expectedOutputCostUsdPerM,
-        latencyMs: candidate.latencyMs,
-        reliability: candidate.reliability,
-        toolValidity: candidate.toolValidity,
-        throughput: candidate.throughput,
-        qualityByTask: candidate.qualityByTask,
-        preference: candidate.preference,
-        fallbackOnly: !!candidate.fallbackOnly,
-    })).sort((left, right) => deploymentKey(left.deployment).localeCompare(deploymentKey(right.deployment)));
-    return (0, crypto_1.createHash)('sha256').update(JSON.stringify(snapshot)).digest('hex');
-}
-function feedbackKey(deployment, taskClass) {
-    return `${deploymentKey(deployment)}\u0000${taskClass}`;
+exports.RouteExecutionController = RouteExecutionController;
+function validationEvidenceIsStale(checkedAt, now = Date.now()) {
+    const checked = Date.parse(String(checkedAt || ''));
+    if (!Number.isFinite(checked))
+        return true;
+    return now - checked > VALIDATION_TTL_MS;
 }
 function trimHealth(health, now) {
     health.events = health.events.filter(event => now - event.at <= HEALTH_WINDOW_MS).slice(-100);

package/dist/core/computerUseSession.js

@@ -21,7 +21,7 @@ class ComputerUseSessionRegistry {
         if (normalizedAction === 'takeover_stop')
             return null;
         const enabled = this.enabledByRuntime.get(runtimeKey) !== false;
-        const readOnly = normalizedAction === 'observe' || normalizedAction === 'app_list' || normalizedAction === 'app_observe' || normalizedAction === 'wait';
+        const readOnly = normalizedAction === 'observe' || normalizedAction === 'app_list' || normalizedAction === 'app_observe' || normalizedAction === 'wait' || normalizedAction === 'wait_for';
         if (!enabled && !readOnly && !dryRun && normalizedAction !== 'takeover_start') {
             return JSON.stringify({
                 ok: false,

package/dist/core/config.js

@@ -90,9 +90,10 @@ class ConfigManager {
                 }
                 const providerIdsMigrated = migrateProviderIdsInConfig(normalized);
                 const marqueeConfigRemoved = removeDeprecatedMarqueeConfig(normalized);
-                if (providerIdsMigrated || marqueeConfigRemoved) {
-                    // Provider ids are routing identities, so legacy/malformed catalogs must
-                    // not wait for an unrelated settings save before becoming collision-safe.
+                const legacyJevRuntimeRemoved = removeLegacyJevDecisionRuntimeConfig(normalized);
+                if (providerIdsMigrated || marqueeConfigRemoved || legacyJevRuntimeRemoved) {
+                    // Persist migrations immediately so obsolete JEV slot settings are not
+                    // left in the active user config until an unrelated settings save.
                     try {
                         if (!this.readOnly)
                             fs.writeFileSync(cp, JSON.stringify(normalized, null, 2), 'utf-8');
@@ -356,13 +357,7 @@ class ConfigManager {
     autoSwitchEnabled() { return this.getBool('models', 'auto_switch'); }
     autoSwitchPreference() {
         const value = this.getStr('models', 'auto_switch_preference');
-        if (value === 'performance')
-            return 'quality';
-        if (value === 'cheap_save')
-            return 'cost';
-        if (value === 'speed')
-            return 'speed';
-        return value === 'quality' || value === 'cost' || value === 'balanced' ? value : 'balanced';
+        return value === 'conservative' || value === 'balanced' || value === 'aggressive' ? value : 'balanced';
     }
     autoSwitchScope() { return this.getStr('models', 'auto_switch_scope') || 'all'; }
     autoSwitchAnchorProvider() { return this.getStr('models', 'auto_switch_anchor_provider'); }
@@ -478,6 +473,25 @@ function removeDeprecatedMarqueeConfig(config) {
     }
     return changed;
 }
+function removeLegacyJevDecisionRuntimeConfig(config) {
+    const models = config.models;
+    if (!models)
+        return false;
+    const retiredKeys = [
+        'jev_runtime', 'jev_model_id', 'jev_model_path',
+        'jev_fallback_strategy', 'jev_fallback_provider_id', 'jev_fallback_model_id',
+    ];
+    const hadRetiredInterface = retiredKeys.some(key => Object.prototype.hasOwnProperty.call(models, key));
+    if (!hadRetiredInterface)
+        return false;
+    const legacyTimeout = models.jev_timeout_ms?.value;
+    if (legacyTimeout === 100 && models.jev_timeout_ms) {
+        models.jev_timeout_ms.value = 3000;
+    }
+    for (const key of retiredKeys)
+        delete models[key];
+    return true;
+}
 function isConfigEntry(value) {
     return !!value && typeof value === 'object' && !Array.isArray(value) && Object.prototype.hasOwnProperty.call(value, 'value');
 }
@@ -879,7 +893,7 @@ function defaultConfig() {
             default_intelligence: { _description: "Default reasoning effort", _type: "choice", _values: ["low", "medium", "high", "xhigh", "max", "ultra"], value: "medium" },
             agent_engine: { _description: "Agent engine", _type: "choice", _values: ["builtin", "codex", "opencode"], value: "builtin" },
             auto_switch: { _description: "Auto-switch models", _type: "boolean", value: false },
-            auto_switch_preference: { _description: "Auto route policy", _type: "choice", _values: ["quality", "balanced", "cost", "speed"], value: "balanced" },
+            auto_switch_preference: { _description: "Qualitative tendency for current-model Auto decisions", _type: "choice", _values: ["conservative", "balanced", "aggressive"], value: "balanced" },
             auto_switch_scope: { _description: "Auto-switch scope", _type: "choice", _values: ["all", "provider"], value: "all" },
             auto_switch_anchor_provider: { _description: "Stable provider id anchor for provider-scoped Auto routing", _type: "string", value: "" },
             auto_switch_subset: { _description: "Explicit deployment allowlist for Auto; new catalog models are never added automatically", _type: "array", value: [] },
@@ -888,6 +902,8 @@ function defaultConfig() {
             auto_data_region: { _description: "Required Auto data region; empty disables the region filter", _type: "string", value: "" },
             auto_required_protocol_parameters: { _description: "Protocol parameters every Auto candidate must support", _type: "array", value: [] },
             auto_max_expected_cost_usd: { _description: "Hard expected request cost ceiling; 0 disables the ceiling", _type: "number", value: 0 },
+            jev_enabled: { _description: "Use the current eligible conversation model for strict-JSON Auto decisions", _type: "boolean", value: true },
+            jev_timeout_ms: { _description: "Hard timeout for one current-model JEV decision in milliseconds", _type: "number", value: 3000 },
             fallback_on_unavailable: { _description: "Fallback when model unavailable", _type: "boolean", value: true },
             openai_api_mode: { _description: "OpenAI-compatible API mode", _type: "choice", _values: ["chat_stream", "chat", "responses"], value: "chat_stream" },
             openai_streaming: { _description: "Legacy streaming flag for OpenAI-compatible chat completions", _type: "boolean", value: true },

package/dist/core/continuation/contracts.d.ts

@@ -146,7 +146,7 @@ export interface FinalRecord {
 export interface ContinuationEvent {
     eventId: string;
     cursor: number;
-    type: 'BuildAccepted' | 'AttemptStarted' | 'AttemptHeartbeat' | 'BuildCommitted' | 'BuildFailed' | 'BuildCancelled' | 'BranchForked' | 'BranchPaused' | 'BranchResumed' | 'BuildQueueRepaired' | 'StaleWriteRejected';
+    type: 'BuildAccepted' | 'AttemptStarted' | 'AttemptHeartbeat' | 'BuildCommitted' | 'BuildFailed' | 'BuildCancelled' | 'BranchForked' | 'BranchPaused' | 'BranchResumed' | 'BuildQueueRepaired' | 'AttemptOrphanRecovered' | 'StaleWriteRejected';
     workspaceId: WorkspaceId;
     rootId: RootId;
     branchId: BranchId;