newmark-agent 0.5.15 → 0.6.1

package/dist/core/agent.d.ts

@@ -1097,6 +1097,7 @@ export declare class Agent {
     evaluateAndSwitch(task: string, override?: AgentPromptMessage['routePolicy']): Promise<boolean>;
     shouldExposeToolInterface(): boolean;
     modelIsUnavailable(modelName: string): boolean;
+    private failedDeploymentsThisRun;
     switchToFallbackModel(errorText?: string): string | null;
     isLlmErrorText(text: string): boolean;
     private scopedSwitchModels;
@@ -1109,8 +1110,8 @@ export declare class Agent {
     /**
      * Final visual safety net: OCR each submitted image and ask a text-only
      * request to conservatively repair the OCR. This is intentionally callable
-     * only after a visual-input refusal and after same-provider vision routing
-     * has been exhausted; the original image is never sent again.
+     * only after a real visual-input refusal. Cached capabilities never bypass
+     * the initial image request, and OCR does not disable later image attempts.
      */
     finalVisualFallback(errorText: string, signal?: AbortSignal): Promise<string | null>;
     engineModel(cache?: BuildProviderCache): LLMProvider | null;

package/dist/core/agent.js

@@ -35,6 +35,7 @@ var __importStar = (this && this.__importStar) || (function () {
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.Agent = exports.ROOT_AGENT_ACTOR_ID = void 0;
 exports.normalizeIntelligenceTier = normalizeIntelligenceTier;
+const modelResponseHealth_1 = require("./modelResponseHealth");
 const fs = __importStar(require("fs"));
 const path = __importStar(require("path"));
 const crypto = __importStar(require("crypto"));
@@ -939,13 +940,12 @@ class Agent {
         return this.modelSelectionValue();
     }
     defaultModelCandidate() {
-        const rejected = new Set(['unavailable', 'auth_error', 'invalid_config']);
         return this.config.allModels()
             .map((model, index) => {
             const validationStatus = effectiveModelValidationStatus(model);
             const evaluationStatus = validationStatus === 'degraded' ? 'degraded' : String(model.evaluation?.status || '').toLowerCase();
             const level = String(model.validation?.level || '').toLowerCase();
-            if (model.enabled === false || rejected.has(validationStatus) || rejected.has(evaluationStatus) || evaluationStatus.startsWith('error')) {
+            if (model.enabled === false) {
                 return null;
             }
             let score = 0;
@@ -1420,8 +1420,8 @@ class Agent {
         const reference = parsed.image;
         const hydrated = this.hydrateDisplayImage(reference);
         const model = this.activeModelConfig();
-        const provider = model?.vision ? this.engineModel() : null;
-        if (!provider || !model?.vision || !hydrated?.dataUrl || !hydrated.sha256)
+        const provider = this.engineModel();
+        if (!provider || !hydrated?.dataUrl || !hydrated.sha256)
             return raw;
         throwIfAgentAborted(signal);
         const deployment = this.activeDeployment();
@@ -3905,7 +3905,9 @@ class Agent {
             // The title request uses the same transport policy and cancellation owner
             // as the formal response. A healthy slow model must not be turned into
             // five failed requests by an unrelated 15-second title deadline.
-            const generated = await this.chatWithConversationUsage(provider, modelName, [{ role: 'user', content: prompt }], system, temperature, 64, signal, reasoningEffort);
+            // Reasoning tokens share the completion budget. A 64-token cap can
+            // truncate even a short title and block every first conversation request.
+            const generated = await this.chatWithConversationUsage(provider, modelName, [{ role: 'user', content: prompt }], system, temperature, 2048, signal, reasoningEffort);
             if (signal?.aborted)
                 return '';
             const title = this.normalizeConversationRenameTitle(generated);
@@ -6183,10 +6185,13 @@ class Agent {
             if (active)
                 return active;
         }
-        const byName = this.config.findModel(modelName);
+        const parsed = parseDeploymentSelectionValue(modelName);
+        const byName = parsed ? this.config.findDeployment(parsed) : this.config.findModel(modelName);
         if (byName)
             return byName;
-        return this.config.findModel(this.config.getStr('models', 'default_model'));
+        const configured = this.config.getStr('models', 'default_model');
+        const configuredRef = parseDeploymentSelectionValue(configured);
+        return configuredRef ? this.config.findDeployment(configuredRef) : this.config.findModel(configured);
     }
     contextMaxTokens(modelName = this.model) {
         const model = this.resolveWindowModel(modelName);
@@ -6978,7 +6983,9 @@ class Agent {
         const active = this.activeModelConfig();
         if (active?.provider_id)
             return active.provider_id;
-        const fallback = this.config.findModel(this.config.getStr('models', 'default_model'));
+        const defaultSelection = this.config.getStr('models', 'default_model');
+        const defaultRef = parseDeploymentSelectionValue(defaultSelection);
+        const fallback = defaultRef ? this.config.findDeployment(defaultRef) : this.config.findModel(defaultSelection);
         return fallback?.provider_id || this.config.providers()[0]?.id || '';
     }
     autoSwitchSubset() {
@@ -7096,7 +7103,7 @@ class Agent {
                 preference: Number.isFinite(Number(routeMetadata.route_preference)) ? Number(routeMetadata.route_preference) : undefined,
                 fallbackOnly: !!model.fallback_only,
             };
-        }).filter(candidate => !this.isBalanceBlockedDeployment(candidate.deployment));
+        });
     }
     persistRouteDecision(decision) {
         try {
@@ -7143,15 +7150,11 @@ class Agent {
         catch { }
     }
     allModelNames() {
-        const names = this.config.allModels().filter(m => {
-            const status = String(m.evaluation?.status || 'unvalidated');
-            const validationStatus = m.validation?.status;
-            return status === 'available' || status === 'unvalidated' || validationStatus === 'verified' || validationStatus === 'degraded';
-        }).map(m => {
+        const names = this.config.allModels().filter(m => m.enabled !== false).map(m => {
             const label = m.display || m.name;
             return `${m.provider} / ${label}`;
         });
-        return this.config.autoSwitchEnabled() && this.config.allModels().length > 0 ? ['auto', ...names] : names;
+        return this.config.autoSwitchEnabled() && names.length > 0 ? ['auto', ...names] : names;
     }
     async evaluateAndSwitch(task, override = undefined) {
         if (!this.config.autoSwitchEnabled() || this.model !== 'auto')
@@ -7241,39 +7244,22 @@ class Agent {
         return !!this.resolvedDeployment && before !== deploymentIdentity(this.resolvedDeployment);
     }
     shouldExposeToolInterface() {
-        // Fixed selections and pre-Standard legacy configurations keep the
-        // historical behavior. Auto can safely suppress schemas because its
-        // eligible deployments have explicit Standard/Extended evidence.
-        if (this.model !== 'auto')
-            return true;
-        const model = this.activeModelConfig();
-        if (!model)
-            return false;
-        const validation = model.validation;
-        if (validation?.level !== 'standard' && validation?.level !== 'extended')
-            return true;
-        return validation.capabilities?.tool_use === true || validation.capabilities?.tools === true;
+        return true;
     }
     modelIsUnavailable(modelName) {
         const model = modelName === this.model || modelName === 'auto'
-            ? this.activeModelConfig()
-            : this.config.findModel(modelName);
-        if (!model)
-            return true;
-        const validationStatus = effectiveModelValidationStatus(model);
-        // `discovered` means unvalidated, not a failed endpoint. Explicit fixed
-        // selections remain usable for backwards compatibility; only an executed
-        // Basic/Standard/Extended validation may pre-emptively mark them bad.
-        if (model.validation?.level !== 'discovered'
-            && (validationStatus === 'unavailable' || validationStatus === 'auth_error' || validationStatus === 'invalid_config'))
-            return true;
-        const status = validationStatus === 'degraded' ? 'degraded' : String(model?.evaluation?.status || '').toLowerCase();
-        return status === 'unavailable' || status.startsWith('error');
+            ? this.activeModelConfig() : (parseDeploymentSelectionValue(modelName) ? this.config.findDeployment(parseDeploymentSelectionValue(modelName)) : this.config.findModel(modelName));
+        // Only configuration/user switches can prevent a request. Response and
+        // validation observations are labels, never permission to disable a model.
+        return !model || model.enabled === false;
     }
+    failedDeploymentsThisRun = new Set();
     switchToFallbackModel(errorText = 'transport failure') {
         const fallbackEnabled = this.config.getBool('models', 'fallback_on_unavailable');
         const observedFailure = (0, autoRouter_1.classifyRouteFailure)(errorText);
         const observedDeployment = this.activeDeployment();
+        if (observedDeployment)
+            this.failedDeploymentsThisRun.add(deploymentIdentity(observedDeployment));
         // Keep balance exhaustion scoped to the deployment that actually failed.
         // Provider adapters normally record this before returning an error, but
         // fallback callers are also a public recovery boundary and must not rely
@@ -7345,7 +7331,7 @@ class Agent {
             || deploymentIdentity(this.deploymentRef(m)) !== deploymentIdentity(currentDeployment));
         if (!all.length)
             return null;
-        const usable = all.filter(m => !this.isBalanceBlockedDeployment(this.deploymentRef(m)) && !modelConfigIsUnavailable(m));
+        const usable = all.filter(m => !modelConfigIsUnavailable(m) && !this.failedDeploymentsThisRun.has(deploymentIdentity(this.deploymentRef(m))));
         if (!usable.length)
             return null;
         const pref = this.config.autoSwitchPreference();
@@ -7377,7 +7363,9 @@ class Agent {
         // which is exactly the same-name cross-provider leak this guard exists to
         // prevent. When no provider can be established unambiguously, return an
         // empty pool: failing the switch is safe, crossing providers is not.
-        const current = currentModelName === 'auto' ? this.activeModelConfig() : this.config.findModel(currentModelName);
+        const currentRef = parseDeploymentSelectionValue(currentModelName);
+        const current = currentModelName === 'auto' ? this.activeModelConfig()
+            : (currentRef ? this.config.findDeployment(currentRef) : this.config.findModel(currentModelName));
         const providerId = current?.provider_id
             || this.fixedDeployment?.providerId
             || (currentModelName !== 'auto' ? this.activeDeployment()?.providerId : undefined);
@@ -7577,8 +7565,8 @@ class Agent {
     /**
      * Final visual safety net: OCR each submitted image and ask a text-only
      * request to conservatively repair the OCR. This is intentionally callable
-     * only after a visual-input refusal and after same-provider vision routing
-     * has been exhausted; the original image is never sent again.
+     * only after a real visual-input refusal. Cached capabilities never bypass
+     * the initial image request, and OCR does not disable later image attempts.
      */
     async finalVisualFallback(errorText, signal) {
         if (!/(?:vision|image|multimodal|image_url|input_image).*(?:not supported|unsupported|拒绝|不支持|failed|failure|invalid)|(?:not supported|unsupported|拒绝|不支持).*(?:vision|image|multimodal|image_url|input_image)/i.test(String(errorText || '')))
@@ -7586,11 +7574,6 @@ class Agent {
         const current = this.activeModelConfig();
         if (!current)
             return null;
-        const alternateVision = this.config.allModels().some(model => model.enabled !== false && model.provider_id === current.provider_id &&
-            model.name !== current.name && !!model.vision && !!model.api_key && !!model.provider_url &&
-            !['unavailable', 'auth_error', 'invalid_config'].includes(String(model.evaluation?.status || model.validation?.status || '').toLowerCase()));
-        if (alternateVision)
-            return null;
         const latest = [...this.history].reverse().find(item => item?.role === 'user');
         const parts = latest?.content && Array.isArray(latest.content) ? latest.content : [];
         const images = parts.map(part => {
@@ -7606,11 +7589,16 @@ class Agent {
                 if (result.ok && result.text.trim())
                     ocr.push({ index: index + 1, text: result.text.slice(0, 50_000), confidence: result.confidence });
             }
-            catch { }
+            catch (error) {
+                if (signal?.aborted)
+                    throw error;
+            }
         }
+        if (signal?.aborted)
+            throw signal.reason || new Error('Aborted');
         if (!ocr.length)
             return JSON.stringify({ ok: false, fallback: 'mini_ocr_llm', error: 'Local OCR returned no readable text; no visual content was fabricated.' });
-        const task = typeof latest?.content === 'string' ? latest.content : '';
+        const task = typeof latest?.content === 'string' ? latest.content : parts.filter(part => part.type === 'text').map(part => String(part.text || '')).join('\n');
         const evidence = ocr.map(item => `Image ${item.index} (OCR confidence ${item.confidence.toFixed(1)}):\n${item.text}`).join('\n\n');
         const prompt = `The provider rejected image input. Answer the user's task using only this approximate OCR evidence. Correct obvious character, spacing, and line-break errors only when supported by context. Preserve [uncertain] markers for ambiguity and never invent missing visual content.\nUser task:\n${task.slice(0, 12_000)}\nOCR evidence:\n${evidence}`;
         let corrected = '';
@@ -7619,12 +7607,15 @@ class Agent {
             if (provider)
                 corrected = String(await this.chatWithConversationUsage(provider, this.activeModelName(), [{ role: 'user', content: prompt }], 'You are a text-only OCR correction assistant. Be conservative and explicit about uncertainty.', 0.05, 3000, signal) || '').trim();
         }
-        catch { }
+        catch (error) {
+            if (signal?.aborted)
+                throw error;
+        }
         return JSON.stringify({
             ok: !!(corrected || ocr.length),
             fallback: 'mini_ocr_llm',
             approximate: true,
-            warning: '视觉输入被拒绝；以下内容来自本地 OCR，并经文本模型保守校正，可能不完整。',
+            warning: corrected ? '视觉输入被拒绝；以下内容来自本地 OCR，并经文本模型保守校正，可能不完整。' : '视觉输入被拒绝，文本校正请求也未成功；以下为未经校正的本地 OCR 结果，可能不完整。',
             raw_ocr: ocr,
             corrected: corrected || ocr.map(item => item.text).join('\n\n'),
             uncertainty: corrected ? 'preserved' : 'raw_ocr_only',
@@ -7657,7 +7648,9 @@ class Agent {
         const adapters = this.config.contextFlag('provider_adapters_v2');
         const thinkingMaps = this.modelThinkingTierMaps(m);
         const proxy = this.providerProxyConfig();
-        const create = () => new provider_1.LLMProvider(m.provider, m.provider_url, m.api_key, m.provider_protocol, apiMode, adapters, undefined, thinkingMaps, proxy);
+        const create = () => (0, modelResponseHealth_1.observeModelResponses)(new provider_1.LLMProvider(m.provider, m.provider_url, m.api_key, m.provider_protocol, apiMode, adapters, undefined, thinkingMaps, proxy), updates => (0, modelResponseHealth_1.recordModelResponseHealth)(this.config.rootPath, {
+            providerId: m.provider_id, modelId: m.name, endpoint: m.provider_url, protocol: m.provider_protocol, credential: m.api_key,
+        }, updates));
         if (!cache)
             return create();
         // Match the provider's configured-proxy/environment precedence. A changed
@@ -7974,6 +7967,7 @@ class Agent {
             return [{ type: 'text', text: '[Workspace required] Select or create a workspace before starting a conversation.' }];
         }
         if (this.processDepth === 0) {
+            this.failedDeploymentsThisRun.clear();
             this.processingConversationId = this.activeConversationId || 'default';
             this.activeProcessAbortController = new AbortController();
             this.subagents.resumeScheduling();
@@ -7992,27 +7986,6 @@ class Agent {
         this.fileDiffs = [];
         this.pendingOptions = [];
         try {
-            if (this.model === 'auto') {
-                // Auto routing re-resolves each turn; drop a stale blocked deployment
-                // so the router can pick an unblocked candidate instead of failing here.
-                if (this.resolvedDeployment && this.isBalanceBlockedDeployment(this.resolvedDeployment)) {
-                    this.resolvedDeployment = null;
-                    this.lastRouteDecision = null;
-                    this.pendingAutoAttempts = [];
-                }
-            }
-            else {
-                const blockedMs = this.balanceBlockedMs();
-                if (blockedMs > 0) {
-                    const waitSeconds = Math.max(1, Math.ceil(blockedMs / 1000));
-                    const deployment = this.activeDeployment();
-                    const label = deployment ? `${deployment.providerId}/${deployment.modelId}` : this.model;
-                    const message = `Provider balance exhausted (HTTP 402) for ${label}. Requests on this deployment are paused for ${waitSeconds}s; switch provider or model to continue immediately.`;
-                    this.status = 'error';
-                    this.emitWorkEvent({ type: 'error', content: message });
-                    throw new Error(message);
-                }
-            }
             let text = typeof input === 'string' ? input : String(input.text || '');
             const inputEnvelope = typeof input === 'string' ? null : input;
             let hiddenUserInput = inputEnvelope?.hiddenUserInput === true;
@@ -8109,16 +8082,6 @@ class Agent {
                 await this.evaluateAndSwitch(`${text}\n[image attachment]`, inputEnvelope?.routePolicy);
                 autoRouteEvaluated = true;
             }
-            const selectedModel = this.activeModelConfig();
-            if (images.length && !selectedModel?.vision) {
-                // Give the normal route planner first chance to select another
-                // same-provider vision deployment. If none is available, the kernel
-                // preflight invokes the final mini-OCR + text-only correction path.
-                const hasSameProviderVision = selectedModel && this.config.allModels().some(model => model.enabled !== false && model.provider_id === selectedModel.provider_id &&
-                    model.name !== selectedModel.name && !!model.vision);
-                if (hasSameProviderVision)
-                    this.switchToFallbackModel('vision input not supported by the selected model');
-            }
             const now = this.nowLabel();
             const visibleUserInput = inputEnvelope?.visibleUserInput === undefined
                 ? text
@@ -9217,9 +9180,7 @@ class Agent {
     }
     subagentToolDefinitions(defs) {
         const modelCapabilities = this.activeModelConfig();
-        const visionFiltered = modelCapabilities?.vision
-            ? defs
-            : defs.filter((tool) => tool.function?.name !== 'image_inspect');
+        const visionFiltered = defs;
         const withImageGeneration = modelCapabilities?.image_output
             ? [...visionFiltered, {
                     type: 'function',
@@ -9263,8 +9224,6 @@ class Agent {
         }
     }
     async handleImageInspect(args) {
-        if (!this.activeModelConfig()?.vision)
-            return '[Image inspect unavailable] The selected model has not passed vision validation.';
         let input = {};
         try {
             input = JSON.parse(args);
@@ -10562,7 +10521,9 @@ function routeProviderFingerprint(provider) {
     });
 }
 function modelConfigurationFingerprint(model) {
-    const { validation, evaluation, _previous_name, previous_name, ...configuration } = model;
+    const { validation, evaluation, response_health, enabled, _previous_name, previous_name, ...configuration } = model;
+    void response_health;
+    void enabled;
     void validation;
     void evaluation;
     void _previous_name;
@@ -10604,12 +10565,7 @@ function resetEditedModelValidationEvidence(incomingProviders, existingProviders
     }
 }
 function modelConfigIsUnavailable(model) {
-    const validationStatus = effectiveModelValidationStatus(model);
-    if (model.validation?.level !== 'discovered'
-        && (validationStatus === 'unavailable' || validationStatus === 'auth_error' || validationStatus === 'invalid_config'))
-        return true;
-    const evaluationStatus = validationStatus === 'degraded' ? 'degraded' : String(model.evaluation?.status || '').toLowerCase();
-    return evaluationStatus === 'unavailable' || evaluationStatus.startsWith('error');
+    return model.enabled === false;
 }
 function parseDeploymentSelectionValue(value) {
     const marker = String(value || '').trim();

package/dist/core/agentKernelRunner.js

@@ -480,15 +480,18 @@ async function runAgentKernel(agent) {
     try {
         const linkedPlanRevisionBeforeRun = agent.getLinkedPlan().revision;
         const modelBeforeKernelRun = agent.model;
-        const preflightVisualFallback = !agent.activeModelConfig()?.vision
-            ? await agent.finalVisualFallback('vision input not supported by the selected model', processSignal)
-            : null;
-        let lastTurn = preflightVisualFallback
-            ? { text: preflightVisualFallback, stopReason: 'stop', errorMessage: '' }
-            : await runWithCompressionResume([], false);
-        if (preflightVisualFallback) {
-            tokens.push({ type: 'text', text: preflightVisualFallback });
-            agent.recordWorkStatus('Final visual fallback used: local mini OCR plus conservative text correction.');
+        let lastTurn = await runWithCompressionResume([], false);
+        if (kernelTurnFailed(agent, lastTurn)) {
+            const visualFallback = await agent.finalVisualFallback(lastTurn.errorMessage || lastTurn.text, processSignal);
+            if (visualFallback) {
+                tokens.push({ type: 'text', text: visualFallback });
+                agent.emitWorkEvent({ type: 'final_response', content: visualFallback });
+                agent.chatMessages.push({ role: 'assistant', content: visualFallback, mode: agent.modeName(), model: agent.model, timestamp: agent.nowLabel(), runId: agent.currentWorkRunId() || undefined });
+                agent.history.push({ role: 'assistant', content: visualFallback, run_id: agent.currentWorkRunId() || undefined });
+                agent.saveWorkspaceConversationState();
+                agent.recordWorkStatus('Final visual fallback used: local mini OCR plus conservative text correction.');
+                lastTurn = { ...lastTurn, text: visualFallback, errorMessage: '', stopReason: 'stop' };
+            }
         }
         if (modelBeforeKernelRun && modelBeforeKernelRun !== agent.model && !tokens.some(t => t.text?.includes('[Model fallback]'))) {
             const notice = `[Model fallback] ${modelBeforeKernelRun} unavailable; switched to ${agent.model}.`;
@@ -601,14 +604,6 @@ async function runAgentKernel(agent) {
             await agent.waitForPlannedRouteRetry();
             lastTurn = await runWithCompressionResume([], false);
         }
-        if (kernelTurnFailed(agent, lastTurn)) {
-            const visualFallback = await agent.finalVisualFallback(lastTurn.errorMessage || lastTurn.text, processSignal);
-            if (visualFallback) {
-                tokens.push({ type: 'text', text: visualFallback });
-                agent.recordWorkStatus('Final visual fallback used: local mini OCR plus conservative text correction.');
-                lastTurn = { ...lastTurn, text: visualFallback, errorMessage: '', stopReason: 'stop' };
-            }
-        }
         if (kernelTurnFailed(agent, lastTurn)) {
             throw new ProviderRunError(normalizePublicProviderError(lastTurn.errorMessage || lastTurn.text, [agent.activeModelConfig()?.api_key]));
         }
@@ -1167,7 +1162,7 @@ function toKernelModel(agent) {
         provider: m?.provider || 'newmark',
         baseUrl: m?.provider_url || '',
         reasoning: !!m?.thinking,
-        input: m?.vision ? ['text', 'image'] : ['text'],
+        input: ['text', 'image'],
         cost: {
             input: Number(m?.cost_per_1k_input || 0) * 1000,
             output: Number(m?.cost_per_1k_output || 0) * 1000,
@@ -1752,8 +1747,6 @@ function visualFallbackImageInput(agent, name, text) {
     if (name !== 'screen_capture' && name !== 'computer_use' && name !== 'browser_use' && name !== 'pdf_read')
         return {};
     const model = agent.activeModelConfig();
-    if (!model?.vision)
-        return {};
     try {
         const parsed = JSON.parse(text);
         const nested = name === 'pdf_read' && parsed.result && typeof parsed.result === 'object'
@@ -1917,8 +1910,7 @@ async function executeNewmarkTool(agent, name, args, inputSchema, signal, onSett
             actorId: agent.runtimeActorId,
             workspaceId: (0, terminalTakeover_1.terminalTakeoverWorkspaceId)(wsDir),
             backend: process.env.NEWMARK_WSL_DISTRO ? 'wsl' : (process.platform === 'win32' ? 'windows' : process.platform),
-            allowEphemeralVisionImage: (name === 'screen_capture' || name === 'computer_use' || name === 'browser_use' || name === 'pdf_read' || name === 'ocr_read')
-                && !!agent.activeModelConfig()?.vision,
+            allowEphemeralVisionImage: (name === 'screen_capture' || name === 'computer_use' || name === 'browser_use' || name === 'pdf_read' || name === 'ocr_read'),
             signal,
         });
         if (signal?.aborted)

package/dist/core/autoRouter.js

@@ -34,19 +34,6 @@ function inSubset(deployment, subset) {
 function inScope(deployment, scope) {
     return scope.kind === 'global' || deployment.providerId === scope.providerId;
 }
-function validationEligible(candidate, now) {
-    const reasons = [];
-    if (candidate.validation.level !== 'standard' && candidate.validation.level !== 'extended') {
-        reasons.push(`validation_level:${candidate.validation.level}`);
-    }
-    if (candidate.validation.status !== 'verified' && candidate.validation.status !== 'degraded') {
-        reasons.push(`validation_status:${candidate.validation.status}`);
-    }
-    const checkedAt = Date.parse(candidate.validation.checkedAt);
-    if (!Number.isFinite(checkedAt) || now - checkedAt > VALIDATION_TTL_MS)
-        reasons.push('validation_expired');
-    return reasons;
-}
 function normalizeAutoPreference(value) {
     switch (String(value || '').toLowerCase()) {
         case 'performance':
@@ -191,7 +178,6 @@ class AutoRouter {
             this.claimEndpointAttempt(fixed.deployment);
             return decision;
         }
-        const requiredCapabilities = new Set([...policy.requiredCapabilities, ...request.requiredCapabilities].map(item => String(item).toLowerCase()));
         const eligible = [];
         for (const candidate of candidates) {
             const reasons = [];
@@ -205,13 +191,8 @@ class AutoRouter {
                 reasons.push('fallback_only');
             if (candidate.preview && !policy.allowPreview)
                 reasons.push('preview_disallowed');
-            reasons.push(...validationEligible(candidate, now));
             if (request.estimatedInputTokens + request.expectedOutputTokens > Math.max(0, candidate.maxContextTokens || 0))
                 reasons.push('context_too_small');
-            const capabilities = new Set(candidate.capabilities.map(item => String(item).toLowerCase()));
-            for (const capability of requiredCapabilities)
-                if (!capabilities.has(capability))
-                    reasons.push(`missing_capability:${capability}`);
             if (policy.privacy !== 'default' && !candidate.privacy.includes(policy.privacy))
                 reasons.push(`privacy:${policy.privacy}`);
             if (policy.dataRegion) {
@@ -230,8 +211,6 @@ class AutoRouter {
             if (policy.maxExpectedCostUsd !== undefined && (expectedCost === undefined || expectedCost > policy.maxExpectedCostUsd)) {
                 reasons.push(expectedCost === undefined ? 'unknown_cost' : 'budget_exceeded');
             }
-            if (this.circuitState(candidate.deployment, now, false) === 'open')
-                reasons.push('circuit_open');
             if (reasons.length)
                 decision.excludedCandidates.push({ deployment: { ...candidate.deployment }, reasons: [...new Set(reasons)] });
             else
@@ -325,9 +304,7 @@ class AutoRouter {
             && inSubset(candidate.deployment, subset)
             && !sameDeployment(candidate.deployment, current)
             && !attemptedDeployments.some(attempted => sameDeployment(candidate.deployment, attempted))
-            && validationEligible(candidate, now).length === 0
-            && this.passedInitialHardFilters(decision, candidate)
-            && this.circuitState(candidate.deployment, now, false) !== 'open');
+            && this.passedInitialHardFilters(decision, candidate));
         const equivalent = currentGroup
             ? eligible.find(candidate => candidate.deployment.logicalModelGroupId === currentGroup && !candidate.fallbackOnly)
             : undefined;

package/dist/core/config.d.ts

@@ -1,3 +1,4 @@
+import { ModelResponseHealth } from './modelResponseHealth';
 export interface JsonValue {
     [key: string]: unknown;
 }
@@ -20,6 +21,7 @@ export interface ProviderConfig {
     models: ModelConfig[];
 }
 export interface ModelConfig {
+    response_health?: ModelResponseHealth;
     name: string;
     display: string;
     description: string;

package/dist/core/config.js

@@ -44,6 +44,7 @@ exports.mergeProviderSecrets = mergeProviderSecrets;
 exports.defaultModelConfig = defaultModelConfig;
 exports.stableProviderId = stableProviderId;
 exports.defaultConfig = defaultConfig;
+const modelResponseHealth_1 = require("./modelResponseHealth");
 const fs = __importStar(require("fs"));
 const path = __importStar(require("path"));
 const crypto_1 = require("crypto");
@@ -191,7 +192,12 @@ class ConfigManager {
         this.workspaceOverrides.clear();
     }
     providers() {
-        return this.normalizeProviders((this.getGlobal('models', 'providers')) || []);
+        return this.normalizeProviders((this.getGlobal('models', 'providers')) || []).map(provider => ({
+            ...provider,
+            models: provider.models.map(model => ({ ...model, response_health: (0, modelResponseHealth_1.readModelResponseHealth)(this.rootPath, {
+                    providerId: provider.id, modelId: model.name, endpoint: provider.base_url, protocol: provider.protocol, credential: provider.api_key,
+                }) })),
+        }));
     }
     allModels() {
         const models = [];

package/dist/core/conversationKernel.js

@@ -384,7 +384,7 @@ class ConversationKernel {
     queueAction(target, action, input = {}) {
         const normalized = this.normalizeTarget(target);
         const runtime = this.findRuntime(normalized) || this.runtime(normalized, {
-            mode: this.host.mode, model: this.host.model, intelligence: this.host.intelligence,
+            mode: this.host.mode, model: this.host.modelSelectionValue(), intelligence: this.host.intelligence,
             inputMode: this.host.inputMode, engine: this.host.engine,
         });
         runtime.runId ||= (0, crypto_1.randomUUID)();
@@ -474,7 +474,7 @@ class ConversationKernel {
         let runtime = this.findRuntime(normalized);
         const runner = runtime?.runner || this.createRunner(normalized);
         if (!runtime && runner.conversationContinuations().length) {
-            runtime = this.runtime(normalized, { mode: runner.mode, model: runner.model, intelligence: runner.intelligence, inputMode: runner.inputMode, engine: runner.engine }, runner);
+            runtime = this.runtime(normalized, { mode: runner.mode, model: runner.modelSelectionValue(), intelligence: runner.intelligence, inputMode: runner.inputMode, engine: runner.engine }, runner);
             // Recovered user input remains visible and manageable until its owner
             // explicitly restores the persisted queue policy or resumes the queue.
             runtime.queuePaused = true;
@@ -491,7 +491,7 @@ class ConversationKernel {
             workEvents: this.events(normalized),
             runtime: this.runtimeState(normalized),
             mode: runner.mode,
-            model: runner.model,
+            model: runner.modelSelectionValue(),
             intelligence: runner.intelligence,
             status: runner.status,
             goal: conversationSnapshot.goal,
@@ -783,7 +783,9 @@ class ConversationKernel {
         const runner = runtime?.runner || this.createRunner(normalized);
         if (!runtime || !runtime.activePromise) {
             // No Build block is running: the selection applies immediately.
-            runner.setModel(model);
+            runner.setModel(model, true);
+            if (runtime)
+                runtime.options.model = runner.modelSelectionValue();
         }
         else {
             // A Build block is running. The in-flight block keeps its current model
@@ -793,7 +795,7 @@ class ConversationKernel {
             runtime.options.model = model;
         }
         runner.saveWorkspaceConversationState(true);
-        return runner.model;
+        return runner.modelSelectionValue();
     }
     async toggleGoalPause(target) {
         const normalized = this.normalizeTarget(target);
@@ -802,7 +804,7 @@ class ConversationKernel {
             const runner = this.createRunner(normalized);
             runtime = this.runtime(normalized, {
                 mode: runner.mode,
-                model: runner.model,
+                model: runner.modelSelectionValue(),
                 intelligence: runner.intelligence,
                 inputMode: runner.inputMode,
                 engine: runner.engine,
@@ -1577,7 +1579,12 @@ class ConversationKernel {
         // Host-global state can belong to a different foreground conversation.
         const restoredMode = agent.getConversationSnapshot(agent.activeConversationId).mode;
         agent.setMode(restoredMode || options.mode);
-        agent.setModel(options.model);
+        // Older clients may echo a bare display name from a snapshot. Preserve
+        // an already restored exact deployment in that case; never guess between
+        // providers for an unbound or different ambiguous model name.
+        if (options.model !== agent.model || !agent.activeDeployment()) {
+            agent.setModel(options.model);
+        }
         agent.setIntelligence(options.intelligence);
         agent.inputMode = options.inputMode;
         agent.engine = options.engine;
@@ -1603,7 +1610,7 @@ class ConversationKernel {
                 newContent: d.newContent,
             })),
             mode: runtime.runner.mode,
-            model: runtime.runner.model,
+            model: runtime.runner.modelSelectionValue(),
             status: runtime.runner.status,
             goal: runtime.runner.goal ? { objective: runtime.runner.goal.objective, paused: runtime.runner.goal.paused } : null,
             options: runtime.runner.pendingOptions,

package/dist/core/modelResponseHealth.d.ts

@@ -0,0 +1,17 @@
+import type { LLMProvider } from '../llm/provider';
+export type ResponseFacet = 'text' | 'vision' | 'tools';
+export type ModelResponseHealth = Partial<Record<ResponseFacet, {
+    ok: boolean;
+    at: string;
+}>>;
+export interface ResponseHealthIdentity {
+    providerId: string;
+    modelId: string;
+    endpoint: string;
+    protocol: string;
+    credential: string;
+}
+export declare function readModelResponseHealth(root: string, identity: ResponseHealthIdentity): ModelResponseHealth;
+export declare function recordModelResponseHealth(root: string, identity: ResponseHealthIdentity, updates: Partial<Record<ResponseFacet, boolean>>): void;
+export declare function observeModelResponses(provider: LLMProvider, record: (updates: Partial<Record<ResponseFacet, boolean>>) => void): LLMProvider;
+//# sourceMappingURL=modelResponseHealth.d.ts.map