@juspay/neurolink 9.50.0 → 9.50.2

package/dist/lib/server/routes/claudeProxyRoutes.js

@@ -15,7 +15,7 @@ import { join } from "node:path";
 import { buildStableClaudeCodeBillingHeader, CLAUDE_CLI_USER_AGENT, CLAUDE_CODE_OAUTH_BETAS, getOrCreateClaudeCodeIdentity, parseClaudeCodeUserId, } from "../../auth/anthropicOAuth.js";
 import { parseQuotaHeaders, saveAccountQuota, } from "../../proxy/accountQuota.js";
 import { buildClaudeError, ClaudeStreamSerializer, generateToolUseId, parseClaudeRequest, serializeClaudeResponse, } from "../../proxy/claudeFormat.js";
-import { ProxyTracer } from "../../proxy/proxyTracer.js";
+import { ProxyTracer, recordFallbackAttempt } from "../../proxy/proxyTracer.js";
 import { createRawStreamCapture } from "../../proxy/rawStreamCapture.js";
 import { logBodyCapture, logRequest, logRequestAttempt, logStreamError, } from "../../proxy/requestLogger.js";
 import { createSSEInterceptor } from "../../proxy/sseInterceptor.js";
@@ -1246,43 +1246,64 @@ async function executeClaudeFallbackTranslation(args) {
     if (body.stream) {
         const streamResult = await ctx.neurolink.stream(options);
         const serializer = new ClaudeStreamSerializer(body.model, 0);
-        async function* sseGenerator() {
-            for (const frame of serializer.start()) {
-                yield frame;
-            }
-            let collectedText = "";
-            for await (const chunk of streamResult.stream) {
-                const text = extractText(chunk);
-                if (text) {
-                    collectedText += text;
-                    for (const frame of serializer.pushDelta(text)) {
-                        yield frame;
-                    }
+        // Eagerly consume stream so errors fire synchronously and the
+        // fallback loop in tryConfiguredClaudeFallbackChain can catch them.
+        const frames = [];
+        let collectedText = "";
+        for (const frame of serializer.start()) {
+            frames.push(frame);
+        }
+        for await (const chunk of streamResult.stream) {
+            const text = extractText(chunk);
+            if (text) {
+                collectedText += text;
+                for (const frame of serializer.pushDelta(text)) {
+                    frames.push(frame);
                 }
             }
-            const toolCalls = streamResult.toolCalls ?? [];
-            if (!hasTranslatedOutput(collectedText, toolCalls)) {
-                throw new Error(`Translated provider ${providerLabel} returned no content or tool calls`);
-            }
-            if (toolCalls.length) {
-                for (const toolCall of toolCalls) {
-                    const toolName = toolCall.toolName ??
-                        toolCall.name ??
-                        "unknown";
-                    for (const frame of serializer.pushToolUse(generateToolUseId(), toolName, extractToolArgs(toolCall))) {
-                        yield frame;
-                    }
+        }
+        const toolCalls = streamResult.toolCalls ?? [];
+        if (!hasTranslatedOutput(collectedText, toolCalls)) {
+            throw new Error(`Translated provider ${providerLabel} returned no content or tool calls`);
+        }
+        if (toolCalls.length) {
+            for (const toolCall of toolCalls) {
+                const toolName = toolCall.toolName ??
+                    toolCall.name ??
+                    "unknown";
+                for (const frame of serializer.pushToolUse(generateToolUseId(), toolName, extractToolArgs(toolCall))) {
+                    frames.push(frame);
                 }
             }
-            const reason = streamResult.finishReason ?? "end_turn";
-            const resolvedUsage = extractUsageFromStreamResult(streamResult.usage);
-            for (const frame of serializer.finish(resolvedUsage.output, reason)) {
-                yield frame;
-            }
         }
+        const reason = streamResult.finishReason ?? "end_turn";
+        const resolvedUsage = extractUsageFromStreamResult(streamResult.usage);
+        for (const frame of serializer.finish(resolvedUsage.output, reason)) {
+            frames.push(frame);
+        }
+        // Telemetry AFTER validation — not before like the old lazy path
         tracer?.end(200, Date.now() - requestStartTime);
         recordFinalSuccess();
-        logFinalRequest(200, "", providerLabel);
+        logFinalRequest(200, "", providerLabel, undefined, undefined, {
+            inputTokens: resolvedUsage.input,
+            outputTokens: resolvedUsage.output,
+        });
+        const bufferedBody = frames.join("");
+        logProxyBody({
+            phase: "client_response",
+            headers: { "content-type": "text/event-stream" },
+            body: bufferedBody,
+            bodySize: Buffer.byteLength(bufferedBody, "utf8"),
+            contentType: "text/event-stream",
+            responseStatus: 200,
+            durationMs: Date.now() - requestStartTime,
+        });
+        // Return generator that yields pre-buffered frames
+        async function* sseGenerator() {
+            for (const frame of frames) {
+                yield frame;
+            }
+        }
         return sseGenerator();
     }
     const streamResult = await ctx.neurolink.stream(options);
@@ -1346,6 +1367,11 @@ async function tryConfiguredClaudeFallbackChain(args) {
             : "auto-provider";
         logger.always(`[proxy] skipping fallback ${label}: ${skipped.reason}`);
     }
+    tracer?.setFallbackInfo({
+        triggered: true,
+        attemptCount: fallbackPlan.attempts.slice(1).length,
+        reason: fallbackPolicyReason ?? "all_anthropic_accounts_exhausted",
+    });
     for (const fallback of fallbackPlan.attempts.slice(1)) {
         if (!fallback.provider || !fallback.model) {
             continue;
@@ -1354,6 +1380,7 @@ async function tryConfiguredClaudeFallbackChain(args) {
         if (!availability.available) {
             logger.always(`[proxy] fallback ${fallback.provider}/${fallback.model} health-check failed (${availability.reason ?? "provider unavailable"}), attempting anyway`);
         }
+        const fallbackStart = Date.now();
         try {
             logger.always(`[proxy] fallback → ${fallback.provider}/${fallback.model}`);
             const options = buildProxyFallbackOptions(parsedFallbackRequest, {
@@ -1370,13 +1397,57 @@ async function tryConfiguredClaudeFallbackChain(args) {
                 options: options,
                 providerLabel: fallback.provider,
             });
+            recordFallbackAttempt({
+                provider: fallback.provider,
+                model: fallback.model,
+                status: "success",
+                durationMs: Date.now() - fallbackStart,
+            });
+            tracer?.setFallbackInfo({
+                triggered: true,
+                provider: fallback.provider,
+                model: fallback.model,
+                attemptCount: fallbackPlan.attempts.slice(1).length,
+                reason: "fallback_success",
+            });
             return {
                 response,
                 fallbackPolicyReason,
             };
         }
         catch (fallbackErr) {
-            logger.always(`[proxy] fallback ${fallback.provider}/${fallback.model} failed: ${fallbackErr instanceof Error ? fallbackErr.message : String(fallbackErr)}`);
+            const errMsg = fallbackErr instanceof Error
+                ? fallbackErr.message
+                : String(fallbackErr);
+            let errorClass = "unknown";
+            if (errMsg.includes("Rate limit") ||
+                errMsg.includes("rate_limit") ||
+                errMsg.includes("max_parallel_requests")) {
+                errorClass = "rate_limit";
+            }
+            else if (errMsg.includes("context length") ||
+                errMsg.includes("ContextWindowExceeded")) {
+                errorClass = "context_overflow";
+            }
+            else if (errMsg.includes("no content or tool calls") ||
+                errMsg.includes("NoOutputGenerated")) {
+                errorClass = "empty_response";
+            }
+            else if (errMsg.includes("thinking_level") ||
+                errMsg.includes("Field required")) {
+                errorClass = "schema_mismatch";
+            }
+            else if (errMsg.includes("Resource exhausted")) {
+                errorClass = "provider_quota";
+            }
+            logger.always(`[proxy] fallback ${fallback.provider}/${fallback.model} failed [${errorClass}]: ${errMsg}`);
+            recordFallbackAttempt({
+                provider: fallback.provider,
+                model: fallback.model,
+                status: "failure",
+                errorMessage: `[${errorClass}] ${errMsg}`,
+                durationMs: Date.now() - fallbackStart,
+            });
         }
     }
     return {
@@ -3541,7 +3612,15 @@ function shouldOmitImagesForTarget(provider, model) {
     return provider === "litellm" && model === "open-large";
 }
 function shouldOmitThinkingConfigForTarget(provider, model) {
-    return provider === "vertex" && model === "gemini-2.5-flash";
+    if (provider === "litellm") {
+        return true;
+    }
+    if (provider !== "vertex") {
+        return false;
+    }
+    // Only Gemini 2.5+ and 3.x support thinking_level on Vertex.
+    const m = model?.toLowerCase() ?? "";
+    return !/gemini-(2\.5|3)/.test(m);
 }
 function extractToolArgs(toolCall) {
     return (toolCall.args ??

package/dist/lib/services/server/ai/observability/instrumentation.js

@@ -445,7 +445,45 @@ function initializeExternalOpenTelemetryMode(config, resource, otlpEndpoint, ser
             const provider = globalProvider;
             if (globalProvider && typeof provider.addSpanProcessor === "function") {
                 provider.addSpanProcessor(new ContextEnricher());
-                const skipLangfuse = config.skipLangfuseSpanProcessor === true || !langfuseProcessor;
+                // Auto-detect: skip if consumer already registered a LangfuseSpanProcessor.
+                //
+                // Detection strategy (ordered by robustness):
+                // 1. `instanceof LangfuseSpanProcessor` — reliable when both sides use
+                //    the same @langfuse/otel package instance (same module identity).
+                // 2. Duck-type check for Langfuse-specific public member
+                //    (`langfuseClient` property) — survives minification.
+                // 3. `constructor.name === "LangfuseSpanProcessor"` — last resort,
+                //    brittle under minification or bundler renaming.
+                //
+                // NOTE: `_registeredSpanProcessors` is an internal OpenTelemetry field.
+                // If the OTel SDK removes or renames it, the array defaults to [] and
+                // `hasExistingLangfuse` is false — NeuroLink registers its own processor
+                // (same behavior as before this check). Consumers can always force skip
+                // via `skipLangfuseSpanProcessor: true`.
+                const existingProcessors = provider
+                    ._registeredSpanProcessors ?? [];
+                const hasExistingLangfuse = existingProcessors.some((p) => {
+                    if (p === null || p === undefined || typeof p !== "object") {
+                        return false;
+                    }
+                    // Prefer instanceof — works when same @langfuse/otel package is shared
+                    if (p instanceof LangfuseSpanProcessor) {
+                        return true;
+                    }
+                    // Duck-type: Langfuse processor exposes a langfuseClient property
+                    if ("langfuseClient" in p) {
+                        return true;
+                    }
+                    // Fallback: constructor name (brittle under minification)
+                    return (p.constructor?.name ===
+                        "LangfuseSpanProcessor");
+                });
+                const skipLangfuse = config.skipLangfuseSpanProcessor === true ||
+                    !langfuseProcessor ||
+                    hasExistingLangfuse;
+                if (hasExistingLangfuse && !config.skipLangfuseSpanProcessor) {
+                    logger.info(`${LOG_PREFIX} Auto-detected existing LangfuseSpanProcessor — skipping SDK registration to avoid duplicates`);
+                }
                 if (!skipLangfuse && langfuseProcessor) {
                     provider.addSpanProcessor(langfuseProcessor);
                 }

package/dist/lib/types/externalMcp.d.ts

@@ -228,6 +228,7 @@ export type ExternalMCPServerEvents = {
     /** Server status changed */
     statusChanged: {
         serverId: string;
+        serverName: string;
         oldStatus: ExternalMCPServerStatus;
         newStatus: ExternalMCPServerStatus;
         timestamp: Date;
@@ -235,24 +236,28 @@ export type ExternalMCPServerEvents = {
     /** Server connected successfully */
     connected: {
         serverId: string;
+        serverName: string;
         toolCount: number;
         timestamp: Date;
     };
     /** Server disconnected */
     disconnected: {
         serverId: string;
+        serverName: string;
         reason?: string;
         timestamp: Date;
     };
     /** Server failed */
     failed: {
         serverId: string;
+        serverName: string;
         error: string;
         timestamp: Date;
     };
     /** Tool discovered */
     toolDiscovered: {
         serverId: string;
+        serverName: string;
         toolName: string;
         toolInfo: ExternalMCPToolInfo;
         timestamp: Date;
@@ -260,12 +265,14 @@ export type ExternalMCPServerEvents = {
     /** Tool removed */
     toolRemoved: {
         serverId: string;
+        serverName: string;
         toolName: string;
         timestamp: Date;
     };
     /** Health check completed */
     healthCheck: {
         serverId: string;
+        serverName: string;
         health: ExternalMCPServerHealth;
         timestamp: Date;
     };

package/dist/mcp/externalServerManager.d.ts

@@ -36,6 +36,11 @@ export declare class ExternalServerManager extends EventEmitter {
      * Get current HITL manager
      */
     getHITLManager(): HITLManager | undefined;
+    /**
+     * Resolve the human-readable server name for an event payload.
+     * Falls back to serverId if the instance or config.name isn't available.
+     */
+    getServerName(serverId: string): string;
     /**
      * Load MCP server configurations from .mcp-config.json file with parallel loading support
      * Automatically registers servers found in the configuration

package/dist/mcp/externalServerManager.js

@@ -194,10 +194,16 @@ export class ExternalServerManager extends EventEmitter {
         this.toolDiscovery = new ToolDiscoveryService();
         // Forward tool discovery events
         this.toolDiscovery.on("toolRegistered", (event) => {
-            this.emit("toolDiscovered", event);
+            this.emit("toolDiscovered", {
+                ...event,
+                serverName: this.getServerName(event.serverId),
+            });
         });
         this.toolDiscovery.on("toolUnregistered", (event) => {
-            this.emit("toolRemoved", event);
+            this.emit("toolRemoved", {
+                ...event,
+                serverName: this.getServerName(event.serverId),
+            });
         });
         // Handle process cleanup
         process.on("SIGINT", () => this.shutdown());
@@ -223,6 +229,14 @@ export class ExternalServerManager extends EventEmitter {
     getHITLManager() {
         return this.hitlManager;
     }
+    /**
+     * Resolve the human-readable server name for an event payload.
+     * Falls back to serverId if the instance or config.name isn't available.
+     */
+    getServerName(serverId) {
+        const instance = this.servers.get(serverId);
+        return instance?.config?.name || serverId;
+    }
     /**
      * Load MCP server configurations from .mcp-config.json file with parallel loading support
      * Automatically registers servers found in the configuration
@@ -712,6 +726,8 @@ export class ExternalServerManager extends EventEmitter {
                 };
             }
             mcpLogger.info(`[ExternalServerManager] Removing server: ${serverId}`);
+            // Capture name before deletion removes the instance
+            const serverName = this.getServerName(serverId);
             // Stop the server
             await this.stopServer(serverId);
             // Remove from registry
@@ -719,6 +735,7 @@ export class ExternalServerManager extends EventEmitter {
             // Emit event
             this.emit("disconnected", {
                 serverId,
+                serverName,
                 reason: "Manually removed",
                 timestamp: new Date(),
             });
@@ -816,6 +833,7 @@ export class ExternalServerManager extends EventEmitter {
             // Emit connected event
             this.emit("connected", {
                 serverId,
+                serverName: this.getServerName(serverId),
                 toolCount: instance.toolsMap.size,
                 timestamp: new Date(),
             });
@@ -921,6 +939,7 @@ export class ExternalServerManager extends EventEmitter {
         // Emit status change event
         this.emit("statusChanged", {
             serverId,
+            serverName: this.getServerName(serverId),
             oldStatus,
             newStatus,
             timestamp: new Date(),
@@ -941,6 +960,7 @@ export class ExternalServerManager extends EventEmitter {
         // Emit failed event
         this.emit("failed", {
             serverId,
+            serverName: this.getServerName(serverId),
             error: error.message,
             timestamp: new Date(),
         });
@@ -965,6 +985,7 @@ export class ExternalServerManager extends EventEmitter {
         // Emit disconnected event
         this.emit("disconnected", {
             serverId,
+            serverName: this.getServerName(serverId),
             reason,
             timestamp: new Date(),
         });
@@ -1078,6 +1099,7 @@ export class ExternalServerManager extends EventEmitter {
             // Emit health check event
             this.emit("healthCheck", {
                 serverId,
+                serverName: this.getServerName(serverId),
                 health,
                 timestamp: new Date(),
             });

package/dist/neurolink.js

@@ -50,7 +50,7 @@ import { createMemoryRetrievalTools } from "./memory/memoryRetrievalTools.js";
 import { getMetricsAggregator, MetricsAggregator, } from "./observability/metricsAggregator.js";
 import { SpanStatus, SpanType } from "./observability/types/spanTypes.js";
 import { SpanSerializer } from "./observability/utils/spanSerializer.js";
-import { flushOpenTelemetry, getLangfuseHealthStatus, initializeOpenTelemetry, isOpenTelemetryInitialized, setLangfuseContext, shutdownOpenTelemetry, } from "./services/server/ai/observability/instrumentation.js";
+import { flushOpenTelemetry, getLangfuseHealthStatus, initializeOpenTelemetry, isOpenTelemetryInitialized, runWithCurrentLangfuseContext, setLangfuseContext, shutdownOpenTelemetry, } from "./services/server/ai/observability/instrumentation.js";
 import { TaskManager } from "./tasks/taskManager.js";
 import { createTaskTools } from "./tasks/tools/taskTools.js";
 import { ATTR } from "./telemetry/attributes.js";
@@ -1129,7 +1129,10 @@ Current user's request: ${currentInput}`;
      * Supports additional users with per-user prompt and maxWords overrides.
      */
     storeMemoryInBackground(originalPrompt, responseContent, userId, additionalUsers) {
-        setImmediate(async () => {
+        // Preserve AsyncLocalStorage context across setImmediate boundary so that
+        // memory writes appear under the originating Langfuse trace instead of
+        // becoming orphan spans.
+        const wrappedMemoryWrite = runWithCurrentLangfuseContext(async () => {
             try {
                 const client = this.ensureMemoryReady();
                 if (!client) {
@@ -1145,12 +1148,18 @@ Current user's request: ${currentInput}`;
                         : undefined;
                     writeOps.push(client.add(user.userId, content, addOptions));
                 }
-                await Promise.all(writeOps);
+                // withTimeout races against Promise.all — if the timeout fires, the
+                // await resolves with an error but the underlying client.add() calls
+                // may still complete in the background. This is acceptable: the memory
+                // client API (Mem0) doesn't support AbortSignal, and these are
+                // fire-and-forget background writes where a stale completion is harmless.
+                await withTimeout(Promise.all(writeOps), 30_000, new Error("Background memory write timed out after 30s"));
             }
             catch (error) {
                 logger.warn("Memory storage failed:", error);
             }
         });
+        setImmediate(wrappedMemoryWrite);
     }
     /**
      * Set up HITL event forwarding to main emitter
@@ -3723,6 +3732,21 @@ Current user's request: ${currentInput}`;
             conversationMessageCount: conversationMessages.length,
             shouldCompact: budgetResult.shouldCompact,
         });
+        // Scale timeout for large contexts if caller didn't set one explicitly.
+        // Providers read options.timeout via getTimeout(), so setting it here
+        // propagates to any downstream provider call.
+        if (options.timeout === undefined &&
+            budgetResult.estimatedInputTokens > 100_000) {
+            // >100K → 1.5x, >200K → 2x, >300K → 2.5x (capped at 4x) of 60s base
+            const scale = 1 + Math.floor((budgetResult.estimatedInputTokens - 1) / 100_000) * 0.5;
+            const scaledMs = Math.round(60_000 * Math.min(scale, 4));
+            options.timeout = scaledMs;
+            logger.info("[TokenBudget] Scaled timeout for large context", {
+                requestId,
+                estimatedTokens: budgetResult.estimatedInputTokens,
+                scaledTimeoutMs: scaledMs,
+            });
+        }
         const compactionSessionId = this.getCompactionSessionId(options);
         const lastCompactionCount = this.lastCompactionMessageCount.get(compactionSessionId) ?? 0;
         if (!budgetResult.shouldCompact ||
@@ -3798,6 +3822,8 @@ Current user's request: ${currentInput}`;
             toolDefinitions: availableTools,
         });
         if (!finalBudget.withinBudget) {
+            // Clear watermark so handleContextOverflow recovery can re-compact
+            this.lastCompactionMessageCount.delete(compactionSessionId);
             throw new ContextBudgetExceededError(`Context exceeds model budget after all compaction stages. ` +
                 `Estimated: ${finalBudget.estimatedInputTokens} tokens, ` +
                 `Budget: ${finalBudget.availableInputTokens} tokens. ` +
@@ -3993,6 +4019,8 @@ Current user's request: ${currentInput}`;
                                 : undefined,
                         });
                         if (!finalBudget.withinBudget) {
+                            // Clear watermark so handleContextOverflow recovery can re-compact
+                            this.lastCompactionMessageCount.delete(dpgCompactionSessionId);
                             throw new ContextBudgetExceededError(`Context exceeds model budget after all compaction stages. ` +
                                 `Estimated: ${finalBudget.estimatedInputTokens} tokens, ` +
                                 `Budget: ${finalBudget.availableInputTokens} tokens.`, {
@@ -5016,6 +5044,8 @@ Current user's request: ${currentInput}`;
                     toolDefinitions: availableTools,
                 });
                 if (!finalBudget.withinBudget) {
+                    // Clear watermark so handleContextOverflow recovery can re-compact
+                    this.lastCompactionMessageCount.delete(streamCompactionSessionId);
                     throw new ContextBudgetExceededError(`Stream context exceeds model budget after all compaction stages. ` +
                         `Estimated: ${finalBudget.estimatedInputTokens} tokens, ` +
                         `Budget: ${finalBudget.availableInputTokens} tokens.`, {
@@ -7508,6 +7538,7 @@ Current user's request: ${currentInput}`;
                 // Emit server added event
                 this.emitter.emit("externalMCP:serverAdded", {
                     serverId,
+                    serverName: config.name || serverId,
                     config,
                     toolCount: result.metadata?.toolsDiscovered || 0,
                     timestamp: Date.now(),
@@ -7535,12 +7566,15 @@ Current user's request: ${currentInput}`;
         this.invalidateToolCache(); // Invalidate cache when an external server is removed
         try {
             mcpLogger.info(`[NeuroLink] Removing external MCP server: ${serverId}`);
+            // Capture the configured name before removal destroys the instance
+            const serverName = this.externalServerManager.getServerName(serverId);
             const result = await this.externalServerManager.removeServer(serverId);
             if (result.success) {
                 mcpLogger.info(`[NeuroLink] External MCP server removed successfully: ${serverId}`);
                 // Emit server removed event
                 this.emitter.emit("externalMCP:serverRemoved", {
                     serverId,
+                    serverName,
                     timestamp: Date.now(),
                 });
             }

package/dist/providers/litellm.js

@@ -356,8 +356,8 @@ export class LiteLLMProvider extends BaseProvider {
         }
         catch (streamError) {
             if (NoOutputGeneratedError.isInstance(streamError)) {
-                logger.warn("LiteLLM: Stream produced no output (NoOutputGeneratedError)");
-                return;
+                logger.warn("LiteLLM: Stream produced no output (NoOutputGeneratedError) — propagating to fallback chain");
+                throw streamError;
             }
             throw streamError;
         }

package/dist/proxy/proxyTracer.d.ts

@@ -90,6 +90,13 @@ declare class ProxyTracer {
      * Sets span attributes and increments the substitution metric counter.
      */
     setModelSubstitution(requestedModel: string, actualModel: string): void;
+    setFallbackInfo(info: {
+        triggered: boolean;
+        provider?: string;
+        model?: string;
+        attemptCount: number;
+        reason: string;
+    }): void;
     /** Log the incoming client request body (redacted). */
     logRequestBody(body: string): void;
     /** Log the incoming client request headers (redacted). */
@@ -129,5 +136,12 @@ declare class ProxyTracer {
      */
     getTraceHeaders(): Record<string, string>;
 }
+export declare function recordFallbackAttempt(attrs: {
+    provider: string;
+    model: string;
+    status: "success" | "failure";
+    errorMessage?: string;
+    durationMs: number;
+}): void;
 export { ProxyTracer };
 export type { ProxyRequestContext, AccountSelectionContext, UpstreamAttemptContext, UsageContext, };

package/dist/proxy/proxyTracer.js

@@ -79,6 +79,18 @@ function getProxyMetrics() {
             description: "Response body size in bytes received from upstream",
             unit: "By",
         }),
+        fallbackAttemptsTotal: meter.createCounter("proxy_fallback_attempts_total", {
+            description: "Total fallback provider attempts",
+            unit: "{attempt}",
+        }),
+        fallbackSuccessTotal: meter.createCounter("proxy_fallback_success_total", {
+            description: "Total successful fallback provider responses",
+            unit: "{success}",
+        }),
+        fallbackFailureTotal: meter.createCounter("proxy_fallback_failure_total", {
+            description: "Total failed fallback provider responses",
+            unit: "{failure}",
+        }),
     };
     _metrics = createdMetrics;
     return createdMetrics;
@@ -396,6 +408,18 @@ class ProxyTracer {
             actual_model: actualModel,
         });
     }
+    setFallbackInfo(info) {
+        if (!this.rootSpan) {
+            return;
+        }
+        this.rootSpan.setAttributes({
+            "proxy.fallback.triggered": info.triggered,
+            ...(info.provider ? { "proxy.fallback.provider": info.provider } : {}),
+            ...(info.model ? { "proxy.fallback.model": info.model } : {}),
+            "proxy.fallback.attempt_count": info.attemptCount,
+            "proxy.fallback.reason": info.reason,
+        });
+    }
     // -------------------------------------------------------------------------
     // Log payloads as span events
     // -------------------------------------------------------------------------
@@ -641,4 +665,23 @@ class ProxyTracer {
         return this.bridge.injectContext({}, trace.setSpan(context.active(), this.rootSpan));
     }
 }
+export function recordFallbackAttempt(attrs) {
+    try {
+        const m = getProxyMetrics();
+        const labels = { provider: attrs.provider, model: attrs.model };
+        m.fallbackAttemptsTotal.add(1, labels);
+        if (attrs.status === "success") {
+            m.fallbackSuccessTotal.add(1, labels);
+        }
+        else {
+            m.fallbackFailureTotal.add(1, {
+                ...labels,
+                error: attrs.errorMessage?.slice(0, 100) ?? "unknown",
+            });
+        }
+    }
+    catch {
+        // metrics are best-effort
+    }
+}
 export { ProxyTracer };