@juspay/neurolink 12.37.2 → 12.37.4

package/dist/constants/contextWindows.js

@@ -118,6 +118,7 @@ export const MODEL_CONTEXT_WINDOWS = {
         // Claude 5.5 / 5.1 (Sep 2026) — 1M context window, confirmed on
         // platform.claude.com/docs/en/about-claude/models/overview
         "claude-opus-5-5": 1_000_000,
+        "claude-sonnet-5-5": 1_000_000,
         "claude-fable-5-1": 1_000_000,
         // Claude 5 (mid 2026) — 1M context window
         "claude-sonnet-5": 1_000_000,
@@ -216,6 +217,7 @@ export const MODEL_CONTEXT_WINDOWS = {
         _default: 1_048_576,
         // Claude on Vertex
         "claude-opus-5-5": 1_000_000,
+        "claude-sonnet-5-5": 1_000_000,
         "claude-fable-5-1": 1_000_000,
         "claude-sonnet-5": 1_000_000,
         "claude-opus-4-6": 1_000_000,
@@ -255,6 +257,7 @@ export const MODEL_CONTEXT_WINDOWS = {
         _default: 200_000,
         // Claude 5.5 / 5.1 (Sep 2026)
         "anthropic.claude-opus-5-5": 1_000_000,
+        "anthropic.claude-sonnet-5-5": 1_000_000,
         "anthropic.claude-fable-5-1": 1_000_000,
         // Claude 4.6
         "anthropic.claude-opus-4-6-v1": 1_000_000,

package/dist/constants/enums.d.ts

@@ -366,6 +366,7 @@ export declare enum AnthropicModels {
     CLAUDE_SONNET_5 = "claude-sonnet-5",
     CLAUDE_FABLE_5 = "claude-fable-5",
     CLAUDE_OPUS_5_5 = "claude-opus-5-5",
+    CLAUDE_SONNET_5_5 = "claude-sonnet-5-5",
     CLAUDE_FABLE_5_1 = "claude-fable-5-1",
     CLAUDE_OPUS_4_8 = "claude-opus-4-8",
     CLAUDE_OPUS_4_7 = "claude-opus-4-7",

package/dist/constants/enums.js

@@ -502,6 +502,7 @@ export var AnthropicModels;
     // Claude 5.5 / 5.1 Series (September 2026) — confirmed on
     // platform.claude.com/docs/en/about-claude/models/overview
     AnthropicModels["CLAUDE_OPUS_5_5"] = "claude-opus-5-5";
+    AnthropicModels["CLAUDE_SONNET_5_5"] = "claude-sonnet-5-5";
     AnthropicModels["CLAUDE_FABLE_5_1"] = "claude-fable-5-1";
     // Claude 4.7 / 4.8 Series
     AnthropicModels["CLAUDE_OPUS_4_8"] = "claude-opus-4-8";

package/dist/core/geminiLoopAdapter.js

@@ -37,6 +37,14 @@ export function createGeminiLoopAdapter(config) {
         }
         return original;
     };
+    // Per-turn monotonic counter for `toolCallId` generation. `executeStep` is
+    // called once per loop step, and a fresh `createGeminiLoopAdapter` closure
+    // is made per turn (per generate()/stream() call), so this resets at the
+    // right lifetime while staying unique across every step of one turn —
+    // unlike the step-local array index it replaces, which restarted at 0 on
+    // every step and let two calls to the same tool in different steps collide
+    // on one id once both landed in the same flattened per-turn list.
+    let nextCallIndex = 0;
     return {
         providerLabel: config.providerLabel,
         maxSteps: config.maxSteps,
@@ -138,9 +146,11 @@ export function createGeminiLoopAdapter(config) {
             // projects the next request from.
             config.noteUsage?.(collected.inputTokens, collected.outputTokens);
             const text = extractTextFromParts(collected.rawResponseParts);
-            // Names cross the engine boundary in their ORIGINAL form.
-            const allCalls = collected.stepFunctionCalls.map((call, index) => ({
-                id: `${config.providerLabel}_${index}_${call.name}`,
+            // Names cross the engine boundary in their ORIGINAL form. `id` uses the
+            // turn-scoped counter above, not a step-local map index, so two calls
+            // to the same tool in different steps of one turn get distinct ids.
+            const allCalls = collected.stepFunctionCalls.map((call) => ({
+                id: `${config.providerLabel}_${nextCallIndex++}_${call.name}`,
                 name: originalNameFor(call.name),
                 args: call.args,
             }));

package/dist/models/manifests/anthropic.js

@@ -37,6 +37,20 @@ export const anthropicManifest = {
             jsonMode: true,
             samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /opus[-_.]?5(?![0-9])/i
         },
+        "claude-sonnet-5-5": {
+            aliases: ["sonnet-5.5"],
+            displayName: "Claude Sonnet 5.5",
+            contextWindow: 1_000_000,
+            maxOutputTokens: 128_000,
+            // No pricingPerMTok DELIBERATELY: PRICING.anthropic already carries
+            // real rates for this id (input 2.0/output 10.0 per MTok) — see the
+            // claude-sonnet-5 comment below for why this defers to that table.
+            vision: true,
+            functionCalling: true,
+            reasoning: true,
+            jsonMode: true,
+            samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /sonnet[-_.]?5(?![0-9])/i
+        },
         "claude-fable-5-1": {
             aliases: ["fable-5.1"],
             displayName: "Claude Fable 5.1",

package/dist/models/modelRegistry.js

@@ -1174,6 +1174,48 @@ const LEGACY_MODEL_REGISTRY = {
         releaseDate: "2026-06-30",
         category: "general",
     },
+    // "claude-sonnet-latest" stays on CLAUDE_SONNET_5 until someone decides to
+    // move it; this entry sits after that one so resolver ties still go to it.
+    [AnthropicModels.CLAUDE_SONNET_5_5]: {
+        id: AnthropicModels.CLAUDE_SONNET_5_5,
+        name: "Claude Sonnet 5.5",
+        provider: AIProviderName.ANTHROPIC,
+        description: "The best combination of speed and intelligence",
+        capabilities: {
+            vision: true,
+            functionCalling: true,
+            codeGeneration: true,
+            reasoning: true,
+            multimodal: true,
+            streaming: true,
+            jsonMode: false,
+        },
+        pricing: {
+            inputCostPer1K: 0.002,
+            outputCostPer1K: 0.01,
+            currency: "USD",
+        },
+        performance: { speed: "fast", quality: "high", accuracy: "high" },
+        limits: {
+            maxContextTokens: 1000000,
+            maxOutputTokens: 128000,
+            maxRequestsPerMinute: 50,
+        },
+        useCases: {
+            coding: 9,
+            creative: 9,
+            analysis: 9,
+            conversation: 9,
+            reasoning: 9,
+            translation: 9,
+            summarization: 9,
+        },
+        aliases: ["sonnet-5.5"],
+        deprecated: false,
+        isLocal: false,
+        releaseDate: "2026-09-28",
+        category: "general",
+    },
     [AnthropicModels.CLAUDE_FABLE_5]: {
         id: AnthropicModels.CLAUDE_FABLE_5,
         name: "Claude Fable 5",

package/dist/providers/googleVertex/client.js

@@ -2595,7 +2595,11 @@ export class GoogleVertexProvider extends BaseProvider {
                 finalText = engineResult.text;
                 lastFinishReason = engineResult.rawStopReason ?? lastFinishReason;
                 for (const call of engineResult.toolCalls) {
-                    allToolCalls.push({ toolName: call.name, args: call.args });
+                    allToolCalls.push({
+                        toolCallId: call.id,
+                        toolName: call.name,
+                        args: call.args,
+                    });
                 }
                 for (const execution of engineResult.toolExecutions) {
                     toolExecutions.push({
@@ -2747,6 +2751,17 @@ export class GoogleVertexProvider extends BaseProvider {
                 reasoningTokens: totalReasoningTokens,
             }),
             responseTime,
+            // The complete attempted-call record (name/args/id), independent of
+            // toolsUsed's success-only names — mirrors the native Anthropic direct
+            // / OpenAI-compatible / SageMaker paths (toolCallsFromSummaries) and
+            // the working Gemini-on-Vertex STREAM path just above. Previously
+            // absent here, so a tool the loop ran internally never reached the
+            // caller's result.toolCalls even though toolsUsed/toolExecutions did.
+            toolCalls: externalToolCalls.map((tc) => ({
+                toolCallId: tc.toolCallId,
+                toolName: tc.toolName,
+                args: tc.args,
+            })),
             toolsUsed: externalToolCalls.map((tc) => tc.toolName),
             toolExecutions: resolveToolExecutionRecords(options, externalToolExecutions),
             enhancedWithTools: externalToolCalls.length > 0,
@@ -4699,7 +4714,11 @@ export class GoogleVertexProvider extends BaseProvider {
             },
             buildToolResultMessages: (conversation, stepResult, toolResults, engineStep) => {
                 for (const result of toolResults) {
-                    allToolCalls.push({ toolName: result.name, args: result.args });
+                    allToolCalls.push({
+                        toolCallId: result.id,
+                        toolName: result.name,
+                        args: result.args,
+                    });
                     // Recorded here as well: `toolExecutions` feeds
                     // resolveToolExecutionRecords and the result's own
                     // toolExecutions, and pushing only to allToolCalls left it empty
@@ -5059,7 +5078,23 @@ export class GoogleVertexProvider extends BaseProvider {
         }
         releaseTurnResources();
         const responseTime = Date.now() - startTime;
-        const externalToolCalls = allToolCalls.filter((tc) => tc.toolName !== "final_result");
+        // Sourced from `engineResult.toolCalls` when the turn produced one, not
+        // the local `allToolCalls` (fed only by the `buildToolResultMessages`
+        // hook below). The engine's own accumulator is pushed to unconditionally
+        // per dispatched batch, before the abortedMidBatch check that skips
+        // `buildToolResultMessages` for that step — so a tool that finished
+        // executing in a batch cut short by an abort is present in
+        // `engineResult.toolCalls` but would otherwise never reach
+        // `allToolCalls`. `allToolCalls` remains the fallback for the rare case
+        // where `resultPromise` itself rejects and `engineResult` never resolves.
+        const engineToolCalls = engineResult
+            ? engineResult.toolCalls.map((call) => ({
+                toolCallId: call.id,
+                toolName: call.name,
+                args: call.args,
+            }))
+            : allToolCalls;
+        const externalToolCalls = engineToolCalls.filter((tc) => tc.toolName !== "final_result");
         const externalToolExecutions = toolExecutions.filter((te) => te.name !== "final_result");
         // Absolute backstop — no code path may surface empty content on a turn
         // that ran tools (the consumer-facing "empty response" incident shape).
@@ -5136,6 +5171,17 @@ export class GoogleVertexProvider extends BaseProvider {
                 ...(totalReasoningTokens > 0 && { reasoning: totalReasoningTokens }),
             },
             responseTime,
+            // The complete attempted-call record (name/args/id), independent of
+            // toolsUsed's success-only names — mirrors the native Anthropic direct
+            // / OpenAI-compatible / SageMaker paths (toolCallsFromSummaries) and
+            // the working Gemini-on-Vertex STREAM path. Previously absent here, so
+            // a tool the loop ran internally never reached the caller's
+            // result.toolCalls even though toolsUsed/toolExecutions did.
+            toolCalls: externalToolCalls.map((tc) => ({
+                toolCallId: tc.toolCallId,
+                toolName: tc.toolName,
+                args: tc.args,
+            })),
             toolsUsed: externalToolCalls.map((tc) => tc.toolName),
             toolExecutions: resolveToolExecutionRecords(options, externalToolExecutions),
             enhancedWithTools: externalToolCalls.length > 0,

package/dist/utils/pricing.js

@@ -100,6 +100,13 @@ const PRICING = {
             cacheRead: 0.2 / 1_000_000,
             cacheCreation: 5.0 / 1_000_000,
         },
+        // Sonnet 5.5 bills exactly as Sonnet 5 does, cache rates included.
+        "claude-sonnet-5-5": {
+            input: 2.0 / 1_000_000,
+            output: 10.0 / 1_000_000,
+            cacheRead: 0.2 / 1_000_000,
+            cacheCreation: 2.5 / 1_000_000,
+        },
         "claude-fable-5": {
             input: 10.0 / 1_000_000,
             output: 50.0 / 1_000_000,
@@ -281,6 +288,13 @@ const PRICING = {
             cacheRead: 0.2 / 1_000_000,
             cacheCreation: 5.0 / 1_000_000,
         },
+        // Sonnet 5.5 bills exactly as Sonnet 5 does, cache rates included.
+        "claude-sonnet-5-5": {
+            input: 2.0 / 1_000_000,
+            output: 10.0 / 1_000_000,
+            cacheRead: 0.2 / 1_000_000,
+            cacheCreation: 2.5 / 1_000_000,
+        },
         "claude-opus-5": {
             input: 5.0 / 1_000_000,
             output: 25.0 / 1_000_000,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.37.2",
+  "version": "12.37.4",
   "packageManager": "pnpm@10.15.1",
   "description": "The pipe layer of an AI nervous system: one interface connecting provider neurons to your application, across three inference types — generate, stream and decide. `decide` returns typed, calibrated judgements from a non-generative model (~400ms, ~$0.00002/call) for routing, tool selection and context budgeting. MCP-native (4 transports), voice TTS/STT/realtime, RAG, agents, memory, compaction, 9 observability exporters. OpenAI · Anthropic · Gemini · Bedrock · Azure · Ollama · TypeSafe Jev and more.",
   "author": {