@juspay/neurolink 12.43.1 → 12.43.3

package/dist/providers/anthropic/client.js

@@ -1153,6 +1153,10 @@ export class AnthropicProvider extends BaseProvider {
                 }
                 // Extended thinking passthrough (providerOptions.anthropic.thinking).
                 const thinking = options.providerOptions?.anthropic?.thinking;
+                // `disabled` is sent so a model that thinks by default (claude-sonnet-5
+                // does, unasked) is told not to, but it fixes nothing about sampling: only a
+                // block that turns thinking ON makes Anthropic reject temperature.
+                const thinkingOn = thinking?.type === "enabled";
                 // Close the stable prefix with a breakpoint on the last tool. The
                 // stream path does the same, just before its own marker count.
                 tools = withLastToolCacheBreakpoint(tools);
@@ -1184,7 +1188,7 @@ export class AnthropicProvider extends BaseProvider {
                 }, "anthropic.doGenerate");
                 // Dropping a caller's explicit sampling parameters is exactly the
                 // kind of silent discard this change fixes elsewhere, so say so.
-                if (thinking &&
+                if (thinkingOn &&
                     (samplingParams.temperature !== undefined ||
                         samplingParams.topP !== undefined)) {
                     logger.debug("[anthropic] extended thinking is enabled, so temperature/top_p are omitted — Anthropic rejects any temperature but 1 while thinking is set");
@@ -1200,10 +1204,10 @@ export class AnthropicProvider extends BaseProvider {
                     // forwarding it alongside thinking turns a call that used to work
                     // into a 400. Drop the sampling knobs for exactly those turns and
                     // let Anthropic's thinking defaults stand.
-                    ...(!thinking && samplingParams.temperature !== undefined
+                    ...(!thinkingOn && samplingParams.temperature !== undefined
                         ? { temperature: samplingParams.temperature }
                         : {}),
-                    ...(!thinking && samplingParams.topP !== undefined
+                    ...(!thinkingOn && samplingParams.topP !== undefined
                         ? { top_p: samplingParams.topP }
                         : {}),
                     ...(options.stopSequences && options.stopSequences.length > 0
@@ -1534,6 +1538,12 @@ export class AnthropicProvider extends BaseProvider {
                 budget_tokens: options.thinkingConfig.budgetTokens,
             };
         }
+        else if (options.thinkingConfig?.type === "disabled") {
+            // Without this the request carries no `thinking` field at all, which on
+            // a model that thinks by default means "think" — the explicit
+            // `type: "disabled"` the type already offered was silently dropped.
+            anthropicNamespace.thinking = { type: "disabled" };
+        }
         // The per-call `timeout` keeps its per-MODEL-CALL meaning once
         // `turnTimeoutMs` owns the whole-turn deadline, and it reaches the model
         // layer only through providerOptions.neurolink. Without it each step fell
@@ -1751,7 +1761,11 @@ export class AnthropicProvider extends BaseProvider {
                 type: "enabled",
                 budget_tokens: options.thinkingConfig.budgetTokens,
             }
-            : undefined;
+            : options.thinkingConfig?.type === "disabled"
+                ? { type: "disabled" }
+                : undefined;
+        // See the generate path: `disabled` is sent, but only ON constrains sampling.
+        const thinkingOn = thinking?.type === "enabled";
         // Wrap the native stream in an OTel span to capture provider-level
         // latency and token usage (same span name as the pre-migration path so
         // dashboards stay continuous).
@@ -1956,7 +1970,7 @@ export class AnthropicProvider extends BaseProvider {
                 const streamSamplingParams = resolveSamplingParams("anthropic", modelId, options.temperature !== undefined && options.temperature !== null
                     ? { temperature: options.temperature }
                     : {}, "anthropic.executeStream");
-                if (thinking && streamSamplingParams.temperature !== undefined) {
+                if (thinkingOn && streamSamplingParams.temperature !== undefined) {
                     logger.debug("[anthropic] extended thinking is enabled, so temperature is omitted on the stream path — Anthropic rejects any temperature but 1 while thinking is set");
                 }
                 return {
@@ -1970,7 +1984,7 @@ export class AnthropicProvider extends BaseProvider {
                     ...(payload.system ? { system: payload.system } : {}),
                     // Same constraint on the streaming path: a temperature alongside
                     // `thinking` is rejected outright.
-                    ...(!thinking && streamSamplingParams.temperature !== undefined
+                    ...(!thinkingOn && streamSamplingParams.temperature !== undefined
                         ? { temperature: streamSamplingParams.temperature }
                         : {}),
                     ...(cachedTools && cachedTools.length > 0

package/dist/proxy/vertexAnthropicFallback.js

@@ -356,10 +356,22 @@ function fitRequestToTarget(payload, targetModel) {
         logger.warn(`[vertex-anthropic] ${targetModel} rejects tool_choice "${choice.type}"; sending "auto" instead`);
     }
     const thinking = payload.thinking;
-    if (record(thinking) && thinking.display === "updates") {
+    if (record(thinking) &&
+        thinking.type === "enabled" &&
+        targetModel === "claude-sonnet-5-5") {
+        // Sonnet 5.5 rejects fixed thinking budgets. Keep the requested display,
+        // but use its adaptive mode; a fixed token budget has no Vertex equivalent.
+        payload.thinking = {
+            type: "adaptive",
+            ...(thinking.display !== undefined ? { display: thinking.display } : {}),
+        };
+        logger.warn(`[vertex-anthropic] ${targetModel} rejects thinking "enabled"; sending adaptive without a fixed budget`);
+    }
+    const targetThinking = payload.thinking;
+    if (record(targetThinking) && targetThinking.display === "updates") {
         // Claude Code requests live thinking updates, but Vertex accepts only
         // summarized or omitted adaptive thinking display.
-        payload.thinking = { ...thinking, display: "summarized" };
+        payload.thinking = { ...targetThinking, display: "summarized" };
     }
     if (record(thinking) && thinking.type === "disabled") {
         const outputConfig = payload.output_config;

package/docs-site/static/search-index.json

@@ -4479,7 +4479,7 @@
 {"objectID":"c94689eb04fe3a8bb13cdf4da5967e12e54dc05d23065bd3e3524af01cd4f971","title":"Feature Implementation Workflow","url":"/docs/features/task-manager#feature-implementation-workflow","content":"This example automates: write doc → review → revise → implement → create PR → resolve comments → push.\n\nHow it plays out:\n\n| Run | AI Behavior                                                                            |\n| --- | -------------------------------------------------------------------------------------- |\n| 1   | Writes docs/auth-design.md using writeFile tool. \"Step 1 complete. Next: review.\"  |\n| 2   | Reads the doc, identifies gaps. \"Step 2 complete: found 3 issues. Next: revise.\"       |\n| 3   | Rewrites sections. \"Step 3 complete. Doc is ready. Next: implement.\"                   |\n| 4   | Reads doc, writes code across multiple files. \"Step 4 complete. Next: create PR.\"      |\n| 5   | Uses GitHub MCP to create PR. \"Step 5 complete. PR #42 created. Next: check comments.\" |\n| 6   | Reads PR comments via GitHub MCP. \"2 comments found. Next: address them.\"              |\n| 7   | Pushes fixes, re-checks. \"All comments resolved. ALL STEPS COMPLETE.\"                  |\n| —   | onSuccess callback detects completion, pauses the task.                              |\n\nBecause this is continuation mode, the AI has full context of every previous run — it knows what it wrote, what was reviewed, and what comments were left. No external state management needed.","hierarchy":{"lvl0":"Features","lvl1":"TaskManager - Scheduled & Self-Running Tasks","lvl2":"Feature Implementation Workflow","lvl3":""}},
 {"objectID":"1ed0cf5cb87ec5d23bf366e3b3d3c81407802f490fb50d3fef44230ef5c3fc6f","title":"Existing Code Impact","url":"/docs/features/task-manager#existing-code-impact","content":"TaskManager is implemented as a new module (src/lib/tasks/). Minimal changes to existing code:\n\n| Existing File                  | Change                              | Lines |\n| ------------------------------ | ----------------------------------- | ----- |\n| src/lib/neurolink.ts         | Add tasks getter property         | ~10   |\n| src/lib/types/index.ts       | Re-export task types                | ~2    |\n| src/lib/agent/directTools.ts | Import and spread task tools        | ~3    |\n| src/cli/index.ts             | Register task CLI command         | ~1    |\n| package.json                 | Add bullmq, croner dependencies | ~2    |\n\nEverything else is new files in src/lib/tasks/. No refactoring, no restructuring of existing code.","hierarchy":{"lvl0":"Features","lvl1":"TaskManager - Scheduled & Self-Running Tasks","lvl2":"Existing Code Impact","lvl3":""}},
 {"objectID":"e937a3f76bbe1a5494c610c87139c64ed299320dc8364078c0849a6231a023b8","title":"FAQ","url":"/docs/features/task-manager#faq","content":"Q: Do I need Redis to use TaskManager?\nA: No. Set backend: \"node-timeout\" for a zero-dependency setup. Redis is only required for the BullMQ backend (the default). When using BullMQ, Redis stores both job scheduling state and task definitions/run logs — no file I/O at all.\n\nQ: Will tasks stay in Redis forever?\nA: No. Active/paused tasks persist as long as they're running. Once a task reaches a terminal state (completed, failed, cancelled), it auto-expires based on taskRetention config — defaults: 30 days for completed, 7 days for failed/cancelled. Run logs are capped at maxRunLogs (default 2000) per task via LTRIM, and individual entries can have a TTL. You can also manually delete tasks via tasks.delete() or neurolink task delete.\n\nQ: Can the AI schedule tasks without user intervention?\nA: Yes. The built-in createTask tool allows the AI to self-schedule tasks during any conversation. If HITL is enabled, the user will be prompted for approval.\n\nQ: How does continuation mode handle growing context?\nA: It uses NeuroLink's existing context compaction system. When conversation history exceeds the model's context budget, BudgetChecker triggers automatic summarization.\n\nQ: Can I use TaskManager in a serverless environment?\nA: The BullMQ backend works in serverless with a persistent Redis instance — no local filesystem needed. The NodeTimeout backend requires a long-running process with filesystem access.\n\nQ: What happens when I switch backends?\nA: Tasks stored in Redis (BullMQ) and tasks stored on disk (NodeTimeout) are independent. Switching backends does not migrate data. If you need to migrate, use tasks.list() on the old backend and tasks.create() on the new one.\n\nQ: How do I monitor task health?\nA: Use neurolink.tasks.list() / neurolink task list for status overview, neurolink.tasks.runs(taskId) / neurolink task logs <id> for run history, and subscribe to task:* events for real-time monitoring.","hierarchy":{"lvl0":"Features","lvl1":"TaskManager - Scheduled & Self-Running Tasks","lvl2":"FAQ","lvl3":""}},
-{"objectID":"5597f3996aaaa278b6fc9cb9310ea7f35b65373b35043abf05f0b33ebe8d5d82","title":"Extended Thinking Configuration","url":"/docs/features/thinking-configuration","content":"Extended Thinking Configuration\n\nEnable extended thinking/reasoning modes for AI models that support deeper reasoning capabilities. This feature allows models to \"think through\" complex problems before providing a response.\n\nOverview\n\nNeuroLink supports extended thinking/reasoning configuration for models that provide this capability. Extended thinking enables models to perform more thorough reasoning, particularly useful for complex tasks like mathematical proofs, coding problems, and multi-step analysis.\n\nSupported Models\n\nGemini 3 Models (Google Vertex AI / AI Studio)\ngemini-3.1-pro - Full thinking support with high token budgets (up to 100,000)\ngemini-3-flash-preview - Fast thinking with support for \"minimal\" level (up to 50,000)\n\nGemini 2.5 Models (Google Vertex AI / AI Studio)\ngemini-2.5-pro - Supports thinking configuration (up to 32,000 tokens)\ngemini-2.5-flash - Supports thinking configuration (up to 32,000 tokens)\n\nClaude Models (Anthropic)\n\nAll Claude 4.0+ models support extended thinking via budget tokens:\nclaude-sonnet-4-20250514 (Claude Sonnet 4)\nclaude-opus-4-20250514 (Claude Opus 4)\nclaude-opus-4-1-20250805 (Claude Opus 4.1)\nclaude-sonnet-4-5-20250929 (Claude Sonnet 4.5)\nclaude-opus-4-5-20251101 (Claude Opus 4.5)\nclaude-haiku-4-5-20251001 (Claude Haiku 4.5)\nclaude-sonnet-4-6 (Claude Sonnet 4.6)\nclaude-opus-4-6 (Claude Opus 4.6)\n\nQuick Start\n\nGemini 3 Thinking Configuration\n\nFor Gemini 3 models, use thinkingLevel to control reasoning depth:\n\nThinking Levels\n\n| Level     | Description                            | Best For                        |\n| --------- | -------------------------------------- | ------------------------------- |\n| minimal | Near-zero thinking (Flash models only) | Simple queries requiring speed  |\n| low     | Fast reasoning for simple tasks        | Quick analysis, summaries       |\n| medium  | Balanced reasoning/latency trade-off   | General-purpose tasks           |\n| high    | Maximum reasoning depth                | Complex reasoning, math, coding |\n\nMaximum Token Budgets by Model\n\n| Model                 | Max Thinking Budget |\n| --------------------- | ------------------- |\n| gemini-3-pro-*      | 100,000 tokens      |\n| gemini-3-flash-*    | 50,000 tokens       |\n| gemini-2.5-*        | 32,000 tokens       |\n| claude-opus-4-6     | 100,000 tokens      |\n| claude-sonnet-4-6   | 100,000 tokens      |\n| claude-opus-4-5-*   | 100,000 tokens      |\n| claude-sonnet-4-5-* | 100,000 tokens      |\n| claude-haiku-4-5-*  | 100,000 tokens      |\n| claude-opus-4-1-*   | 100,000 tokens      |\n| claude-opus-4-*     | 100,000 tokens      |\n| claude-sonnet-4-*   | 100,000 tokens      |\n\nAnthropic Claude Thinking Configuration\n\nFor Claude models, use budgetTokens to set the thinking token budget:\n\nBudget Token Guidelines\nMinimum: 5,000 tokens\nMaximum: 100,000 tokens\nRecommended for simple tasks: 5,000-10,000 tokens\nRecommended for complex reasoning: 20,000-50,000 tokens\nMaximum depth: 50,000-100,000 tokens\n\nConfiguration Options\n\nThe thinkingConfig object supports the following options:\n\nCLI Usage\n\nExtended thinking is also available via the CLI:\n\nCLI Options\n\n| Option             | Description                                           | Default |\n| ------------------ | ----------------------------------------------------- | ------- |\n| --thinking       | Enable extended thinking                              | false   |\n| --thinkingBudget | Token budget (Anthropic: 5000-100000)                 | 10000   |\n| --thinkingLevel  | Thinking level (Gemini 3: minimal, low, medium, high) | medium  |\n\nBest Practices\n\nWhen to Use High Thinking\nComplex mathematical proofs and calculations\nMulti-step coding problems and debugging\nDetailed analysis requiring multiple considerations\nTasks where accuracy is more important than speed\n\nWhen to Use Low/Minimal Thinking\nSimple queries where speed matters\nStraightforward information retrieval\nQuick summaries and formatting tasks\nHigh-volume, latency-sensitive applications\n\nGeneral Guidelines\nStart with medium: Use medium as your default and adjust based on results\nMatch model to task: Use Pro models for complex tasks, Flash for speed\nMonitor token usage: Higher thinking levels consume more tokens\nTest performance: Compare response quality vs. latency for your use case\n\nExample: Complex Reasoning Task\n\nModel Detection Utilities\n\nNeuroLink provides utilities to check thinking support:\n\nImportant Notes\nProvider compatibility: Thinking configuration is provider-specific. Gemini uses thinkingLevel, Claude uses budgetTokens\nToken consumption: Extended thinking uses additional tokens beyond the response\nLatency impact: Higher thinking levels increase response time\nNot all models support thinking: Check supportsThinkingConfig() before enabling\nStreaming support: Thinking configuration works with both generate() and stream()\n\nSee Also\nAPI Reference\nProvider Configuration\nStreaming","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"","lvl3":""}},
+{"objectID":"5597f3996aaaa278b6fc9cb9310ea7f35b65373b35043abf05f0b33ebe8d5d82","title":"Extended Thinking Configuration","url":"/docs/features/thinking-configuration","content":"Extended Thinking Configuration\n\nEnable extended thinking/reasoning modes for AI models that support deeper reasoning capabilities. This feature allows models to \"think through\" complex problems before providing a response.\n\nOverview\n\nNeuroLink supports extended thinking/reasoning configuration for models that provide this capability. Extended thinking enables models to perform more thorough reasoning, particularly useful for complex tasks like mathematical proofs, coding problems, and multi-step analysis.\n\nSupported Models\n\nGemini 3 Models (Google Vertex AI / AI Studio)\ngemini-3.1-pro - Full thinking support with high token budgets (up to 100,000)\ngemini-3-flash-preview - Fast thinking with support for \"minimal\" level (up to 50,000)\n\nGemini 2.5 Models (Google Vertex AI / AI Studio)\ngemini-2.5-pro - Supports thinking configuration (up to 32,000 tokens)\ngemini-2.5-flash - Supports thinking configuration (up to 32,000 tokens)\n\nClaude Models (Anthropic)\n\nAll Claude 4.0+ models support extended thinking via budget tokens:\nclaude-sonnet-4-20250514 (Claude Sonnet 4)\nclaude-opus-4-20250514 (Claude Opus 4)\nclaude-opus-4-1-20250805 (Claude Opus 4.1)\nclaude-sonnet-4-5-20250929 (Claude Sonnet 4.5)\nclaude-opus-4-5-20251101 (Claude Opus 4.5)\nclaude-haiku-4-5-20251001 (Claude Haiku 4.5)\nclaude-sonnet-4-6 (Claude Sonnet 4.6)\nclaude-opus-4-6 (Claude Opus 4.6)\n\nQuick Start\n\nGemini 3 Thinking Configuration\n\nFor Gemini 3 models, use thinkingLevel to control reasoning depth:\n\nThinking Levels\n\n| Level     | Description                            | Best For                        |\n| --------- | -------------------------------------- | ------------------------------- |\n| minimal | Near-zero thinking (Flash models only) | Simple queries requiring speed  |\n| low     | Fast reasoning for simple tasks        | Quick analysis, summaries       |\n| medium  | Balanced reasoning/latency trade-off   | General-purpose tasks           |\n| high    | Maximum reasoning depth                | Complex reasoning, math, coding |\n\nMaximum Token Budgets by Model\n\n| Model                 | Max Thinking Budget |\n| --------------------- | ------------------- |\n| gemini-3-pro-*      | 100,000 tokens      |\n| gemini-3-flash-*    | 50,000 tokens       |\n| gemini-2.5-*        | 32,000 tokens       |\n| claude-opus-4-6     | 100,000 tokens      |\n| claude-sonnet-4-6   | 100,000 tokens      |\n| claude-opus-4-5-*   | 100,000 tokens      |\n| claude-sonnet-4-5-* | 100,000 tokens      |\n| claude-haiku-4-5-*  | 100,000 tokens      |\n| claude-opus-4-1-*   | 100,000 tokens      |\n| claude-opus-4-*     | 100,000 tokens      |\n| claude-sonnet-4-*   | 100,000 tokens      |\n\nAnthropic Claude Thinking Configuration\n\nFor Claude models, use budgetTokens to set the thinking token budget:\n\nBudget Token Guidelines\nMinimum: 5,000 tokens\nMaximum: 100,000 tokens\nRecommended for simple tasks: 5,000-10,000 tokens\nRecommended for complex reasoning: 20,000-50,000 tokens\nMaximum depth: 50,000-100,000 tokens\n\nTurning thinking off\n\nSome Claude models, claude-sonnet-5 among them, think before answering even when\nno thinkingConfig is given, and that thinking counts against the request's\nmax_tokens. To turn it off for a call, pass type: \"disabled\":\n\nNeuroLink sends thinking: { type: \"disabled\" } on both generate() and\nstream(). Turning thinking off does not change which sampling parameters a\nmodel accepts. On a model that takes a temperature, NeuroLink still sends it\nwith thinking disabled; it is dropped only while thinking is on. Some models,\nclaude-sonnet-5 among them, do not accept a temperature at all, and NeuroLink\nleaves it off for them whatever the thinking setting.\n\nenabled: false, or no thinkingConfig at all, sends nothing, so the model's own\ndefault applies. If enabled: true and budgetTokens are also set, thinking is\nenabled, as before.\n\nConfiguration Options\n\nThe thinkingConfig object supports the following options:\n\nCLI Usage\n\nExtended thinking is also available via the CLI:\n\nCLI Options\n\n| Option             | Description                                           | Default |\n| ------------------ | ----------------------------------------------------- | ------- |\n| --thinking       | Enable extended thinking                              | false   |\n| --thinkingBudget | Token budget (Anthropic: 5000-100000)                 | 10000   |\n| --thinkingLevel  | Thinking level (Gemini 3: minimal, low, medium, high) | medium  |\n\nBest Practices\n\nWhen to Use High Thinking\nComplex mathematical proofs and calculations\nMulti-step coding problems and debugging\nDetailed analysis requiring multiple considerations\nTasks where accuracy is more important than speed\n\nWhen to Use Low/Minimal Thinking\nSimple queries where speed matters\nStraightforward information retrieval\nQuick summaries and formatting tasks\nHigh-volume, latency-sensitive applications\n\nGeneral Guidelines\nStart with medium: Use medium as your default and adjust based on results\nMatch model to task: Use Pro models for complex tasks, ","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"","lvl3":""}},
 {"objectID":"b1a3fc87278e92aa39a0271b6f2eb465b62008b51d064acce63cf3af3f5c9673","title":"Extended Thinking Configuration","url":"/docs/features/thinking-configuration#extended-thinking-configuration","content":"Enable extended thinking/reasoning modes for AI models that support deeper reasoning capabilities. This feature allows models to \"think through\" complex problems before providing a response.","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Extended Thinking Configuration","lvl3":""}},
 {"objectID":"5e898a7f44534e174e959788be122e3ba4d1ee14def3b6214209393edf08cbae","title":"Overview","url":"/docs/features/thinking-configuration#overview","content":"NeuroLink supports extended thinking/reasoning configuration for models that provide this capability. Extended thinking enables models to perform more thorough reasoning, particularly useful for complex tasks like mathematical proofs, coding problems, and multi-step analysis.","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Overview","lvl3":""}},
 {"objectID":"3c3340ac962c9488fd2f05b80b29b47a3a272754a135742aa678ed7dc1845462","title":"Gemini 3 Models (Google Vertex AI / AI Studio)","url":"/docs/features/thinking-configuration#gemini-3-models-google-vertex-ai-ai-studio","content":"gemini-3.1-pro - Full thinking support with high token budgets (up to 100,000)\ngemini-3-flash-preview - Fast thinking with support for \"minimal\" level (up to 50,000)","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Gemini 3 Models (Google Vertex AI / AI Studio)","lvl3":""}},
@@ -4490,6 +4490,7 @@
 {"objectID":"d29afbd5be885a8dfae9576498c586034610e698e65e0020c32f504a6ffb0b4d","title":"Maximum Token Budgets by Model","url":"/docs/features/thinking-configuration#maximum-token-budgets-by-model","content":"| Model                 | Max Thinking Budget |\n| --------------------- | ------------------- |\n| gemini-3-pro-*      | 100,000 tokens      |\n| gemini-3-flash-*    | 50,000 tokens       |\n| gemini-2.5-*        | 32,000 tokens       |\n| claude-opus-4-6     | 100,000 tokens      |\n| claude-sonnet-4-6   | 100,000 tokens      |\n| claude-opus-4-5-*   | 100,000 tokens      |\n| claude-sonnet-4-5-* | 100,000 tokens      |\n| claude-haiku-4-5-*  | 100,000 tokens      |\n| claude-opus-4-1-*   | 100,000 tokens      |\n| claude-opus-4-*     | 100,000 tokens      |\n| claude-sonnet-4-*   | 100,000 tokens      |","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Maximum Token Budgets by Model","lvl3":""}},
 {"objectID":"a1b9be5a0a0a16cff259e6051df668ddc07f75a6ae77ebba4121f649746eb25e","title":"Anthropic Claude Thinking Configuration","url":"/docs/features/thinking-configuration#anthropic-claude-thinking-configuration","content":"For Claude models, use budgetTokens to set the thinking token budget:","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Anthropic Claude Thinking Configuration","lvl3":""}},
 {"objectID":"942ce04a601afc6269d1765ad4da7295a802070a5b036c93dd360e92478e3be3","title":"Budget Token Guidelines","url":"/docs/features/thinking-configuration#budget-token-guidelines","content":"Minimum: 5,000 tokens\nMaximum: 100,000 tokens\nRecommended for simple tasks: 5,000-10,000 tokens\nRecommended for complex reasoning: 20,000-50,000 tokens\nMaximum depth: 50,000-100,000 tokens","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Budget Token Guidelines","lvl3":""}},
+{"objectID":"43d4815c3e1af5de7858131d9b25bcf7e59ef2dee4327c46cf613cbdd0f704d9","title":"Turning thinking off","url":"/docs/features/thinking-configuration#turning-thinking-off","content":"Some Claude models, claude-sonnet-5 among them, think before answering even when\nno thinkingConfig is given, and that thinking counts against the request's\nmax_tokens. To turn it off for a call, pass type: \"disabled\":\n\nNeuroLink sends thinking: { type: \"disabled\" } on both generate() and\nstream(). Turning thinking off does not change which sampling parameters a\nmodel accepts. On a model that takes a temperature, NeuroLink still sends it\nwith thinking disabled; it is dropped only while thinking is on. Some models,\nclaude-sonnet-5 among them, do not accept a temperature at all, and NeuroLink\nleaves it off for them whatever the thinking setting.\n\nenabled: false, or no thinkingConfig at all, sends nothing, so the model's own\ndefault applies. If enabled: true and budgetTokens are also set, thinking is\nenabled, as before.","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Turning thinking off","lvl3":""}},
 {"objectID":"2b5576316a6208f5cfec1c81bc1e6fef5b0183103738313b707f4f56e5cfa57a","title":"Configuration Options","url":"/docs/features/thinking-configuration#configuration-options","content":"The thinkingConfig object supports the following options:","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Configuration Options","lvl3":""}},
 {"objectID":"a8d5868d4d30b359f7c8882f700ef9b1930d1be223f54728849a4634314aa07f","title":"CLI Usage","url":"/docs/features/thinking-configuration#cli-usage","content":"Extended thinking is also available via the CLI:\n\n`bash","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"CLI Usage","lvl3":""}},
 {"objectID":"89ef76b888510f16b5214e2da51842acf293f9c6b751ba484a0a1261f5a82bbd","title":"Enable thinking with default settings","url":"/docs/features/thinking-configuration#enable-thinking-with-default-settings","content":"neurolink generate \"Solve this problem\" --thinking","hierarchy":{"lvl0":"Features","lvl1":"Extended Thinking Configuration","lvl2":"Enable thinking with default settings","lvl3":""}},

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "12.43.1",
+  "version": "12.43.3",
   "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": {
@@ -102,6 +102,7 @@
     "test:openai-compat-streaming-retry": "pnpm exec tsx test/continuous-test-suite-openai-compat-streaming-retry.ts",
     "test:anthropic-streaming-retry": "pnpm exec tsx test/continuous-test-suite-anthropic-streaming-retry.ts",
     "test:anthropic-silent-truncation": "pnpm exec tsx test/continuous-test-suite-anthropic-silent-truncation.ts",
+    "test:anthropic-thinking-disabled": "pnpm exec tsx test/continuous-test-suite-anthropic-thinking-disabled.ts",
     "test:cache-cost-accounting": "pnpm exec tsx test/continuous-test-suite-cache-cost-accounting.ts",
     "test:adjust-body-after-400": "pnpm exec tsx test/continuous-test-suite-adjust-body-after-400.ts",
     "test:zod3-schema-native-path": "pnpm exec tsx test/continuous-test-suite-zod3-schema-native-path.ts",