@juspay/neurolink 12.37.4 → 12.37.5

package/dist/core/baseProvider.d.ts

@@ -262,12 +262,21 @@ export declare abstract class BaseProvider implements AIProvider {
      * and a retired default model stops degrading to the next live one in the
      * catalog — which is exactly what the provider contract gate checks.
      */
-    protected runGenerateWithModelFallback(attempt: () => Promise<EnhancedGenerateResult | null>, callerOwnsFallback: boolean): Promise<EnhancedGenerateResult | null>;
+    protected runGenerateWithModelFallback(attempt: () => Promise<EnhancedGenerateResult | null>, callerOwnsFallback: boolean, options?: TextGenerationOptions): Promise<EnhancedGenerateResult | null>;
     /**
      * Model ids to try when this provider rejects its current model as invalid.
      * Empty by default; catalog-driven providers return their `fallbacks`.
      */
     protected getModelFallbacks(): string[];
+    /**
+     * The fallbacks an invalid-model retry may try for this request. A catalog
+     * orders `fallbacks` for text, so a request carrying an image keeps only the
+     * vision-capable ones: a text-only candidate fails the vision guard, and
+     * that error names the fallback instead of the retired model the caller
+     * asked for. With no vision-capable candidate left, the caller gets the
+     * original invalid-model error.
+     */
+    protected modelFallbacksFor(options: TextGenerationOptions | StreamOptions, requestedModel: string): string[];
     /**
      * Alias for generate method - implements AIProvider interface
      */

package/dist/core/baseProvider.js

@@ -46,6 +46,7 @@ import { StreamHandler } from "./modules/StreamHandler.js";
 import { TelemetryHandler } from "./modules/TelemetryHandler.js";
 import { ToolsManager } from "./modules/ToolsManager.js";
 import { Utilities } from "./modules/Utilities.js";
+import { ProviderImageAdapter } from "../adapters/providerImageAdapter.js";
 import { generateOnceNative } from "../utils/nativeSingleShot.js";
 import { validateExecutionControl } from "../utils/parameterValidation.js";
 import { extractTokenUsage } from "../utils/tokenUtils.js";
@@ -107,6 +108,18 @@ function isProviderErrorClassified(error) {
     }
     return error[PROVIDER_ERROR_CLASSIFIED] === true;
 }
+function requestCarriesImages(options) {
+    const input = options.input;
+    if (!input) {
+        return false;
+    }
+    if ((input.images?.length ?? 0) > 0) {
+        return true;
+    }
+    return ("content" in input &&
+        Array.isArray(input.content) &&
+        input.content.some((part) => part.type === "image"));
+}
 export class BaseProvider {
     // Not `readonly` because providers that auto-discover the model from a
     // /v1/models endpoint (lm-studio, llamacpp) need to update modelName after
@@ -523,9 +536,7 @@ export class BaseProvider {
                     throw error;
                 }
                 const requestedModel = provider.modelName;
-                const candidates = provider
-                    .getModelFallbacks()
-                    .filter((model) => model !== requestedModel);
+                const candidates = provider.modelFallbacksFor(options, requestedModel);
                 for (const candidate of candidates) {
                     logger.warn(`[${provider.providerName}] model "${requestedModel}" was rejected as invalid — retrying stream with fallback "${candidate}". This provider's catalog entry is stale; run "pnpm run check:models".`);
                     provider.refreshHandlersForModel(candidate);
@@ -671,7 +682,7 @@ export class BaseProvider {
             return undefined;
         }
         const requestedModel = this.modelName;
-        const candidates = this.getModelFallbacks().filter((model) => model !== requestedModel);
+        const candidates = this.modelFallbacksFor(options, requestedModel);
         for (const candidate of candidates) {
             logger.warn(`[${this.providerName}] model "${requestedModel}" was rejected as invalid — retrying stream with fallback "${candidate}". This provider's catalog entry is stale; run "pnpm run check:models".`);
             this.refreshHandlersForModel(candidate);
@@ -1493,7 +1504,7 @@ export class BaseProvider {
         // or a router that retries on its own) pass the flag on both paths;
         // TextGenerationOptions declares it, so a plain read is enough here.
         const callerOwnsFallback = options.disableInternalFallback === true;
-        return await this.runGenerateWithModelFallback(attempt, callerOwnsFallback);
+        return await this.runGenerateWithModelFallback(attempt, callerOwnsFallback, options);
     }
     /**
      * Models the catalog names are the models a vendor served when the entry was
@@ -1520,7 +1531,7 @@ export class BaseProvider {
      * and a retired default model stops degrading to the next live one in the
      * catalog — which is exactly what the provider contract gate checks.
      */
-    async runGenerateWithModelFallback(attempt, callerOwnsFallback) {
+    async runGenerateWithModelFallback(attempt, callerOwnsFallback, options) {
         const requestedModel = this.modelName;
         try {
             return await attempt();
@@ -1529,7 +1540,9 @@ export class BaseProvider {
             if (callerOwnsFallback || !isInvalidModelError(error)) {
                 throw error;
             }
-            const candidates = this.getModelFallbacks().filter((model) => model !== requestedModel);
+            const candidates = options
+                ? this.modelFallbacksFor(options, requestedModel)
+                : this.getModelFallbacks().filter((model) => model !== requestedModel);
             if (candidates.length === 0) {
                 throw error;
             }
@@ -1563,6 +1576,21 @@ export class BaseProvider {
     getModelFallbacks() {
         return [];
     }
+    /**
+     * The fallbacks an invalid-model retry may try for this request. A catalog
+     * orders `fallbacks` for text, so a request carrying an image keeps only the
+     * vision-capable ones: a text-only candidate fails the vision guard, and
+     * that error names the fallback instead of the retired model the caller
+     * asked for. With no vision-capable candidate left, the caller gets the
+     * original invalid-model error.
+     */
+    modelFallbacksFor(options, requestedModel) {
+        const candidates = this.getModelFallbacks().filter((model) => model !== requestedModel);
+        if (!requestCarriesImages(options)) {
+            return candidates;
+        }
+        return candidates.filter((model) => ProviderImageAdapter.supportsVision(this.providerName, model));
+    }
     /**
      * Alias for generate method - implements AIProvider interface
      */

package/dist/models/manifests/anthropic.js

@@ -38,7 +38,7 @@ export const anthropicManifest = {
             samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /opus[-_.]?5(?![0-9])/i
         },
         "claude-sonnet-5-5": {
-            aliases: ["sonnet-5.5"],
+            aliases: ["sonnet-5.5", "claude-sonnet"],
             displayName: "Claude Sonnet 5.5",
             contextWindow: 1_000_000,
             maxOutputTokens: 128_000,
@@ -66,7 +66,7 @@ export const anthropicManifest = {
             samplingParams: false, // matches SAMPLING_PARAM_REJECTING_FAMILIES /fable/i
         },
         "claude-sonnet-5": {
-            aliases: ["sonnet-5", "claude-sonnet"],
+            aliases: ["sonnet-5"],
             displayName: "Claude Sonnet 5",
             contextWindow: 1_000_000,
             // Was 64_000; corrected to 128_000 to match the official spec

package/dist/models/modelRegistry.d.ts

@@ -4,7 +4,7 @@
  * Part of Phase 4.1 - Models Command System
  */
 import { AIProviderName } from "../constants/enums.js";
-import type { JsonValue, ModelCapabilities, ModelInfo } from "../types/index.js";
+import type { ClaudeDisabledThinkingReplacement, JsonValue, ModelCapabilities, ModelInfo } from "../types/index.js";
 /**
  * Public MODEL_REGISTRY: the manifest-derived entries merged over the
  * hand-authored LEGACY_MODEL_REGISTRY base. Spread order matters — manifest
@@ -90,6 +90,17 @@ export declare function resolveSamplingParams(provider: string, model: string |
     temperature?: number;
     topP?: number;
 };
+/**
+ * Whether a Claude model accepts a forced `tool_choice` (`any` or a named
+ * `tool`). Unknown models default to accepting it.
+ */
+export declare function modelSupportsForcedToolChoice(model: string | undefined): boolean;
+/**
+ * What to send instead of `thinking: {type: "disabled"}` to a Claude model.
+ * `between_tools` is refused at `xhigh` and `max` effort, so those fall back
+ * to adaptive thinking (`omit`).
+ */
+export declare function claudeDisabledThinkingReplacement(model: string | undefined, effort: string | undefined): ClaudeDisabledThinkingReplacement;
 /**
  * Get models by provider
  */

package/dist/models/modelRegistry.js

@@ -1168,14 +1168,14 @@ const LEGACY_MODEL_REGISTRY = {
             translation: 9,
             summarization: 9,
         },
-        aliases: ["sonnet-5", "claude-sonnet-latest"],
+        aliases: ["sonnet-5"],
         deprecated: false,
         isLocal: false,
         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.
+    // Owns "claude-sonnet-latest", but sits after CLAUDE_SONNET_5 so resolver
+    // score ties still go to Sonnet 5.
     [AnthropicModels.CLAUDE_SONNET_5_5]: {
         id: AnthropicModels.CLAUDE_SONNET_5_5,
         name: "Claude Sonnet 5.5",
@@ -1210,7 +1210,7 @@ const LEGACY_MODEL_REGISTRY = {
             translation: 9,
             summarization: 9,
         },
-        aliases: ["sonnet-5.5"],
+        aliases: ["sonnet-5.5", "claude-sonnet-latest"],
         deprecated: false,
         isLocal: false,
         releaseDate: "2026-09-28",
@@ -1374,7 +1374,7 @@ const LEGACY_MODEL_REGISTRY = {
             translation: 8,
             summarization: 8,
         },
-        // "claude-sonnet-latest" now belongs to CLAUDE_SONNET_5 — see the note
+        // "claude-sonnet-latest" now belongs to CLAUDE_SONNET_5_5 — see the note
         // on CLAUDE_OPUS_4_5's aliases.
         aliases: ["claude-4.5-sonnet", "sonnet-4.5"],
         deprecated: false,
@@ -2846,6 +2846,58 @@ export function resolveSamplingParams(provider, model, params, site) {
     }
     return {};
 }
+/**
+ * Claude models that answer a forced `tool_choice` (`any` or `tool`) with a
+ * 400: "tool_choice: type "tool" and "any" are not supported for this model."
+ * Sonnet 5.5 and Opus 5.5 were probed live on Vertex; the Fable 5.1 docs state
+ * the same, and Mythos 5.1 shares Fable 5.1's capabilities. The same families
+ * also refuse `thinking: {type: "disabled"}`.
+ */
+const FORCED_TOOL_CHOICE_REJECTING_FAMILIES = [
+    /sonnet[-_.]?5[-_.]5(?![0-9])/i,
+    /opus[-_.]?5[-_.]5(?![0-9])/i,
+    /(?:fable|mythos)[-_.]?5[-_.]1(?![0-9])/i,
+];
+/**
+ * Of those, the ones that accept `thinking: {type: "between_tools"}` as their
+ * lowest setting. The rest think adaptively, always on, and take no
+ * `thinking` field in its place.
+ */
+const BETWEEN_TOOLS_THINKING_FAMILIES = [
+    /sonnet[-_.]?5[-_.]5(?![0-9])/i,
+];
+function matchesClaudeFamily(families, model) {
+    const normalizedModel = (model ?? "").toLowerCase();
+    if (!normalizedModel) {
+        return false;
+    }
+    const modelId = MODEL_ALIASES[normalizedModel] ?? normalizedModel;
+    return families.some((family) => family.test(modelId) || family.test(normalizedModel));
+}
+/**
+ * Whether a Claude model accepts a forced `tool_choice` (`any` or a named
+ * `tool`). Unknown models default to accepting it.
+ */
+export function modelSupportsForcedToolChoice(model) {
+    return !matchesClaudeFamily(FORCED_TOOL_CHOICE_REJECTING_FAMILIES, model);
+}
+/**
+ * What to send instead of `thinking: {type: "disabled"}` to a Claude model.
+ * `between_tools` is refused at `xhigh` and `max` effort, so those fall back
+ * to adaptive thinking (`omit`).
+ */
+export function claudeDisabledThinkingReplacement(model, effort) {
+    if (!matchesClaudeFamily(FORCED_TOOL_CHOICE_REJECTING_FAMILIES, model)) {
+        return "keep";
+    }
+    const level = effort?.toLowerCase();
+    if (matchesClaudeFamily(BETWEEN_TOOLS_THINKING_FAMILIES, model) &&
+        level !== "xhigh" &&
+        level !== "max") {
+        return "between_tools";
+    }
+    return "omit";
+}
 /**
  * Get models by provider
  */

package/dist/providers/amazonSagemaker.js

@@ -126,7 +126,7 @@ export class AmazonSageMakerProvider extends BaseProvider {
             options.disableInternalFallback === true;
         // The native loop bypasses BaseProvider.executeGeneration, so the turn
         // budget has to be composed here or it stops existing for this provider.
-        return this.runGenerateWithModelFallback(() => this.withTurnTimeout({ ...options, tools: mergedTools }, this.getDescriptorGenerateMs(), (timedOptions) => this.executeNativeGenerate(timedOptions)), callerOwnsFallback);
+        return this.runGenerateWithModelFallback(() => this.withTurnTimeout({ ...options, tools: mergedTools }, this.getDescriptorGenerateMs(), (timedOptions) => this.executeNativeGenerate(timedOptions)), callerOwnsFallback, options);
     }
     async executeNativeGenerate(options) {
         const startTime = Date.now();

package/dist/providers/anthropic/client.js

@@ -41,14 +41,14 @@ import { buildNoOutputSentinel, stampNoOutputSpan, } from "../../utils/noOutputS
 import { convertZodToJsonSchema } from "../../utils/schemaConversion.js";
 import { resolveClaudeMaxTokens } from "../../utils/tokenLimits.js";
 import { toAnthropicImageBlock, fileToAnthropicBlock, } from "../anthropicImageBlocks.js";
-import { resolveSamplingParams } from "../../models/modelRegistry.js";
+import { modelSupportsForcedToolChoice, resolveSamplingParams, } from "../../models/modelRegistry.js";
 import { createDeferredAnalytics, stringifyToolInput, } from "../openaiChatCompletionsClient.js";
 import { createStreamChannel } from "../../core/streamChannel.js";
 import { toNativeToolDeclarations } from "../../core/nativeToolFormat.js";
 import { ANTHROPIC_BETA_HEADERS } from "./constants.js";
 import { cacheControlOf, withLastToolCacheBreakpoint } from "./cacheControl.js";
 import { createNativeGenerateGuard } from "../../context/nativeGenerateGuard.js";
-import { appendFinalResultInstruction, appendFinalResultTool, FINAL_RESULT_TOOL_NAME, } from "./structuredOutput.js";
+import { appendFinalResultInstruction, appendFinalResultTool, appendToolAnswerInstruction, FINAL_RESULT_TOOL_NAME, } from "./structuredOutput.js";
 // AnthropicProviderConfig is imported from types/providers.ts
 // Re-export for backward compatibility
 // Configuration helpers - now using consolidated utility
@@ -424,6 +424,23 @@ const toolChoiceToAnthropic = (choice) => {
     }
     return undefined;
 };
+/**
+ * Claude 5.5 / 5.1 models answer a forced choice with a 400, so for them it
+ * becomes `auto`: the tool stays declared and the model still picks it.
+ */
+const relaxForcedToolChoice = (modelId, choice) => {
+    if ((choice?.type !== "any" && choice?.type !== "tool") ||
+        modelSupportsForcedToolChoice(modelId)) {
+        return choice;
+    }
+    logger.warn(`[anthropic] ${modelId} rejects tool_choice "${choice.type}"; sending "auto" instead`);
+    return {
+        type: "auto",
+        ...(choice.disable_parallel_tool_use !== undefined
+            ? { disable_parallel_tool_use: choice.disable_parallel_tool_use }
+            : {}),
+    };
+};
 /** Map Anthropic stop_reason onto the V3 unified finish reason. */
 const mapAnthropicStopReason = (raw) => {
     switch (raw) {
@@ -1107,8 +1124,15 @@ export class AnthropicProvider extends BaseProvider {
                             input_schema: schema,
                         },
                     ];
-                    toolChoice = { type: "tool", name: jsonTool };
+                    if (modelSupportsForcedToolChoice(modelId)) {
+                        toolChoice = { type: "tool", name: jsonTool };
+                    }
+                    else {
+                        toolChoice = undefined;
+                        system = appendToolAnswerInstruction(system, jsonTool);
+                    }
                 }
+                toolChoice = relaxForcedToolChoice(modelId, toolChoice);
                 // Additive structured output: when the caller wants a schema AND real
                 // tools, the forced-json path above cannot be used (it replaces the
                 // tools array), and structured output is excluded for this surface
@@ -1458,7 +1482,7 @@ export class AnthropicProvider extends BaseProvider {
             options.disableInternalFallback === true;
         // The native loop bypasses BaseProvider.executeGeneration, so the turn
         // budget has to be composed here or it stops existing for this provider.
-        return this.runGenerateWithModelFallback(() => this.withTurnTimeout({ ...options, tools: mergedTools }, this.getDescriptorGenerateMs(), (timedOptions) => this.executeNativeGenerate(timedOptions)), callerOwnsFallback);
+        return this.runGenerateWithModelFallback(() => this.withTurnTimeout({ ...options, tools: mergedTools }, this.getDescriptorGenerateMs(), (timedOptions) => this.executeNativeGenerate(timedOptions)), callerOwnsFallback, options);
     }
     async executeNativeGenerate(options) {
         const startTime = Date.now();
@@ -1717,7 +1741,7 @@ export class AnthropicProvider extends BaseProvider {
         }
         const modelId = this.modelName || getDefaultAnthropicModel();
         const anthropicToolChoice = shouldUseTools && anthropicTools && anthropicTools.length > 0
-            ? toolChoiceToAnthropic(resolveToolChoice(options, toolsRecord, shouldUseTools))
+            ? relaxForcedToolChoice(modelId, toolChoiceToAnthropic(resolveToolChoice(options, toolsRecord, shouldUseTools)))
             : undefined;
         // Extended thinking: enabled when the caller supplies an explicit token
         // budget (mirrors the previous experimental_thinking gating). Thinking

package/dist/providers/anthropic/structuredOutput.d.ts

@@ -17,8 +17,10 @@ import type Anthropic from "@anthropic-ai/sdk";
  */
 /** Internal tool name — filtered out of every returned tool call / execution. */
 export declare const FINAL_RESULT_TOOL_NAME = "final_result";
+/** The system-prompt instruction to answer through the named tool. */
+export declare function toolAnswerInstruction(toolName: string): string;
 /** Appended to the system prompt whenever the final_result tool is in play. */
-export declare const FINAL_RESULT_INSTRUCTION = "\n\nIMPORTANT: You MUST call the 'final_result' tool to return your response in the required structured format. Do not respond with plain text - always use the final_result tool.";
+export declare const FINAL_RESULT_INSTRUCTION: string;
 /**
  * Build the `final_result` tool definition from a JSON Schema.
  *
@@ -47,6 +49,7 @@ export declare function appendFinalResultTool(tools: Anthropic.Messages.Tool[] |
  * existing one: rewriting a block that carries a `cache_control` marker would
  * change the cached prefix and invalidate the prompt cache on every turn.
  */
+export declare function appendToolAnswerInstruction(system: string | Anthropic.Messages.TextBlockParam[] | undefined, toolName: string): string | Anthropic.Messages.TextBlockParam[];
 export declare function appendFinalResultInstruction(system: string | Anthropic.Messages.TextBlockParam[] | undefined): string | Anthropic.Messages.TextBlockParam[];
 /**
  * Canonical JSON text for a `final_result` payload.

package/dist/providers/anthropic/structuredOutput.js

@@ -19,8 +19,12 @@ import { inlineJsonSchema } from "../../utils/schemaConversion.js";
 /** Internal tool name — filtered out of every returned tool call / execution. */
 export const FINAL_RESULT_TOOL_NAME = "final_result";
 const FINAL_RESULT_TOOL_DESCRIPTION = "Return the final structured result. You MUST call this tool when you have gathered all information and are ready to provide the final answer. The arguments should contain the structured data matching the expected schema.";
+/** The system-prompt instruction to answer through the named tool. */
+export function toolAnswerInstruction(toolName) {
+    return `\n\nIMPORTANT: You MUST call the '${toolName}' tool to return your response in the required structured format. Do not respond with plain text - always use the ${toolName} tool.`;
+}
 /** Appended to the system prompt whenever the final_result tool is in play. */
-export const FINAL_RESULT_INSTRUCTION = "\n\nIMPORTANT: You MUST call the 'final_result' tool to return your response in the required structured format. Do not respond with plain text - always use the final_result tool.";
+export const FINAL_RESULT_INSTRUCTION = toolAnswerInstruction(FINAL_RESULT_TOOL_NAME);
 /**
  * Build the `final_result` tool definition from a JSON Schema.
  *
@@ -71,14 +75,18 @@ export function appendFinalResultTool(tools, jsonSchema) {
  * existing one: rewriting a block that carries a `cache_control` marker would
  * change the cached prefix and invalidate the prompt cache on every turn.
  */
-export function appendFinalResultInstruction(system) {
+export function appendToolAnswerInstruction(system, toolName) {
+    const instruction = toolAnswerInstruction(toolName);
     if (system === undefined) {
-        return FINAL_RESULT_INSTRUCTION.trim();
+        return instruction.trim();
     }
     if (typeof system === "string") {
-        return system + FINAL_RESULT_INSTRUCTION;
+        return system + instruction;
     }
-    return [...system, { type: "text", text: FINAL_RESULT_INSTRUCTION.trim() }];
+    return [...system, { type: "text", text: instruction.trim() }];
+}
+export function appendFinalResultInstruction(system) {
+    return appendToolAnswerInstruction(system, FINAL_RESULT_TOOL_NAME);
 }
 /**
  * Canonical JSON text for a `final_result` payload.

package/dist/providers/catalog/cerebras.json

@@ -9,7 +9,7 @@
     },
     "models": {
         "default": "gpt-oss-120b",
-        "fallbacks": ["gpt-oss-120b", "gemma-4-31b"],
+        "fallbacks": ["gpt-oss-120b"],
         "defaultContextWindow": 65536,
         "defaultMaxOutputTokens": 8192,
         "catalog": {
@@ -29,13 +29,12 @@
                     "input": 0.99,
                     "output": 1.49
                 },
-                "vision": true,
-                "status": "production",
-                "description": "Google Gemma 4 31B — vision-capable"
+                "vision": false,
+                "status": "retired",
+                "description": "Retired 2026-09 — gone from the authenticated /v1/models roster; the nightly vision test has failed on it since 2026-09-10"
             }
         },
-        "visionModel": "gemma-4-31b",
-        "topModels": ["gpt-oss-120b", "gemma-4-31b"]
+        "topModels": ["gpt-oss-120b"]
     },
     "capabilities": {
         "text": true,
@@ -68,8 +67,8 @@
     },
     "evidence": {
         "rosterVerified": {
-            "date": "2026-08-27",
-            "method": "authenticated GET /v1/models"
+            "date": "2026-09-28",
+            "method": "authenticated GET /v1/models — roster is now gpt-oss-120b and qwen-3.8-27b; gemma-4-31b is gone, so it is retired and Cerebras has no verified vision model. qwen-3.8-27b is not added: the account answers 402 payment_required, so no chat call could verify it"
         },
         "authProbe": {
             "date": "2026-08-26",

package/dist/providers/catalog/fireworks.json

@@ -8,12 +8,11 @@
         "baseURL": "https://api.fireworks.ai/inference/v1"
     },
     "models": {
-        "default": "accounts/fireworks/models/kimi-k2p6",
+        "default": "accounts/fireworks/models/kimi-k3",
         "fallbackModelName": "accounts/fireworks/models/gpt-oss-120b",
         "fallbacks": [
-            "accounts/fireworks/models/kimi-k2p6",
-            "accounts/fireworks/models/gpt-oss-120b",
             "accounts/fireworks/models/kimi-k3",
+            "accounts/fireworks/models/gpt-oss-120b",
             "accounts/fireworks/models/qwen3p8-max",
             "accounts/fireworks/models/glm-5p3"
         ],
@@ -22,9 +21,9 @@
         "catalog": {
             "accounts/fireworks/models/kimi-k2p6": {
                 "enumMember": "KIMI_K2P6",
-                "vision": true,
-                "status": "production",
-                "description": "Recommended - Kimi K2.6, Moonshot flagship; vision-capable"
+                "vision": false,
+                "status": "retired",
+                "description": "Retired 2026-09 — still on the Fireworks roster, but a real chat call returns 404 'not deployed' (since the 2026-09-26 nightly)"
             },
             "accounts/fireworks/models/gpt-oss-120b": {
                 "enumMember": "GPT_OSS_120B",
@@ -36,7 +35,7 @@
                 "enumMember": "KIMI_K3",
                 "vision": true,
                 "status": "production",
-                "description": "Kimi K3 — newest Moonshot generation; vision-capable"
+                "description": "Recommended - Kimi K3, newest Moonshot generation; vision-capable"
             },
             "accounts/fireworks/models/qwen3p8-max": {
                 "enumMember": "QWEN3P8_MAX",
@@ -106,9 +105,8 @@
             }
         },
         "topModels": [
-            "accounts/fireworks/models/kimi-k2p6",
-            "accounts/fireworks/models/gpt-oss-120b",
             "accounts/fireworks/models/kimi-k3",
+            "accounts/fireworks/models/gpt-oss-120b",
             "accounts/fireworks/models/qwen3p8-max"
         ]
     },
@@ -143,12 +141,12 @@
     },
     "evidence": {
         "rosterVerified": {
-            "date": "2026-08-30",
+            "date": "2026-09-28",
             "method": "GET /models against the live account, cross-checked with a real chat call per id"
         },
         "liveMatrix": {
-            "date": "2026-08-30",
-            "result": "The previous default, deepseek-v4-pro, is the reason a roster listing is never sufficient evidence here: it is still listed but a real chat call hangs past 120s, so the default moved to kimi-k2p6 (200 OK). Six ids return 404 'not deployed', including all three old vision models. Vision did not disappear with them: re-tested with a valid 64x64 PNG, kimi-k2p6, kimi-k3 and qwen3p8-max all describe the image correctly, while glm-5p3 answers 'does not support image inputs'. gpt-oss-20b is listed but 404s and was therefore not added."
+            "date": "2026-09-28",
+            "result": "kimi-k2p6 is still listed on /models but a real chat call now returns 404 'not deployed', which failed the nightly model-liveness gate from 2026-09-26 and sent the matrix vision test down the text-only fallback. kimi-k3 answers a text call and a call carrying a PNG with 200, so it becomes the default; gpt-oss-120b, qwen3p8-max and glm-5p3 also answer text calls. Earlier (2026-08-30): deepseek-v4-pro was listed but hung past 120s, six ids returned 404 'not deployed', and glm-5p3 answers 'does not support image inputs'."
         },
         "addedInPR": "https://github.com/juspay/neurolink/pull/1587"
     }

package/dist/providers/googleVertex/client.d.ts

@@ -222,9 +222,9 @@ export declare class GoogleVertexProvider extends BaseProvider {
      * clamped below the ceiling rather than passed through — an over-large
      * budget is a 400 from the API.
      *
-     * `useFinalResultTool` (schema / structured-output mode) always pairs with
-     * a forced `tool_choice:{type:"any"}` in the request built right after this
-     * call — Anthropic hard-rejects `thinking` combined with a `tool_choice`
+     * `forcesToolChoice` (schema mode on a model that accepts a forced choice)
+     * pairs with a forced `tool_choice:{type:"any"}` in the request built right
+     * after this call — Anthropic hard-rejects `thinking` combined with a `tool_choice`
      * that forces tool use: a live 400, "Thinking may not be enabled when
      * tool_choice forces tool use." So `thinking` is omitted entirely in that
      * mode, preserving the prior (pre-thinking-fix) behaviour of silently