@juspay/neurolink 12.38.0 → 12.39.0

package/dist/providers/catalog/synthetic.json

@@ -0,0 +1,205 @@
+{
+    "$schema": "./provider-catalog.schema.json",
+    "id": "synthetic",
+    "displayName": "Synthetic",
+    "aliases": [],
+    "tier": 2,
+    "wire": {
+        "baseURL": "https://api.synthetic.new/openai/v1"
+    },
+    "models": {
+        "default": "syn:large:text",
+        "fallbacks": [
+            "syn:small:text",
+            "syn:large:vision",
+            "syn:small:vision",
+            "hf:zai-org/GLM-5.3-Flash",
+            "hf:openai/gpt-oss-120b",
+            "hf:nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4",
+            "hf:deepseek-ai/DeepSeek-V4.1-Flash",
+            "hf:moonshotai/Kimi-K3",
+            "hf:Qwen/Qwen3.8-27B",
+            "hf:zai-org/GLM-4.7-Flash"
+        ],
+        "defaultContextWindow": 524288,
+        "defaultMaxOutputTokens": 65536,
+        "catalog": {
+            "syn:large:text": {
+                "contextWindow": 524288,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.6,
+                    "output": 1.2,
+                    "cachedInput": 0.03
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Recommended - Synthetic's \"large:text\" convenience alias for deepseek-ai/DeepSeek-V4.1-Flash, described by the vendor as \"an efficient near-Fable model with vision.\" Despite the `:text` suffix it accepts image input (input_modalities: text, image, per the live roster). 512K context, 64K max output. Per unauthenticated GET https://api.synthetic.new/openai/v1/models (2026-09-28, HTTP 200): supported_features = tools, json_mode, structured_outputs, reasoning; reasoning_parameters.efforts = none/low/high/xhigh/max. Pricing (converted from the roster's per-token USD rates): $0.60 / $1.20 per MTok input/output, $0.03 per MTok cached input. Also reachable via the underlying weights id hf:deepseek-ai/DeepSeek-V4.1-Flash (same model, same pricing) — see that catalog entry. Not itself capability-probed with a live key; this is a roster+docs verified entry (see evidence.liveMatrix: null)."
+            },
+            "syn:small:text": {
+                "contextWindow": 196608,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.5,
+                    "cachedInput": 0.02
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Synthetic's \"small:text\" convenience alias for zai-org/GLM-4.7-Flash — the smaller, faster sibling in the ~30B-parameter class, text-only input (no image modality on the live roster). Consumes rate limits more slowly per the vendor's own description. 192K context, 64K max output. supported_features = tools, json_mode, structured_outputs, reasoning (live roster GET, 2026-09-28). Pricing: $0.10 / $0.50 per MTok input/output, $0.02 per MTok cached input. Also reachable via hf:zai-org/GLM-4.7-Flash (same model). Roster+docs verified only; no live key probe."
+            },
+            "syn:large:vision": {
+                "contextWindow": 524288,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 3.0,
+                    "output": 15.0,
+                    "cachedInput": 0.45
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Synthetic's \"large:vision\" convenience alias for moonshotai/Kimi-K3, described by the vendor as \"a massive 2.8T-parameter open-source model ... [that] supports vision on uploaded images.\" 512K context, 64K max output. supported_features = tools, json_mode, structured_outputs, reasoning (live roster GET, 2026-09-28). Pricing: $3.00 / $15.00 per MTok input/output, $0.45 per MTok cached input — the most expensive model in this roster. Also reachable via hf:moonshotai/Kimi-K3 (same model). Roster+docs verified only; no live key probe."
+            },
+            "syn:small:vision": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.45,
+                    "output": 2.2,
+                    "cachedInput": 0.09
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Synthetic's \"small:vision\" convenience alias for Qwen/Qwen3.8-27B — \"a small but capable coding and vision model from Qwen\" that consumes rate limits less quickly per the vendor. 256K context, 64K max output. supported_features = tools, json_mode, structured_outputs, reasoning (live roster GET, 2026-09-28). Pricing: $0.45 / $2.20 per MTok input/output, $0.09 per MTok cached input. Also reachable via hf:Qwen/Qwen3.8-27B (same model). Roster+docs verified only; no live key probe."
+            },
+            "hf:zai-org/GLM-5.3-Flash": {
+                "contextWindow": 524288,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.15,
+                    "output": 0.5,
+                    "cachedInput": 0.04
+                },
+                "vision": true,
+                "status": "production",
+                "description": "zai-org GLM-5.3-Flash, described by the vendor as \"an extremely efficient, Opus-4.8-equivalent model with vision.\" No `syn:` short alias — only reachable via this hf: (Hugging Face weights id) form. 512K context, 64K max output. supported_features = tools, json_mode, structured_outputs, reasoning (live roster GET, 2026-09-28). Pricing: $0.15 / $0.50 per MTok input/output, $0.04 per MTok cached input. Roster+docs verified only; no live key probe."
+            },
+            "hf:openai/gpt-oss-120b": {
+                "contextWindow": 131072,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.1,
+                    "cachedInput": 0.02
+                },
+                "vision": false,
+                "status": "production",
+                "description": "OpenAI's open-weight gpt-oss-120b, described by the vendor as \"OpenAI's largest open-source model. A strong reasoning model with great support for agentic tasks.\" Text-only input on the live roster. 128K context, 64K max output. supported_features = tools, json_mode, structured_outputs, reasoning (live roster GET, 2026-09-28). Pricing: $0.10 / $0.10 per MTok input/output, $0.02 per MTok cached input. Roster+docs verified only; no live key probe."
+            },
+            "hf:nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.3,
+                    "output": 1.0,
+                    "cachedInput": 0.06
+                },
+                "vision": false,
+                "status": "production",
+                "description": "NVIDIA Nemotron-3-Super-120B (NVFP4 quant on the roster id; the vendor's own hugging_face_id field reports the FP8 build), described by the vendor as \"a fast, 120b-parameter model with good support for coding agents. Consumes rate limits more slowly.\" Text-only input on the live roster. 256K context, 64K max output. supported_features = tools, json_mode, structured_outputs, reasoning (live roster GET, 2026-09-28). Pricing: $0.30 / $1.00 per MTok input/output, $0.06 per MTok cached input. Roster+docs verified only; no live key probe."
+            },
+            "hf:deepseek-ai/DeepSeek-V4.1-Flash": {
+                "contextWindow": 524288,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.6,
+                    "output": 1.2,
+                    "cachedInput": 0.03
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Underlying weights id for the model also reachable via the syn:large:text convenience alias — same hugging_face_id, same pricing, same specs. Kept as a distinct catalog/fallback entry because both id forms are valid values for `model` on the live roster. See syn:large:text for the fuller description."
+            },
+            "hf:moonshotai/Kimi-K3": {
+                "contextWindow": 524288,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 3.0,
+                    "output": 15.0,
+                    "cachedInput": 0.45
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Underlying weights id for the model also reachable via the syn:large:vision convenience alias — same hugging_face_id, same pricing, same specs. Kept as a distinct catalog/fallback entry because both id forms are valid values for `model` on the live roster. See syn:large:vision for the fuller description."
+            },
+            "hf:Qwen/Qwen3.8-27B": {
+                "contextWindow": 262144,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.45,
+                    "output": 2.2,
+                    "cachedInput": 0.09
+                },
+                "vision": true,
+                "status": "production",
+                "description": "Underlying weights id for the model also reachable via the syn:small:vision convenience alias — same hugging_face_id (Qwen/Qwen3.8-27B-FP8), same pricing, same specs. Kept as a distinct catalog/fallback entry because both id forms are valid values for `model` on the live roster. See syn:small:vision for the fuller description."
+            },
+            "hf:zai-org/GLM-4.7-Flash": {
+                "contextWindow": 196608,
+                "maxOutputTokens": 65536,
+                "pricingPerMTok": {
+                    "input": 0.1,
+                    "output": 0.5,
+                    "cachedInput": 0.02
+                },
+                "vision": false,
+                "status": "production",
+                "description": "Underlying weights id for the model also reachable via the syn:small:text convenience alias — same hugging_face_id, same pricing, same specs. Kept as a distinct catalog/fallback entry because both id forms are valid values for `model` on the live roster. See syn:small:text for the fuller description."
+            }
+        },
+        "topModels": [
+            "syn:large:text",
+            "syn:small:text",
+            "syn:large:vision",
+            "syn:small:vision",
+            "hf:zai-org/GLM-5.3-Flash"
+        ]
+    },
+    "capabilities": {
+        "text": true,
+        "streaming": true,
+        "tools": true,
+        "toolsWithStreaming": true,
+        "structuredOutput": true,
+        "structuredOutputWithTools": false,
+        "embeddings": false,
+        "thinking": true
+    },
+    "errorRules": [],
+    "setup": {
+        "url": "https://synthetic.new",
+        "apiKeyFormat": null,
+        "billingPolicy": "no-free-tier",
+        "instructions": [
+            "1. Visit: https://synthetic.new and choose Sign up (or Log in if you already have an account)",
+            "2. Subscribe to a Pack ($1/day or $30/month, 500 requests/5hr per the pricing page) or set up enterprise usage-based billing — Synthetic has no free tier or keyless trial",
+            "3. Generate an API key from your account dashboard",
+            "4. Set {apiKeyEnvVar} in your .env file"
+        ],
+        "description": "OpenAI-compatible endpoint at https://api.synthetic.new/openai/v1 (confirmed by both the vendor's own docs and LiteLLM's third-party provider page, which lists /chat/completions as the only supported operation), serving 11 model ids covering 7 unique always-on open-source models (DeepSeek V4.1 Flash, Kimi K3, Qwen 3.8 27B, GLM 5.3 Flash, GLM 4.7 Flash, GPT-OSS-120B, Nemotron-3-Super-120B) under one flat-fee or usage-based subscription. No free tier or keyless trial; billing is a $1/day or $30/month pack (500 req/5hr) or enterprise pay-per-token. The pricing page also lists a separate embeddings model (hf:nomic-ai/nomic-embed-text-v1.5) included in the subscription, but NeuroLink's embeddings capability is not currently wired for Tier-2 catalog providers (only OpenAI, Google AI Studio, Google Vertex and Bedrock embed natively), so capabilities.embeddings is false here regardless of vendor support."
+    },
+    "evidence": {
+        "rosterVerified": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://api.synthetic.new/openai/v1/models — 11 models; credential-free, no key used",
+            "status": 200
+        },
+        "billingProbe": {
+            "date": "2026-09-28",
+            "method": "unauthenticated GET https://synthetic.new/pricing — no free tier; $1/day or $30/month subscription pack (500 req/5hr), or enterprise usage-based pricing; credential-free, no key used",
+            "status": 200
+        },
+        "liveMatrix": null,
+        "addedInPR": "https://github.com/juspay/neurolink/pull/1848"
+    }
+}

package/dist/providers/openRouter/client.js

@@ -203,6 +203,7 @@ export class OpenRouterProvider extends OpenAIChatCompletionsProvider {
             "mistralai/devstral",
             "meta-llama/llama-3.3",
             "meta-llama/llama-3.2",
+            "meta-llama/llama-3.1",
             "qwen/qwen3",
             "nvidia/nemotron",
         ];

package/dist/providers/openaiChatCompletionsBase.js

@@ -1648,6 +1648,12 @@ export class OpenAIChatCompletionsProvider extends BaseProvider {
         // middleware short-circuits the request — the case where there is no
         // turn to build on at all.
         let loopConversation;
+        // Cell-3 identity fix: the last model id the SERVER actually echoed on
+        // the wire (via `onModelObserved`, see `runStreamLoop`), distinct from
+        // the pre-call resolved/requested `modelId` below. Stays undefined for
+        // backends that never echo `model` on a chunk, in which case the
+        // `result.model` getter falls back to `modelId` — unchanged behavior.
+        let observedServerModel;
         const providerNameForLoop = this.providerName;
         const streamBaseModel = {
             specificationVersion: "v3",
@@ -1700,6 +1706,9 @@ export class OpenAIChatCompletionsProvider extends BaseProvider {
                     resolveUsage,
                     resolveFinish,
                     responseFormat,
+                    onModelObserved: (model) => {
+                        observedServerModel = model;
+                    },
                 });
                 const completion = loopPromise.then(() => Promise.all([usagePromise, finishPromise]).then(([usage, reason]) => ({
                     type: "finish",
@@ -1930,6 +1939,24 @@ export class OpenAIChatCompletionsProvider extends BaseProvider {
             toolsUsed,
             metadata: streamMetadata,
         };
+        // Cell-3 identity fix: `model` starts as a plain `modelId` data property
+        // above (satisfying StreamResult's type and giving every reader a value
+        // even before the loop starts), then gets redefined as a live getter.
+        // A plain reassignment (`result.model = observedServerModel`) would not
+        // survive `preserveLiveStreamAccessors`' re-wrapping at the
+        // BaseProvider/NeuroLink boundaries — those boundaries spread `result`
+        // into a new object BEFORE the background loop has produced a single
+        // chunk, which freezes a plain field at its construction-time value.
+        // Re-defining as a getter (the same mechanism already used for
+        // `toolExecutions` below) means every boundary's spread re-applies the
+        // getter itself, so a caller reading `.model` after draining `.stream`
+        // sees whatever the server actually echoed, falling back to the
+        // pre-call resolved `modelId` when the backend never echoed one.
+        Object.defineProperty(result, "model", {
+            enumerable: true,
+            configurable: true,
+            get: () => observedServerModel ?? modelId,
+        });
         // Lazy getter: every read transforms the live `toolExecutionSummaries`
         // through the canonical `transformToolExecutions()` so consumers see
         // `{name, input, output, duration}[]` (codebase convention), while still
@@ -1952,7 +1979,7 @@ export class OpenAIChatCompletionsProvider extends BaseProvider {
         return result;
     }
     async runStreamLoop(args) {
-        const { maxSteps, modelId, url, fetchImpl, abortSignal, options, conversation, openAITools, openAIToolChoice, toolsRecord, toolNameFromWire, emitter, toolsUsed, toolExecutionSummaries, pushChunk, closeChannel, resolveUsage, resolveFinish, responseFormat, } = args;
+        const { maxSteps, modelId, url, fetchImpl, abortSignal, options, conversation, openAITools, openAIToolChoice, toolsRecord, toolNameFromWire, emitter, toolsUsed, toolExecutionSummaries, pushChunk, closeChannel, resolveUsage, resolveFinish, responseFormat, onModelObserved, } = args;
         // Hoisted above the try so the catch can resolve the usage accumulated
         // by steps that completed BEFORE the failure — those steps were billed.
         let stepFinish = null;
@@ -2060,6 +2087,15 @@ export class OpenAIChatCompletionsProvider extends BaseProvider {
                 if (stepResult.usage) {
                     stepUsage = mergeUsage(stepUsage, stepResult.usage);
                 }
+                // Cell-3 identity fix: `stepResult.model` is what the SERVER echoed
+                // on the wire (parseSSEStream captures `chunk.model`), which can
+                // differ from the pre-call resolved `modelId` a gateway/router was
+                // asked for. Surfaced via callback rather than the method's return
+                // value so `executeStream` can update the already-returned
+                // `StreamResult` in place (see `onModelObserved` wiring there).
+                if (stepResult.model) {
+                    onModelObserved?.(stepResult.model);
+                }
                 if (stepResult.toolCalls.size === 0) {
                     break;
                 }

package/dist/types/openaiCompatible.d.ts

@@ -255,6 +255,16 @@ export type StreamLoopArgs = {
      * schema was requested, or when tools suppress it.
      */
     responseFormat?: OpenAICompatResponseFormat;
+    /**
+     * Fired once per step that echoes a `model` field on the wire (SSE
+     * `chunk.model`, captured by `parseSSEStream` into `OpenAICompatSSEResult.model`).
+     * Lets the caller learn what the SERVER actually served, distinct from the
+     * pre-call resolved/requested `modelId` — a gateway or router can rewrite
+     * an alias to a concrete model id. Not called when a step's response omits
+     * `model` (some backends don't echo it), so the caller must keep its own
+     * fallback.
+     */
+    onModelObserved?: (model: string) => void;
 };
 export type OpenAICompatBuildBodyArgs = {
     modelId: string;

package/dist/types/providerCatalog.generated.d.ts

@@ -1,2 +1,2 @@
-export type CatalogProviderName = "api-route" | "baseten" | "cerebras" | "cloudflare" | "deepseek" | "fireworks" | "friendli" | "gmicloud" | "groq" | "huggingface" | "inception-labs" | "io-intelligence" | "mancer" | "mistral" | "morph" | "novita" | "perplexity" | "sambanova" | "together-ai" | "upstage" | "xai";
-export type CatalogCredentialKey = "apiRoute" | "baseten" | "cerebras" | "cloudflare" | "deepseek" | "fireworks" | "friendli" | "gmicloud" | "groq" | "huggingFace" | "inceptionLabs" | "ioIntelligence" | "mancer" | "mistral" | "morph" | "novita" | "perplexity" | "sambanova" | "together" | "upstage" | "xai";
+export type CatalogProviderName = "above-dev" | "aiand" | "ambient" | "api-route" | "baseten" | "bee-heossi" | "cerebras" | "charm-hyper" | "chutes" | "cloudflare" | "deepinfra" | "deepseek" | "empiriolabs" | "featherless-ai" | "fireworks" | "friendli" | "gmicloud" | "groq" | "huggingface" | "inception-labs" | "inco" | "inference-net" | "io-intelligence" | "koscompute" | "lilac" | "llmtech" | "mancer" | "mistral" | "moark" | "morph" | "neuralwatt" | "novita" | "ovhcloud" | "pareto-inference" | "perplexity" | "sambanova" | "sarvam" | "synthetic" | "together-ai" | "upstage" | "xai";
+export type CatalogCredentialKey = "aboveDev" | "aiand" | "ambient" | "apiRoute" | "baseten" | "beeHeossi" | "cerebras" | "charmHyper" | "chutes" | "cloudflare" | "deepinfra" | "deepseek" | "empiriolabs" | "featherlessAi" | "fireworks" | "friendli" | "gmicloud" | "groq" | "huggingFace" | "inceptionLabs" | "inco" | "inferenceNet" | "ioIntelligence" | "koscompute" | "lilac" | "llmtech" | "mancer" | "mistral" | "moark" | "morph" | "neuralwatt" | "novita" | "ovhcloud" | "paretoInference" | "perplexity" | "sambanova" | "sarvam" | "synthetic" | "together" | "upstage" | "xai";

package/dist/types/providers.d.ts

@@ -172,6 +172,18 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    aboveDev?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    aiand?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    ambient?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     apiRoute?: {
         apiKey?: string;
         baseURL?: string;
@@ -180,19 +192,43 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    beeHeossi?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     cerebras?: {
         apiKey?: string;
         baseURL?: string;
     };
+    charmHyper?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    chutes?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     cloudflare?: {
         apiKey?: string;
         baseURL?: string;
         accountId?: string;
     };
+    deepinfra?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     deepseek?: {
         apiKey?: string;
         baseURL?: string;
     };
+    empiriolabs?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    featherlessAi?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     fireworks?: {
         apiKey?: string;
         baseURL?: string;
@@ -217,10 +253,30 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    inco?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    inferenceNet?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     ioIntelligence?: {
         apiKey?: string;
         baseURL?: string;
     };
+    koscompute?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    lilac?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    llmtech?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     mancer?: {
         apiKey?: string;
         baseURL?: string;
@@ -229,14 +285,30 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    moark?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     morph?: {
         apiKey?: string;
         baseURL?: string;
     };
+    neuralwatt?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     novita?: {
         apiKey?: string;
         baseURL?: string;
     };
+    ovhcloud?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    paretoInference?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     perplexity?: {
         apiKey?: string;
         baseURL?: string;
@@ -245,6 +317,14 @@ export type NeurolinkCredentials = {
         apiKey?: string;
         baseURL?: string;
     };
+    sarvam?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
+    synthetic?: {
+        apiKey?: string;
+        baseURL?: string;
+    };
     together?: {
         apiKey?: string;
         baseURL?: string;