@juspay/neurolink 12.25.0 → 12.25.2

package/dist/cli/utils/interactiveSetup.js

@@ -207,7 +207,7 @@ export const PROVIDER_CONFIGS = [
     {
         id: AIProviderName.DEEPSEEK,
         name: "DeepSeek",
-        description: "Cost-efficient frontier models (deepseek-chat V3, deepseek-reasoner R1)",
+        description: "Cost-efficient frontier models (deepseek-flash V4.1, deepseek-v4-pro)",
         envVars: [
             {
                 key: "DEEPSEEK_API_KEY",

package/dist/constants/contextWindows.js

@@ -46,11 +46,6 @@ export const DEFAULT_OUTPUT_RESERVE_RATIO = 0.35;
  * The "_default" key is the fallback for unknown models within a provider.
  */
 export const MODEL_CONTEXT_WINDOWS = {
-    deepseek: {
-        _default: 64_000,
-        "deepseek-chat": 64_000,
-        "deepseek-reasoner": 64_000,
-    },
     "nvidia-nim": {
         _default: 128_000,
         "meta/llama-3.3-70b-instruct": 128_000,

package/dist/constants/enums.d.ts

@@ -796,7 +796,9 @@ export declare enum CloudflareModels {
 }
 export declare enum DeepSeekModels {
     DEEPSEEK_CHAT = "deepseek-chat",
-    DEEPSEEK_REASONER = "deepseek-reasoner"
+    DEEPSEEK_REASONER = "deepseek-reasoner",
+    DEEPSEEK_FLASH = "deepseek-flash",
+    DEEPSEEK_V4_PRO = "deepseek-v4-pro"
 }
 export declare enum FireworksModels {
     KIMI_K2P6 = "accounts/fireworks/models/kimi-k2p6",

package/dist/constants/enums.js

@@ -1029,6 +1029,8 @@ export var DeepSeekModels;
 (function (DeepSeekModels) {
     DeepSeekModels["DEEPSEEK_CHAT"] = "deepseek-chat";
     DeepSeekModels["DEEPSEEK_REASONER"] = "deepseek-reasoner";
+    DeepSeekModels["DEEPSEEK_FLASH"] = "deepseek-flash";
+    DeepSeekModels["DEEPSEEK_V4_PRO"] = "deepseek-v4-pro";
 })(DeepSeekModels || (DeepSeekModels = {}));
 export var FireworksModels;
 (function (FireworksModels) {

package/dist/models/manifests/deepseek.js

@@ -5,12 +5,12 @@
  * metadata consolidation plan.
  */
 export const deepseekManifest = {
-    defaultContextWindow: 64000,
+    defaultContextWindow: 1048576,
     models: {
         _default: {
             aliases: [],
-            contextWindow: 64000,
-            maxOutputTokens: 64000,
+            contextWindow: 1048576,
+            maxOutputTokens: 393216,
             vision: false,
             functionCalling: false,
         },

package/dist/providers/catalog/deepseek.json

@@ -10,23 +10,37 @@
     },
     "models": {
         "default": "deepseek-chat",
-        "fallbacks": ["deepseek-chat", "deepseek-reasoner"],
-        "defaultContextWindow": 64000,
-        "defaultMaxOutputTokens": 64000,
+        "fallbacks": ["deepseek-chat", "deepseek-reasoner", "deepseek-v4-pro"],
+        "defaultContextWindow": 1048576,
+        "defaultMaxOutputTokens": 393216,
         "catalog": {
             "deepseek-chat": {
-                "contextWindow": 64000,
-                "maxOutputTokens": 64000,
-                "vision": false,
+                "contextWindow": 1048576,
+                "maxOutputTokens": 393216,
+                "vision": true,
                 "status": "production",
-                "description": "DeepSeek-V3 general chat"
+                "description": "Alias of deepseek-flash with thinking off (DeepSeek-V4.1-Flash)"
             },
             "deepseek-reasoner": {
-                "contextWindow": 64000,
-                "maxOutputTokens": 64000,
+                "contextWindow": 1048576,
+                "maxOutputTokens": 393216,
+                "vision": true,
+                "status": "production",
+                "description": "Alias of deepseek-flash with thinking on (DeepSeek-V4.1-Flash)"
+            },
+            "deepseek-flash": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 393216,
+                "vision": true,
+                "status": "production",
+                "description": "DeepSeek-V4.1-Flash — thinking on by default, image input"
+            },
+            "deepseek-v4-pro": {
+                "contextWindow": 1048576,
+                "maxOutputTokens": 393216,
                 "vision": false,
                 "status": "production",
-                "description": "DeepSeek-R1 reasoning (slower, deeper)"
+                "description": "DeepSeek-V4-Pro-0813 — thinking on by default, text only"
             }
         }
     },
@@ -54,9 +68,9 @@
             "message": "DeepSeek account has insufficient balance. Top up at https://platform.deepseek.com/usage"
         },
         {
-            "pattern": "model_not_found",
+            "pattern": "supported API model names|model_not_found",
             "class": "invalid-model",
-            "message": "DeepSeek model '{model}' not found. Use 'deepseek-chat' or 'deepseek-reasoner'."
+            "message": "DeepSeek model '{model}' not found. Use 'deepseek-flash' or 'deepseek-v4-pro'."
         }
     ],
     "quirks": {
@@ -75,8 +89,8 @@
     },
     "evidence": {
         "rosterVerified": {
-            "date": "2026-09-22",
-            "method": "Ported from already-committed repo data (DeepSeekModels enum, deepseekManifest.ts, deepseek.ts's own error-rule and quirk comments) on origin/release @ a7fd9bf04 — no live vendor probe performed, per this migration's no-vendor-call constraint. billingPolicy is carried over from general knowledge, not independently re-verified live; flagged for confirmation before merge. The capability flags mirror the hand-written live-matrix row this entry replaces (test/helpers/providerMatrix.ts on origin/release), so the migration leaves the set of live-matrix cells unchanged."
+            "date": "2026-09-24",
+            "method": "Live against api.deepseek.com on 2026-09-24. GET /models lists only deepseek-flash and deepseek-v4-pro, with context_window 1048576 and max_output_tokens 393216 (the pricing page's 1M / 384K). deepseek-chat and deepseek-reasoner still answer 200 but report model deepseek-flash — deepseek-chat without reasoning_content, deepseek-reasoner with it — so both stay production as aliases rather than being marked retired. Vision was probed with a solid-blue PNG: deepseek-flash, deepseek-chat and deepseek-reasoner all named the colour; deepseek-v4-pro answered 200 but described a scene that was not in the image, so it keeps vision false, matching the pricing page. An unknown model id returns 400 'The supported API model names are …' with code invalid_request_error — never 'model_not_found' — so the invalid-model rule now matches that text. tools plus response_format json_object in one request returned 200 with tool_calls, confirming structuredOutputWithTools. billingPolicy is not re-verified."
         },
         "liveMatrix": null,
         "addedInPR": "TBD — integration-tested in a worktree before a PR was opened"

package/dist/providers/catalog/huggingface.json

@@ -17,7 +17,7 @@
         "testModel": "meta-llama/Llama-3.1-8B-Instruct",
         "topModels": [
             "meta-llama/Llama-3.3-70B-Instruct",
-            "mistralai/Mistral-Large-3-675B-Instruct-2512",
+            "zai-org/GLM-5",
             "deepseek-ai/DeepSeek-R1",
             "Qwen/Qwen2.5-72B-Instruct",
             "microsoft/phi-4",
@@ -34,20 +34,20 @@
             },
             "meta-llama/Llama-3.2-1B": {
                 "vision": false,
-                "status": "production",
-                "description": "Meta Llama 3.2 — Llama-3.2-1B",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "LLAMA_3_2_1B"
             },
             "meta-llama/Llama-3.2-3B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Meta Llama 3.2 — Llama-3.2-3B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "LLAMA_3_2_3B_INSTRUCT"
             },
             "meta-llama/Llama-3.1-8B": {
                 "vision": false,
-                "status": "production",
-                "description": "Meta Llama 3.1 — Llama-3.1-8B",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "LLAMA_3_1_8B"
             },
             "meta-llama/Llama-3.1-8B-Instruct": {
@@ -58,74 +58,74 @@
             },
             "meta-llama/Llama-3.1-70B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Meta Llama 3.1 — Llama-3.1-70B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "LLAMA_3_1_70B_INSTRUCT"
             },
             "meta-llama/Llama-3.1-405B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Meta Llama 3.1 — Llama-3.1-405B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "LLAMA_3_1_405B_INSTRUCT"
             },
             "meta-llama/Meta-Llama-3-8B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Meta Llama 3.0 — Meta-Llama-3-8B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "LLAMA_3_8B_INSTRUCT"
             },
             "meta-llama/Meta-Llama-3-70B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Meta Llama 3.0 — Meta-Llama-3-70B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "LLAMA_3_70B_INSTRUCT"
             },
             "mistralai/Mistral-Large-3-675B-Instruct-2512": {
                 "vision": false,
-                "status": "production",
-                "description": "Mistral Large",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "MISTRAL_LARGE_3_675B"
             },
             "mistralai/Mistral-Small-3.1-24B-Instruct-2503": {
                 "vision": false,
-                "status": "production",
-                "description": "Mistral Small — Mistral-Small-3.1-24B-Instruct-2503",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "MISTRAL_SMALL_3_1_24B"
             },
             "mistralai/Mistral-Small-24B-Instruct-2501": {
                 "vision": false,
-                "status": "production",
-                "description": "Mistral Small — Mistral-Small-24B-Instruct-2501",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "MISTRAL_SMALL_24B"
             },
             "mistralai/Mistral-7B-Instruct-v0.2": {
                 "vision": false,
-                "status": "production",
-                "description": "Mistral — Mistral-7B-Instruct-v0.2",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "MISTRAL_7B_INSTRUCT"
             },
             "mistralai/Mixtral-8x7B-Instruct-v0.1": {
                 "vision": false,
-                "status": "production",
-                "description": "Mistral — Mixtral-8x7B-Instruct-v0.1",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "MIXTRAL_8X7B_INSTRUCT"
             },
             "mistralai/Devstral-2": {
                 "vision": false,
-                "status": "production",
-                "description": "Mistral Devstral — Devstral-2",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "DEVSTRAL_2"
             },
             "Qwen/Qwen2.5-7B": {
                 "vision": false,
-                "status": "production",
-                "description": "Qwen 2.5 — Qwen2.5-7B",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "QWEN_2_5_7B"
             },
             "Qwen/Qwen2.5-32B": {
                 "vision": false,
-                "status": "production",
-                "description": "Qwen 2.5 — Qwen2.5-32B",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "QWEN_2_5_32B"
             },
             "Qwen/Qwen2.5-72B-Instruct": {
@@ -136,8 +136,8 @@
             },
             "Qwen/Qwen2.5-Coder-7B": {
                 "vision": false,
-                "status": "production",
-                "description": "Qwen 2.5 Coder — Qwen2.5-Coder-7B",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "QWEN_2_5_CODER_7B"
             },
             "Qwen/Qwen2.5-Coder-32B-Instruct": {
@@ -148,14 +148,14 @@
             },
             "Qwen/QwQ-32B": {
                 "vision": false,
-                "status": "production",
-                "description": "Qwen QwQ — QwQ-32B",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "QWQ_32B"
             },
             "Qwen/Qwen2.5-VL-32B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Qwen 2.5 VL (Multimodal) — Qwen2.5-VL-32B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "QWEN_2_5_VL_32B"
             },
             "deepseek-ai/DeepSeek-R1": {
@@ -190,44 +190,44 @@
             },
             "microsoft/Phi-4-reasoning": {
                 "vision": false,
-                "status": "production",
-                "description": "Microsoft Phi — Phi-4-reasoning",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "PHI_4_REASONING"
             },
             "microsoft/Phi-4-mini-instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Microsoft Phi — Phi-4-mini-instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "PHI_4_MINI_INSTRUCT"
             },
             "microsoft/Phi-4-mini-reasoning": {
                 "vision": false,
-                "status": "production",
-                "description": "Microsoft Phi — Phi-4-mini-reasoning",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "PHI_4_MINI_REASONING"
             },
             "microsoft/Phi-3-mini-128k-instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Microsoft Phi — Phi-3-mini-128k-instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "PHI_3_MINI_128K_INSTRUCT"
             },
             "microsoft/Phi-3-vision-128k-instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Microsoft Phi — Phi-3-vision-128k-instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "PHI_3_VISION_128K_INSTRUCT"
             },
             "google/gemma-3-270m": {
                 "vision": false,
-                "status": "production",
-                "description": "Google Gemma 3 — gemma-3-270m",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "GEMMA_3_270M"
             },
             "google/gemma-3-1b-it": {
                 "vision": false,
-                "status": "production",
-                "description": "Google Gemma 3 — gemma-3-1b-it",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "GEMMA_3_1B_IT"
             },
             "google/gemma-3-4b-it": {
@@ -250,62 +250,62 @@
             },
             "google/gemma-2-9b": {
                 "vision": false,
-                "status": "production",
-                "description": "Google Gemma 2 — gemma-2-9b",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "GEMMA_2_9B"
             },
             "google/gemma-2-27b": {
                 "vision": false,
-                "status": "production",
-                "description": "Google Gemma 2 — gemma-2-27b",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "GEMMA_2_27B"
             },
             "google/gemma-2b": {
                 "vision": false,
-                "status": "production",
-                "description": "Google Gemma 1 — gemma-2b",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "GEMMA_2B"
             },
             "google/gemma-7b": {
                 "vision": false,
-                "status": "production",
-                "description": "Google Gemma 1 — gemma-7b",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "GEMMA_7B"
             },
             "tiiuae/falcon-40b-instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Falcon — falcon-40b-instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "FALCON_40B_INSTRUCT"
             },
             "tiiuae/falcon-180B-chat": {
                 "vision": false,
-                "status": "production",
-                "description": "Falcon — falcon-180B-chat",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "FALCON_180B_CHAT"
             },
             "bigcode/starcoder2-15b": {
                 "vision": false,
-                "status": "production",
-                "description": "Code Models — starcoder2-15b",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "STARCODER2_15B"
             },
             "codellama/CodeLlama-34b-Instruct-hf": {
                 "vision": false,
-                "status": "production",
-                "description": "Code Models — CodeLlama-34b-Instruct-hf",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "CODELLAMA_34B_INSTRUCT"
             },
             "bigscience/bloom-7b1": {
                 "vision": false,
-                "status": "production",
-                "description": "BLOOM — bloom-7b1",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "BLOOM_7B1"
             },
             "bigscience/bloom-1b3": {
                 "vision": false,
-                "status": "production",
-                "description": "BLOOM — bloom-1b3",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "BLOOM_1B3"
             },
             "zai-org/GLM-5": {
@@ -322,26 +322,26 @@
             },
             "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16": {
                 "vision": false,
-                "status": "production",
-                "description": "NVIDIA Nemotron 3 Nano — NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "NEMOTRON_3_NANO_30B"
             },
             "HuggingFaceTB/SmolLM3-3B": {
                 "vision": false,
-                "status": "production",
-                "description": "HuggingFace SmolLM3 — SmolLM3-3B",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "SMOLLM3_3B"
             },
             "tiiuae/Falcon3-7B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Falcon 3 Series — Falcon3-7B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "FALCON_3_7B_INSTRUCT"
             },
             "tiiuae/Falcon3-10B-Instruct": {
                 "vision": false,
-                "status": "production",
-                "description": "Falcon 3 Series — Falcon3-10B-Instruct",
+                "status": "retired",
+                "description": "Retired 2026-09 — absent from the router's roster and rejected by a live call",
                 "enumMember": "FALCON_3_10B_INSTRUCT"
             }
         }
@@ -370,12 +370,12 @@
         {
             "pattern": "model.*not found|not found.*model",
             "class": "invalid-model",
-            "message": "HuggingFace model '{model}' not found.\n\nSuggestions:\n1. Check model name spelling\n2. Ensure model exists on HuggingFace Hub\n3. For tool calling, use: Llama-3.1-8B-Instruct, Hermes-3-Llama-3.2-3B, or CodeLlama-34b-Instruct-hf"
+            "message": "HuggingFace model '{model}' not found.\n\nSuggestions:\n1. Check model name spelling\n2. Ensure an Inference Provider serves the model (https://huggingface.co/inference/models)\n3. For tool calling, use: meta-llama/Llama-3.3-70B-Instruct, zai-org/GLM-5 or Qwen/Qwen3.5-397B-A17B"
         },
         {
             "pattern": "function|tool",
             "class": "provider",
-            "message": "HuggingFace tool calling error: {message}\n\nNotes:\n1. Ensure you're using a tool-capable model (Llama-3.1+, Hermes-3+, CodeLlama)\n2. Check that your model supports function calling\n3. Verify tool schema format is correct"
+            "message": "HuggingFace tool calling error: {message}\n\nNotes:\n1. Ensure you're using a tool-capable model, e.g. meta-llama/Llama-3.3-70B-Instruct, zai-org/GLM-5 or Qwen/Qwen3.5-397B-A17B\n2. Check that your model supports function calling\n3. Verify tool schema format is correct"
         }
     ],
     "timeouts": {
@@ -385,8 +385,8 @@
     "autoSelectPriority": 10,
     "setup": {
         "url": "https://huggingface.co/settings/tokens",
-        "apiKeyFormat": "^hf_[A-Za-z0-9]{37}$",
-        "billingPolicy": "free-with-card",
+        "apiKeyFormat": "^hf_[A-Za-z0-9]{34,}$",
+        "billingPolicy": "free-tier",
         "instructions": [
             "1. Visit: https://huggingface.co/settings/tokens",
             "2. Create or sign in to your HuggingFace account",
@@ -396,8 +396,8 @@
     },
     "evidence": {
         "rosterVerified": {
-            "date": "2026-09-22",
-            "method": "Ported from already-committed repo data (HuggingFaceModels enum, huggingfaceManifest.ts, the hand ProviderDescriptor, and huggingFace/client.ts's own error-rule/model comments) on this worktree's current state — no live vendor probe performed, per this migration's no-vendor-call constraint. models.catalog carries every member of the HuggingFaceModels enum it replaces, with the same enumMember names, plus the Llama-3.1-8B-Instruct fallback; topModels restores the previous picker order and text. Per-model vision is false to mirror huggingfaceManifest, which declares no HuggingFace vision support — it is not a vendor capability claim; billingPolicy is carried over from general knowledge, not independently re-verified live. testModel and the capability flags mirror the hand-written live-matrix row this entry replaces (test/helpers/providerMatrix.ts on origin/release), so the migration leaves the set of live-matrix cells unchanged."
+            "date": "2026-09-24",
+            "method": "Live against router.huggingface.co on 2026-09-24. GET /v1/models returned 138 models; each of the 53 catalog ids was also sent a real chat call. 39 ids are absent from the roster AND rejected with 400 ('does not exist' / 'not supported by any provider'), so they are marked retired and keep their enumMember; the other 14 are on the roster and answered 402 (the test account's monthly credits are spent), which is a billing block, not a rejection, so they stay production. No id had the two signals disagree. topModels swaps the retired Mistral-Large-3 for zai-org/GLM-5, which is on the roster. billingPolicy is free-tier per https://huggingface.co/docs/inference-providers/pricing (free users get monthly credits, no card). apiKeyFormat accepts the hf_ + 34 characters a real token has; the previous hf_ + 37 pattern rejected it. structuredOutput, toolsWithStreaming and per-model vision are left unchanged: /v1/models reports them per serving provider, and the 402 prevented a live round trip that would prove them."
         },
         "liveMatrix": null,
         "addedInPR": "TBD — integration-tested in a worktree before a PR was opened"

package/dist/server/routes/claudeProxyRoutes.js

@@ -289,8 +289,10 @@ function drainAccountAdmissionWaiters(accountKey, state) {
  * removed the waiter's cap — and uncapped arrivals count toward `active`, so
  * that waiter would otherwise never drain. An uncapped request from an older
  * snapshot (one that predates a reload adding the cap) or from the same one
- * leaves waiters queued: their cap is still current. With no generation there
- * is no runtime config store, the cap cannot change, and nothing is admitted.
+ * leaves waiters queued: their cap is still current. With no generation,
+ * which only direct callers such as test hooks pass (the route passes 0 from
+ * its fallback snapshot when no runtime config store is attached), there is
+ * nothing to compare against and nothing is admitted.
  */
 function admitWaitersOfRemovedCap(accountKey, state, generation) {
     if (generation === undefined) {
@@ -7068,8 +7070,8 @@ async function handleAnthropicRoutedClaudeRequest(args) {
         });
     }
     const { accounts, enabledAccounts, orderedAccounts, bodyStr, requestStart, toolCount, url, clientHeaders, isClaudeClientRequest, sessionId, } = loadedAccounts;
-    // Accounts this request sent an upstream attempt to; one skipped only for
-    // its admission cap is not among them.
+    // Accounts this request sent an upstream attempt to or whose token refresh
+    // failed; one skipped only for its admission cap is not among them.
     const attemptedAccountKeys = new Set();
     // Only fill-first over more than one account reads a binding (see
     // selectClaudeProxyAccountOrder), so nothing else records one.
@@ -7233,6 +7235,9 @@ async function handleAnthropicRoutedClaudeRequest(args) {
                     !preparedAttempt.finalBodyStr ||
                     !preparedAttempt.preparedContext ||
                     preparedAttempt.fetchStartMs === undefined) {
+                    // A failed token refresh is this account failing the request, so a
+                    // session bound to it must not stay there when another one serves.
+                    attemptedAccountKeys.add(account.key);
                     if (transportPermit.probe) {
                         providerTransportCoordinator.reportProbeAbandoned(transportPermit);
                     }

package/dist/types/proxy.d.ts

@@ -1143,8 +1143,10 @@ export type QueuedAccountAdmission = {
 /** One queued request waiting for per-account admission capacity. */
 export type AccountAdmissionWaiter = {
     capacity: number;
-    /** Runtime-config generation of the snapshot the request queued under;
-     *  absent when no runtime config store publishes one. */
+    /** Runtime-config generation of the snapshot the request queued under.
+     *  The route always passes one (0 from its fallback snapshot when no
+     *  runtime config store is attached); only direct callers such as test
+     *  hooks leave it absent. */
     generation?: number;
     resolve: (lease: AccountAdmissionLease) => void;
 };