From 80e0bc2dc6861222781f3dc87aa5ab72786dae8f Mon Sep 17 00:00:00 2001
From: ryan-crabbe-berri <ryan@berri.ai>
Date: Sat, 29 Aug 2026 10:58:10 -0700
Subject: [PATCH 01/19] docs(claude.md): require tests to check behavior, not
 code structure

Claude-Session: https://claude.ai/code/session_017nZW6omb93ZuAfqqCzKSU5
---
 CLAUDE.md | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/CLAUDE.md b/CLAUDE.md
index 930825aeb89..d9e9e8f1586 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -23,6 +23,8 @@ When adding new features, add meaningful tests. Don't add tests that don't check
 
 Same thing for bug fixes. The tests should make it so that this specific bug can never happen again without failing tests (i.e., regression)
 
+Never test structure of code only function of it
+
 `tests/test_litellm/` mirrors `litellm/` in a parallel path (see `tests/test_litellm/readme.md`). Name tests `test_<filename>.py`, but always match the existing test file in the directory you touch — many provider dirs use longer descriptive names (e.g. `test_anthropic_chat_transformation.py`) to avoid ambiguity across sibling folders. For bug fixes, extend the existing mapped test file rather than creating a new one. Only create a new test file for a new feature (provider, endpoint, or transformation module) that has no mapped test yet, following that directory's naming convention (or `test_<filename>.py` if you're the first test there). One focused regression test beats many shallow ones
 
 End-to-end tests belong in `tests/e2e/` and must follow the harness conventions documented in that directory's `CLAUDE.md`

From 5024c4d52052e7a04eebb4b33d6615e71ee1cb0c Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Sat, 29 Aug 2026 20:56:21 +0000
Subject: [PATCH 02/19] fix: update stale source URLs in model cost map

119 entries pointed at 404ing or permanently-moved pages (Pylon #7777).
Replaced with verified working equivalents (200-checked or permanent
redirect targets).

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 ...odel_prices_and_context_window_backup.json | 238 +++++++++---------
 model_prices_and_context_window.json          | 238 +++++++++---------
 2 files changed, 238 insertions(+), 238 deletions(-)

diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json
index ec175025b42..b548de07452 100644
--- a/litellm/model_prices_and_context_window_backup.json
+++ b/litellm/model_prices_and_context_window_backup.json
@@ -3726,7 +3726,7 @@
         "output_cost_per_token": 0,
         "litellm_provider": "azure_ai",
         "mode": "chat",
-        "source": "https://azure.microsoft.com/en-us/pricing/details/ai-services/",
+        "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/aoai/",
         "comment": "Flat cost of $0.14 per M input tokens for Azure AI Foundry Model Router infrastructure. Use pattern: azure_ai/model_router/<deployment-name> where deployment-name is your Azure deployment (e.g., azure-model-router)"
     },
     "azure/eu/gpt-4o-2024-08-06": {
@@ -5328,7 +5328,7 @@
         "input_cost_per_second": 0.0002833333333333333,
         "litellm_provider": "azure",
         "mode": "audio_transcription",
-        "source": "https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/gpt-realtime-whisper",
+        "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/gpt-realtime-whisper",
         "supported_endpoints": [
             "/v1/realtime",
             "/v1/realtime/transcription_sessions"
@@ -9107,7 +9107,7 @@
         "mode": "embedding",
         "output_cost_per_token": 0.0,
         "output_vector_size": 1024,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
         "supports_embedding_image_input": true
     },
     "azure_ai/Cohere-embed-v3-multilingual": {
@@ -9118,7 +9118,7 @@
         "mode": "embedding",
         "output_cost_per_token": 0.0,
         "output_vector_size": 1024,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
         "supports_embedding_image_input": true
     },
     "azure_ai/FLUX-1.1-pro": {
@@ -9134,7 +9134,7 @@
         "litellm_provider": "azure_ai",
         "mode": "image_generation",
         "output_cost_per_image": 0.04,
-        "source": "https://azuremarketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
         "supported_endpoints": [
             "/v1/images/generations"
         ]
@@ -9467,7 +9467,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 3.7e-07,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-11b-vision-instruct-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-11b-vision-instruct-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true,
         "supports_vision": true
@@ -9481,7 +9481,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 2.04e-06,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-90b-vision-instruct-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-90b-vision-instruct-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true,
         "supports_vision": true
@@ -9494,7 +9494,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 7.1e-07,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/metagenai.llama-3-3-70b-instruct-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/metagenai.llama-3-3-70b-instruct-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -9543,7 +9543,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 1.6e-05,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-405b-instruct-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-405b-instruct-offer?tab=PlansAndPrice",
         "supports_tool_choice": true
     },
     "azure_ai/Meta-Llama-3.1-70B-Instruct": {
@@ -9554,7 +9554,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 3.54e-06,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-70b-instruct-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-70b-instruct-offer?tab=PlansAndPrice",
         "supports_tool_choice": true
     },
     "azure_ai/Meta-Llama-3.1-8B-Instruct": {
@@ -9566,7 +9566,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 6.1e-07,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-8b-instruct-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-8b-instruct-offer?tab=PlansAndPrice",
         "supports_tool_choice": true
     },
     "azure_ai/Phi-3-medium-128k-instruct": {
@@ -9756,7 +9756,7 @@
         "supported_endpoints": [
             "/v1/ocr"
         ],
-        "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/"
+        "source": "https://ai.azure.com/catalog/models/mistral-document-ai-2512"
     },
     "azure_ai/doc-intelligence/prebuilt-read": {
         "litellm_provider": "azure_ai",
@@ -9967,7 +9967,7 @@
         "mode": "embedding",
         "output_cost_per_token": 0.0,
         "output_vector_size": 3072,
-        "source": "https://azuremarketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
         "supported_endpoints": [
             "/v1/embeddings"
         ],
@@ -10151,7 +10151,7 @@
         "max_tokens": 8192,
         "mode": "chat",
         "output_cost_per_token": 0.00971,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/ai-foundry/models/jais-30b-chat"
+        "source": "https://ai.azure.com/catalog/models/jais-30b-chat"
     },
     "azure_ai/jamba-instruct": {
         "input_cost_per_token": 5e-07,
@@ -10208,7 +10208,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 4e-08,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.ministral-3b-2410-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.ministral-3b-2410-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -10231,7 +10231,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -10243,7 +10243,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -10280,7 +10280,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 1.5e-07,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.mistral-nemo-12b-2407?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-nemo-12b-2407?tab=PlansAndPrice",
         "supports_function_calling": true
     },
     "azure_ai/mistral-small": {
@@ -22796,7 +22796,7 @@
         "output_cost_per_reasoning_token": 3e-06,
         "output_cost_per_token": 3e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -22850,7 +22850,7 @@
         "output_cost_per_reasoning_token": 9e-06,
         "output_cost_per_token": 9e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -22915,7 +22915,7 @@
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -22974,7 +22974,7 @@
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23178,7 +23178,7 @@
         "mode": "chat",
         "output_cost_per_reasoning_token": 3e-06,
         "output_cost_per_token": 3e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23262,7 +23262,7 @@
         "mode": "chat",
         "output_cost_per_reasoning_token": 9e-06,
         "output_cost_per_token": 9e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23325,7 +23325,7 @@
         "output_cost_per_token": 3.75e-06,
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23382,7 +23382,7 @@
         "output_cost_per_token": 3.75e-06,
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -26982,7 +26982,7 @@
         "supports_tool_choice": true,
         "supports_vision": true,
         "supports_web_search": true,
-        "source": "https://platform.openai.com/docs/models/gpt-5.6-cyber",
+        "source": "https://developers.openai.com/api/docs/models/gpt-5.6-cyber",
         "supports_computer_use": true,
         "supports_parallel_function_calling": true
     },
@@ -27021,7 +27021,7 @@
         "supports_tool_choice": true,
         "supports_vision": true,
         "supports_web_search": true,
-        "source": "https://platform.openai.com/docs/models/daybreak-red-latest",
+        "source": "https://developers.openai.com/api/docs/models/daybreak-red-latest",
         "supports_computer_use": true,
         "supports_parallel_function_calling": true
     },
@@ -27061,7 +27061,7 @@
         "supports_tool_choice": true,
         "supports_vision": true,
         "supports_web_search": true,
-        "source": "https://platform.openai.com/docs/models/daybreak-blue-latest",
+        "source": "https://developers.openai.com/api/docs/models/daybreak-blue-latest",
         "supports_parallel_function_calling": true
     },
     "chat-latest": {
@@ -27073,7 +27073,7 @@
         "max_tokens": 128000,
         "mode": "chat",
         "output_cost_per_token": 3e-05,
-        "source": "https://platform.openai.com/docs/models/chat-latest",
+        "source": "https://developers.openai.com/api/docs/models/chat-latest",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses"
@@ -30368,7 +30368,7 @@
             "search_context_size_low": 0.0025,
             "search_context_size_medium": 0.0025
         },
-        "source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
+        "source": "https://ai.developer.meta.com/docs/pricing-rate-limits",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses",
@@ -30409,7 +30409,7 @@
             "search_context_size_low": 0.0025,
             "search_context_size_medium": 0.0025
         },
-        "source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
+        "source": "https://ai.developer.meta.com/docs/pricing-rate-limits",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses",
@@ -30450,7 +30450,7 @@
             "search_context_size_low": 0.0025,
             "search_context_size_medium": 0.0025
         },
-        "source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
+        "source": "https://ai.developer.meta.com/docs/pricing-rate-limits",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses",
@@ -30483,7 +30483,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text"
         ],
@@ -30499,7 +30499,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text"
         ],
@@ -30515,7 +30515,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text",
             "image"
@@ -30532,7 +30532,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text",
             "image"
@@ -32477,7 +32477,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_reasoning": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-R1-0528": {
         "max_tokens": 164000,
@@ -32489,7 +32489,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_reasoning": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-R1-Distill-Llama-70B": {
         "max_tokens": 128000,
@@ -32500,7 +32500,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-V3": {
         "max_tokens": 128000,
@@ -32511,7 +32511,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-V3-0324": {
         "max_tokens": 128000,
@@ -32522,7 +32522,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/google/gemma-3-27b-it": {
         "max_tokens": 128000,
@@ -32534,7 +32534,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Llama-3.3-70B-Instruct": {
         "max_tokens": 128000,
@@ -32545,7 +32545,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Llama-Guard-3-8B": {
         "max_tokens": 128000,
@@ -32555,7 +32555,7 @@
         "output_cost_per_token": 6e-08,
         "litellm_provider": "nebius",
         "mode": "chat",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Meta-Llama-3.1-8B-Instruct": {
         "max_tokens": 128000,
@@ -32566,7 +32566,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Meta-Llama-3.1-70B-Instruct": {
         "max_tokens": 128000,
@@ -32577,7 +32577,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Meta-Llama-3.1-405B-Instruct": {
         "max_tokens": 128000,
@@ -32588,7 +32588,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/mistralai/Mistral-Nemo-Instruct-2407": {
         "max_tokens": 128000,
@@ -32599,7 +32599,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/NousResearch/Hermes-3-Llama-3.1-405B": {
         "max_tokens": 128000,
@@ -32610,7 +32610,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/nvidia/Llama-3.1-Nemotron-Ultra-253B-v1": {
         "max_tokens": 128000,
@@ -32621,7 +32621,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/nvidia/Llama-3.3-Nemotron-Super-49B-v1": {
         "max_tokens": 131072,
@@ -32632,7 +32632,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-235B-A22B": {
         "max_tokens": 262144,
@@ -32643,7 +32643,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-32B": {
         "max_tokens": 32768,
@@ -32654,7 +32654,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-30B-A3B": {
         "max_tokens": 32768,
@@ -32665,7 +32665,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-14B": {
         "max_tokens": 32768,
@@ -32676,7 +32676,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-4B": {
         "max_tokens": 32768,
@@ -32687,7 +32687,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/QwQ-32B": {
         "max_tokens": 32768,
@@ -32699,7 +32699,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_reasoning": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-72B-Instruct": {
         "max_tokens": 128000,
@@ -32710,7 +32710,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-32B-Instruct": {
         "max_tokens": 128000,
@@ -32721,7 +32721,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-Coder-7B": {
         "max_tokens": 32768,
@@ -32732,7 +32732,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-VL-72B-Instruct": {
         "max_tokens": 131072,
@@ -32744,7 +32744,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2-VL-72B-Instruct": {
         "max_tokens": 131072,
@@ -32756,7 +32756,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2-VL-7B-Instruct": {
         "max_tokens": 131072,
@@ -32767,7 +32767,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/BAAI/bge-en-icl": {
         "max_tokens": 32768,
@@ -32776,7 +32776,7 @@
         "output_cost_per_token": 0.0,
         "litellm_provider": "nebius",
         "mode": "embedding",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/BAAI/bge-multilingual-gemma2": {
         "max_tokens": 8192,
@@ -32785,7 +32785,7 @@
         "output_cost_per_token": 0.0,
         "litellm_provider": "nebius",
         "mode": "embedding",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/intfloat/e5-mistral-7b-instruct": {
         "max_tokens": 32768,
@@ -32794,7 +32794,7 @@
         "output_cost_per_token": 0.0,
         "litellm_provider": "nebius",
         "mode": "embedding",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nvidia.nemotron-nano-12b-v2": {
         "input_cost_per_token": 2e-07,
@@ -33535,7 +33535,7 @@
         "max_tokens": 4000,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33548,7 +33548,7 @@
         "max_tokens": 4000,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33561,7 +33561,7 @@
         "max_tokens": 4000,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33618,7 +33618,7 @@
         "max_tokens": 8192,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true,
@@ -33632,7 +33632,7 @@
         "max_tokens": 8192,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": false,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33643,7 +33643,7 @@
         "max_input_tokens": 512,
         "mode": "embedding",
         "output_vector_size": 1024,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_vision": true
     },
     "oci/cohere.command-a-reasoning-08-2025": {
@@ -34540,7 +34540,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 2e-07,
-        "source": "https://openrouter.ai/api/v1/models/bytedance/ui-tars-1.5-7b",
+        "source": "https://openrouter.ai/bytedance/ui-tars-1.5-7b",
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-chat": {
@@ -34775,7 +34775,7 @@
         "output_cost_per_reasoning_token": 3e-06,
         "output_cost_per_token": 3e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -34816,7 +34816,7 @@
         "output_cost_per_reasoning_token": 1.5e-06,
         "output_cost_per_token": 1.5e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -35901,7 +35901,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 6.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/deepseek-r1-distill-llama-70b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -35915,7 +35915,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 1e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/llama-3-1-8b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35928,7 +35928,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 6.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/meta-llama-3-1-70b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": false,
         "supports_tool_choice": false
@@ -35941,7 +35941,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 6.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/meta-llama-3-3-70b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35954,7 +35954,7 @@
         "max_tokens": 127000,
         "mode": "chat",
         "output_cost_per_token": 1e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mistral-7b-instruct-v0-3",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35967,7 +35967,7 @@
         "max_tokens": 118000,
         "mode": "chat",
         "output_cost_per_token": 1.3e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mistral-nemo-instruct-2407",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35980,7 +35980,7 @@
         "max_tokens": 128000,
         "mode": "chat",
         "output_cost_per_token": 2.8e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mistral-small-3-2-24b-instruct-2506",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true,
@@ -35994,7 +35994,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 6.3e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mixtral-8x7b-instruct-v0-1",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false
@@ -36007,7 +36007,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 8.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/qwen2-5-coder-32b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false
@@ -36020,7 +36020,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 9.1e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/qwen2-5-vl-72b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false,
@@ -36034,7 +36034,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 2.3e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/qwen3-32b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -36048,7 +36048,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 4e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/gpt-oss-120b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -36062,7 +36062,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 1.5e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/gpt-oss-20b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -36076,7 +36076,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 2.9e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/llava-next-mistral-7b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false,
@@ -36090,7 +36090,7 @@
         "max_tokens": 256000,
         "mode": "chat",
         "output_cost_per_token": 1.9e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mamba-codestral-7b-v0-1",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false
@@ -38856,7 +38856,7 @@
         "max_input_tokens": 262144,
         "mode": "chat",
         "output_cost_per_token": 3.6e-06,
-        "source": "https://www.together.ai/models/Qwen/Qwen3.5-397B-A17B",
+        "source": "https://www.together.ai/models/qwen3-5-397b-a17b",
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
         "supports_prompt_caching": true,
@@ -43097,7 +43097,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 5e-07,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true,
@@ -43113,7 +43113,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 5e-07,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -43130,7 +43130,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true,
@@ -43146,7 +43146,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -43266,7 +43266,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.4,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -43280,7 +43280,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.15,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -43295,7 +43295,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.4,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -43310,7 +43310,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.15,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -44946,7 +44946,7 @@
         "litellm_provider": "azure",
         "mode": "video_generation",
         "output_cost_per_video_per_second": 0.1,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/video-generation",
+        "source": "https://ai.azure.com/catalog/models/sora-2",
         "supported_modalities": [
             "text"
         ],
@@ -44958,7 +44958,7 @@
         "litellm_provider": "azure",
         "mode": "video_generation",
         "output_cost_per_video_per_second": 0.3,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/video-generation",
+        "source": "https://ai.azure.com/catalog/models/sora-2-pro",
         "supported_modalities": [
             "text"
         ],
@@ -44970,7 +44970,7 @@
         "litellm_provider": "azure",
         "mode": "video_generation",
         "output_cost_per_video_per_second": 0.5,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/video-generation",
+        "source": "https://ai.azure.com/catalog/models/sora-2-pro",
         "supported_modalities": [
             "text"
         ],
@@ -49180,7 +49180,7 @@
         "input_cost_per_second": 0.0002833333333333333,
         "litellm_provider": "openai",
         "mode": "audio_transcription",
-        "source": "https://platform.openai.com/docs/models/gpt-realtime-whisper",
+        "source": "https://developers.openai.com/api/docs/models/gpt-realtime-whisper",
         "supported_endpoints": [
             "/v1/realtime",
             "/v1/realtime/transcription_sessions"
@@ -50432,7 +50432,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -50470,7 +50470,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -50508,7 +50508,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -50546,7 +50546,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -51295,7 +51295,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/qwen3.6-35b-a3b": {
         "max_tokens": 131072,
@@ -51308,7 +51308,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/qwen3-30b-a3b": {
         "max_tokens": 131072,
@@ -51321,7 +51321,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/qwen3-coder-30b-a3b": {
         "max_tokens": 131072,
@@ -51334,7 +51334,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": false,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/deepseek-v4-flash": {
         "max_tokens": 163840,
@@ -51347,7 +51347,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/minimax-m2.7": {
         "max_tokens": 1000192,
@@ -51360,7 +51360,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": false,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "darkbloom/gemma-4-26b": {
         "input_cost_per_token": 3e-08,
@@ -51464,7 +51464,7 @@
         "input_cost_per_second": 7.5e-05,
         "litellm_provider": "openai",
         "mode": "audio_transcription",
-        "source": "https://platform.openai.com/docs/models/gpt-transcribe",
+        "source": "https://developers.openai.com/api/docs/models/gpt-transcribe",
         "supported_endpoints": [
             "/v1/audio/transcriptions",
             "/v1/realtime/transcription_sessions"
@@ -51482,7 +51482,7 @@
         "input_cost_per_second": 0.0002833333333333333,
         "litellm_provider": "openai",
         "mode": "audio_transcription",
-        "source": "https://platform.openai.com/docs/models/gpt-live-transcribe",
+        "source": "https://developers.openai.com/api/docs/models/gpt-live-transcribe",
         "supported_endpoints": [
             "/v1/realtime",
             "/v1/realtime/transcription_sessions"
@@ -51503,7 +51503,7 @@
         "max_output_tokens": 2000,
         "max_tokens": 2000,
         "mode": "realtime",
-        "source": "https://platform.openai.com/docs/models/gpt-realtime-translate",
+        "source": "https://developers.openai.com/api/docs/models/gpt-realtime-translate",
         "supported_modalities": [
             "audio"
         ],
@@ -51531,7 +51531,7 @@
             "search_context_size_low": 0.01,
             "search_context_size_medium": 0.01
         },
-        "source": "https://docs.claude.com/en/docs/about-claude/models/overview",
+        "source": "https://platform.claude.com/docs/en/about-claude/models/overview",
         "supports_adaptive_thinking": true,
         "thinking_always_on": true,
         "supports_mid_conversation_system": true,
@@ -51570,7 +51570,7 @@
             "search_context_size_low": 0.01,
             "search_context_size_medium": 0.01
         },
-        "source": "https://docs.claude.com/en/docs/about-claude/models/overview",
+        "source": "https://platform.claude.com/docs/en/about-claude/models/overview",
         "supports_adaptive_thinking": true,
         "thinking_always_on": true,
         "supports_assistant_prefill": false,
diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json
index ec175025b42..b548de07452 100644
--- a/model_prices_and_context_window.json
+++ b/model_prices_and_context_window.json
@@ -3726,7 +3726,7 @@
         "output_cost_per_token": 0,
         "litellm_provider": "azure_ai",
         "mode": "chat",
-        "source": "https://azure.microsoft.com/en-us/pricing/details/ai-services/",
+        "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/aoai/",
         "comment": "Flat cost of $0.14 per M input tokens for Azure AI Foundry Model Router infrastructure. Use pattern: azure_ai/model_router/<deployment-name> where deployment-name is your Azure deployment (e.g., azure-model-router)"
     },
     "azure/eu/gpt-4o-2024-08-06": {
@@ -5328,7 +5328,7 @@
         "input_cost_per_second": 0.0002833333333333333,
         "litellm_provider": "azure",
         "mode": "audio_transcription",
-        "source": "https://learn.microsoft.com/en-us/azure/ai-foundry/openai/concepts/gpt-realtime-whisper",
+        "source": "https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/gpt-realtime-whisper",
         "supported_endpoints": [
             "/v1/realtime",
             "/v1/realtime/transcription_sessions"
@@ -9107,7 +9107,7 @@
         "mode": "embedding",
         "output_cost_per_token": 0.0,
         "output_vector_size": 1024,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
         "supports_embedding_image_input": true
     },
     "azure_ai/Cohere-embed-v3-multilingual": {
@@ -9118,7 +9118,7 @@
         "mode": "embedding",
         "output_cost_per_token": 0.0,
         "output_vector_size": 1024,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/cohere.cohere-embed-v3-english-offer?tab=PlansAndPrice",
         "supports_embedding_image_input": true
     },
     "azure_ai/FLUX-1.1-pro": {
@@ -9134,7 +9134,7 @@
         "litellm_provider": "azure_ai",
         "mode": "image_generation",
         "output_cost_per_image": 0.04,
-        "source": "https://azuremarketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
         "supported_endpoints": [
             "/v1/images/generations"
         ]
@@ -9467,7 +9467,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 3.7e-07,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-11b-vision-instruct-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-11b-vision-instruct-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true,
         "supports_vision": true
@@ -9481,7 +9481,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 2.04e-06,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-90b-vision-instruct-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/metagenai.meta-llama-3-2-90b-vision-instruct-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true,
         "supports_vision": true
@@ -9494,7 +9494,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 7.1e-07,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/metagenai.llama-3-3-70b-instruct-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/metagenai.llama-3-3-70b-instruct-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -9543,7 +9543,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 1.6e-05,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-405b-instruct-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-405b-instruct-offer?tab=PlansAndPrice",
         "supports_tool_choice": true
     },
     "azure_ai/Meta-Llama-3.1-70B-Instruct": {
@@ -9554,7 +9554,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 3.54e-06,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-70b-instruct-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-70b-instruct-offer?tab=PlansAndPrice",
         "supports_tool_choice": true
     },
     "azure_ai/Meta-Llama-3.1-8B-Instruct": {
@@ -9566,7 +9566,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 6.1e-07,
-        "source": "https://azuremarketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-8b-instruct-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en-us/marketplace/apps/metagenai.meta-llama-3-1-8b-instruct-offer?tab=PlansAndPrice",
         "supports_tool_choice": true
     },
     "azure_ai/Phi-3-medium-128k-instruct": {
@@ -9756,7 +9756,7 @@
         "supported_endpoints": [
             "/v1/ocr"
         ],
-        "source": "https://azure.microsoft.com/en-us/pricing/details/ai-foundry-models/"
+        "source": "https://ai.azure.com/catalog/models/mistral-document-ai-2512"
     },
     "azure_ai/doc-intelligence/prebuilt-read": {
         "litellm_provider": "azure_ai",
@@ -9967,7 +9967,7 @@
         "mode": "embedding",
         "output_cost_per_token": 0.0,
         "output_vector_size": 3072,
-        "source": "https://azuremarketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/pt-br/marketplace/apps/cohere.cohere-embed-4-offer?tab=PlansAndPrice",
         "supported_endpoints": [
             "/v1/embeddings"
         ],
@@ -10151,7 +10151,7 @@
         "max_tokens": 8192,
         "mode": "chat",
         "output_cost_per_token": 0.00971,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/ai-foundry/models/jais-30b-chat"
+        "source": "https://ai.azure.com/catalog/models/jais-30b-chat"
     },
     "azure_ai/jamba-instruct": {
         "input_cost_per_token": 5e-07,
@@ -10208,7 +10208,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 4e-08,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.ministral-3b-2410-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.ministral-3b-2410-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -10231,7 +10231,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -10243,7 +10243,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-ai-large-2407-offer?tab=Overview",
         "supports_function_calling": true,
         "supports_tool_choice": true
     },
@@ -10280,7 +10280,7 @@
         "max_tokens": 4096,
         "mode": "chat",
         "output_cost_per_token": 1.5e-07,
-        "source": "https://azuremarketplace.microsoft.com/en/marketplace/apps/000-000.mistral-nemo-12b-2407?tab=PlansAndPrice",
+        "source": "https://marketplace.microsoft.com/en/marketplace/apps/000-000.mistral-nemo-12b-2407?tab=PlansAndPrice",
         "supports_function_calling": true
     },
     "azure_ai/mistral-small": {
@@ -22796,7 +22796,7 @@
         "output_cost_per_reasoning_token": 3e-06,
         "output_cost_per_token": 3e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -22850,7 +22850,7 @@
         "output_cost_per_reasoning_token": 9e-06,
         "output_cost_per_token": 9e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -22915,7 +22915,7 @@
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -22974,7 +22974,7 @@
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23178,7 +23178,7 @@
         "mode": "chat",
         "output_cost_per_reasoning_token": 3e-06,
         "output_cost_per_token": 3e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23262,7 +23262,7 @@
         "mode": "chat",
         "output_cost_per_reasoning_token": 9e-06,
         "output_cost_per_token": 9e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23325,7 +23325,7 @@
         "output_cost_per_token": 3.75e-06,
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -23382,7 +23382,7 @@
         "output_cost_per_token": 3.75e-06,
         "output_cost_per_token_batches": 1.875e-06,
         "output_cost_per_token_flex": 1.875e-06,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -26982,7 +26982,7 @@
         "supports_tool_choice": true,
         "supports_vision": true,
         "supports_web_search": true,
-        "source": "https://platform.openai.com/docs/models/gpt-5.6-cyber",
+        "source": "https://developers.openai.com/api/docs/models/gpt-5.6-cyber",
         "supports_computer_use": true,
         "supports_parallel_function_calling": true
     },
@@ -27021,7 +27021,7 @@
         "supports_tool_choice": true,
         "supports_vision": true,
         "supports_web_search": true,
-        "source": "https://platform.openai.com/docs/models/daybreak-red-latest",
+        "source": "https://developers.openai.com/api/docs/models/daybreak-red-latest",
         "supports_computer_use": true,
         "supports_parallel_function_calling": true
     },
@@ -27061,7 +27061,7 @@
         "supports_tool_choice": true,
         "supports_vision": true,
         "supports_web_search": true,
-        "source": "https://platform.openai.com/docs/models/daybreak-blue-latest",
+        "source": "https://developers.openai.com/api/docs/models/daybreak-blue-latest",
         "supports_parallel_function_calling": true
     },
     "chat-latest": {
@@ -27073,7 +27073,7 @@
         "max_tokens": 128000,
         "mode": "chat",
         "output_cost_per_token": 3e-05,
-        "source": "https://platform.openai.com/docs/models/chat-latest",
+        "source": "https://developers.openai.com/api/docs/models/chat-latest",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses"
@@ -30368,7 +30368,7 @@
             "search_context_size_low": 0.0025,
             "search_context_size_medium": 0.0025
         },
-        "source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
+        "source": "https://ai.developer.meta.com/docs/pricing-rate-limits",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses",
@@ -30409,7 +30409,7 @@
             "search_context_size_low": 0.0025,
             "search_context_size_medium": 0.0025
         },
-        "source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
+        "source": "https://ai.developer.meta.com/docs/pricing-rate-limits",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses",
@@ -30450,7 +30450,7 @@
             "search_context_size_low": 0.0025,
             "search_context_size_medium": 0.0025
         },
-        "source": "https://dev.meta.ai/docs/getting-started/pricing-rate-limits",
+        "source": "https://ai.developer.meta.com/docs/pricing-rate-limits",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/responses",
@@ -30483,7 +30483,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text"
         ],
@@ -30499,7 +30499,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text"
         ],
@@ -30515,7 +30515,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text",
             "image"
@@ -30532,7 +30532,7 @@
         "max_output_tokens": 4028,
         "max_tokens": 4028,
         "mode": "chat",
-        "source": "https://llama.developer.meta.com/docs/models",
+        "source": "https://ai.developer.meta.com/docs/models",
         "supported_modalities": [
             "text",
             "image"
@@ -32477,7 +32477,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_reasoning": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-R1-0528": {
         "max_tokens": 164000,
@@ -32489,7 +32489,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_reasoning": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-R1-Distill-Llama-70B": {
         "max_tokens": 128000,
@@ -32500,7 +32500,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-V3": {
         "max_tokens": 128000,
@@ -32511,7 +32511,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/deepseek-ai/DeepSeek-V3-0324": {
         "max_tokens": 128000,
@@ -32522,7 +32522,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/google/gemma-3-27b-it": {
         "max_tokens": 128000,
@@ -32534,7 +32534,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Llama-3.3-70B-Instruct": {
         "max_tokens": 128000,
@@ -32545,7 +32545,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Llama-Guard-3-8B": {
         "max_tokens": 128000,
@@ -32555,7 +32555,7 @@
         "output_cost_per_token": 6e-08,
         "litellm_provider": "nebius",
         "mode": "chat",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Meta-Llama-3.1-8B-Instruct": {
         "max_tokens": 128000,
@@ -32566,7 +32566,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Meta-Llama-3.1-70B-Instruct": {
         "max_tokens": 128000,
@@ -32577,7 +32577,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/meta-llama/Meta-Llama-3.1-405B-Instruct": {
         "max_tokens": 128000,
@@ -32588,7 +32588,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/mistralai/Mistral-Nemo-Instruct-2407": {
         "max_tokens": 128000,
@@ -32599,7 +32599,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/NousResearch/Hermes-3-Llama-3.1-405B": {
         "max_tokens": 128000,
@@ -32610,7 +32610,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/nvidia/Llama-3.1-Nemotron-Ultra-253B-v1": {
         "max_tokens": 128000,
@@ -32621,7 +32621,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/nvidia/Llama-3.3-Nemotron-Super-49B-v1": {
         "max_tokens": 131072,
@@ -32632,7 +32632,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-235B-A22B": {
         "max_tokens": 262144,
@@ -32643,7 +32643,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-32B": {
         "max_tokens": 32768,
@@ -32654,7 +32654,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-30B-A3B": {
         "max_tokens": 32768,
@@ -32665,7 +32665,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-14B": {
         "max_tokens": 32768,
@@ -32676,7 +32676,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen3-4B": {
         "max_tokens": 32768,
@@ -32687,7 +32687,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/QwQ-32B": {
         "max_tokens": 32768,
@@ -32699,7 +32699,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_reasoning": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-72B-Instruct": {
         "max_tokens": 128000,
@@ -32710,7 +32710,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-32B-Instruct": {
         "max_tokens": 128000,
@@ -32721,7 +32721,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-Coder-7B": {
         "max_tokens": 32768,
@@ -32732,7 +32732,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_function_calling": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2.5-VL-72B-Instruct": {
         "max_tokens": 131072,
@@ -32744,7 +32744,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2-VL-72B-Instruct": {
         "max_tokens": 131072,
@@ -32756,7 +32756,7 @@
         "mode": "chat",
         "supports_function_calling": true,
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/Qwen/Qwen2-VL-7B-Instruct": {
         "max_tokens": 131072,
@@ -32767,7 +32767,7 @@
         "litellm_provider": "nebius",
         "mode": "chat",
         "supports_vision": true,
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/BAAI/bge-en-icl": {
         "max_tokens": 32768,
@@ -32776,7 +32776,7 @@
         "output_cost_per_token": 0.0,
         "litellm_provider": "nebius",
         "mode": "embedding",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/BAAI/bge-multilingual-gemma2": {
         "max_tokens": 8192,
@@ -32785,7 +32785,7 @@
         "output_cost_per_token": 0.0,
         "litellm_provider": "nebius",
         "mode": "embedding",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nebius/intfloat/e5-mistral-7b-instruct": {
         "max_tokens": 32768,
@@ -32794,7 +32794,7 @@
         "output_cost_per_token": 0.0,
         "litellm_provider": "nebius",
         "mode": "embedding",
-        "source": "https://nebius.com/prices-ai-studio"
+        "source": "https://nebius.com/prices"
     },
     "nvidia.nemotron-nano-12b-v2": {
         "input_cost_per_token": 2e-07,
@@ -33535,7 +33535,7 @@
         "max_tokens": 4000,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33548,7 +33548,7 @@
         "max_tokens": 4000,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33561,7 +33561,7 @@
         "max_tokens": 4000,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33618,7 +33618,7 @@
         "max_tokens": 8192,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": true,
         "supports_response_schema": false,
         "supports_native_streaming": true,
@@ -33632,7 +33632,7 @@
         "max_tokens": 8192,
         "mode": "chat",
         "output_cost_per_token": 1.56e-06,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_function_calling": false,
         "supports_response_schema": false,
         "supports_native_streaming": true
@@ -33643,7 +33643,7 @@
         "max_input_tokens": 512,
         "mode": "embedding",
         "output_vector_size": 1024,
-        "source": "https://www.oracle.com/cloud/ai/generative-ai/pricing/",
+        "source": "https://www.oracle.com/artificial-intelligence/enterprise-ai/cost-estimator/",
         "supports_vision": true
     },
     "oci/cohere.command-a-reasoning-08-2025": {
@@ -34540,7 +34540,7 @@
         "max_tokens": 2048,
         "mode": "chat",
         "output_cost_per_token": 2e-07,
-        "source": "https://openrouter.ai/api/v1/models/bytedance/ui-tars-1.5-7b",
+        "source": "https://openrouter.ai/bytedance/ui-tars-1.5-7b",
         "supports_tool_choice": true
     },
     "openrouter/deepseek/deepseek-chat": {
@@ -34775,7 +34775,7 @@
         "output_cost_per_reasoning_token": 3e-06,
         "output_cost_per_token": 3e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -34816,7 +34816,7 @@
         "output_cost_per_reasoning_token": 1.5e-06,
         "output_cost_per_token": 1.5e-06,
         "rpm": 2000,
-        "source": "https://ai.google.dev/pricing/gemini-3",
+        "source": "https://ai.google.dev/gemini-api/docs/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/completions",
@@ -35901,7 +35901,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 6.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/deepseek-r1-distill-llama-70b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -35915,7 +35915,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 1e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/llama-3-1-8b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35928,7 +35928,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 6.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/meta-llama-3-1-70b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": false,
         "supports_tool_choice": false
@@ -35941,7 +35941,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 6.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/meta-llama-3-3-70b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35954,7 +35954,7 @@
         "max_tokens": 127000,
         "mode": "chat",
         "output_cost_per_token": 1e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mistral-7b-instruct-v0-3",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35967,7 +35967,7 @@
         "max_tokens": 118000,
         "mode": "chat",
         "output_cost_per_token": 1.3e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mistral-nemo-instruct-2407",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true
@@ -35980,7 +35980,7 @@
         "max_tokens": 128000,
         "mode": "chat",
         "output_cost_per_token": 2.8e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mistral-small-3-2-24b-instruct-2506",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true,
@@ -35994,7 +35994,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 6.3e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mixtral-8x7b-instruct-v0-1",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false
@@ -36007,7 +36007,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 8.7e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/qwen2-5-coder-32b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false
@@ -36020,7 +36020,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 9.1e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/qwen2-5-vl-72b-instruct",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false,
@@ -36034,7 +36034,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 2.3e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/qwen3-32b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -36048,7 +36048,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 4e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/gpt-oss-120b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -36062,7 +36062,7 @@
         "max_tokens": 131000,
         "mode": "chat",
         "output_cost_per_token": 1.5e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/gpt-oss-20b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -36076,7 +36076,7 @@
         "max_tokens": 32000,
         "mode": "chat",
         "output_cost_per_token": 2.9e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/llava-next-mistral-7b",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false,
@@ -36090,7 +36090,7 @@
         "max_tokens": 256000,
         "mode": "chat",
         "output_cost_per_token": 1.9e-07,
-        "source": "https://endpoints.ai.cloud.ovh.net/models/mamba-codestral-7b-v0-1",
+        "source": "https://www.ovhcloud.com/en/public-cloud/ai-endpoints/catalog/",
         "supports_function_calling": false,
         "supports_response_schema": true,
         "supports_tool_choice": false
@@ -38856,7 +38856,7 @@
         "max_input_tokens": 262144,
         "mode": "chat",
         "output_cost_per_token": 3.6e-06,
-        "source": "https://www.together.ai/models/Qwen/Qwen3.5-397B-A17B",
+        "source": "https://www.together.ai/models/qwen3-5-397b-a17b",
         "supports_function_calling": true,
         "supports_parallel_function_calling": true,
         "supports_prompt_caching": true,
@@ -43097,7 +43097,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 5e-07,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true,
@@ -43113,7 +43113,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 5e-07,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -43130,7 +43130,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_response_schema": true,
         "supports_tool_choice": true,
@@ -43146,7 +43146,7 @@
         "max_tokens": 2000000,
         "mode": "chat",
         "output_cost_per_token": 6e-06,
-        "source": "https://docs.x.ai/docs/models (Vertex AI Model Garden)",
+        "source": "https://docs.x.ai/developers/models",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_response_schema": true,
@@ -43266,7 +43266,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.4,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -43280,7 +43280,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.15,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -43295,7 +43295,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.4,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -43310,7 +43310,7 @@
         "max_tokens": 1024,
         "mode": "video_generation",
         "output_cost_per_second": 0.15,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo",
+        "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/veo/3-1-generate",
         "supported_modalities": [
             "text"
         ],
@@ -44946,7 +44946,7 @@
         "litellm_provider": "azure",
         "mode": "video_generation",
         "output_cost_per_video_per_second": 0.1,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/video-generation",
+        "source": "https://ai.azure.com/catalog/models/sora-2",
         "supported_modalities": [
             "text"
         ],
@@ -44958,7 +44958,7 @@
         "litellm_provider": "azure",
         "mode": "video_generation",
         "output_cost_per_video_per_second": 0.3,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/video-generation",
+        "source": "https://ai.azure.com/catalog/models/sora-2-pro",
         "supported_modalities": [
             "text"
         ],
@@ -44970,7 +44970,7 @@
         "litellm_provider": "azure",
         "mode": "video_generation",
         "output_cost_per_video_per_second": 0.5,
-        "source": "https://azure.microsoft.com/en-us/products/ai-services/video-generation",
+        "source": "https://ai.azure.com/catalog/models/sora-2-pro",
         "supported_modalities": [
             "text"
         ],
@@ -49180,7 +49180,7 @@
         "input_cost_per_second": 0.0002833333333333333,
         "litellm_provider": "openai",
         "mode": "audio_transcription",
-        "source": "https://platform.openai.com/docs/models/gpt-realtime-whisper",
+        "source": "https://developers.openai.com/api/docs/models/gpt-realtime-whisper",
         "supported_endpoints": [
             "/v1/realtime",
             "/v1/realtime/transcription_sessions"
@@ -50432,7 +50432,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -50470,7 +50470,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -50508,7 +50508,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -50546,7 +50546,7 @@
         "max_output_tokens": 128000,
         "max_tokens": 128000,
         "mode": "chat",
-        "source": "https://www.volcengine.com/docs/82379/1330310",
+        "source": "https://docs.volcengine.com/docs/82379/1330310",
         "supports_function_calling": true,
         "supports_reasoning": true,
         "supports_tool_choice": false,
@@ -51295,7 +51295,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/qwen3.6-35b-a3b": {
         "max_tokens": 131072,
@@ -51308,7 +51308,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/qwen3-30b-a3b": {
         "max_tokens": 131072,
@@ -51321,7 +51321,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/qwen3-coder-30b-a3b": {
         "max_tokens": 131072,
@@ -51334,7 +51334,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": false,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/deepseek-v4-flash": {
         "max_tokens": 163840,
@@ -51347,7 +51347,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": true,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "pinstripes/ps/minimax-m2.7": {
         "max_tokens": 1000192,
@@ -51360,7 +51360,7 @@
         "supports_function_calling": true,
         "supports_assistant_prefill": true,
         "supports_reasoning": false,
-        "source": "https://pinstripes.io/pricing"
+        "source": "https://pinstripes.io/"
     },
     "darkbloom/gemma-4-26b": {
         "input_cost_per_token": 3e-08,
@@ -51464,7 +51464,7 @@
         "input_cost_per_second": 7.5e-05,
         "litellm_provider": "openai",
         "mode": "audio_transcription",
-        "source": "https://platform.openai.com/docs/models/gpt-transcribe",
+        "source": "https://developers.openai.com/api/docs/models/gpt-transcribe",
         "supported_endpoints": [
             "/v1/audio/transcriptions",
             "/v1/realtime/transcription_sessions"
@@ -51482,7 +51482,7 @@
         "input_cost_per_second": 0.0002833333333333333,
         "litellm_provider": "openai",
         "mode": "audio_transcription",
-        "source": "https://platform.openai.com/docs/models/gpt-live-transcribe",
+        "source": "https://developers.openai.com/api/docs/models/gpt-live-transcribe",
         "supported_endpoints": [
             "/v1/realtime",
             "/v1/realtime/transcription_sessions"
@@ -51503,7 +51503,7 @@
         "max_output_tokens": 2000,
         "max_tokens": 2000,
         "mode": "realtime",
-        "source": "https://platform.openai.com/docs/models/gpt-realtime-translate",
+        "source": "https://developers.openai.com/api/docs/models/gpt-realtime-translate",
         "supported_modalities": [
             "audio"
         ],
@@ -51531,7 +51531,7 @@
             "search_context_size_low": 0.01,
             "search_context_size_medium": 0.01
         },
-        "source": "https://docs.claude.com/en/docs/about-claude/models/overview",
+        "source": "https://platform.claude.com/docs/en/about-claude/models/overview",
         "supports_adaptive_thinking": true,
         "thinking_always_on": true,
         "supports_mid_conversation_system": true,
@@ -51570,7 +51570,7 @@
             "search_context_size_low": 0.01,
             "search_context_size_medium": 0.01
         },
-        "source": "https://docs.claude.com/en/docs/about-claude/models/overview",
+        "source": "https://platform.claude.com/docs/en/about-claude/models/overview",
         "supports_adaptive_thinking": true,
         "thinking_always_on": true,
         "supports_assistant_prefill": false,

From e2ffb6b01c52764fb31d9e931c64f4c53a14a747 Mon Sep 17 00:00:00 2001
From: mubashir1osmani <mubashir.osmani777@gmail.com>
Date: Wed, 26 Aug 2026 14:42:08 -0400
Subject: [PATCH 03/19] feat(ci): close duplicate issues after a 3-day grace
 period

Duplicate detection already labelled and commented on new issues, and then
closed them outright at 0.85 title similarity. That gave the reporter no chance
to push back, and a title-similarity match is not strong enough evidence to
close on its own.

Detection now only flags. A new daily sweep closes a flagged issue three days
later, and only if nobody engaged with the flag. Replying to it, thumbs-downing
it, or applying an opt-out label all keep the issue open. The notice says all of
that up front, so the reporter knows what happens and how to stop it.

The two workflows hand off through an HTML marker in the comment body rather
than its prose, so rewording the notice cannot silently break the sweep. The
sweep lists by label instead of walking the whole backlog: 1663 open issues
against 23 carrying the label meant a comments request each, which would burn
the Actions token's hourly budget for a handful of matches.

Candidates are taken as the lowest issue number, not the first one listed. The
detector orders by score rather than age, so the first candidate can be newer
than the issue being closed, and folding an original report into a later one is
backwards. An issue whose only candidates are newer is skipped.

Closures use state_reason=duplicate rather than not_planned, which reads as
"see the other issue" instead of "we are not doing this".

Also drops {{html_url}} from the notice. The detection action only exposes
number, title and accuracy, so that placeholder had been rendering empty and
every "similar issue" link in the comment pointed nowhere.
---
 .github/workflows/check_duplicate_issues.yml  |  38 ++---
 .../close_stale_duplicate_issues.yml          | 148 ++++++++++++++++++
 2 files changed, 158 insertions(+), 28 deletions(-)
 create mode 100644 .github/workflows/close_stale_duplicate_issues.yml

diff --git a/.github/workflows/check_duplicate_issues.yml b/.github/workflows/check_duplicate_issues.yml
index 78198b2c7bb..a087007bff3 100644
--- a/.github/workflows/check_duplicate_issues.yml
+++ b/.github/workflows/check_duplicate_issues.yml
@@ -1,5 +1,10 @@
 name: Check Duplicate Issues
 
+# Flags newly opened issues that look like existing ones. Flagging only: the actual
+# close happens 3 days later in "Close Stale Duplicate Issues", and only if nobody
+# replied to the comment posted here. The HTML marker below is the handshake between
+# the two workflows, so keep it in the template.
+
 on:
   issues:
     types: [opened, edited]
@@ -19,35 +24,12 @@ jobs:
           threshold: 0.6
           reaction: eyes
           comment: |
-            **⚠️ Potential duplicate detected**
+            <!-- litellm:potential-duplicate -->
+            **Potential duplicate detected**
 
-            This issue appears similar to existing issue(s):
+            This looks similar to:
             {{#issues}}
-            - [#{{number}}]({{html_url}}) - {{title}} ({{accuracy}}% similar)
+            - #{{number}} - {{title}} ({{accuracy}}% similar)
             {{/issues}}
 
-            Please review the linked issue(s) to see if they address your concern. If this is not a duplicate, please provide additional context to help us understand the difference.
-
-      - name: Checkout close script
-        if: github.event.action == 'opened'
-        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
-        with:
-          sparse-checkout: .github/scripts
-          persist-credentials: false
-
-      - name: Set up Python
-        if: github.event.action == 'opened'
-        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
-        with:
-          python-version: "3.12"
-
-      - name: Auto-close if high-confidence duplicate
-        if: github.event.action == 'opened'
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          python3 .github/scripts/close_duplicate_issues.py \
-            --issue-number ${{ github.event.issue.number }} \
-            --repo ${{ github.repository }} \
-            --threshold 0.85 \
-            --close
+            This issue will close automatically in 3 days unless someone responds. If it is a duplicate, please 👍 the existing issue and follow along there. If it is not, comment here or 👎 this comment and it stays open.
diff --git a/.github/workflows/close_stale_duplicate_issues.yml b/.github/workflows/close_stale_duplicate_issues.yml
new file mode 100644
index 00000000000..8bc1af90454
--- /dev/null
+++ b/.github/workflows/close_stale_duplicate_issues.yml
@@ -0,0 +1,148 @@
+name: Close Stale Duplicate Issues
+
+# Closes issues that "Check Duplicate Issues" flagged and that nobody acknowledged
+# within the grace period. Replying to the flag, thumbs-downing it, or applying an
+# opt-out label all keep an issue open.
+#
+# Dry-run preview (touches nothing):
+#   gh workflow run "Close Stale Duplicate Issues" -f dry_run=true
+
+on:
+  schedule:
+    # Daily at 09:30 UTC, after the midnight stale sweep and off the hour.
+    - cron: "30 9 * * *"
+  workflow_dispatch:
+    inputs:
+      dry_run:
+        description: "Report what would close without touching any issue."
+        required: false
+        default: "false"
+        type: choice
+        options:
+          - "true"
+          - "false"
+      grace_period_days:
+        description: "Days to wait after the duplicate flag before closing."
+        required: false
+        default: "3"
+      limit:
+        description: "Maximum number of issues to close in a single run."
+        required: false
+        default: "50"
+
+permissions:
+  contents: read
+  issues: write
+
+jobs:
+  close-stale-duplicates:
+    if: github.repository == 'BerriAI/litellm'
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - name: Close unacknowledged duplicates
+        uses: actions/github-script@f28e40c7f34bde8b3046d885e986cb6290c5673b # v7.1.0
+        env:
+          DRY_RUN: ${{ github.event.inputs.dry_run || 'false' }}
+          GRACE_PERIOD_DAYS: ${{ github.event.inputs.grace_period_days || '3' }}
+          LIMIT: ${{ github.event.inputs.limit || '50' }}
+        with:
+          script: |
+            const FLAG_MARKER = '<!-- litellm:potential-duplicate -->';
+            const FLAG_LABEL = 'potential-duplicate';
+            const OPTOUT_LABELS = ['do not close', 'keep open', 'not a duplicate'];
+
+            const dryRun = process.env.DRY_RUN === 'true';
+            const graceDays = Number(process.env.GRACE_PERIOD_DAYS);
+            const limit = Number(process.env.LIMIT);
+            const cutoff = Date.now() - graceDays * 86400000;
+            const { owner, repo } = context.repo;
+
+            // The oldest issue the flag points at, excluding the issue itself. The
+            // detector orders candidates by score, not age, so the first one listed
+            // can be newer than the original report.
+            const canonicalTarget = (body, self) => {
+              const refs = new Set();
+              for (const [, n] of body.matchAll(/#(\d+)/g)) refs.add(Number(n));
+              for (const [, n] of body.matchAll(/github\.com\/[^/\s]+\/[^/\s]+\/issues\/(\d+)/g)) refs.add(Number(n));
+              refs.delete(self);
+              return refs.size ? Math.min(...refs) : null;
+            };
+
+            // Only issues the detector labelled: scanning the whole open backlog would
+            // cost one comments request each and exhaust the token's hourly budget.
+            const issues = await github.paginate(github.rest.issues.listForRepo, {
+              owner, repo, state: 'open', labels: FLAG_LABEL, per_page: 100,
+            });
+            core.info(`Scanning ${issues.length} open issues labelled '${FLAG_LABEL}' in ${owner}/${repo}.`);
+
+            const closures = [];
+            for (const issue of issues) {
+              const skip = (reason) => core.info(`  #${issue.number}: skip, ${reason}`);
+              const labels = issue.labels.map((l) => (l.name || l).toLowerCase());
+              const blocking = OPTOUT_LABELS.find((l) => labels.includes(l));
+              if (blocking) { skip(`carries opt-out label '${blocking}'`); continue; }
+
+              const comments = await github.paginate(github.rest.issues.listComments, {
+                owner, repo, issue_number: issue.number, per_page: 100,
+              });
+
+              // Author filter matters: "Quote reply" carries the marker into a human
+              // comment, and treating that as a fresh flag would restart the clock.
+              const flags = comments.filter((c) => c.user?.type === 'Bot' && c.body?.includes(FLAG_MARKER));
+              if (!flags.length) { skip('never flagged as a potential duplicate'); continue; }
+
+              const flag = flags.reduce((a, b) => (new Date(a.created_at) > new Date(b.created_at) ? a : b));
+              const flaggedAt = new Date(flag.created_at).getTime();
+              if (flaggedAt > cutoff) { skip(`flagged less than ${graceDays}d ago`); continue; }
+
+              if (comments.some((c) => new Date(c.created_at).getTime() > flaggedAt)) {
+                skip('someone replied after the flag went up'); continue;
+              }
+
+              const reactions = await github.paginate(github.rest.reactions.listForIssueComment, {
+                owner, repo, comment_id: flag.id, per_page: 100,
+              });
+              if (reactions.some((r) => r.content === '-1' && r.user?.login === issue.user?.login)) {
+                skip('author thumbs-downed the flag'); continue;
+              }
+
+              const target = canonicalTarget(flag.body, issue.number);
+              if (target === null) { skip('flag comment names no other issue number'); continue; }
+              if (target > issue.number) { skip(`only candidate #${target} is newer than this issue`); continue; }
+
+              core.info(`  #${issue.number}: unacknowledged duplicate of #${target}`);
+              closures.push({ number: issue.number, title: issue.title, target });
+            }
+
+            const actionable = closures.slice(0, limit);
+            if (closures.length > limit) {
+              core.info(`Reached limit ${limit}; ${closures.length - limit} further match(es) left for the next run.`);
+            }
+
+            for (const { number, target } of actionable) {
+              if (dryRun) { core.info(`  WOULD close #${number} as duplicate of #${target}`); continue; }
+              core.info(`  closing #${number} as duplicate of #${target}`);
+              await github.rest.issues.createComment({
+                owner, repo, issue_number: number,
+                body: `Closing as a duplicate of #${target}.\n\nThe duplicate notice on this issue went `
+                  + `unanswered for ${graceDays} days, so it is being closed automatically. If that call is `
+                  + `wrong, reopen the issue and say how it differs from #${target}, and we will pick it back `
+                  + `up.\n\n<!-- litellm:closed-as-duplicate -->`,
+              });
+              await github.rest.issues.addLabels({ owner, repo, issue_number: number, labels: ['duplicate'] });
+              await github.rest.issues.update({
+                owner, repo, issue_number: number, state: 'closed', state_reason: 'duplicate',
+              });
+            }
+
+            const heading = dryRun ? 'Would close as duplicates (dry run)' : 'Closed as duplicates';
+            const rows = actionable.length
+              ? ['| Issue | Duplicate of |', '| --- | --- |',
+                 ...actionable.map((c) => `| [#${c.number}](https://github.com/${owner}/${repo}/issues/${c.number}) ${c.title} | #${c.target} |`)]
+              : ['No issue reached the end of its grace period unacknowledged.'];
+            await core.summary
+              .addRaw([`## ${heading}`, '', ...rows, '', `Scanned ${issues.length} flagged issues.`].join('\n'))
+              .write();
+
+            core.info(`\n${dryRun ? 'Would close' : 'Closed'}: ${actionable.length}`);

From af340c02402a3c0f989c388f217d5d974603809f Mon Sep 17 00:00:00 2001
From: mubashir1osmani <mubashir.osmani777@gmail.com>
Date: Wed, 26 Aug 2026 15:04:35 -0400
Subject: [PATCH 04/19] fix(ci): read duplicate candidates from the marker, not
 the notice prose

The notice interpolates each candidate's title, and the sweep scanned the whole
comment for issue references and took the lowest. Titles are attacker-controlled,
so filing a candidate titled "... see #1" redirected the closure: any later report
matching that candidate would be closed as a duplicate of #1 instead.

The detector now emits the candidate numbers as a digits-only field inside the
marker, built from the API's number field, and the sweep reads only that. Prose is
never parsed, so nothing a reporter can type reaches the target selection.
---
 .github/workflows/check_duplicate_issues.yml    |  2 +-
 .../workflows/close_stale_duplicate_issues.yml  | 17 ++++++++++-------
 2 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/.github/workflows/check_duplicate_issues.yml b/.github/workflows/check_duplicate_issues.yml
index a087007bff3..71d2a3b75eb 100644
--- a/.github/workflows/check_duplicate_issues.yml
+++ b/.github/workflows/check_duplicate_issues.yml
@@ -24,7 +24,7 @@ jobs:
           threshold: 0.6
           reaction: eyes
           comment: |
-            <!-- litellm:potential-duplicate -->
+            <!-- litellm:potential-duplicate candidates={{#issues}}{{number}},{{/issues}} -->
             **Potential duplicate detected**
 
             This looks similar to:
diff --git a/.github/workflows/close_stale_duplicate_issues.yml b/.github/workflows/close_stale_duplicate_issues.yml
index 8bc1af90454..c5f904d2b88 100644
--- a/.github/workflows/close_stale_duplicate_issues.yml
+++ b/.github/workflows/close_stale_duplicate_issues.yml
@@ -48,7 +48,8 @@ jobs:
           LIMIT: ${{ github.event.inputs.limit || '50' }}
         with:
           script: |
-            const FLAG_MARKER = '<!-- litellm:potential-duplicate -->';
+            const FLAG_MARKER = '<!-- litellm:potential-duplicate';
+            const CANDIDATES = /<!-- litellm:potential-duplicate candidates=([\d,]*) -->/;
             const FLAG_LABEL = 'potential-duplicate';
             const OPTOUT_LABELS = ['do not close', 'keep open', 'not a duplicate'];
 
@@ -58,13 +59,15 @@ jobs:
             const cutoff = Date.now() - graceDays * 86400000;
             const { owner, repo } = context.repo;
 
-            // The oldest issue the flag points at, excluding the issue itself. The
-            // detector orders candidates by score, not age, so the first one listed
-            // can be newer than the original report.
+            // Read candidates from the marker's digits-only field, never from the prose.
+            // Titles are user-controlled and get interpolated into this same comment, so
+            // scanning the body would let an issue titled "... see #1" redirect a closure
+            // onto an unrelated report. Take the lowest: the detector orders by score, not
+            // age, so the first candidate listed can be newer than the original report.
             const canonicalTarget = (body, self) => {
-              const refs = new Set();
-              for (const [, n] of body.matchAll(/#(\d+)/g)) refs.add(Number(n));
-              for (const [, n] of body.matchAll(/github\.com\/[^/\s]+\/[^/\s]+\/issues\/(\d+)/g)) refs.add(Number(n));
+              const field = body.match(CANDIDATES);
+              if (!field) return null;
+              const refs = new Set(field[1].split(',').filter(Boolean).map(Number));
               refs.delete(self);
               return refs.size ? Math.min(...refs) : null;
             };

From 3ea11b64e65628f07b47a2f0e0853e79b1ff8334 Mon Sep 17 00:00:00 2001
From: mubashir1osmani <mubashir.osmani777@gmail.com>
Date: Wed, 26 Aug 2026 15:18:06 -0400
Subject: [PATCH 05/19] refactor(ci): run the duplicate sweep as a Bun
 TypeScript script

Moves the sweep out of inline workflow JavaScript and into scripts/, following
the layout anthropics/claude-code uses for the same job: a checked-out repo, a
sha-pinned setup-bun step, and `bun run scripts/auto-close-duplicates.ts`.

The script mirrors that repo's file shape, keeping the same request helper,
interfaces, per-issue debug logging, and top-level catch, so the two read the
same way side by side.

Two things stay deliberately different. Candidates come from the notice marker's
digits-only field rather than a regex over the comment prose, because titles are
attacker-controlled and are interpolated into that same comment. The label is
also added on its own endpoint instead of alongside the state change, since
sending labels with a PATCH replaces every label already on the issue.
---
 .github/workflows/auto-close-duplicates.yml   |  33 ++
 .../close_stale_duplicate_issues.yml          | 151 ---------
 scripts/auto-close-duplicates.ts              | 308 ++++++++++++++++++
 3 files changed, 341 insertions(+), 151 deletions(-)
 create mode 100644 .github/workflows/auto-close-duplicates.yml
 delete mode 100644 .github/workflows/close_stale_duplicate_issues.yml
 create mode 100644 scripts/auto-close-duplicates.ts

diff --git a/.github/workflows/auto-close-duplicates.yml b/.github/workflows/auto-close-duplicates.yml
new file mode 100644
index 00000000000..886aeaaa8e6
--- /dev/null
+++ b/.github/workflows/auto-close-duplicates.yml
@@ -0,0 +1,33 @@
+name: Auto-close duplicate issues
+description: Auto-closes issues that are duplicates of existing issues
+on:
+  schedule:
+    - cron: "0 9 * * *"
+  workflow_dispatch:
+
+jobs:
+  auto-close-duplicates:
+    if: github.repository == 'BerriAI/litellm'
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          persist-credentials: false
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0 (sha-pinned)
+        with:
+          bun-version: latest
+
+      - name: Auto-close duplicate issues
+        run: bun run scripts/auto-close-duplicates.ts
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_REPOSITORY_OWNER: ${{ github.repository_owner }}
+          GITHUB_REPOSITORY_NAME: ${{ github.event.repository.name }}
diff --git a/.github/workflows/close_stale_duplicate_issues.yml b/.github/workflows/close_stale_duplicate_issues.yml
deleted file mode 100644
index c5f904d2b88..00000000000
--- a/.github/workflows/close_stale_duplicate_issues.yml
+++ /dev/null
@@ -1,151 +0,0 @@
-name: Close Stale Duplicate Issues
-
-# Closes issues that "Check Duplicate Issues" flagged and that nobody acknowledged
-# within the grace period. Replying to the flag, thumbs-downing it, or applying an
-# opt-out label all keep an issue open.
-#
-# Dry-run preview (touches nothing):
-#   gh workflow run "Close Stale Duplicate Issues" -f dry_run=true
-
-on:
-  schedule:
-    # Daily at 09:30 UTC, after the midnight stale sweep and off the hour.
-    - cron: "30 9 * * *"
-  workflow_dispatch:
-    inputs:
-      dry_run:
-        description: "Report what would close without touching any issue."
-        required: false
-        default: "false"
-        type: choice
-        options:
-          - "true"
-          - "false"
-      grace_period_days:
-        description: "Days to wait after the duplicate flag before closing."
-        required: false
-        default: "3"
-      limit:
-        description: "Maximum number of issues to close in a single run."
-        required: false
-        default: "50"
-
-permissions:
-  contents: read
-  issues: write
-
-jobs:
-  close-stale-duplicates:
-    if: github.repository == 'BerriAI/litellm'
-    runs-on: ubuntu-latest
-    timeout-minutes: 20
-    steps:
-      - name: Close unacknowledged duplicates
-        uses: actions/github-script@f28e40c7f34bde8b3046d885e986cb6290c5673b # v7.1.0
-        env:
-          DRY_RUN: ${{ github.event.inputs.dry_run || 'false' }}
-          GRACE_PERIOD_DAYS: ${{ github.event.inputs.grace_period_days || '3' }}
-          LIMIT: ${{ github.event.inputs.limit || '50' }}
-        with:
-          script: |
-            const FLAG_MARKER = '<!-- litellm:potential-duplicate';
-            const CANDIDATES = /<!-- litellm:potential-duplicate candidates=([\d,]*) -->/;
-            const FLAG_LABEL = 'potential-duplicate';
-            const OPTOUT_LABELS = ['do not close', 'keep open', 'not a duplicate'];
-
-            const dryRun = process.env.DRY_RUN === 'true';
-            const graceDays = Number(process.env.GRACE_PERIOD_DAYS);
-            const limit = Number(process.env.LIMIT);
-            const cutoff = Date.now() - graceDays * 86400000;
-            const { owner, repo } = context.repo;
-
-            // Read candidates from the marker's digits-only field, never from the prose.
-            // Titles are user-controlled and get interpolated into this same comment, so
-            // scanning the body would let an issue titled "... see #1" redirect a closure
-            // onto an unrelated report. Take the lowest: the detector orders by score, not
-            // age, so the first candidate listed can be newer than the original report.
-            const canonicalTarget = (body, self) => {
-              const field = body.match(CANDIDATES);
-              if (!field) return null;
-              const refs = new Set(field[1].split(',').filter(Boolean).map(Number));
-              refs.delete(self);
-              return refs.size ? Math.min(...refs) : null;
-            };
-
-            // Only issues the detector labelled: scanning the whole open backlog would
-            // cost one comments request each and exhaust the token's hourly budget.
-            const issues = await github.paginate(github.rest.issues.listForRepo, {
-              owner, repo, state: 'open', labels: FLAG_LABEL, per_page: 100,
-            });
-            core.info(`Scanning ${issues.length} open issues labelled '${FLAG_LABEL}' in ${owner}/${repo}.`);
-
-            const closures = [];
-            for (const issue of issues) {
-              const skip = (reason) => core.info(`  #${issue.number}: skip, ${reason}`);
-              const labels = issue.labels.map((l) => (l.name || l).toLowerCase());
-              const blocking = OPTOUT_LABELS.find((l) => labels.includes(l));
-              if (blocking) { skip(`carries opt-out label '${blocking}'`); continue; }
-
-              const comments = await github.paginate(github.rest.issues.listComments, {
-                owner, repo, issue_number: issue.number, per_page: 100,
-              });
-
-              // Author filter matters: "Quote reply" carries the marker into a human
-              // comment, and treating that as a fresh flag would restart the clock.
-              const flags = comments.filter((c) => c.user?.type === 'Bot' && c.body?.includes(FLAG_MARKER));
-              if (!flags.length) { skip('never flagged as a potential duplicate'); continue; }
-
-              const flag = flags.reduce((a, b) => (new Date(a.created_at) > new Date(b.created_at) ? a : b));
-              const flaggedAt = new Date(flag.created_at).getTime();
-              if (flaggedAt > cutoff) { skip(`flagged less than ${graceDays}d ago`); continue; }
-
-              if (comments.some((c) => new Date(c.created_at).getTime() > flaggedAt)) {
-                skip('someone replied after the flag went up'); continue;
-              }
-
-              const reactions = await github.paginate(github.rest.reactions.listForIssueComment, {
-                owner, repo, comment_id: flag.id, per_page: 100,
-              });
-              if (reactions.some((r) => r.content === '-1' && r.user?.login === issue.user?.login)) {
-                skip('author thumbs-downed the flag'); continue;
-              }
-
-              const target = canonicalTarget(flag.body, issue.number);
-              if (target === null) { skip('flag comment names no other issue number'); continue; }
-              if (target > issue.number) { skip(`only candidate #${target} is newer than this issue`); continue; }
-
-              core.info(`  #${issue.number}: unacknowledged duplicate of #${target}`);
-              closures.push({ number: issue.number, title: issue.title, target });
-            }
-
-            const actionable = closures.slice(0, limit);
-            if (closures.length > limit) {
-              core.info(`Reached limit ${limit}; ${closures.length - limit} further match(es) left for the next run.`);
-            }
-
-            for (const { number, target } of actionable) {
-              if (dryRun) { core.info(`  WOULD close #${number} as duplicate of #${target}`); continue; }
-              core.info(`  closing #${number} as duplicate of #${target}`);
-              await github.rest.issues.createComment({
-                owner, repo, issue_number: number,
-                body: `Closing as a duplicate of #${target}.\n\nThe duplicate notice on this issue went `
-                  + `unanswered for ${graceDays} days, so it is being closed automatically. If that call is `
-                  + `wrong, reopen the issue and say how it differs from #${target}, and we will pick it back `
-                  + `up.\n\n<!-- litellm:closed-as-duplicate -->`,
-              });
-              await github.rest.issues.addLabels({ owner, repo, issue_number: number, labels: ['duplicate'] });
-              await github.rest.issues.update({
-                owner, repo, issue_number: number, state: 'closed', state_reason: 'duplicate',
-              });
-            }
-
-            const heading = dryRun ? 'Would close as duplicates (dry run)' : 'Closed as duplicates';
-            const rows = actionable.length
-              ? ['| Issue | Duplicate of |', '| --- | --- |',
-                 ...actionable.map((c) => `| [#${c.number}](https://github.com/${owner}/${repo}/issues/${c.number}) ${c.title} | #${c.target} |`)]
-              : ['No issue reached the end of its grace period unacknowledged.'];
-            await core.summary
-              .addRaw([`## ${heading}`, '', ...rows, '', `Scanned ${issues.length} flagged issues.`].join('\n'))
-              .write();
-
-            core.info(`\n${dryRun ? 'Would close' : 'Closed'}: ${actionable.length}`);
diff --git a/scripts/auto-close-duplicates.ts b/scripts/auto-close-duplicates.ts
new file mode 100644
index 00000000000..1e94b009a1e
--- /dev/null
+++ b/scripts/auto-close-duplicates.ts
@@ -0,0 +1,308 @@
+#!/usr/bin/env bun
+
+declare global {
+  var process: {
+    env: Record<string, string | undefined>;
+  };
+}
+
+interface GitHubIssue {
+  number: number;
+  title: string;
+  user: { login: string };
+  labels: { name: string }[];
+}
+
+interface GitHubComment {
+  id: number;
+  body: string;
+  created_at: string;
+  user: { type: string };
+}
+
+interface GitHubReaction {
+  user: { login: string };
+  content: string;
+}
+
+const FLAG_LABEL = "potential-duplicate";
+const FLAG_MARKER = "<!-- litellm:potential-duplicate";
+const CANDIDATES = /<!-- litellm:potential-duplicate candidates=([\d,]*) -->/;
+const GRACE_PERIOD_DAYS = 3;
+
+async function githubRequest<T>(
+  endpoint: string,
+  token: string,
+  method: string = "GET",
+  body?: any,
+): Promise<T> {
+  const response = await fetch(`https://api.github.com${endpoint}`, {
+    method,
+    headers: {
+      Authorization: `Bearer ${token}`,
+      Accept: "application/vnd.github.v3+json",
+      "User-Agent": "auto-close-duplicates-script",
+      ...(body && { "Content-Type": "application/json" }),
+    },
+    ...(body && { body: JSON.stringify(body) }),
+  });
+
+  if (!response.ok) {
+    throw new Error(
+      `GitHub API request failed: ${response.status} ${response.statusText}`,
+    );
+  }
+
+  return response.json();
+}
+
+function extractDuplicateIssueNumber(
+  commentBody: string,
+  issueNumber: number,
+): number | null {
+  // Read candidates from the marker's digits-only field, never from the prose.
+  // Titles are user-controlled and are interpolated into this same comment, so
+  // scanning the body would let an issue titled "... see #1" redirect a closure
+  // onto an unrelated report.
+  const field = commentBody.match(CANDIDATES);
+  if (!field) {
+    return null;
+  }
+
+  const candidates = field[1]
+    .split(",")
+    .filter((value) => value !== "")
+    .map(Number)
+    .filter((value) => value !== issueNumber);
+
+  // The detector orders candidates by score, not age, so the first one listed can
+  // be newer than the original report. Duplicates fold into the earliest issue.
+  return candidates.length > 0 ? Math.min(...candidates) : null;
+}
+
+async function closeIssueAsDuplicate(
+  owner: string,
+  repo: string,
+  issueNumber: number,
+  duplicateOfNumber: number,
+  token: string,
+): Promise<void> {
+  await githubRequest(
+    `/repos/${owner}/${repo}/issues/${issueNumber}/comments`,
+    token,
+    "POST",
+    {
+      body: `This issue has been automatically closed as a duplicate of #${duplicateOfNumber}.
+
+The duplicate notice went unanswered for ${GRACE_PERIOD_DAYS} days. If this is incorrect, please re-open this issue and say how it differs from #${duplicateOfNumber}.
+
+<!-- litellm:closed-as-duplicate -->`,
+    },
+  );
+
+  // Added on its own endpoint rather than in the PATCH below, because sending
+  // `labels` with the state change replaces every label on the issue.
+  await githubRequest(
+    `/repos/${owner}/${repo}/issues/${issueNumber}/labels`,
+    token,
+    "POST",
+    { labels: ["duplicate"] },
+  );
+
+  await githubRequest(
+    `/repos/${owner}/${repo}/issues/${issueNumber}`,
+    token,
+    "PATCH",
+    { state: "closed", state_reason: "duplicate" },
+  );
+}
+
+async function autoCloseDuplicates(): Promise<void> {
+  console.log("[DEBUG] Starting auto-close duplicates script");
+
+  const token = process.env.GITHUB_TOKEN;
+  if (!token) {
+    throw new Error("GITHUB_TOKEN environment variable is required");
+  }
+  console.log("[DEBUG] GitHub token found");
+
+  const owner = process.env.GITHUB_REPOSITORY_OWNER || "BerriAI";
+  const repo = process.env.GITHUB_REPOSITORY_NAME || "litellm";
+  console.log(`[DEBUG] Repository: ${owner}/${repo}`);
+
+  const threeDaysAgo = new Date();
+  threeDaysAgo.setDate(threeDaysAgo.getDate() - GRACE_PERIOD_DAYS);
+  console.log(
+    `[DEBUG] Checking for duplicate comments older than: ${threeDaysAgo.toISOString()}`,
+  );
+
+  // Only issues the detector labelled. Walking the whole open backlog would cost a
+  // comments request per issue, which on a four-figure backlog exhausts the Actions
+  // token's hourly rate limit for a handful of matches.
+  console.log(`[DEBUG] Fetching open issues labelled '${FLAG_LABEL}'...`);
+  const allIssues: GitHubIssue[] = [];
+  let page = 1;
+  const perPage = 100;
+
+  while (true) {
+    const pageIssues: GitHubIssue[] = await githubRequest(
+      `/repos/${owner}/${repo}/issues?state=open&labels=${FLAG_LABEL}&per_page=${perPage}&page=${page}`,
+      token,
+    );
+
+    if (pageIssues.length === 0) break;
+
+    allIssues.push(...pageIssues);
+    page++;
+
+    // Safety limit to avoid infinite loops
+    if (page > 20) break;
+  }
+
+  const issues = allIssues;
+  console.log(`[DEBUG] Found ${issues.length} flagged issues`);
+
+  let processedCount = 0;
+  let candidateCount = 0;
+
+  for (const issue of issues) {
+    processedCount++;
+    console.log(
+      `[DEBUG] Processing issue #${issue.number} (${processedCount}/${issues.length}): ${issue.title}`,
+    );
+
+    console.log(`[DEBUG] Fetching comments for issue #${issue.number}...`);
+    const comments: GitHubComment[] = await githubRequest(
+      `/repos/${owner}/${repo}/issues/${issue.number}/comments?per_page=100`,
+      token,
+    );
+    console.log(
+      `[DEBUG] Issue #${issue.number} has ${comments.length} comments`,
+    );
+
+    // The author filter matters: GitHub's "Quote reply" carries the HTML marker into
+    // a human comment, and treating that as a fresh notice restarts the clock.
+    const dupeComments = comments.filter(
+      (comment) =>
+        comment.body.includes(FLAG_MARKER) && comment.user.type === "Bot",
+    );
+    console.log(
+      `[DEBUG] Issue #${issue.number} has ${dupeComments.length} duplicate detection comments`,
+    );
+
+    if (dupeComments.length === 0) {
+      console.log(
+        `[DEBUG] Issue #${issue.number} - no duplicate comments found, skipping`,
+      );
+      continue;
+    }
+
+    const lastDupeComment = dupeComments[dupeComments.length - 1];
+    const dupeCommentDate = new Date(lastDupeComment.created_at);
+    console.log(
+      `[DEBUG] Issue #${issue.number} - most recent duplicate comment from: ${dupeCommentDate.toISOString()}`,
+    );
+
+    if (dupeCommentDate > threeDaysAgo) {
+      console.log(
+        `[DEBUG] Issue #${issue.number} - duplicate comment is too recent, skipping`,
+      );
+      continue;
+    }
+    console.log(
+      `[DEBUG] Issue #${issue.number} - duplicate comment is old enough (${Math.floor(
+        (Date.now() - dupeCommentDate.getTime()) / (1000 * 60 * 60 * 24),
+      )} days)`,
+    );
+
+    const commentsAfterDupe = comments.filter(
+      (comment) => new Date(comment.created_at) > dupeCommentDate,
+    );
+    console.log(
+      `[DEBUG] Issue #${issue.number} - ${commentsAfterDupe.length} comments after duplicate detection`,
+    );
+
+    if (commentsAfterDupe.length > 0) {
+      console.log(
+        `[DEBUG] Issue #${issue.number} - has activity after duplicate comment, skipping`,
+      );
+      continue;
+    }
+
+    console.log(
+      `[DEBUG] Issue #${issue.number} - checking reactions on duplicate comment...`,
+    );
+    const reactions: GitHubReaction[] = await githubRequest(
+      `/repos/${owner}/${repo}/issues/comments/${lastDupeComment.id}/reactions?per_page=100`,
+      token,
+    );
+    console.log(
+      `[DEBUG] Issue #${issue.number} - duplicate comment has ${reactions.length} reactions`,
+    );
+
+    const authorThumbsDown = reactions.some(
+      (reaction) =>
+        reaction.user.login === issue.user.login && reaction.content === "-1",
+    );
+    console.log(
+      `[DEBUG] Issue #${issue.number} - author thumbs down reaction: ${authorThumbsDown}`,
+    );
+
+    if (authorThumbsDown) {
+      console.log(
+        `[DEBUG] Issue #${issue.number} - author disagreed with duplicate detection, skipping`,
+      );
+      continue;
+    }
+
+    const duplicateIssueNumber = extractDuplicateIssueNumber(
+      lastDupeComment.body,
+      issue.number,
+    );
+    if (!duplicateIssueNumber) {
+      console.log(
+        `[DEBUG] Issue #${issue.number} - could not extract duplicate issue number from comment, skipping`,
+      );
+      continue;
+    }
+
+    if (duplicateIssueNumber > issue.number) {
+      console.log(
+        `[DEBUG] Issue #${issue.number} - only candidate #${duplicateIssueNumber} is newer, skipping`,
+      );
+      continue;
+    }
+
+    candidateCount++;
+    const issueUrl = `https://github.com/${owner}/${repo}/issues/${issue.number}`;
+
+    try {
+      console.log(
+        `[INFO] Auto-closing issue #${issue.number} as duplicate of #${duplicateIssueNumber}: ${issueUrl}`,
+      );
+      await closeIssueAsDuplicate(
+        owner,
+        repo,
+        issue.number,
+        duplicateIssueNumber,
+        token,
+      );
+      console.log(
+        `[SUCCESS] Successfully closed issue #${issue.number} as duplicate of #${duplicateIssueNumber}`,
+      );
+    } catch (error) {
+      console.error(
+        `[ERROR] Failed to close issue #${issue.number} as duplicate: ${error}`,
+      );
+    }
+  }
+
+  console.log(
+    `[DEBUG] Script completed. Processed ${processedCount} issues, found ${candidateCount} candidates for auto-close`,
+  );
+}
+
+autoCloseDuplicates().catch(console.error);
+
+// Make it a module
+export {};

From f4542d960511368eaeab50f68580a89aa13903e6 Mon Sep 17 00:00:00 2001
From: mubashir1osmani <mubashir.osmani777@gmail.com>
Date: Wed, 26 Aug 2026 15:26:02 -0400
Subject: [PATCH 06/19] fix(ci): pin the Bun runtime instead of tracking latest

The setup step ran `bun-version: latest`, carried over from the upstream layout,
and the step after it holds an issues: write token. A compromised Bun release
would have executed privileged in that job and could rewrite or close issues.

Pinned to 1.4.0, the release the passing runs already resolved to. setup-bun
takes no checksum input, so pinning the action by sha and the runtime by exact
version is as far as this can be hardened without hand-rolling the download.
---
 .github/workflows/auto-close-duplicates.yml | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/auto-close-duplicates.yml b/.github/workflows/auto-close-duplicates.yml
index 886aeaaa8e6..ff3b5eff7c2 100644
--- a/.github/workflows/auto-close-duplicates.yml
+++ b/.github/workflows/auto-close-duplicates.yml
@@ -23,7 +23,10 @@ jobs:
       - name: Setup Bun
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0 (sha-pinned)
         with:
-          bun-version: latest
+          # Exact version, never latest: the next step holds an issues: write token,
+          # so a compromised Bun release would run privileged here. setup-bun exposes
+          # no checksum input, so pinning the action and the version is the ceiling.
+          bun-version: "1.4.0"
 
       - name: Auto-close duplicate issues
         run: bun run scripts/auto-close-duplicates.ts

From f118511f5562eff67d0473346056d3bd6bbf06e1 Mon Sep 17 00:00:00 2001
From: mubashir1osmani <mubashir.osmani777@gmail.com>
Date: Wed, 26 Aug 2026 15:35:36 -0400
Subject: [PATCH 07/19] fix(ci): honour a duplicate-notice thumbs down from
 anyone

The notice tells every reader that a thumbs down keeps the issue open, but the
sweep only counted the reaction when it came from the issue author. A maintainer
or another affected user could follow the instruction exactly and still watch the
issue close, which made the notice a promise the sweep did not keep.

Any thumbs down now spares the issue. That buys back nothing an abuser did not
already have: a plain comment stops the clock for anyone, so restricting the
reaction only ever penalised people who did what they were told.

Drops the issue and reaction author fields, since nothing reads them now.
---
 scripts/auto-close-duplicates.ts | 16 +++++++---------
 1 file changed, 7 insertions(+), 9 deletions(-)

diff --git a/scripts/auto-close-duplicates.ts b/scripts/auto-close-duplicates.ts
index 1e94b009a1e..6e361af3e24 100644
--- a/scripts/auto-close-duplicates.ts
+++ b/scripts/auto-close-duplicates.ts
@@ -9,7 +9,6 @@ declare global {
 interface GitHubIssue {
   number: number;
   title: string;
-  user: { login: string };
   labels: { name: string }[];
 }
 
@@ -21,7 +20,6 @@ interface GitHubComment {
 }
 
 interface GitHubReaction {
-  user: { login: string };
   content: string;
 }
 
@@ -240,17 +238,17 @@ async function autoCloseDuplicates(): Promise<void> {
       `[DEBUG] Issue #${issue.number} - duplicate comment has ${reactions.length} reactions`,
     );
 
-    const authorThumbsDown = reactions.some(
-      (reaction) =>
-        reaction.user.login === issue.user.login && reaction.content === "-1",
-    );
+    // Any thumbs down, not just the author's. The notice tells every reader that a
+    // 👎 keeps the issue open, and anyone can already stop the clock by commenting,
+    // so honouring only the author would make the notice a lie without buying safety.
+    const thumbsDown = reactions.some((reaction) => reaction.content === "-1");
     console.log(
-      `[DEBUG] Issue #${issue.number} - author thumbs down reaction: ${authorThumbsDown}`,
+      `[DEBUG] Issue #${issue.number} - thumbs down reaction: ${thumbsDown}`,
     );
 
-    if (authorThumbsDown) {
+    if (thumbsDown) {
       console.log(
-        `[DEBUG] Issue #${issue.number} - author disagreed with duplicate detection, skipping`,
+        `[DEBUG] Issue #${issue.number} - someone disagreed with duplicate detection, skipping`,
       );
       continue;
     }

From 539bc8ef929e93603e2f99d8b5529cb5eb13a14c Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 14:02:07 -0700
Subject: [PATCH 08/19] fix(ci): close only identical-title duplicates, dry-run
 the sweep, reopen on reply

The merged detector's 0.6 flag threshold had become the close bar, and 6 of the 7 real flagged pairs at 85% or more were not duplicates. The sweep now closes only when an older open issue has the identical normalized title, measures the grace period from the latest bot notice, and leaves the issue open when anyone replies or gives the notice a thumbs down.

A reporter cannot reopen an issue the bot closed, so a reporter comment after the automatic close reopens it, drops the duplicate label, and asks for a human look. Manual dispatch defaults to a dry run and takes a grace_period_days input, the runner supplies the repository, the dead python closer is gone, and the decision core has bun tests on a PR-triggered job.
---
 .github/scripts/close_duplicate_issues.py    | 230 --------
 .github/workflows/auto-close-duplicates.yml  |  59 +-
 .github/workflows/check_duplicate_issues.yml |  14 +-
 scripts/auto-close-duplicates.test.ts        | 327 +++++++++++
 scripts/auto-close-duplicates.ts             | 543 +++++++++----------
 5 files changed, 647 insertions(+), 526 deletions(-)
 delete mode 100755 .github/scripts/close_duplicate_issues.py
 create mode 100644 scripts/auto-close-duplicates.test.ts

diff --git a/.github/scripts/close_duplicate_issues.py b/.github/scripts/close_duplicate_issues.py
deleted file mode 100755
index ec522af4f88..00000000000
--- a/.github/scripts/close_duplicate_issues.py
+++ /dev/null
@@ -1,230 +0,0 @@
-#!/usr/bin/env python3
-"""
-Detect and close duplicate GitHub issues using title similarity.
-
-Modes:
-  --scan           Compare all open issues against each other (batch)
-  --issue-number N Check a single issue against older open issues
-
-Requires the `gh` CLI to be authenticated.
-"""
-
-import argparse
-import difflib
-import json
-import re
-import subprocess
-import sys
-
-
-def normalize_title(title: str) -> str:
-    """Strip common prefixes, lowercase, and collapse whitespace."""
-    title = re.sub(
-        r"^\[?(bug|feature request|enhancement|question|docs)[:\]]?\s*",
-        "",
-        title,
-        flags=re.IGNORECASE,
-    )
-    return " ".join(title.lower().split())
-
-
-def gh(*args: str) -> str:
-    """Run a gh CLI command and return stdout."""
-    result = subprocess.run(
-        ["gh", *args],
-        capture_output=True,
-        text=True,
-        check=True,
-    )
-    return result.stdout
-
-
-def fetch_open_issues(repo: str | None) -> list[dict]:
-    """Fetch all open issues (excluding PRs) via gh api --paginate."""
-    if repo:
-        endpoint = (
-            f"repos/{repo}/issues?state=open&per_page=100&sort=created&direction=asc"
-        )
-    else:
-        endpoint = "repos/{owner}/{repo}/issues?state=open&per_page=100&sort=created&direction=asc"
-    cmd = ["api", "--paginate", endpoint]
-
-    raw = gh(*cmd)
-    # gh --paginate concatenates JSON arrays, so we may get multiple arrays
-    issues = []
-    for line in raw.strip().splitlines():
-        line = line.strip()
-        if not line:
-            continue
-        parsed = json.loads(line)
-        if isinstance(parsed, list):
-            issues.extend(parsed)
-        else:
-            issues.append(parsed)
-
-    # Filter out pull requests (they also appear in the issues endpoint)
-    return [i for i in issues if "pull_request" not in i]
-
-
-def close_as_duplicate(
-    issue_number: int, duplicate_of: int, repo: str | None, dry_run: bool
-) -> None:
-    """Close an issue as duplicate of another, adding a comment and label."""
-    repo_args = ["--repo", repo] if repo else []
-
-    if dry_run:
-        print(
-            f"  [DRY RUN] Would close #{issue_number} as duplicate of #{duplicate_of}"
-        )
-        return
-
-    # Add comment
-    comment_body = (
-        f"Closing as duplicate of #{duplicate_of}.\n\n"
-        "If you believe this is not a duplicate, please reopen and add context "
-        "explaining how this differs."
-    )
-    gh("issue", "comment", str(issue_number), "--body", comment_body, *repo_args)
-
-    # Add label
-    gh("issue", "edit", str(issue_number), "--add-label", "duplicate", *repo_args)
-
-    # Close with not_planned reason
-    gh(
-        "api",
-        f"repos/{repo or '{owner}/{repo}'}/issues/{issue_number}",
-        "-X",
-        "PATCH",
-        "-f",
-        "state=closed",
-        "-f",
-        "state_reason=not_planned",
-    )
-
-    print(f"  Closed #{issue_number} as duplicate of #{duplicate_of}")
-
-
-def find_duplicate(
-    issue: dict, candidates: list[dict], threshold: float
-) -> dict | None:
-    """Return the first candidate whose normalized title is above threshold."""
-    norm = normalize_title(issue["title"])
-    for candidate in candidates:
-        if candidate["number"] == issue["number"]:
-            continue
-        cand_norm = normalize_title(candidate["title"])
-        ratio = difflib.SequenceMatcher(None, norm, cand_norm).ratio()
-        if ratio >= threshold:
-            return candidate
-    return None
-
-
-def scan_all(
-    issues: list[dict], threshold: float, repo: str | None, dry_run: bool
-) -> int:
-    """Compare every issue against all older issues. Returns count of duplicates found."""
-    # Sort oldest first
-    issues.sort(key=lambda i: i["number"])
-    closed_count = 0
-
-    for idx, issue in enumerate(issues):
-        older = issues[:idx]
-        if not older:
-            continue
-        dup = find_duplicate(issue, older, threshold)
-        if dup:
-            ratio = difflib.SequenceMatcher(
-                None,
-                normalize_title(issue["title"]),
-                normalize_title(dup["title"]),
-            ).ratio()
-            print(
-                f"#{issue['number']}: \"{issue['title']}\"\n"
-                f"  -> duplicate of #{dup['number']}: \"{dup['title']}\" "
-                f"({ratio:.0%} similar)"
-            )
-            close_as_duplicate(issue["number"], dup["number"], repo, dry_run)
-            closed_count += 1
-
-    return closed_count
-
-
-def check_single(
-    issue_number: int,
-    issues: list[dict],
-    threshold: float,
-    repo: str | None,
-    dry_run: bool,
-) -> bool:
-    """Check a single issue against all older open issues. Returns True if duplicate found."""
-    target = None
-    for i in issues:
-        if i["number"] == issue_number:
-            target = i
-            break
-
-    if target is None:
-        print(f"Issue #{issue_number} not found among open issues.")
-        return False
-
-    older = [i for i in issues if i["number"] < issue_number]
-    dup = find_duplicate(target, older, threshold)
-    if dup:
-        ratio = difflib.SequenceMatcher(
-            None,
-            normalize_title(target["title"]),
-            normalize_title(dup["title"]),
-        ).ratio()
-        print(
-            f"#{target['number']}: \"{target['title']}\"\n"
-            f"  -> duplicate of #{dup['number']}: \"{dup['title']}\" "
-            f"({ratio:.0%} similar)"
-        )
-        close_as_duplicate(issue_number, dup["number"], repo, dry_run)
-        return True
-
-    print(f"#{issue_number}: no duplicate found above threshold {threshold}")
-    return False
-
-
-def main() -> None:
-    parser = argparse.ArgumentParser(
-        description="Detect and close duplicate GitHub issues"
-    )
-    mode = parser.add_mutually_exclusive_group(required=True)
-    mode.add_argument("--scan", action="store_true", help="Scan all open issues")
-    mode.add_argument("--issue-number", type=int, help="Check a single issue number")
-    parser.add_argument(
-        "--threshold", type=float, default=0.85, help="Similarity threshold (0-1)"
-    )
-    parser.add_argument(
-        "--close",
-        action="store_true",
-        help="Actually close duplicates (default is dry-run)",
-    )
-    parser.add_argument(
-        "--repo", type=str, help="Repository (owner/repo). Auto-detected if omitted."
-    )
-    args = parser.parse_args()
-
-    dry_run = not args.close
-
-    if dry_run:
-        print("=== DRY RUN MODE (pass --close to actually close issues) ===\n")
-
-    print("Fetching open issues...")
-    issues = fetch_open_issues(args.repo)
-    print(f"Found {len(issues)} open issues.\n")
-
-    if args.scan:
-        count = scan_all(issues, args.threshold, args.repo, dry_run)
-        print(f"\nTotal duplicates {'found' if dry_run else 'closed'}: {count}")
-    else:
-        found = check_single(
-            args.issue_number, issues, args.threshold, args.repo, dry_run
-        )
-        sys.exit(0 if found else 0)  # Always exit 0; finding no dup is not an error
-
-
-if __name__ == "__main__":
-    main()
diff --git a/.github/workflows/auto-close-duplicates.yml b/.github/workflows/auto-close-duplicates.yml
index ff3b5eff7c2..d8256917805 100644
--- a/.github/workflows/auto-close-duplicates.yml
+++ b/.github/workflows/auto-close-duplicates.yml
@@ -1,19 +1,33 @@
 name: Auto-close duplicate issues
-description: Auto-closes issues that are duplicates of existing issues
+
 on:
   schedule:
     - cron: "0 9 * * *"
   workflow_dispatch:
+    inputs:
+      dry_run:
+        description: Log which issues would close without closing anything
+        type: boolean
+        default: true
+      grace_period_days:
+        description: Days a duplicate notice must go unanswered before the close
+        type: number
+        default: 3
+  pull_request:
+    paths:
+      - .github/workflows/auto-close-duplicates.yml
+      - scripts/auto-close-duplicates.ts
+      - scripts/auto-close-duplicates.test.ts
+
+permissions: {}
 
 jobs:
-  auto-close-duplicates:
-    if: github.repository == 'BerriAI/litellm'
+  test:
+    if: github.event_name == 'pull_request'
     runs-on: ubuntu-latest
-    timeout-minutes: 10
+    timeout-minutes: 5
     permissions:
       contents: read
-      issues: write
-
     steps:
       - name: Checkout repository
         uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
@@ -21,16 +35,35 @@ jobs:
           persist-credentials: false
 
       - name: Setup Bun
-        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0 (sha-pinned)
+        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
-          # Exact version, never latest: the next step holds an issues: write token,
-          # so a compromised Bun release would run privileged here. setup-bun exposes
-          # no checksum input, so pinning the action and the version is the ceiling.
           bun-version: "1.4.0"
 
-      - name: Auto-close duplicate issues
+      - name: Test the sweep
+        run: bun test scripts/auto-close-duplicates.test.ts
+
+  sweep:
+    if: github.event_name != 'pull_request' && github.repository == 'BerriAI/litellm'
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
+        with:
+          persist-credentials: false
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
+        with:
+          # Exact version, never latest: the next step holds an issues: write token
+          bun-version: "1.4.0"
+
+      - name: Close unanswered duplicates, reopen ones the reporter answered
         run: bun run scripts/auto-close-duplicates.ts
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          GITHUB_REPOSITORY_OWNER: ${{ github.repository_owner }}
-          GITHUB_REPOSITORY_NAME: ${{ github.event.repository.name }}
+          DRY_RUN: ${{ inputs.dry_run == true }}
+          GRACE_PERIOD_DAYS: ${{ inputs.grace_period_days }}
diff --git a/.github/workflows/check_duplicate_issues.yml b/.github/workflows/check_duplicate_issues.yml
index 71d2a3b75eb..41ec43a1d9b 100644
--- a/.github/workflows/check_duplicate_issues.yml
+++ b/.github/workflows/check_duplicate_issues.yml
@@ -1,17 +1,19 @@
 name: Check Duplicate Issues
 
-# Flags newly opened issues that look like existing ones. Flagging only: the actual
-# close happens 3 days later in "Close Stale Duplicate Issues", and only if nobody
-# replied to the comment posted here. The HTML marker below is the handshake between
-# the two workflows, so keep it in the template.
+# Flagging only. "Auto-close duplicate issues" closes a flagged issue 3 days later,
+# and only when its title is identical to an older open issue and nobody replied.
+# The HTML marker below is the handshake between the two, so keep it in the template.
 
 on:
   issues:
     types: [opened, edited]
 
+permissions: {}
+
 jobs:
   check-duplicate:
     runs-on: ubuntu-latest
+    timeout-minutes: 5
     permissions:
       issues: write
       contents: read
@@ -29,7 +31,7 @@ jobs:
 
             This looks similar to:
             {{#issues}}
-            - #{{number}} - {{title}} ({{accuracy}}% similar)
+            - #{{number}} - {{title}}
             {{/issues}}
 
-            This issue will close automatically in 3 days unless someone responds. If it is a duplicate, please 👍 the existing issue and follow along there. If it is not, comment here or 👎 this comment and it stays open.
+            If this is a duplicate, add a thumbs-up reaction to the existing issue and follow along there. When the title is identical to an older open issue, this issue closes automatically in 3 days unless someone responds. If it is not a duplicate, comment here or add a thumbs-down reaction to this comment and it stays open.
diff --git a/scripts/auto-close-duplicates.test.ts b/scripts/auto-close-duplicates.test.ts
new file mode 100644
index 00000000000..6a7a3a507bd
--- /dev/null
+++ b/scripts/auto-close-duplicates.test.ts
@@ -0,0 +1,327 @@
+import { describe, expect, test } from "bun:test";
+
+import {
+  CLOSED_MARKER,
+  REOPEN_COMMENT,
+  candidateNumbers,
+  duplicateTarget,
+  normalizeTitle,
+  pendingNotice,
+  readConfig,
+  reopenTarget,
+  sweepClosedIssue,
+  sweepIssue,
+  type Comment,
+  type GitHubApi,
+  type Issue,
+  type SweepConfig,
+} from "./auto-close-duplicates";
+
+const NOW = new Date("2026-09-04T09:00:00Z");
+const DAY_MS = 24 * 60 * 60 * 1000;
+const daysAgo = (days: number): string => new Date(NOW.getTime() - days * DAY_MS).toISOString();
+
+const issue = (number: number, title: string, overrides: Partial<Issue> = {}): Issue => ({
+  number,
+  title,
+  state: "open",
+  user: { login: "reporter" },
+  ...overrides,
+});
+
+const notice = (candidates: readonly number[], createdAt: string, overrides: Partial<Comment> = {}): Comment => ({
+  id: 900,
+  body: `<!-- litellm:potential-duplicate candidates=${candidates.join(",")}, -->\n**Potential duplicate detected**`,
+  created_at: createdAt,
+  user: { type: "Bot", login: "github-actions[bot]" },
+  ...overrides,
+});
+
+const humanComment = (createdAt: string, body = "It is not the same thing", login = "reporter"): Comment => ({
+  id: 901,
+  body,
+  created_at: createdAt,
+  user: { type: "User", login },
+});
+
+const config: SweepConfig = { repo: "BerriAI/litellm", graceDays: 3, dryRun: false, now: NOW };
+
+describe("normalizeTitle", () => {
+  test("drops the template prefix, case, and punctuation", () => {
+    expect(normalizeTitle("[Bug]: Gemma 4-e4b fails on Vertex!")).toBe("gemma 4 e4b fails on vertex");
+    expect(normalizeTitle("[Feature]:   ")).toBe("");
+  });
+});
+
+describe("candidateNumbers", () => {
+  test("reads only the marker field, keeps older issues, sorted ascending and deduplicated", () => {
+    const body = "<!-- litellm:potential-duplicate candidates=40,10,30,10, -->\n- #1 - see #1 (100% similar)";
+    expect(candidateNumbers(body, 35)).toEqual([10, 30]);
+  });
+
+  test("returns nothing without the marker", () => {
+    expect(candidateNumbers("- #1 - looks like #1", 35)).toEqual([]);
+  });
+});
+
+describe("pendingNotice", () => {
+  test("waits out the grace period from the latest notice", () => {
+    const fresh = pendingNotice(issue(35, "t"), [notice([10], daysAgo(2.9))], config);
+    expect(fresh.kind).toBe("skip");
+    const aged = pendingNotice(issue(35, "t"), [notice([10], daysAgo(3.1))], config);
+    expect(aged.kind).toBe("pending");
+    const reposted = pendingNotice(
+      issue(35, "t"),
+      [notice([10], daysAgo(6)), notice([10], daysAgo(1), { id: 902 })],
+      config,
+    );
+    expect(reposted.kind).toBe("skip");
+  });
+
+  test("a zero-day grace period acts on the notice at once", () => {
+    const verdict = pendingNotice(issue(35, "t"), [notice([10], daysAgo(0.01))], { ...config, graceDays: 0 });
+    expect(verdict.kind).toBe("pending");
+  });
+
+  test("a human reply after the notice keeps the issue open, a bot reply does not", () => {
+    const human = pendingNotice(issue(35, "t"), [notice([10], daysAgo(5)), humanComment(daysAgo(4))], config);
+    expect(human).toEqual({ kind: "skip", reason: "someone replied after the notice" });
+    const bot = pendingNotice(
+      issue(35, "t"),
+      [notice([10], daysAgo(5)), { id: 903, body: "triage", created_at: daysAgo(4), user: { type: "Bot", login: "triage[bot]" } }],
+      config,
+    );
+    expect(bot.kind).toBe("pending");
+  });
+
+  test("a human quoting the marker is not a notice", () => {
+    const quoted = pendingNotice(issue(35, "t"), [notice([10], daysAgo(5), { user: { type: "User", login: "reporter" } })], config);
+    expect(quoted).toEqual({ kind: "skip", reason: "carries no duplicate notice" });
+  });
+
+  test("never closes an issue twice: a reopened issue is left alone", () => {
+    const reopened = pendingNotice(
+      issue(35, "t"),
+      [notice([10], daysAgo(9)), { id: 904, body: `Closed automatically\n\n${CLOSED_MARKER}`, created_at: daysAgo(5), user: { type: "Bot", login: "github-actions[bot]" } }],
+      config,
+    );
+    expect(reopened).toEqual({ kind: "skip", reason: "was reopened after an automatic close" });
+  });
+
+  test("skips pull requests and issues whose only candidates are newer", () => {
+    expect(pendingNotice(issue(35, "t", { pull_request: {} }), [notice([10], daysAgo(5))], config).kind).toBe("skip");
+    expect(pendingNotice(issue(35, "t"), [notice([40], daysAgo(5))], config)).toEqual({
+      kind: "skip",
+      reason: "no candidate is older than this issue",
+    });
+  });
+});
+
+describe("duplicateTarget", () => {
+  const reporter = issue(35, "[Bug]: Gemma 4-e4b fails on Vertex");
+
+  test("closes only against the earliest open issue with the identical normalized title", () => {
+    const verdict = duplicateTarget(
+      reporter,
+      [issue(10, "[Bug]: Gemma 4-e4n fails on Vertex"), issue(20, "[bug]: gemma 4-e4b fails on vertex"), issue(30, "[Bug]: Gemma 4-e4b fails on Vertex")],
+      [],
+    );
+    expect(verdict).toEqual({ kind: "close", duplicateOf: 20 });
+  });
+
+  test("a near miss in the title is not a duplicate", () => {
+    const verdict = duplicateTarget(reporter, [issue(10, "[Bug]: Gemma 4-e4n fails on Vertex")], []);
+    expect(verdict).toEqual({ kind: "skip", reason: "no older open issue has the identical title" });
+  });
+
+  test("bare template titles never match each other", () => {
+    const verdict = duplicateTarget(issue(35, "[Bug]: "), [issue(10, "[Bug]: ")], []);
+    expect(verdict.kind).toBe("skip");
+    expect(verdict.kind === "skip" && verdict.reason).toContain("too short");
+  });
+
+  test("a closed candidate or a pull request is never the target", () => {
+    expect(duplicateTarget(reporter, [issue(10, reporter.title, { state: "closed" })], []).kind).toBe("skip");
+    expect(duplicateTarget(reporter, [issue(10, reporter.title, { pull_request: {} })], []).kind).toBe("skip");
+  });
+
+  test("a thumbs down on the notice keeps the issue open", () => {
+    const verdict = duplicateTarget(reporter, [issue(10, reporter.title)], [{ content: "+1" }, { content: "-1" }]);
+    expect(verdict).toEqual({ kind: "skip", reason: "someone gave the notice a thumbs down" });
+  });
+});
+
+describe("sweepIssue", () => {
+  const reporter = issue(35, "[Bug]: Gemma 4-e4b fails on Vertex");
+  const original = issue(10, "[Bug]: Gemma 4-e4b fails on Vertex");
+
+  function fakeApi(): { readonly api: GitHubApi; readonly writes: readonly string[] } {
+    const writes: string[] = [];
+    const api: GitHubApi = {
+      request: async <T>(method: string, path: string, body?: object): Promise<T> => {
+        if (method !== "GET") {
+          writes.push(`${method} ${path} ${JSON.stringify(body)}`);
+          return {} as T;
+        }
+        if (path.startsWith("/repos/BerriAI/litellm/issues/35/comments")) {
+          return [notice([10], daysAgo(5))] as T;
+        }
+        if (path.startsWith("/repos/BerriAI/litellm/issues/comments/900/reactions")) {
+          return [] as T;
+        }
+        if (path === "/repos/BerriAI/litellm/issues/10") {
+          return original as T;
+        }
+        throw new Error(`unexpected GET ${path}`);
+      },
+    };
+    return { api, writes };
+  }
+
+  test("a dry run reports the close and writes nothing", async () => {
+    const { api, writes } = fakeApi();
+    const verdict = await sweepIssue(api, { ...config, dryRun: true }, reporter);
+    expect(verdict).toEqual({ kind: "close", duplicateOf: 10 });
+    expect(writes).toEqual([]);
+  });
+
+  test("a real run comments, labels, then closes with the duplicate reason", async () => {
+    const { api, writes } = fakeApi();
+    const verdict = await sweepIssue(api, config, reporter);
+    expect(verdict).toEqual({ kind: "close", duplicateOf: 10 });
+    expect(writes.map((write) => write.split(" ").slice(0, 2).join(" "))).toEqual([
+      "POST /repos/BerriAI/litellm/issues/35/comments",
+      "POST /repos/BerriAI/litellm/issues/35/labels",
+      "PATCH /repos/BerriAI/litellm/issues/35",
+    ]);
+    expect(writes[0]).toContain("duplicate of #10");
+    expect(writes[0]).toContain("unanswered for 3 days");
+    expect(writes[0]).toContain(CLOSED_MARKER);
+    expect(writes[1]).toContain('{"labels":["duplicate"]}');
+    expect(writes[2]).toContain('{"state":"closed","state_reason":"duplicate"}');
+  });
+});
+
+describe("reopenTarget", () => {
+  const closedByBot = (overrides: Partial<Issue> = {}): Issue =>
+    issue(35, "t", { state: "closed", closed_by: { type: "Bot" }, ...overrides });
+  const closeMarker = (createdAt: string): Comment => ({
+    id: 905,
+    body: `Closed automatically as a duplicate of #10.\n\n${CLOSED_MARKER}`,
+    created_at: createdAt,
+    user: { type: "Bot", login: "github-actions[bot]" },
+  });
+
+  test("a reporter reply after the automatic close reopens", () => {
+    const verdict = reopenTarget(closedByBot(), [closeMarker(daysAgo(2)), humanComment(daysAgo(1))]);
+    expect(verdict).toEqual({ kind: "reopen" });
+  });
+
+  test("an issue closed by a person stays closed", () => {
+    const verdict = reopenTarget(closedByBot({ closed_by: { type: "User" } }), [
+      closeMarker(daysAgo(2)),
+      humanComment(daysAgo(1)),
+    ]);
+    expect(verdict).toEqual({ kind: "skip", reason: "was closed by a person" });
+  });
+
+  test("without the automatic-close marker nothing reopens", () => {
+    const verdict = reopenTarget(closedByBot(), [humanComment(daysAgo(1))]);
+    expect(verdict).toEqual({ kind: "skip", reason: "carries no automatic-close marker" });
+  });
+
+  test("a maintainer reply alone does not reopen", () => {
+    const verdict = reopenTarget(closedByBot(), [
+      closeMarker(daysAgo(2)),
+      humanComment(daysAgo(1), "Confirmed duplicate", "maintainer"),
+    ]);
+    expect(verdict).toEqual({ kind: "skip", reason: "the reporter has not replied since the close" });
+  });
+
+  test("a reporter comment from before the close does not reopen", () => {
+    const verdict = reopenTarget(closedByBot(), [humanComment(daysAgo(3)), closeMarker(daysAgo(2))]);
+    expect(verdict).toEqual({ kind: "skip", reason: "the reporter has not replied since the close" });
+  });
+
+  test("a pull request never reopens", () => {
+    const verdict = reopenTarget(closedByBot({ pull_request: {} }), [closeMarker(daysAgo(2)), humanComment(daysAgo(1))]);
+    expect(verdict).toEqual({ kind: "skip", reason: "is a pull request" });
+  });
+});
+
+describe("sweepClosedIssue", () => {
+  function fakeApi(issueBody: Issue, comments: readonly Comment[]): { readonly api: GitHubApi; readonly writes: readonly string[] } {
+    const writes: string[] = [];
+    const api: GitHubApi = {
+      request: async <T>(method: string, path: string, body?: object): Promise<T> => {
+        if (method !== "GET") {
+          writes.push(`${method} ${path} ${JSON.stringify(body)}`);
+          return {} as T;
+        }
+        if (path.startsWith("/repos/BerriAI/litellm/issues/35/comments")) {
+          return comments as T;
+        }
+        if (path === "/repos/BerriAI/litellm/issues/35") {
+          return issueBody as T;
+        }
+        throw new Error(`unexpected GET ${path}`);
+      },
+    };
+    return { api, writes };
+  }
+
+  const closedByBot = issue(35, "t", { state: "closed", closed_by: { type: "Bot" } });
+  const closeMarker: Comment = {
+    id: 905,
+    body: `Closed automatically as a duplicate of #10.\n\n${CLOSED_MARKER}`,
+    created_at: daysAgo(2),
+    user: { type: "Bot", login: "github-actions[bot]" },
+  };
+
+  test("a real run unlabels, reopens, then explains", async () => {
+    const { api, writes } = fakeApi(closedByBot, [closeMarker, humanComment(daysAgo(1))]);
+    const verdict = await sweepClosedIssue(api, config, 35);
+    expect(verdict).toEqual({ kind: "reopen" });
+    expect(writes).toEqual([
+      "DELETE /repos/BerriAI/litellm/issues/35/labels/duplicate undefined",
+      'PATCH /repos/BerriAI/litellm/issues/35 {"state":"open"}',
+      `POST /repos/BerriAI/litellm/issues/35/comments {"body":"${REOPEN_COMMENT}"}`,
+    ]);
+  });
+
+  test("a dry run reports the reopen and writes nothing", async () => {
+    const { api, writes } = fakeApi(closedByBot, [closeMarker, humanComment(daysAgo(1))]);
+    const verdict = await sweepClosedIssue(api, { ...config, dryRun: true }, 35);
+    expect(verdict).toEqual({ kind: "reopen" });
+    expect(writes).toEqual([]);
+  });
+});
+
+describe("readConfig", () => {
+  test("defaults to a real run with a 3-day grace period", () => {
+    const parsed = readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "BerriAI/litellm" }, NOW);
+    expect(parsed).toEqual({ token: "t", repo: "BerriAI/litellm", graceDays: 3, dryRun: false, now: NOW });
+  });
+
+  test("honors DRY_RUN and GRACE_PERIOD_DAYS overrides", () => {
+    const parsed = readConfig(
+      { GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "o/r", DRY_RUN: "true", GRACE_PERIOD_DAYS: "0" },
+      NOW,
+    );
+    expect(parsed.dryRun).toBe(true);
+    expect(parsed.graceDays).toBe(0);
+  });
+
+  test("an empty GRACE_PERIOD_DAYS, as a schedule run renders it, means the default", () => {
+    const parsed = readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "o/r", GRACE_PERIOD_DAYS: "" }, NOW);
+    expect(parsed.graceDays).toBe(3);
+  });
+
+  test("refuses a missing token, a malformed repository, or a bad grace period", () => {
+    expect(() => readConfig({ GITHUB_REPOSITORY: "o/r" }, NOW)).toThrow("GITHUB_TOKEN");
+    expect(() => readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "litellm" }, NOW)).toThrow("owner/repo");
+    expect(() => readConfig({ GITHUB_TOKEN: "t", GITHUB_REPOSITORY: "o/r", GRACE_PERIOD_DAYS: "-1" }, NOW)).toThrow(
+      "GRACE_PERIOD_DAYS",
+    );
+  });
+});
diff --git a/scripts/auto-close-duplicates.ts b/scripts/auto-close-duplicates.ts
index 6e361af3e24..941f281efe6 100644
--- a/scripts/auto-close-duplicates.ts
+++ b/scripts/auto-close-duplicates.ts
@@ -1,306 +1,295 @@
 #!/usr/bin/env bun
 
-declare global {
-  var process: {
-    env: Record<string, string | undefined>;
-  };
+declare const process: { readonly env: Readonly<Record<string, string | undefined>> };
+
+export interface Issue {
+  readonly number: number;
+  readonly title: string;
+  readonly state: string;
+  readonly user: { readonly login: string };
+  readonly closed_by?: { readonly type: string } | null;
+  readonly pull_request?: unknown;
 }
 
-interface GitHubIssue {
-  number: number;
-  title: string;
-  labels: { name: string }[];
+export interface Comment {
+  readonly id: number;
+  readonly body: string;
+  readonly created_at: string;
+  readonly user: { readonly type: string; readonly login: string };
 }
 
-interface GitHubComment {
-  id: number;
-  body: string;
-  created_at: string;
-  user: { type: string };
+export interface Reaction {
+  readonly content: string;
 }
 
-interface GitHubReaction {
-  content: string;
+export interface GitHubApi {
+  readonly request: <T>(method: "GET" | "POST" | "PATCH" | "DELETE", path: string, body?: object) => Promise<T>;
 }
 
-const FLAG_LABEL = "potential-duplicate";
-const FLAG_MARKER = "<!-- litellm:potential-duplicate";
-const CANDIDATES = /<!-- litellm:potential-duplicate candidates=([\d,]*) -->/;
-const GRACE_PERIOD_DAYS = 3;
-
-async function githubRequest<T>(
-  endpoint: string,
-  token: string,
-  method: string = "GET",
-  body?: any,
-): Promise<T> {
-  const response = await fetch(`https://api.github.com${endpoint}`, {
-    method,
-    headers: {
-      Authorization: `Bearer ${token}`,
-      Accept: "application/vnd.github.v3+json",
-      "User-Agent": "auto-close-duplicates-script",
-      ...(body && { "Content-Type": "application/json" }),
-    },
-    ...(body && { body: JSON.stringify(body) }),
-  });
-
-  if (!response.ok) {
-    throw new Error(
-      `GitHub API request failed: ${response.status} ${response.statusText}`,
-    );
-  }
-
-  return response.json();
+export interface SweepConfig {
+  readonly repo: string;
+  readonly graceDays: number;
+  readonly dryRun: boolean;
+  readonly now: Date;
 }
 
-function extractDuplicateIssueNumber(
-  commentBody: string,
-  issueNumber: number,
-): number | null {
-  // Read candidates from the marker's digits-only field, never from the prose.
-  // Titles are user-controlled and are interpolated into this same comment, so
-  // scanning the body would let an issue titled "... see #1" redirect a closure
-  // onto an unrelated report.
-  const field = commentBody.match(CANDIDATES);
+export type NoticeVerdict =
+  | { readonly kind: "pending"; readonly notice: Comment; readonly candidates: readonly number[] }
+  | { readonly kind: "skip"; readonly reason: string };
+
+export type CloseVerdict =
+  | { readonly kind: "close"; readonly duplicateOf: number }
+  | { readonly kind: "skip"; readonly reason: string };
+
+export type ReopenVerdict =
+  | { readonly kind: "reopen" }
+  | { readonly kind: "skip"; readonly reason: string };
+
+export const FLAG_LABEL = "potential-duplicate";
+export const CLOSED_MARKER = "<!-- litellm:closed-as-duplicate -->";
+export const DEFAULT_GRACE_DAYS = 3;
+export const REOPEN_COMMENT =
+  "Reopened automatically: the reporter replied after the duplicate close, so this needs a human look.";
+const NOTICE_MARKER = /<!-- litellm:potential-duplicate candidates=([\d,]*) -->/;
+const MIN_TITLE_WORDS = 3;
+const PAGE_SIZE = 100;
+const DAY_MS = 24 * 60 * 60 * 1000;
+const REOPEN_LOOKBACK_DAYS = 30;
+
+const skip = (reason: string): { readonly kind: "skip"; readonly reason: string } => ({ kind: "skip", reason });
+
+export function normalizeTitle(title: string): string {
+  return title
+    .toLowerCase()
+    .replace(/^\s*\[[^\]]*\]\s*:?/, "")
+    .replace(/[^a-z0-9]+/g, " ")
+    .trim();
+}
+
+export function candidateNumbers(noticeBody: string, issueNumber: number): readonly number[] {
+  const field = noticeBody.match(NOTICE_MARKER);
   if (!field) {
-    return null;
+    return [];
   }
-
-  const candidates = field[1]
+  const older = field[1]
     .split(",")
     .filter((value) => value !== "")
     .map(Number)
-    .filter((value) => value !== issueNumber);
-
-  // The detector orders candidates by score, not age, so the first one listed can
-  // be newer than the original report. Duplicates fold into the earliest issue.
-  return candidates.length > 0 ? Math.min(...candidates) : null;
+    .filter((candidate) => candidate < issueNumber);
+  return [...new Set(older)].sort((a, b) => a - b);
 }
 
-async function closeIssueAsDuplicate(
-  owner: string,
-  repo: string,
+export function pendingNotice(
+  issue: Issue,
+  comments: readonly Comment[],
+  config: Pick<SweepConfig, "graceDays" | "now">,
+): NoticeVerdict {
+  if (issue.pull_request !== undefined) {
+    return skip("is a pull request");
+  }
+  if (comments.some((comment) => comment.body.includes(CLOSED_MARKER))) {
+    return skip("was reopened after an automatic close");
+  }
+  const notices = comments.filter((comment) => comment.user.type === "Bot" && NOTICE_MARKER.test(comment.body));
+  const notice = notices[notices.length - 1];
+  if (notice === undefined) {
+    return skip("carries no duplicate notice");
+  }
+  const noticeAt = new Date(notice.created_at);
+  const ageDays = (config.now.getTime() - noticeAt.getTime()) / DAY_MS;
+  if (ageDays < config.graceDays) {
+    return skip(`notice is ${ageDays.toFixed(1)} days old, grace period is ${config.graceDays}`);
+  }
+  if (comments.some((comment) => comment.user.type !== "Bot" && new Date(comment.created_at) > noticeAt)) {
+    return skip("someone replied after the notice");
+  }
+  const candidates = candidateNumbers(notice.body, issue.number);
+  if (candidates.length === 0) {
+    return skip("no candidate is older than this issue");
+  }
+  return { kind: "pending", notice, candidates };
+}
+
+export function duplicateTarget(
+  issue: Issue,
+  candidates: readonly Issue[],
+  reactions: readonly Reaction[],
+): CloseVerdict {
+  if (reactions.some((reaction) => reaction.content === "-1")) {
+    return skip("someone gave the notice a thumbs down");
+  }
+  const title = normalizeTitle(issue.title);
+  if (title.split(" ").length < MIN_TITLE_WORDS) {
+    return skip(`title "${issue.title}" is too short to match on`);
+  }
+  const original = candidates.find(
+    (candidate) =>
+      candidate.state === "open" && candidate.pull_request === undefined && normalizeTitle(candidate.title) === title,
+  );
+  if (original === undefined) {
+    return skip("no older open issue has the identical title");
+  }
+  return { kind: "close", duplicateOf: original.number };
+}
+
+export function reopenTarget(issue: Issue, comments: readonly Comment[]): ReopenVerdict {
+  if (issue.pull_request !== undefined) {
+    return skip("is a pull request");
+  }
+  if (issue.closed_by?.type !== "Bot") {
+    return skip("was closed by a person");
+  }
+  const marker = comments.find((comment) => comment.body.includes(CLOSED_MARKER));
+  if (marker === undefined) {
+    return skip("carries no automatic-close marker");
+  }
+  const markerAt = new Date(marker.created_at);
+  if (!comments.some((comment) => comment.user.login === issue.user.login && new Date(comment.created_at) > markerAt)) {
+    return skip("the reporter has not replied since the close");
+  }
+  return { kind: "reopen" };
+}
+
+export function closingComment(duplicateOf: number, graceDays: number): string {
+  return `Closed automatically as a duplicate of #${duplicateOf}. Its title is identical to that older open issue and the duplicate notice above went unanswered for ${graceDays} days. If this is wrong, comment here with how it differs from #${duplicateOf} and this issue will be reopened automatically within a day.
+
+${CLOSED_MARKER}`;
+}
+
+async function listAll<T>(api: GitHubApi, path: string, page = 1): Promise<readonly T[]> {
+  const separator = path.includes("?") ? "&" : "?";
+  const batch = await api.request<readonly T[]>("GET", `${path}${separator}per_page=${PAGE_SIZE}&page=${page}`);
+  return batch.length < PAGE_SIZE ? batch : [...batch, ...(await listAll<T>(api, path, page + 1))];
+}
+
+async function closeAsDuplicate(
+  api: GitHubApi,
+  config: SweepConfig,
   issueNumber: number,
-  duplicateOfNumber: number,
-  token: string,
+  duplicateOf: number,
 ): Promise<void> {
-  await githubRequest(
-    `/repos/${owner}/${repo}/issues/${issueNumber}/comments`,
-    token,
-    "POST",
-    {
-      body: `This issue has been automatically closed as a duplicate of #${duplicateOfNumber}.
+  const issuePath = `/repos/${config.repo}/issues/${issueNumber}`;
+  await api.request("POST", `${issuePath}/comments`, { body: closingComment(duplicateOf, config.graceDays) });
+  await api.request("POST", `${issuePath}/labels`, { labels: ["duplicate"] });
+  await api.request("PATCH", issuePath, { state: "closed", state_reason: "duplicate" });
+}
 
-The duplicate notice went unanswered for ${GRACE_PERIOD_DAYS} days. If this is incorrect, please re-open this issue and say how it differs from #${duplicateOfNumber}.
+async function reopenForReporter(api: GitHubApi, config: SweepConfig, issueNumber: number): Promise<void> {
+  const issuePath = `/repos/${config.repo}/issues/${issueNumber}`;
+  await api.request("DELETE", `${issuePath}/labels/duplicate`);
+  await api.request("PATCH", issuePath, { state: "open" });
+  await api.request("POST", `${issuePath}/comments`, { body: REOPEN_COMMENT });
+}
 
-<!-- litellm:closed-as-duplicate -->`,
+export async function sweepClosedIssue(api: GitHubApi, config: SweepConfig, issueNumber: number): Promise<ReopenVerdict> {
+  const issue = await api.request<Issue>("GET", `/repos/${config.repo}/issues/${issueNumber}`);
+  const comments = await listAll<Comment>(api, `/repos/${config.repo}/issues/${issueNumber}/comments`);
+  const verdict = reopenTarget(issue, comments);
+  if (verdict.kind === "reopen" && !config.dryRun) {
+    await reopenForReporter(api, config, issueNumber);
+  }
+  return verdict;
+}
+
+export async function sweepIssue(api: GitHubApi, config: SweepConfig, issue: Issue): Promise<CloseVerdict> {
+  const comments = await listAll<Comment>(api, `/repos/${config.repo}/issues/${issue.number}/comments`);
+  const pending = pendingNotice(issue, comments, config);
+  if (pending.kind === "skip") {
+    return pending;
+  }
+  const reactions = await listAll<Reaction>(api, `/repos/${config.repo}/issues/comments/${pending.notice.id}/reactions`);
+  const candidates = await Promise.all(
+    pending.candidates.map((candidate) => api.request<Issue>("GET", `/repos/${config.repo}/issues/${candidate}`)),
+  );
+  const verdict = duplicateTarget(issue, candidates, reactions);
+  if (verdict.kind === "close" && !config.dryRun) {
+    await closeAsDuplicate(api, config, issue.number, verdict.duplicateOf);
+  }
+  return verdict;
+}
+
+function describe(issue: Issue, verdict: CloseVerdict, dryRun: boolean): string {
+  if (verdict.kind === "skip") {
+    return `#${issue.number}: skipped, ${verdict.reason}`;
+  }
+  return `#${issue.number}: ${dryRun ? "would close" : "closed"} as a duplicate of #${verdict.duplicateOf}`;
+}
+
+export async function sweep(api: GitHubApi, config: SweepConfig): Promise<readonly CloseVerdict[]> {
+  const issues = await listAll<Issue>(api, `/repos/${config.repo}/issues?state=open&labels=${FLAG_LABEL}`);
+  console.log(`${issues.length} open issues carry the ${FLAG_LABEL} label in ${config.repo}${config.dryRun ? " (dry run)" : ""}`);
+  return issues.reduce<Promise<readonly CloseVerdict[]>>(async (previous, issue) => {
+    const verdicts = await previous;
+    const verdict = await sweepIssue(api, config, issue);
+    console.log(describe(issue, verdict, config.dryRun));
+    return [...verdicts, verdict];
+  }, Promise.resolve([]));
+}
+
+function describeReopen(issueNumber: number, verdict: ReopenVerdict, dryRun: boolean): string {
+  if (verdict.kind === "skip") {
+    return `#${issueNumber}: skipped, ${verdict.reason}`;
+  }
+  return `#${issueNumber}: ${dryRun ? "would reopen" : "reopened"} for the reporter's reply`;
+}
+
+export async function reopenSweep(api: GitHubApi, config: SweepConfig): Promise<readonly ReopenVerdict[]> {
+  const since = new Date(config.now.getTime() - REOPEN_LOOKBACK_DAYS * DAY_MS).toISOString();
+  const closedPath = `/repos/${config.repo}/issues?state=closed&labels=duplicate,${FLAG_LABEL}&since=${encodeURIComponent(since)}`;
+  const issues = await listAll<Issue>(api, closedPath);
+  console.log(`${issues.length} recently closed issues carry the duplicate and ${FLAG_LABEL} labels in ${config.repo}${config.dryRun ? " (dry run)" : ""}`);
+  return issues.reduce<Promise<readonly ReopenVerdict[]>>(async (previous, issue) => {
+    const verdicts = await previous;
+    const verdict = await sweepClosedIssue(api, config, issue.number);
+    console.log(describeReopen(issue.number, verdict, config.dryRun));
+    return [...verdicts, verdict];
+  }, Promise.resolve([]));
+}
+
+export function readConfig(env: Readonly<Record<string, string | undefined>>, now: Date): SweepConfig & { readonly token: string } {
+  const token = env.GITHUB_TOKEN;
+  const repo = env.GITHUB_REPOSITORY;
+  if (!token || !repo || !/^[\w.-]+\/[\w.-]+$/.test(repo)) {
+    throw new Error("GITHUB_TOKEN and GITHUB_REPOSITORY (owner/repo) are required");
+  }
+  const rawGraceDays = env.GRACE_PERIOD_DAYS?.trim();
+  const graceDays = rawGraceDays === undefined || rawGraceDays === "" ? DEFAULT_GRACE_DAYS : Number(rawGraceDays);
+  if (!Number.isFinite(graceDays) || graceDays < 0) {
+    throw new Error(`GRACE_PERIOD_DAYS must be a non-negative number, got "${env.GRACE_PERIOD_DAYS}"`);
+  }
+  return { token, repo, graceDays, dryRun: env.DRY_RUN === "true", now };
+}
+
+export function githubApi(token: string): GitHubApi {
+  return {
+    request: async <T>(method: "GET" | "POST" | "PATCH" | "DELETE", path: string, body?: object): Promise<T> => {
+      const response = await fetch(`https://api.github.com${path}`, {
+        method,
+        headers: {
+          Authorization: `Bearer ${token}`,
+          Accept: "application/vnd.github+json",
+          "X-GitHub-Api-Version": "2022-11-28",
+          "User-Agent": "litellm-auto-close-duplicates",
+          ...(body === undefined ? {} : { "Content-Type": "application/json" }),
+        },
+        body: body === undefined ? undefined : JSON.stringify(body),
+      });
+      if (!response.ok) {
+        throw new Error(`${method} ${path} failed: ${response.status} ${response.statusText}`);
+      }
+      return (await response.json()) as T;
     },
-  );
-
-  // Added on its own endpoint rather than in the PATCH below, because sending
-  // `labels` with the state change replaces every label on the issue.
-  await githubRequest(
-    `/repos/${owner}/${repo}/issues/${issueNumber}/labels`,
-    token,
-    "POST",
-    { labels: ["duplicate"] },
-  );
-
-  await githubRequest(
-    `/repos/${owner}/${repo}/issues/${issueNumber}`,
-    token,
-    "PATCH",
-    { state: "closed", state_reason: "duplicate" },
-  );
+  };
 }
 
-async function autoCloseDuplicates(): Promise<void> {
-  console.log("[DEBUG] Starting auto-close duplicates script");
-
-  const token = process.env.GITHUB_TOKEN;
-  if (!token) {
-    throw new Error("GITHUB_TOKEN environment variable is required");
-  }
-  console.log("[DEBUG] GitHub token found");
-
-  const owner = process.env.GITHUB_REPOSITORY_OWNER || "BerriAI";
-  const repo = process.env.GITHUB_REPOSITORY_NAME || "litellm";
-  console.log(`[DEBUG] Repository: ${owner}/${repo}`);
-
-  const threeDaysAgo = new Date();
-  threeDaysAgo.setDate(threeDaysAgo.getDate() - GRACE_PERIOD_DAYS);
+if (import.meta.main) {
+  const { token, ...config } = readConfig(process.env, new Date());
+  const api = githubApi(token);
+  const closeVerdicts = await sweep(api, config);
+  const reopenVerdicts = await reopenSweep(api, config);
+  const closed = closeVerdicts.filter((verdict) => verdict.kind === "close").length;
+  const reopened = reopenVerdicts.filter((verdict) => verdict.kind === "reopen").length;
   console.log(
-    `[DEBUG] Checking for duplicate comments older than: ${threeDaysAgo.toISOString()}`,
-  );
-
-  // Only issues the detector labelled. Walking the whole open backlog would cost a
-  // comments request per issue, which on a four-figure backlog exhausts the Actions
-  // token's hourly rate limit for a handful of matches.
-  console.log(`[DEBUG] Fetching open issues labelled '${FLAG_LABEL}'...`);
-  const allIssues: GitHubIssue[] = [];
-  let page = 1;
-  const perPage = 100;
-
-  while (true) {
-    const pageIssues: GitHubIssue[] = await githubRequest(
-      `/repos/${owner}/${repo}/issues?state=open&labels=${FLAG_LABEL}&per_page=${perPage}&page=${page}`,
-      token,
-    );
-
-    if (pageIssues.length === 0) break;
-
-    allIssues.push(...pageIssues);
-    page++;
-
-    // Safety limit to avoid infinite loops
-    if (page > 20) break;
-  }
-
-  const issues = allIssues;
-  console.log(`[DEBUG] Found ${issues.length} flagged issues`);
-
-  let processedCount = 0;
-  let candidateCount = 0;
-
-  for (const issue of issues) {
-    processedCount++;
-    console.log(
-      `[DEBUG] Processing issue #${issue.number} (${processedCount}/${issues.length}): ${issue.title}`,
-    );
-
-    console.log(`[DEBUG] Fetching comments for issue #${issue.number}...`);
-    const comments: GitHubComment[] = await githubRequest(
-      `/repos/${owner}/${repo}/issues/${issue.number}/comments?per_page=100`,
-      token,
-    );
-    console.log(
-      `[DEBUG] Issue #${issue.number} has ${comments.length} comments`,
-    );
-
-    // The author filter matters: GitHub's "Quote reply" carries the HTML marker into
-    // a human comment, and treating that as a fresh notice restarts the clock.
-    const dupeComments = comments.filter(
-      (comment) =>
-        comment.body.includes(FLAG_MARKER) && comment.user.type === "Bot",
-    );
-    console.log(
-      `[DEBUG] Issue #${issue.number} has ${dupeComments.length} duplicate detection comments`,
-    );
-
-    if (dupeComments.length === 0) {
-      console.log(
-        `[DEBUG] Issue #${issue.number} - no duplicate comments found, skipping`,
-      );
-      continue;
-    }
-
-    const lastDupeComment = dupeComments[dupeComments.length - 1];
-    const dupeCommentDate = new Date(lastDupeComment.created_at);
-    console.log(
-      `[DEBUG] Issue #${issue.number} - most recent duplicate comment from: ${dupeCommentDate.toISOString()}`,
-    );
-
-    if (dupeCommentDate > threeDaysAgo) {
-      console.log(
-        `[DEBUG] Issue #${issue.number} - duplicate comment is too recent, skipping`,
-      );
-      continue;
-    }
-    console.log(
-      `[DEBUG] Issue #${issue.number} - duplicate comment is old enough (${Math.floor(
-        (Date.now() - dupeCommentDate.getTime()) / (1000 * 60 * 60 * 24),
-      )} days)`,
-    );
-
-    const commentsAfterDupe = comments.filter(
-      (comment) => new Date(comment.created_at) > dupeCommentDate,
-    );
-    console.log(
-      `[DEBUG] Issue #${issue.number} - ${commentsAfterDupe.length} comments after duplicate detection`,
-    );
-
-    if (commentsAfterDupe.length > 0) {
-      console.log(
-        `[DEBUG] Issue #${issue.number} - has activity after duplicate comment, skipping`,
-      );
-      continue;
-    }
-
-    console.log(
-      `[DEBUG] Issue #${issue.number} - checking reactions on duplicate comment...`,
-    );
-    const reactions: GitHubReaction[] = await githubRequest(
-      `/repos/${owner}/${repo}/issues/comments/${lastDupeComment.id}/reactions?per_page=100`,
-      token,
-    );
-    console.log(
-      `[DEBUG] Issue #${issue.number} - duplicate comment has ${reactions.length} reactions`,
-    );
-
-    // Any thumbs down, not just the author's. The notice tells every reader that a
-    // 👎 keeps the issue open, and anyone can already stop the clock by commenting,
-    // so honouring only the author would make the notice a lie without buying safety.
-    const thumbsDown = reactions.some((reaction) => reaction.content === "-1");
-    console.log(
-      `[DEBUG] Issue #${issue.number} - thumbs down reaction: ${thumbsDown}`,
-    );
-
-    if (thumbsDown) {
-      console.log(
-        `[DEBUG] Issue #${issue.number} - someone disagreed with duplicate detection, skipping`,
-      );
-      continue;
-    }
-
-    const duplicateIssueNumber = extractDuplicateIssueNumber(
-      lastDupeComment.body,
-      issue.number,
-    );
-    if (!duplicateIssueNumber) {
-      console.log(
-        `[DEBUG] Issue #${issue.number} - could not extract duplicate issue number from comment, skipping`,
-      );
-      continue;
-    }
-
-    if (duplicateIssueNumber > issue.number) {
-      console.log(
-        `[DEBUG] Issue #${issue.number} - only candidate #${duplicateIssueNumber} is newer, skipping`,
-      );
-      continue;
-    }
-
-    candidateCount++;
-    const issueUrl = `https://github.com/${owner}/${repo}/issues/${issue.number}`;
-
-    try {
-      console.log(
-        `[INFO] Auto-closing issue #${issue.number} as duplicate of #${duplicateIssueNumber}: ${issueUrl}`,
-      );
-      await closeIssueAsDuplicate(
-        owner,
-        repo,
-        issue.number,
-        duplicateIssueNumber,
-        token,
-      );
-      console.log(
-        `[SUCCESS] Successfully closed issue #${issue.number} as duplicate of #${duplicateIssueNumber}`,
-      );
-    } catch (error) {
-      console.error(
-        `[ERROR] Failed to close issue #${issue.number} as duplicate: ${error}`,
-      );
-    }
-  }
-
-  console.log(
-    `[DEBUG] Script completed. Processed ${processedCount} issues, found ${candidateCount} candidates for auto-close`,
+    `${config.dryRun ? "Would close" : "Closed"} ${closed} of ${closeVerdicts.length} flagged issues, ${config.dryRun ? "would reopen" : "reopened"} ${reopened}`,
   );
 }
-
-autoCloseDuplicates().catch(console.error);
-
-// Make it a module
-export {};

From ed5761daef4ae17152446d182c860630c38b7268 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sat, 29 Aug 2026 14:28:47 -0700
Subject: [PATCH 09/19] fix(ci): keep earlier objections when the duplicate
 notice is re-posted

The detector fires on issue edits and posts a fresh notice each time, so the sweep now counts replies from the first notice on and a thumbs down on any notice
---
 scripts/auto-close-duplicates.test.ts | 29 +++++++++++++++++++++++----
 scripts/auto-close-duplicates.ts      | 23 ++++++++++++---------
 2 files changed, 39 insertions(+), 13 deletions(-)

diff --git a/scripts/auto-close-duplicates.test.ts b/scripts/auto-close-duplicates.test.ts
index 6a7a3a507bd..b49bf05cbc2 100644
--- a/scripts/auto-close-duplicates.test.ts
+++ b/scripts/auto-close-duplicates.test.ts
@@ -14,6 +14,7 @@ import {
   type Comment,
   type GitHubApi,
   type Issue,
+  type Reaction,
   type SweepConfig,
 } from "./auto-close-duplicates";
 
@@ -78,6 +79,15 @@ describe("pendingNotice", () => {
     expect(reposted.kind).toBe("skip");
   });
 
+  test("an objection posted before a re-posted notice still keeps the issue open", () => {
+    const verdict = pendingNotice(
+      issue(35, "t"),
+      [notice([10], daysAgo(10)), humanComment(daysAgo(7)), notice([10], daysAgo(4), { id: 902 })],
+      config,
+    );
+    expect(verdict).toEqual({ kind: "skip", reason: "someone replied after the notice" });
+  });
+
   test("a zero-day grace period acts on the notice at once", () => {
     const verdict = pendingNotice(issue(35, "t"), [notice([10], daysAgo(0.01))], { ...config, graceDays: 0 });
     expect(verdict.kind).toBe("pending");
@@ -155,7 +165,10 @@ describe("sweepIssue", () => {
   const reporter = issue(35, "[Bug]: Gemma 4-e4b fails on Vertex");
   const original = issue(10, "[Bug]: Gemma 4-e4b fails on Vertex");
 
-  function fakeApi(): { readonly api: GitHubApi; readonly writes: readonly string[] } {
+  function fakeApi(
+    comments: readonly Comment[] = [notice([10], daysAgo(5))],
+    reactionsByNotice: Readonly<Record<number, readonly Reaction[]>> = {},
+  ): { readonly api: GitHubApi; readonly writes: readonly string[] } {
     const writes: string[] = [];
     const api: GitHubApi = {
       request: async <T>(method: string, path: string, body?: object): Promise<T> => {
@@ -164,10 +177,11 @@ describe("sweepIssue", () => {
           return {} as T;
         }
         if (path.startsWith("/repos/BerriAI/litellm/issues/35/comments")) {
-          return [notice([10], daysAgo(5))] as T;
+          return comments as T;
         }
-        if (path.startsWith("/repos/BerriAI/litellm/issues/comments/900/reactions")) {
-          return [] as T;
+        const reactionsPath = path.match(/^\/repos\/BerriAI\/litellm\/issues\/comments\/(\d+)\/reactions/);
+        if (reactionsPath) {
+          return (reactionsByNotice[Number(reactionsPath[1])] ?? []) as T;
         }
         if (path === "/repos/BerriAI/litellm/issues/10") {
           return original as T;
@@ -185,6 +199,13 @@ describe("sweepIssue", () => {
     expect(writes).toEqual([]);
   });
 
+  test("a thumbs down on an earlier notice still keeps the issue open", async () => {
+    const { api, writes } = fakeApi([notice([10], daysAgo(9)), notice([10], daysAgo(5), { id: 902 })], { 900: [{ content: "-1" }] });
+    const verdict = await sweepIssue(api, config, reporter);
+    expect(verdict).toEqual({ kind: "skip", reason: "someone gave the notice a thumbs down" });
+    expect(writes).toEqual([]);
+  });
+
   test("a real run comments, labels, then closes with the duplicate reason", async () => {
     const { api, writes } = fakeApi();
     const verdict = await sweepIssue(api, config, reporter);
diff --git a/scripts/auto-close-duplicates.ts b/scripts/auto-close-duplicates.ts
index 941f281efe6..c595104d886 100644
--- a/scripts/auto-close-duplicates.ts
+++ b/scripts/auto-close-duplicates.ts
@@ -34,7 +34,7 @@ export interface SweepConfig {
 }
 
 export type NoticeVerdict =
-  | { readonly kind: "pending"; readonly notice: Comment; readonly candidates: readonly number[] }
+  | { readonly kind: "pending"; readonly notices: readonly Comment[]; readonly candidates: readonly number[] }
   | { readonly kind: "skip"; readonly reason: string };
 
 export type CloseVerdict =
@@ -91,23 +91,24 @@ export function pendingNotice(
     return skip("was reopened after an automatic close");
   }
   const notices = comments.filter((comment) => comment.user.type === "Bot" && NOTICE_MARKER.test(comment.body));
-  const notice = notices[notices.length - 1];
-  if (notice === undefined) {
+  const first = notices[0];
+  const latest = notices[notices.length - 1];
+  if (first === undefined || latest === undefined) {
     return skip("carries no duplicate notice");
   }
-  const noticeAt = new Date(notice.created_at);
-  const ageDays = (config.now.getTime() - noticeAt.getTime()) / DAY_MS;
+  const ageDays = (config.now.getTime() - new Date(latest.created_at).getTime()) / DAY_MS;
   if (ageDays < config.graceDays) {
     return skip(`notice is ${ageDays.toFixed(1)} days old, grace period is ${config.graceDays}`);
   }
-  if (comments.some((comment) => comment.user.type !== "Bot" && new Date(comment.created_at) > noticeAt)) {
+  const firstNoticeAt = new Date(first.created_at);
+  if (comments.some((comment) => comment.user.type !== "Bot" && new Date(comment.created_at) > firstNoticeAt)) {
     return skip("someone replied after the notice");
   }
-  const candidates = candidateNumbers(notice.body, issue.number);
+  const candidates = candidateNumbers(latest.body, issue.number);
   if (candidates.length === 0) {
     return skip("no candidate is older than this issue");
   }
-  return { kind: "pending", notice, candidates };
+  return { kind: "pending", notices, candidates };
 }
 
 export function duplicateTarget(
@@ -197,7 +198,11 @@ export async function sweepIssue(api: GitHubApi, config: SweepConfig, issue: Iss
   if (pending.kind === "skip") {
     return pending;
   }
-  const reactions = await listAll<Reaction>(api, `/repos/${config.repo}/issues/comments/${pending.notice.id}/reactions`);
+  const reactions = (
+    await Promise.all(
+      pending.notices.map((notice) => listAll<Reaction>(api, `/repos/${config.repo}/issues/comments/${notice.id}/reactions`)),
+    )
+  ).flat();
   const candidates = await Promise.all(
     pending.candidates.map((candidate) => api.request<Issue>("GET", `/repos/${config.repo}/issues/${candidate}`)),
   );

From 4fb1440747d704140beb558b54a21520aa3c57d3 Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Sun, 30 Aug 2026 07:40:56 +0000
Subject: [PATCH 10/19] docs(proxy): clarify spend semantics on /v2/user/info
 and /user/daily/activity

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../internal_user_endpoints.py                    | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/litellm/proxy/management_endpoints/internal_user_endpoints.py b/litellm/proxy/management_endpoints/internal_user_endpoints.py
index 9a98bdbb6b1..73e993b37a1 100644
--- a/litellm/proxy/management_endpoints/internal_user_endpoints.py
+++ b/litellm/proxy/management_endpoints/internal_user_endpoints.py
@@ -997,6 +997,13 @@ async def user_info_v2(
     This is the v2 replacement for /user/info, designed to avoid the "god endpoint" problem
     where the old endpoint loaded all keys and teams into memory.
 
+    Note on `spend`: this is the user's running budget counter, which is zeroed by the
+    budget reset job whenever `budget_reset_at` elapses (see `budget_duration`). It is NOT
+    lifetime or per-period historical spend. For historical spend over a date range, use
+    `/user/daily/activity` or `/user/daily/activity/aggregated`, which read immutable daily
+    spend records that are never reset. The two values are expected to diverge once a
+    budget reset has occurred within the queried period.
+
     Access control:
     - Proxy admins can query any user
     - Team admins can query users within their teams
@@ -2687,6 +2694,10 @@ async def get_user_daily_activity(
 
     Meant to optimize querying spend data for analytics for a user.
 
+    Reads immutable daily spend records, which are never affected by budget resets.
+    This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
+    is a running budget counter zeroed on every budget reset.
+
     Returns:
     (by date)
     - spend
@@ -2800,6 +2811,10 @@ async def get_user_daily_activity_aggregated(
     """
     Aggregated analytics for a user's daily activity without pagination.
     Returns the same response shape as the paginated endpoint with page metadata set to single-page.
+
+    Reads immutable daily spend records, which are never affected by budget resets.
+    This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
+    is a running budget counter zeroed on every budget reset.
     """
     from litellm.proxy.proxy_server import prisma_client
 

From 36c53e1288051aaef506f8bbfde0722aabc3fa3b Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Sun, 30 Aug 2026 07:46:45 +0000
Subject: [PATCH 11/19] chore(ui): regenerate schema.d.ts for updated endpoint
 descriptions

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 ui/litellm-dashboard/src/lib/http/schema.d.ts | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 137c67e837c..b4526834b98 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -16210,6 +16210,10 @@ export interface paths {
          *
          *     Meant to optimize querying spend data for analytics for a user.
          *
+         *     Reads immutable daily spend records, which are never affected by budget resets.
+         *     This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
+         *     is a running budget counter zeroed on every budget reset.
+         *
          *     Returns:
          *     (by date)
          *     - spend
@@ -16241,6 +16245,10 @@ export interface paths {
          * Get User Daily Activity Aggregated
          * @description Aggregated analytics for a user's daily activity without pagination.
          *     Returns the same response shape as the paginated endpoint with page metadata set to single-page.
+         *
+         *     Reads immutable daily spend records, which are never affected by budget resets.
+         *     This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
+         *     is a running budget counter zeroed on every budget reset.
          */
         get: operations["get_user_daily_activity_aggregated_user_daily_activity_aggregated_get"];
         put?: never;
@@ -21004,6 +21012,13 @@ export interface paths {
          *     This is the v2 replacement for /user/info, designed to avoid the "god endpoint" problem
          *     where the old endpoint loaded all keys and teams into memory.
          *
+         *     Note on `spend`: this is the user's running budget counter, which is zeroed by the
+         *     budget reset job whenever `budget_reset_at` elapses (see `budget_duration`). It is NOT
+         *     lifetime or per-period historical spend. For historical spend over a date range, use
+         *     `/user/daily/activity` or `/user/daily/activity/aggregated`, which read immutable daily
+         *     spend records that are never reset. The two values are expected to diverge once a
+         *     budget reset has occurred within the queried period.
+         *
          *     Access control:
          *     - Proxy admins can query any user
          *     - Team admins can query users within their teams

From cafdfda8ba586429dcaf0c6c1dcf7e7c5caf70a7 Mon Sep 17 00:00:00 2001
From: Cursor Agent <cursoragent@cursor.com>
Date: Mon, 31 Aug 2026 05:52:17 +0000
Subject: [PATCH 12/19] feat(cli): set ENABLE_TOOL_SEARCH=true for lite claude

Claude Code turns tool search off when ANTHROPIC_BASE_URL is a proxy.
lite claude, lite up, login --config-claude, and autoroute now force
ENABLE_TOOL_SEARCH=true so MCP tools stay deferred through the proxy

Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
---
 litellm/proxy/client/cli/README.md                |  6 +++---
 litellm/proxy/client/cli/commands/agents.py       |  7 ++++++-
 .../client/cli/commands/autoroute/settings.py     |  3 +++
 .../proxy/client/cli/commands/claude_settings.py  | 15 +++++++++++----
 .../proxy/client/cli/autoroute/test_commands.py   |  1 +
 .../proxy/client/cli/autoroute/test_settings.py   |  7 +++++++
 .../test_litellm/proxy/client/cli/test_agents.py  | 14 ++++++++++++++
 .../proxy/client/cli/test_auth_commands.py        |  1 +
 .../proxy/client/cli/test_claude_settings.py      |  1 +
 .../proxy/client/cli/test_up_commands.py          | 12 +++++++++++-
 10 files changed, 58 insertions(+), 9 deletions(-)

diff --git a/litellm/proxy/client/cli/README.md b/litellm/proxy/client/cli/README.md
index fe417396317..0da219df700 100644
--- a/litellm/proxy/client/cli/README.md
+++ b/litellm/proxy/client/cli/README.md
@@ -489,7 +489,7 @@ lite codex exec "summarize the repo"
 
 Each command resolves your LiteLLM key (logging in via SSO when none is stored and you are at a terminal; otherwise it expects `LITELLM_PROXY_API_KEY` or `--api-key`), checks the key against the proxy so bad credentials fail immediately instead of deep inside the agent, exports the environment variables the agent reads, then replaces itself with the agent process.
 
-The right variables are picked per agent. Claude Code gets `ANTHROPIC_BASE_URL` (the proxy root, so it appends `/v1/messages`) and `ANTHROPIC_AUTH_TOKEN`, with any stray `ANTHROPIC_API_KEY` cleared so the proxy token wins. Codex and OpenCode get `OPENAI_BASE_URL` (the proxy plus `/v1`) and `OPENAI_API_KEY`. Codex ignores `OPENAI_BASE_URL`, so it is additionally pointed at the proxy through a custom provider passed as `-c` config overrides (HTTP/SSE Responses transport, since the proxy does not speak the Responses WebSocket protocol).
+The right variables are picked per agent. Claude Code gets `ANTHROPIC_BASE_URL` (the proxy root, so it appends `/v1/messages`) and `ANTHROPIC_AUTH_TOKEN`, with any stray `ANTHROPIC_API_KEY` cleared so the proxy token wins, and `ENABLE_TOOL_SEARCH=true` so Claude Code keeps tool search on even though the base URL is a proxy rather than a first-party Anthropic host. Codex and OpenCode get `OPENAI_BASE_URL` (the proxy plus `/v1`) and `OPENAI_API_KEY`. Codex ignores `OPENAI_BASE_URL`, so it is additionally pointed at the proxy through a custom provider passed as `-c` config overrides (HTTP/SSE Responses transport, since the proxy does not speak the Responses WebSocket protocol).
 
 Options (these belong to the wrapper, so put them before the agent's own flags):
 
@@ -505,7 +505,7 @@ The credential is short-lived by design (default 24h, configurable via `LITELLM_
 
 ### Route Every Claude Code Session Through the Proxy
 
-`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
+`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL, `env.ENABLE_TOOL_SEARCH` to `true`, and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
 
 Two things need to already be true: you've run `lite login` (or `lite login --pkce`, whose key the helper renews on its own), since the apiKeyHelper depends on that stored token, and the proxy is already reachable, since `lite up` does not start one for you.
 
@@ -529,7 +529,7 @@ Cursor is not supported: it has no equivalent file-based config to hot-patch thi
 lite --base-url https://your-proxy.example.com login --config-claude
 ```
 
-It writes the same two settings `lite up` does, `env.ANTHROPIC_BASE_URL` and `apiKeyHelper`, but persistently: there is no backup, nothing to restore, and no foreground process to keep alive. Every other key in `~/.claude/settings.json` is preserved, the file is created if it does not exist, and it is written atomically with owner-only permissions. Plain `lite login` is unchanged; nothing happens to your Claude Code config unless you pass the flag.
+It writes the same settings `lite up` does, `env.ANTHROPIC_BASE_URL`, `env.ENABLE_TOOL_SEARCH`, and `apiKeyHelper`, but persistently: there is no backup, nothing to restore, and no foreground process to keep alive. Every other key in `~/.claude/settings.json` is preserved, the file is created if it does not exist, and it is written atomically with owner-only permissions. Plain `lite login` is unchanged; nothing happens to your Claude Code config unless you pass the flag.
 
 Because the credential is reached through `apiKeyHelper` rather than copied into the file, a later `lite login` refreshes it with no further action: Claude Code re-runs the helper on every request and picks up whatever token the most recent login stored. Nothing secret is written to `settings.json`.
 
diff --git a/litellm/proxy/client/cli/commands/agents.py b/litellm/proxy/client/cli/commands/agents.py
index e05e85ae483..4d3a441272e 100644
--- a/litellm/proxy/client/cli/commands/agents.py
+++ b/litellm/proxy/client/cli/commands/agents.py
@@ -13,6 +13,8 @@ from .auth import context_secret_vault, get_stored_api_key, login
 ANTHROPIC_BASE_URL_ENV: Final = "ANTHROPIC_BASE_URL"
 ANTHROPIC_AUTH_TOKEN_ENV: Final = "ANTHROPIC_AUTH_TOKEN"
 ANTHROPIC_API_KEY_ENV: Final = "ANTHROPIC_API_KEY"
+ENABLE_TOOL_SEARCH_ENV: Final = "ENABLE_TOOL_SEARCH"
+ENABLE_TOOL_SEARCH_VALUE: Final = "true"
 OPENAI_BASE_URL_ENV: Final = "OPENAI_BASE_URL"
 OPENAI_API_KEY_ENV: Final = "OPENAI_API_KEY"
 
@@ -61,7 +63,9 @@ def build_agent_env(
     Anthropic clients (Claude Code) append /v1/messages to ANTHROPIC_BASE_URL,
     so it stays the bare proxy root; OpenAI clients (Codex, OpenCode) expect the
     /v1 suffix on OPENAI_BASE_URL. ANTHROPIC_API_KEY is dropped so a stray
-    Anthropic key cannot win over the bearer token we set.
+    Anthropic key cannot win over the bearer token we set. ENABLE_TOOL_SEARCH is
+    forced on because Claude Code turns tool search off when ANTHROPIC_BASE_URL
+    is not a first-party Anthropic host.
     """
     env: Final = dict(base_env)
     root: Final = base_url.rstrip("/")
@@ -69,6 +73,7 @@ def build_agent_env(
         env[ANTHROPIC_BASE_URL_ENV] = root
         env[ANTHROPIC_AUTH_TOKEN_ENV] = api_key
         env.pop(ANTHROPIC_API_KEY_ENV, None)
+        env[ENABLE_TOOL_SEARCH_ENV] = ENABLE_TOOL_SEARCH_VALUE
     if PROFILE_OPENAI in profiles:
         env[OPENAI_BASE_URL_ENV] = root + "/v1"
         env[OPENAI_API_KEY_ENV] = api_key
diff --git a/litellm/proxy/client/cli/commands/autoroute/settings.py b/litellm/proxy/client/cli/commands/autoroute/settings.py
index 9fcb11a585b..8331d41ae78 100644
--- a/litellm/proxy/client/cli/commands/autoroute/settings.py
+++ b/litellm/proxy/client/cli/commands/autoroute/settings.py
@@ -9,6 +9,8 @@ API_KEY_HELPER_KEY: Final = "apiKeyHelper"
 ANTHROPIC_API_KEY_KEY: Final = "ANTHROPIC_API_KEY"
 ANTHROPIC_AUTH_TOKEN_KEY: Final = "ANTHROPIC_AUTH_TOKEN"
 ANTHROPIC_BASE_URL_KEY: Final = "ANTHROPIC_BASE_URL"
+ENABLE_TOOL_SEARCH_KEY: Final = "ENABLE_TOOL_SEARCH"
+ENABLE_TOOL_SEARCH_VALUE: Final = "true"
 # Force every one of Claude Code's own model tiers to request the auto-router by name.
 # Router's auto-router registry is keyed by the literal requested model string
 # (litellm/router.py:10711-10717) with no wildcard/pattern resolution, so a bare "*"
@@ -37,6 +39,7 @@ def merge_claude_settings_static_token(
         **base_env,
         ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
         ANTHROPIC_AUTH_TOKEN_KEY: auth_token,
+        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
         **{key: AUTOROUTER_MODEL_NAME for key in ANTHROPIC_DEFAULT_MODEL_ENV_KEYS},
     }
     env.pop(ANTHROPIC_API_KEY_KEY, None)
diff --git a/litellm/proxy/client/cli/commands/claude_settings.py b/litellm/proxy/client/cli/commands/claude_settings.py
index e18e5b1b7ee..6d88c69bbd4 100644
--- a/litellm/proxy/client/cli/commands/claude_settings.py
+++ b/litellm/proxy/client/cli/commands/claude_settings.py
@@ -21,6 +21,8 @@ ENV_KEY: Final = "env"
 API_KEY_HELPER_KEY: Final = "apiKeyHelper"
 ANTHROPIC_BASE_URL_KEY: Final = "ANTHROPIC_BASE_URL"
 ANTHROPIC_API_KEY_KEY: Final = "ANTHROPIC_API_KEY"
+ENABLE_TOOL_SEARCH_KEY: Final = "ENABLE_TOOL_SEARCH"
+ENABLE_TOOL_SEARCH_VALUE: Final = "true"
 
 CLAUDE_SETTINGS_PATH: Final = Path.home() / ".claude" / "settings.json"
 BACKUP_PATH: Final = Path.home() / ".litellm" / "claude_settings_backup.json"
@@ -68,16 +70,19 @@ def merge_claude_settings(
 ) -> dict[str, JsonValue]:
     """Return a new settings dict wired to route Claude Code through the proxy.
 
-    Only env.ANTHROPIC_BASE_URL and the top-level apiKeyHelper are overridden; a
-    stray env.ANTHROPIC_API_KEY is dropped so it cannot outrank the helper-issued
-    token (same reasoning as build_agent_env in agents.py). Every other key is
-    preserved untouched.
+    Only env.ANTHROPIC_BASE_URL, env.ENABLE_TOOL_SEARCH, and the top-level
+    apiKeyHelper are overridden; a stray env.ANTHROPIC_API_KEY is dropped so it
+    cannot outrank the helper-issued token (same reasoning as build_agent_env
+    in agents.py). ENABLE_TOOL_SEARCH is forced on because Claude Code turns
+    tool search off when ANTHROPIC_BASE_URL is not a first-party Anthropic host.
+    Every other key is preserved untouched.
     """
     raw_env: Final = settings.get(ENV_KEY, {})
     base_env: Final = raw_env if isinstance(raw_env, dict) else {}
     env: Final = {
         **{key: value for key, value in base_env.items() if key != ANTHROPIC_API_KEY_KEY},
         ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
+        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
     }
     return {**settings, ENV_KEY: env, API_KEY_HELPER_KEY: api_key_helper}
 
@@ -141,6 +146,8 @@ __all__ = (
     "ANTHROPIC_API_KEY_KEY",
     "ANTHROPIC_BASE_URL_KEY",
     "API_KEY_HELPER_KEY",
+    "ENABLE_TOOL_SEARCH_KEY",
+    "ENABLE_TOOL_SEARCH_VALUE",
     "AUTOROUTE_BACKUP_PATH",
     "BACKUP_PATH",
     "CLAUDE_SETTINGS_PATH",
diff --git a/tests/test_litellm/proxy/client/cli/autoroute/test_commands.py b/tests/test_litellm/proxy/client/cli/autoroute/test_commands.py
index 74bf1c95777..4a3b3ef22c4 100644
--- a/tests/test_litellm/proxy/client/cli/autoroute/test_commands.py
+++ b/tests/test_litellm/proxy/client/cli/autoroute/test_commands.py
@@ -157,6 +157,7 @@ class TestUpCommand:
         assert captured["settings"]["theme"] == "dark"
         assert captured["settings"]["env"]["ANTHROPIC_BASE_URL"] == "http://127.0.0.1:5483"
         assert captured["settings"]["env"]["ANTHROPIC_AUTH_TOKEN"] == "fixed-master-key"
+        assert captured["settings"]["env"]["ENABLE_TOOL_SEARCH"] == "true"
         assert "apiKeyHelper" not in captured["settings"]
         assert captured["settings_mode"] == 0o600
 
diff --git a/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py b/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py
index 40d3e7f2aee..ded8f0808a0 100644
--- a/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py
+++ b/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py
@@ -19,6 +19,13 @@ def test_sets_base_url_and_auth_token():
     merged = merge_claude_settings_static_token({}, "http://127.0.0.1:4000/", "token-abc")
     assert merged["env"]["ANTHROPIC_BASE_URL"] == "http://127.0.0.1:4000"
     assert merged["env"]["ANTHROPIC_AUTH_TOKEN"] == "token-abc"
+    assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
+
+
+def test_forces_tool_search_on():
+    settings = {"env": {"ENABLE_TOOL_SEARCH": "false"}}
+    merged = merge_claude_settings_static_token(settings, "http://127.0.0.1:4000", "token-abc")
+    assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
 
 
 def test_drops_stray_api_key():
diff --git a/tests/test_litellm/proxy/client/cli/test_agents.py b/tests/test_litellm/proxy/client/cli/test_agents.py
index 32dfb8d521d..d03c5d9db2d 100644
--- a/tests/test_litellm/proxy/client/cli/test_agents.py
+++ b/tests/test_litellm/proxy/client/cli/test_agents.py
@@ -77,9 +77,19 @@ class TestBuildAgentEnv:
         )
         assert env["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
         assert env["ANTHROPIC_AUTH_TOKEN"] == "sk-key"
+        assert env["ENABLE_TOOL_SEARCH"] == "true"
         assert "OPENAI_BASE_URL" not in env
         assert "OPENAI_API_KEY" not in env
 
+    def test_anthropic_profile_forces_tool_search_on(self):
+        env = build_agent_env(
+            {"ENABLE_TOOL_SEARCH": "false"},
+            "http://localhost:4000",
+            "sk-key",
+            frozenset({"anthropic"}),
+        )
+        assert env["ENABLE_TOOL_SEARCH"] == "true"
+
     def test_anthropic_profile_drops_existing_api_key(self):
         env = build_agent_env(
             {"ANTHROPIC_API_KEY": "real-key"},
@@ -96,6 +106,7 @@ class TestBuildAgentEnv:
         assert env["OPENAI_BASE_URL"] == "http://localhost:4000/v1"
         assert env["OPENAI_API_KEY"] == "sk-key"
         assert "ANTHROPIC_BASE_URL" not in env
+        assert "ENABLE_TOOL_SEARCH" not in env
 
     def test_both_profiles_set_everything(self):
         env = build_agent_env(
@@ -105,6 +116,7 @@ class TestBuildAgentEnv:
         assert env["OPENAI_BASE_URL"] == "http://localhost:4000/v1"
         assert env["ANTHROPIC_AUTH_TOKEN"] == "sk-key"
         assert env["OPENAI_API_KEY"] == "sk-key"
+        assert env["ENABLE_TOOL_SEARCH"] == "true"
 
     def test_preserves_unrelated_env_and_does_not_mutate_input(self):
         base = {"PATH": "/usr/bin", "ANTHROPIC_API_KEY": "real-key"}
@@ -201,6 +213,7 @@ class TestRunAgent:
         env = calls["env"]
         assert env["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
         assert env["ANTHROPIC_AUTH_TOKEN"] == "sk-key"
+        assert env["ENABLE_TOOL_SEARCH"] == "true"
         assert "ANTHROPIC_API_KEY" not in env
         assert "OPENAI_BASE_URL" not in env
 
@@ -218,6 +231,7 @@ class TestRunAgent:
         assert calls["env"]["OPENAI_BASE_URL"] == "http://localhost:4000/v1"
         assert calls["env"]["OPENAI_API_KEY"] == "sk-key"
         assert "ANTHROPIC_BASE_URL" not in calls["env"]
+        assert "ENABLE_TOOL_SEARCH" not in calls["env"]
 
     def test_codex_injects_proxy_provider_args_before_user_args(self):
         calls = {}
diff --git a/tests/test_litellm/proxy/client/cli/test_auth_commands.py b/tests/test_litellm/proxy/client/cli/test_auth_commands.py
index 85a4d90abf9..1d0a99b8e0a 100644
--- a/tests/test_litellm/proxy/client/cli/test_auth_commands.py
+++ b/tests/test_litellm/proxy/client/cli/test_auth_commands.py
@@ -1373,6 +1373,7 @@ class TestLoginConfigClaude:
         assert result.exit_code == 0
         written = json.loads(settings_path.read_text())
         assert written["env"]["ANTHROPIC_BASE_URL"] == "https://test.example.com"
+        assert written["env"]["ENABLE_TOOL_SEARCH"] == "true"
         assert written["apiKeyHelper"] == "/usr/local/bin/lite --base-url https://test.example.com auth print-token"
         assert "Configured Claude Code" in result.output
 
diff --git a/tests/test_litellm/proxy/client/cli/test_claude_settings.py b/tests/test_litellm/proxy/client/cli/test_claude_settings.py
index 9010fb4c022..898f9ab1ed7 100644
--- a/tests/test_litellm/proxy/client/cli/test_claude_settings.py
+++ b/tests/test_litellm/proxy/client/cli/test_claude_settings.py
@@ -48,6 +48,7 @@ class TestWriteClaudeSettings:
 
         written = json.loads(settings_path.read_text())
         assert written["env"]["ANTHROPIC_BASE_URL"] == "https://proxy.example.com"
+        assert written["env"]["ENABLE_TOOL_SEARCH"] == "true"
         assert written["apiKeyHelper"] == "/usr/local/bin/lite --base-url https://proxy.example.com auth print-token"
 
     def test_updates_an_existing_file_preserving_unrelated_settings(self, paths, lite_on_path):
diff --git a/tests/test_litellm/proxy/client/cli/test_up_commands.py b/tests/test_litellm/proxy/client/cli/test_up_commands.py
index 9958286884b..a1aa9eeeb8a 100644
--- a/tests/test_litellm/proxy/client/cli/test_up_commands.py
+++ b/tests/test_litellm/proxy/client/cli/test_up_commands.py
@@ -55,8 +55,14 @@ class TestMergeClaudeSettings:
         }
         merged = merge_claude_settings(settings, "http://localhost:4000/", "new-helper")
         assert merged["env"]["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
+        assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
         assert merged["apiKeyHelper"] == "new-helper"
 
+    def test_forces_tool_search_on(self):
+        settings = {"env": {"ENABLE_TOOL_SEARCH": "false"}}
+        merged = merge_claude_settings(settings, "http://localhost:4000", "helper")
+        assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
+
     def test_drops_stray_api_key(self):
         settings = {"env": {"ANTHROPIC_API_KEY": "leaked-key"}}
         merged = merge_claude_settings(settings, "http://localhost:4000", "helper")
@@ -64,7 +70,10 @@ class TestMergeClaudeSettings:
 
     def test_works_from_empty_settings(self):
         merged = merge_claude_settings({}, "http://localhost:4000", "helper")
-        assert merged["env"] == {"ANTHROPIC_BASE_URL": "http://localhost:4000"}
+        assert merged["env"] == {
+            "ANTHROPIC_BASE_URL": "http://localhost:4000",
+            "ENABLE_TOOL_SEARCH": "true",
+        }
         assert merged["apiKeyHelper"] == "helper"
 
     def test_does_not_mutate_input(self):
@@ -486,6 +495,7 @@ class TestUpCommand:
         assert captured["backup_existed"] is True
         assert captured["settings"]["theme"] == "dark"
         assert captured["settings"]["env"]["ANTHROPIC_BASE_URL"] == "http://localhost:4000"
+        assert captured["settings"]["env"]["ENABLE_TOOL_SEARCH"] == "true"
         assert captured["settings"]["apiKeyHelper"] == "/usr/local/bin/lite auth print-token"
         assert json.loads(settings_path.read_text()) == original
         assert not backup_path.exists()

From 096e016baf04fc833ce617025eace2f8937cefe2 Mon Sep 17 00:00:00 2001
From: Cursor Agent <cursoragent@cursor.com>
Date: Mon, 31 Aug 2026 05:58:00 +0000
Subject: [PATCH 13/19] fix(cli): sort claude_settings __all__ for ruff RUF022

Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
---
 litellm/proxy/client/cli/commands/claude_settings.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/litellm/proxy/client/cli/commands/claude_settings.py b/litellm/proxy/client/cli/commands/claude_settings.py
index 6d88c69bbd4..41ba25447e1 100644
--- a/litellm/proxy/client/cli/commands/claude_settings.py
+++ b/litellm/proxy/client/cli/commands/claude_settings.py
@@ -146,11 +146,11 @@ __all__ = (
     "ANTHROPIC_API_KEY_KEY",
     "ANTHROPIC_BASE_URL_KEY",
     "API_KEY_HELPER_KEY",
-    "ENABLE_TOOL_SEARCH_KEY",
-    "ENABLE_TOOL_SEARCH_VALUE",
     "AUTOROUTE_BACKUP_PATH",
     "BACKUP_PATH",
     "CLAUDE_SETTINGS_PATH",
+    "ENABLE_TOOL_SEARCH_KEY",
+    "ENABLE_TOOL_SEARCH_VALUE",
     "ENV_KEY",
     "SETTINGS_FILE_OWNERS",
     "ClaudeSettingsError",

From 43c64e24715cdd1aa7595544f1d4aba55f7c2512 Mon Sep 17 00:00:00 2001
From: Cursor Agent <cursoragent@cursor.com>
Date: Mon, 31 Aug 2026 06:00:45 +0000
Subject: [PATCH 14/19] fix(cli): keep an existing ENABLE_TOOL_SEARCH value

Default remains true so lite claude turns tool search back on through
a proxy. An explicit false or auto in the env or settings is left alone

Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>
---
 litellm/proxy/client/cli/README.md                 |  4 ++--
 litellm/proxy/client/cli/commands/agents.py        | 10 ++++++----
 .../client/cli/commands/autoroute/settings.py      |  2 +-
 .../proxy/client/cli/commands/claude_settings.py   | 14 +++++++-------
 .../proxy/client/cli/autoroute/test_settings.py    |  4 ++--
 tests/test_litellm/proxy/client/cli/test_agents.py |  4 ++--
 .../proxy/client/cli/test_up_commands.py           |  4 ++--
 7 files changed, 22 insertions(+), 20 deletions(-)

diff --git a/litellm/proxy/client/cli/README.md b/litellm/proxy/client/cli/README.md
index 0da219df700..3ddce35b53d 100644
--- a/litellm/proxy/client/cli/README.md
+++ b/litellm/proxy/client/cli/README.md
@@ -489,7 +489,7 @@ lite codex exec "summarize the repo"
 
 Each command resolves your LiteLLM key (logging in via SSO when none is stored and you are at a terminal; otherwise it expects `LITELLM_PROXY_API_KEY` or `--api-key`), checks the key against the proxy so bad credentials fail immediately instead of deep inside the agent, exports the environment variables the agent reads, then replaces itself with the agent process.
 
-The right variables are picked per agent. Claude Code gets `ANTHROPIC_BASE_URL` (the proxy root, so it appends `/v1/messages`) and `ANTHROPIC_AUTH_TOKEN`, with any stray `ANTHROPIC_API_KEY` cleared so the proxy token wins, and `ENABLE_TOOL_SEARCH=true` so Claude Code keeps tool search on even though the base URL is a proxy rather than a first-party Anthropic host. Codex and OpenCode get `OPENAI_BASE_URL` (the proxy plus `/v1`) and `OPENAI_API_KEY`. Codex ignores `OPENAI_BASE_URL`, so it is additionally pointed at the proxy through a custom provider passed as `-c` config overrides (HTTP/SSE Responses transport, since the proxy does not speak the Responses WebSocket protocol).
+The right variables are picked per agent. Claude Code gets `ANTHROPIC_BASE_URL` (the proxy root, so it appends `/v1/messages`) and `ANTHROPIC_AUTH_TOKEN`, with any stray `ANTHROPIC_API_KEY` cleared so the proxy token wins, and `ENABLE_TOOL_SEARCH=true` (unless you already set it) so Claude Code keeps tool search on even though the base URL is a proxy rather than a first-party Anthropic host. Codex and OpenCode get `OPENAI_BASE_URL` (the proxy plus `/v1`) and `OPENAI_API_KEY`. Codex ignores `OPENAI_BASE_URL`, so it is additionally pointed at the proxy through a custom provider passed as `-c` config overrides (HTTP/SSE Responses transport, since the proxy does not speak the Responses WebSocket protocol).
 
 Options (these belong to the wrapper, so put them before the agent's own flags):
 
@@ -505,7 +505,7 @@ The credential is short-lived by design (default 24h, configurable via `LITELLM_
 
 ### Route Every Claude Code Session Through the Proxy
 
-`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL, `env.ENABLE_TOOL_SEARCH` to `true`, and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
+`lite claude` wraps a single invocation, but `lite up` goes further: it patches `~/.claude/settings.json`, Claude Code's own config file, so that every Claude Code session started afterward -- from any terminal, launched normally with just `claude`, no wrapper needed -- routes through your LiteLLM proxy. It sets `env.ANTHROPIC_BASE_URL` to the proxy URL, `env.ENABLE_TOOL_SEARCH` to `true` when that key is missing, and `apiKeyHelper` to a `lite auth print-token` invocation, drops any stray static `ANTHROPIC_API_KEY` so the helper-issued token wins, and leaves every other setting in the file untouched. It backs up the original file before patching it.
 
 Two things need to already be true: you've run `lite login` (or `lite login --pkce`, whose key the helper renews on its own), since the apiKeyHelper depends on that stored token, and the proxy is already reachable, since `lite up` does not start one for you.
 
diff --git a/litellm/proxy/client/cli/commands/agents.py b/litellm/proxy/client/cli/commands/agents.py
index 4d3a441272e..45e05d353fb 100644
--- a/litellm/proxy/client/cli/commands/agents.py
+++ b/litellm/proxy/client/cli/commands/agents.py
@@ -63,9 +63,10 @@ def build_agent_env(
     Anthropic clients (Claude Code) append /v1/messages to ANTHROPIC_BASE_URL,
     so it stays the bare proxy root; OpenAI clients (Codex, OpenCode) expect the
     /v1 suffix on OPENAI_BASE_URL. ANTHROPIC_API_KEY is dropped so a stray
-    Anthropic key cannot win over the bearer token we set. ENABLE_TOOL_SEARCH is
-    forced on because Claude Code turns tool search off when ANTHROPIC_BASE_URL
-    is not a first-party Anthropic host.
+    Anthropic key cannot win over the bearer token we set. ENABLE_TOOL_SEARCH
+    defaults to true because Claude Code turns tool search off when
+    ANTHROPIC_BASE_URL is not a first-party Anthropic host; a value already in
+    the environment is left alone.
     """
     env: Final = dict(base_env)
     root: Final = base_url.rstrip("/")
@@ -73,7 +74,8 @@ def build_agent_env(
         env[ANTHROPIC_BASE_URL_ENV] = root
         env[ANTHROPIC_AUTH_TOKEN_ENV] = api_key
         env.pop(ANTHROPIC_API_KEY_ENV, None)
-        env[ENABLE_TOOL_SEARCH_ENV] = ENABLE_TOOL_SEARCH_VALUE
+        if ENABLE_TOOL_SEARCH_ENV not in env:
+            env[ENABLE_TOOL_SEARCH_ENV] = ENABLE_TOOL_SEARCH_VALUE
     if PROFILE_OPENAI in profiles:
         env[OPENAI_BASE_URL_ENV] = root + "/v1"
         env[OPENAI_API_KEY_ENV] = api_key
diff --git a/litellm/proxy/client/cli/commands/autoroute/settings.py b/litellm/proxy/client/cli/commands/autoroute/settings.py
index 8331d41ae78..60729b5410d 100644
--- a/litellm/proxy/client/cli/commands/autoroute/settings.py
+++ b/litellm/proxy/client/cli/commands/autoroute/settings.py
@@ -36,10 +36,10 @@ def merge_claude_settings_static_token(
     raw_env: Final = settings.get(ENV_KEY, {})
     base_env: Final = raw_env if isinstance(raw_env, dict) else {}
     env: Final[dict[str, JsonValue]] = {
+        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
         **base_env,
         ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
         ANTHROPIC_AUTH_TOKEN_KEY: auth_token,
-        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
         **{key: AUTOROUTER_MODEL_NAME for key in ANTHROPIC_DEFAULT_MODEL_ENV_KEYS},
     }
     env.pop(ANTHROPIC_API_KEY_KEY, None)
diff --git a/litellm/proxy/client/cli/commands/claude_settings.py b/litellm/proxy/client/cli/commands/claude_settings.py
index 41ba25447e1..ea1fa019c83 100644
--- a/litellm/proxy/client/cli/commands/claude_settings.py
+++ b/litellm/proxy/client/cli/commands/claude_settings.py
@@ -70,19 +70,19 @@ def merge_claude_settings(
 ) -> dict[str, JsonValue]:
     """Return a new settings dict wired to route Claude Code through the proxy.
 
-    Only env.ANTHROPIC_BASE_URL, env.ENABLE_TOOL_SEARCH, and the top-level
-    apiKeyHelper are overridden; a stray env.ANTHROPIC_API_KEY is dropped so it
-    cannot outrank the helper-issued token (same reasoning as build_agent_env
-    in agents.py). ENABLE_TOOL_SEARCH is forced on because Claude Code turns
-    tool search off when ANTHROPIC_BASE_URL is not a first-party Anthropic host.
-    Every other key is preserved untouched.
+    Only env.ANTHROPIC_BASE_URL and the top-level apiKeyHelper are overridden; a
+    stray env.ANTHROPIC_API_KEY is dropped so it cannot outrank the helper-issued
+    token (same reasoning as build_agent_env in agents.py). ENABLE_TOOL_SEARCH
+    defaults to true because Claude Code turns tool search off when
+    ANTHROPIC_BASE_URL is not a first-party Anthropic host; an existing value is
+    left alone. Every other key is preserved untouched.
     """
     raw_env: Final = settings.get(ENV_KEY, {})
     base_env: Final = raw_env if isinstance(raw_env, dict) else {}
     env: Final = {
+        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
         **{key: value for key, value in base_env.items() if key != ANTHROPIC_API_KEY_KEY},
         ANTHROPIC_BASE_URL_KEY: base_url.rstrip("/"),
-        ENABLE_TOOL_SEARCH_KEY: ENABLE_TOOL_SEARCH_VALUE,
     }
     return {**settings, ENV_KEY: env, API_KEY_HELPER_KEY: api_key_helper}
 
diff --git a/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py b/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py
index ded8f0808a0..87a33c79a79 100644
--- a/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py
+++ b/tests/test_litellm/proxy/client/cli/autoroute/test_settings.py
@@ -22,10 +22,10 @@ def test_sets_base_url_and_auth_token():
     assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
 
 
-def test_forces_tool_search_on():
+def test_preserves_existing_tool_search():
     settings = {"env": {"ENABLE_TOOL_SEARCH": "false"}}
     merged = merge_claude_settings_static_token(settings, "http://127.0.0.1:4000", "token-abc")
-    assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
+    assert merged["env"]["ENABLE_TOOL_SEARCH"] == "false"
 
 
 def test_drops_stray_api_key():
diff --git a/tests/test_litellm/proxy/client/cli/test_agents.py b/tests/test_litellm/proxy/client/cli/test_agents.py
index d03c5d9db2d..0191dad3d94 100644
--- a/tests/test_litellm/proxy/client/cli/test_agents.py
+++ b/tests/test_litellm/proxy/client/cli/test_agents.py
@@ -81,14 +81,14 @@ class TestBuildAgentEnv:
         assert "OPENAI_BASE_URL" not in env
         assert "OPENAI_API_KEY" not in env
 
-    def test_anthropic_profile_forces_tool_search_on(self):
+    def test_anthropic_profile_preserves_existing_tool_search(self):
         env = build_agent_env(
             {"ENABLE_TOOL_SEARCH": "false"},
             "http://localhost:4000",
             "sk-key",
             frozenset({"anthropic"}),
         )
-        assert env["ENABLE_TOOL_SEARCH"] == "true"
+        assert env["ENABLE_TOOL_SEARCH"] == "false"
 
     def test_anthropic_profile_drops_existing_api_key(self):
         env = build_agent_env(
diff --git a/tests/test_litellm/proxy/client/cli/test_up_commands.py b/tests/test_litellm/proxy/client/cli/test_up_commands.py
index a1aa9eeeb8a..053c90c36b4 100644
--- a/tests/test_litellm/proxy/client/cli/test_up_commands.py
+++ b/tests/test_litellm/proxy/client/cli/test_up_commands.py
@@ -58,10 +58,10 @@ class TestMergeClaudeSettings:
         assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
         assert merged["apiKeyHelper"] == "new-helper"
 
-    def test_forces_tool_search_on(self):
+    def test_preserves_existing_tool_search(self):
         settings = {"env": {"ENABLE_TOOL_SEARCH": "false"}}
         merged = merge_claude_settings(settings, "http://localhost:4000", "helper")
-        assert merged["env"]["ENABLE_TOOL_SEARCH"] == "true"
+        assert merged["env"]["ENABLE_TOOL_SEARCH"] == "false"
 
     def test_drops_stray_api_key(self):
         settings = {"env": {"ANTHROPIC_API_KEY": "leaked-key"}}

From 4291afbfa507e08a0ea270b77966b08e1a735b6b Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Mon, 31 Aug 2026 13:08:06 +0000
Subject: [PATCH 15/19] fix(registry): correct OpenAI preview shutdown dates,
 add whisper/transcribe and Bedrock/Vertex deprecation dates

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 ...odel_prices_and_context_window_backup.json | 29 +++++++++++--------
 model_prices_and_context_window.json          | 29 +++++++++++--------
 2 files changed, 34 insertions(+), 24 deletions(-)

diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json
index 9def855bc51..7071eaa0807 100644
--- a/litellm/model_prices_and_context_window_backup.json
+++ b/litellm/model_prices_and_context_window_backup.json
@@ -12315,7 +12315,8 @@
         "output_cost_per_token": 1e-05,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "claude-haiku-4-5-20251001": {
         "deprecation_date": "2026-10-15",
@@ -17767,7 +17768,8 @@
         "supports_response_schema": true,
         "supports_tool_choice": true,
         "supports_vision": true,
-        "prompt_cache_min_tokens": 1024
+        "prompt_cache_min_tokens": 1024,
+        "deprecation_date": "2027-01-08"
     },
     "eu.anthropic.claude-opus-4-20250514-v1:0": {
         "cache_creation_input_token_cost": 1.875e-05,
@@ -25327,7 +25329,7 @@
         "supports_vision": true
     },
     "gpt-4o-audio-preview": {
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 2.5e-06,
         "litellm_provider": "openai",
@@ -25650,7 +25652,7 @@
         "supports_vision": true
     },
     "gpt-4o-mini-audio-preview": {
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_token": 1.5e-07,
         "litellm_provider": "openai",
@@ -25688,7 +25690,7 @@
     "gpt-4o-mini-realtime-preview": {
         "cache_creation_input_audio_token_cost": 3e-07,
         "cache_read_input_token_cost": 3e-07,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_token": 6e-07,
         "litellm_provider": "openai",
@@ -25787,7 +25789,8 @@
         "output_cost_per_token": 5e-06,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "gpt-4o-mini-tts": {
         "input_cost_per_token": 2.5e-06,
@@ -25809,7 +25812,7 @@
     },
     "gpt-4o-realtime-preview": {
         "cache_read_input_token_cost": 2.5e-06,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 5e-06,
         "litellm_provider": "openai",
@@ -25828,7 +25831,7 @@
     },
     "gpt-4o-realtime-preview-2024-12-17": {
         "cache_read_input_token_cost": 2.5e-06,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 5e-06,
         "litellm_provider": "openai",
@@ -25847,7 +25850,7 @@
     },
     "gpt-4o-realtime-preview-2025-06-03": {
         "cache_read_input_token_cost": 2.5e-06,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 5e-06,
         "litellm_provider": "openai",
@@ -25926,7 +25929,8 @@
         "output_cost_per_token": 1e-05,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "gpt-image-1.5": {
         "cache_read_input_token_cost": 1.25e-06,
@@ -38321,7 +38325,7 @@
         "source": "https://docs.mistral.ai/capabilities/code_generation/"
     },
     "text-embedding-004": {
-        "deprecation_date": "2026-01-14",
+        "deprecation_date": "2027-04-01",
         "input_cost_per_character": 2.5e-08,
         "input_cost_per_token": 1e-07,
         "litellm_provider": "vertex_ai-embedding-models",
@@ -44100,7 +44104,8 @@
         "output_cost_per_second": 0.0001,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "xai/grok-3": {
         "cache_read_input_token_cost": 2e-07,
diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json
index 9def855bc51..7071eaa0807 100644
--- a/model_prices_and_context_window.json
+++ b/model_prices_and_context_window.json
@@ -12315,7 +12315,8 @@
         "output_cost_per_token": 1e-05,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "claude-haiku-4-5-20251001": {
         "deprecation_date": "2026-10-15",
@@ -17767,7 +17768,8 @@
         "supports_response_schema": true,
         "supports_tool_choice": true,
         "supports_vision": true,
-        "prompt_cache_min_tokens": 1024
+        "prompt_cache_min_tokens": 1024,
+        "deprecation_date": "2027-01-08"
     },
     "eu.anthropic.claude-opus-4-20250514-v1:0": {
         "cache_creation_input_token_cost": 1.875e-05,
@@ -25327,7 +25329,7 @@
         "supports_vision": true
     },
     "gpt-4o-audio-preview": {
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 2.5e-06,
         "litellm_provider": "openai",
@@ -25650,7 +25652,7 @@
         "supports_vision": true
     },
     "gpt-4o-mini-audio-preview": {
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_token": 1.5e-07,
         "litellm_provider": "openai",
@@ -25688,7 +25690,7 @@
     "gpt-4o-mini-realtime-preview": {
         "cache_creation_input_audio_token_cost": 3e-07,
         "cache_read_input_token_cost": 3e-07,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 1e-05,
         "input_cost_per_token": 6e-07,
         "litellm_provider": "openai",
@@ -25787,7 +25789,8 @@
         "output_cost_per_token": 5e-06,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "gpt-4o-mini-tts": {
         "input_cost_per_token": 2.5e-06,
@@ -25809,7 +25812,7 @@
     },
     "gpt-4o-realtime-preview": {
         "cache_read_input_token_cost": 2.5e-06,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 5e-06,
         "litellm_provider": "openai",
@@ -25828,7 +25831,7 @@
     },
     "gpt-4o-realtime-preview-2024-12-17": {
         "cache_read_input_token_cost": 2.5e-06,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 5e-06,
         "litellm_provider": "openai",
@@ -25847,7 +25850,7 @@
     },
     "gpt-4o-realtime-preview-2025-06-03": {
         "cache_read_input_token_cost": 2.5e-06,
-        "deprecation_date": "2027-01-20",
+        "deprecation_date": "2026-05-07",
         "input_cost_per_audio_token": 4e-05,
         "input_cost_per_token": 5e-06,
         "litellm_provider": "openai",
@@ -25926,7 +25929,8 @@
         "output_cost_per_token": 1e-05,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "gpt-image-1.5": {
         "cache_read_input_token_cost": 1.25e-06,
@@ -38321,7 +38325,7 @@
         "source": "https://docs.mistral.ai/capabilities/code_generation/"
     },
     "text-embedding-004": {
-        "deprecation_date": "2026-01-14",
+        "deprecation_date": "2027-04-01",
         "input_cost_per_character": 2.5e-08,
         "input_cost_per_token": 1e-07,
         "litellm_provider": "vertex_ai-embedding-models",
@@ -44100,7 +44104,8 @@
         "output_cost_per_second": 0.0001,
         "supported_endpoints": [
             "/v1/audio/transcriptions"
-        ]
+        ],
+        "deprecation_date": "2027-02-26"
     },
     "xai/grok-3": {
         "cache_read_input_token_cost": 2e-07,

From f079e4061bf986d6bb368da864f6c9dec22a0cac Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Mon, 31 Aug 2026 09:22:36 -0700
Subject: [PATCH 16/19] fix(proxy): deliver budget alerts on webhook-only
 alerting and accept ALERTING_WEBHOOK_URL (#38441)

* fix(proxy): deliver budget alerts on webhook-only alerting and accept ALERTING_WEBHOOK_URL

ProxyLogging.budget_alerts forwarded to the alerting pipeline only when
'slack' was in general_settings.alerting, so alerting: ['webhook'] plus
WEBHOOK_URL silently never delivered a budget alert (the config
/health/services?service=webhook exists to test). Forward when 'webhook'
is present too; SlackAlerting.send_alert already fans out per channel.

Also accept a provider-neutral ALERTING_WEBHOOK_URL env fallback for the
Slack-format channel (any Slack-compatible receiver works), mark it as a
sensitive var, and de-brand the admin UI alerting copy.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(ui): format settings.tsx with prettier

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* chore(ui): regenerate schema.d.ts for updated alerting description

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* ci: retrigger checks after ALERTING_WEBHOOK_URL docs merged

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: yassin <yassin@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 litellm/constants.py                          |  1 +
 .../SlackAlerting/slack_alerting.py           |  8 +--
 litellm/proxy/_types.py                       |  2 +-
 litellm/proxy/proxy_server.py                 |  3 +-
 litellm/proxy/utils.py                        |  6 +-
 .../SlackAlerting/test_slack_alerting.py      | 55 ++++++++++++++++++-
 .../test_slack_alerting_digest.py             | 17 ++++++
 .../utils/proxy_logging/test_alerting.py      | 29 ++++++++++
 .../src/components/settings.tsx               |  5 +-
 ui/litellm-dashboard/src/lib/http/schema.d.ts |  2 +-
 10 files changed, 116 insertions(+), 12 deletions(-)

diff --git a/litellm/constants.py b/litellm/constants.py
index cc6db6c10cc..9872783bfab 100644
--- a/litellm/constants.py
+++ b/litellm/constants.py
@@ -1728,6 +1728,7 @@ SENTRY_DENYLIST: Final = [
     "jwt_token",
     "private_key",
     "SLACK_WEBHOOK_URL",
+    "ALERTING_WEBHOOK_URL",
     "webhook_url",
     "LANGFUSE_SECRET_KEY",
     # Email Configuration
diff --git a/litellm/integrations/SlackAlerting/slack_alerting.py b/litellm/integrations/SlackAlerting/slack_alerting.py
index d7d06387d85..94d734546be 100644
--- a/litellm/integrations/SlackAlerting/slack_alerting.py
+++ b/litellm/integrations/SlackAlerting/slack_alerting.py
@@ -1485,9 +1485,9 @@ Model Info:
             elif self.default_webhook_url is not None:
                 _digest_webhook = self.default_webhook_url
             else:
-                _digest_webhook = os.getenv("SLACK_WEBHOOK_URL", None)
+                _digest_webhook = os.getenv("SLACK_WEBHOOK_URL") or os.getenv("ALERTING_WEBHOOK_URL")
             if _digest_webhook is None:
-                raise ValueError("Missing SLACK_WEBHOOK_URL from environment")
+                raise ValueError("Missing SLACK_WEBHOOK_URL / ALERTING_WEBHOOK_URL from environment")
 
             digest_key: Final = f"{alert_type_name_str}:{request_model or ''}:{api_base or ''}"
 
@@ -1516,10 +1516,10 @@ Model Info:
         elif self.default_webhook_url is not None:
             slack_webhook_url = self.default_webhook_url
         else:
-            slack_webhook_url = os.getenv("SLACK_WEBHOOK_URL", None)
+            slack_webhook_url = os.getenv("SLACK_WEBHOOK_URL") or os.getenv("ALERTING_WEBHOOK_URL")
 
         if slack_webhook_url is None:
-            raise ValueError("Missing SLACK_WEBHOOK_URL from environment")
+            raise ValueError("Missing SLACK_WEBHOOK_URL / ALERTING_WEBHOOK_URL from environment")
         payload: Final = {"text": formatted_message}
         headers: Final = {"Content-type": "application/json"}
 
diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py
index 5ba5e8fa1aa..0f2d97b8b1c 100644
--- a/litellm/proxy/_types.py
+++ b/litellm/proxy/_types.py
@@ -2541,7 +2541,7 @@ class ConfigGeneralSettings(LiteLLMPydanticObjectBase):
     )
     alerting: list | None = Field(
         None,
-        description="List of alerting integrations. Today, just slack - `alerting: ['slack']`",
+        description="List of alerting integrations - e.g. `alerting: ['slack', 'webhook', 'email']`. 'slack' posts Slack-format messages to any Slack-compatible webhook (Slack, Rocket.Chat, Mattermost); 'webhook' posts structured JSON budget alerts to WEBHOOK_URL",
     )
     alert_types: list[AlertType] | None = Field(
         None,
diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index 887716a383a..da2e09bd7f6 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -1363,7 +1363,7 @@ _OPENAPI_HTTP_METHODS: Final = {
 # the UI. Kept here at module scope to match the analogous descriptor
 # `is_secret` flags in litellm.proxy.config_resolvers and the
 # `_CACHE_SENSITIVE_FIELDS` constant in the cache endpoint file.
-_ALERTING_SENSITIVE_VARS: Final[set[str]] = {"SLACK_WEBHOOK_URL", "SMTP_PASSWORD"}
+_ALERTING_SENSITIVE_VARS: Final[set[str]] = {"ALERTING_WEBHOOK_URL", "SLACK_WEBHOOK_URL", "SMTP_PASSWORD"}
 
 
 def _strip_operation_id_method_suffix(operation_id: str) -> str:
@@ -16566,6 +16566,7 @@ async def create_config_audit_log(
 
 _EXTRA_SECRET_CALLBACK_ENV_VARS: Final = frozenset(
     {
+        "ALERTING_WEBHOOK_URL",
         "GALILEO_USERNAME",
         "GENERIC_LOGGER_HEADERS",
         "OTEL_HEADERS",
diff --git a/litellm/proxy/utils.py b/litellm/proxy/utils.py
index cf56fc0b1dd..051d36c4d0f 100644
--- a/litellm/proxy/utils.py
+++ b/litellm/proxy/utils.py
@@ -645,7 +645,7 @@ class ProxyLogging:
         self.max_parallel_request_limiter = _PROXY_MaxParallelRequestsHandler(self.internal_usage_cache)
         self.max_budget_limiter = _PROXY_MaxBudgetLimiter()
         self.cache_control_check = _PROXY_CacheControlCheck()
-        self.alerting: list | None = None
+        self.alerting: list[str] | None = None
         self.alerting_threshold: float = 300  # default to 5 min. threshold
         self.alert_types: list[AlertType] = DEFAULT_ALERT_TYPES
         self.alert_to_webhook_url: dict | None = None
@@ -2364,7 +2364,9 @@ class ProxyLogging:
             # do nothing if alerting is not switched on (unless it's a soft_budget alert with team-specific emails)
             return
 
-        if self.alerting is not None and ("slack" in self.alerting or "ms_teams" in self.alerting):
+        if self.alerting is not None and (
+            "slack" in self.alerting or "ms_teams" in self.alerting or "webhook" in self.alerting
+        ):
             if self.slack_alerting_instance is not None:
                 await self.slack_alerting_instance.budget_alerts(
                     type=type,
diff --git a/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py b/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py
index cfbd3e76a88..55e2dcdc270 100644
--- a/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py
+++ b/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py
@@ -12,7 +12,7 @@ import litellm
 from litellm.caching.caching import DualCache
 from litellm.integrations.SlackAlerting.slack_alerting import SlackAlerting
 from litellm.proxy._types import CallInfo, Litellm_EntityType
-from litellm.types.integrations.slack_alerting import SlackAlertingCacheKeys
+from litellm.types.integrations.slack_alerting import AlertType, SlackAlertingCacheKeys
 
 
 class TestSlackAlerting(unittest.TestCase):
@@ -366,3 +366,56 @@ async def test_scheduled_daily_report_threads_the_pod_lock_manager_through():
 
     _, kwargs = slack_alerting._run_scheduler_helper.await_args
     assert kwargs["pod_lock_manager"] is pod_lock_manager
+
+
+def _slack_alerting_with_env_resolution() -> SlackAlerting:
+    slack_alerting: Final = SlackAlerting(alerting=["slack"], internal_usage_cache=DualCache())
+    slack_alerting.periodic_started = True
+    return slack_alerting
+
+
+@pytest.mark.asyncio
+async def test_send_alert_falls_back_to_alerting_webhook_url_env(monkeypatch):
+    monkeypatch.delenv("SLACK_WEBHOOK_URL", raising=False)
+    monkeypatch.setenv("ALERTING_WEBHOOK_URL", "https://chat.example.com/hooks/abc")
+    slack_alerting: Final = _slack_alerting_with_env_resolution()
+
+    await slack_alerting.send_alert(
+        message="budget crossed",
+        level="High",
+        alert_type=AlertType.budget_alerts,
+        alerting_metadata={},
+    )
+
+    assert slack_alerting.log_queue[0]["url"] == "https://chat.example.com/hooks/abc"
+
+
+@pytest.mark.asyncio
+async def test_send_alert_prefers_slack_webhook_url_over_fallback(monkeypatch):
+    monkeypatch.setenv("SLACK_WEBHOOK_URL", "https://hooks.slack.com/services/T0/B0/X0")
+    monkeypatch.setenv("ALERTING_WEBHOOK_URL", "https://chat.example.com/hooks/abc")
+    slack_alerting: Final = _slack_alerting_with_env_resolution()
+
+    await slack_alerting.send_alert(
+        message="budget crossed",
+        level="High",
+        alert_type=AlertType.budget_alerts,
+        alerting_metadata={},
+    )
+
+    assert slack_alerting.log_queue[0]["url"] == "https://hooks.slack.com/services/T0/B0/X0"
+
+
+@pytest.mark.asyncio
+async def test_send_alert_raises_when_no_webhook_url_configured(monkeypatch):
+    monkeypatch.delenv("SLACK_WEBHOOK_URL", raising=False)
+    monkeypatch.delenv("ALERTING_WEBHOOK_URL", raising=False)
+    slack_alerting: Final = _slack_alerting_with_env_resolution()
+
+    with pytest.raises(ValueError, match="SLACK_WEBHOOK_URL / ALERTING_WEBHOOK_URL"):
+        await slack_alerting.send_alert(
+            message="budget crossed",
+            level="High",
+            alert_type=AlertType.budget_alerts,
+            alerting_metadata={},
+        )
diff --git a/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting_digest.py b/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting_digest.py
index edce5c5f3a2..d614823c0ef 100644
--- a/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting_digest.py
+++ b/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting_digest.py
@@ -79,6 +79,23 @@ class TestDigestMode(unittest.IsolatedAsyncioTestCase):
 
         self.assertEqual(len(self.slack_alerting.digest_buckets), 2)
 
+    async def test_digest_falls_back_to_alerting_webhook_url_env(self):
+        """With SLACK_WEBHOOK_URL unset, the digest entry resolves ALERTING_WEBHOOK_URL instead."""
+        env = {k: v for k, v in os.environ.items() if k != "SLACK_WEBHOOK_URL"}
+        env["ALERTING_WEBHOOK_URL"] = "https://chat.example.com/hooks/abc"
+        with unittest.mock.patch.dict(os.environ, env, clear=True):
+            await self.slack_alerting.send_alert(
+                message="`Requests are hanging`",
+                level="Medium",
+                alert_type=AlertType.llm_requests_hanging,
+                alerting_metadata={},
+                request_model="gemini-2.5-flash",
+                api_base="None",
+            )
+
+        bucket = list(self.slack_alerting.digest_buckets.values())[0]
+        self.assertEqual(bucket["webhook_url"], "https://chat.example.com/hooks/abc")
+
     async def test_non_digest_alert_goes_to_queue(self):
         """Alert types without digest enabled should go straight to the log queue."""
         message = "Budget exceeded"
diff --git a/tests/test_litellm/proxy/utils/proxy_logging/test_alerting.py b/tests/test_litellm/proxy/utils/proxy_logging/test_alerting.py
index cede859cb38..77c0f71dbf9 100644
--- a/tests/test_litellm/proxy/utils/proxy_logging/test_alerting.py
+++ b/tests/test_litellm/proxy/utils/proxy_logging/test_alerting.py
@@ -115,6 +115,35 @@ async def test_budget_alerts_slack_when_slack_alerting(proxy_logging):
     assert snapshot == {"type": "user_budget", "user_info_is_callinfo": True, "user_id": "u1"}
 
 
+@pytest.mark.asyncio
+async def test_budget_alerts_webhook_only_forwards_to_slack_alerting_instance(proxy_logging):
+    proxy_logging.alerting = ["webhook"]
+    captured: Dict[str, Any] = {}
+
+    async def fake_alert(**kwargs):
+        captured.update(kwargs)
+
+    proxy_logging.slack_alerting_instance = MagicMock(budget_alerts=fake_alert)
+    proxy_logging.email_logging_instance = None
+    await proxy_logging.budget_alerts(type="user_budget", user_info=_user_info())
+    snapshot = {
+        "type": captured["type"],
+        "user_info_is_callinfo": isinstance(captured["user_info"], CallInfo),
+        "user_id": captured["user_info"].user_id,
+    }
+    assert snapshot == {"type": "user_budget", "user_info_is_callinfo": True, "user_id": "u1"}
+
+
+@pytest.mark.asyncio
+async def test_budget_alerts_email_only_skips_slack_alerting_instance(proxy_logging):
+    proxy_logging.alerting = ["email"]
+    proxy_logging.slack_alerting_instance = MagicMock(budget_alerts=AsyncMock())
+    proxy_logging.email_logging_instance = MagicMock(budget_alerts=AsyncMock())
+    await proxy_logging.budget_alerts(type="user_budget", user_info=_user_info())
+    proxy_logging.slack_alerting_instance.budget_alerts.assert_not_called()
+    proxy_logging.email_logging_instance.budget_alerts.assert_called_once()
+
+
 @pytest.mark.asyncio
 async def test_budget_alerts_soft_budget_with_alert_emails_bypasses_global(proxy_logging):
     proxy_logging.alerting = None
diff --git a/ui/litellm-dashboard/src/components/settings.tsx b/ui/litellm-dashboard/src/components/settings.tsx
index 05e66985e6f..72da01d0918 100644
--- a/ui/litellm-dashboard/src/components/settings.tsx
+++ b/ui/litellm-dashboard/src/components/settings.tsx
@@ -522,7 +522,8 @@ const Settings: React.FC<SettingsPageProps> = ({ accessToken, userRole, userID,
           <TabsContent value="alerting-types" keepMounted>
             <Card className="p-6">
               <p className="my-2">
-                Alerts are only supported for Slack Webhook URLs. Get your webhook urls from{" "}
+                Alerts are sent to any Slack-compatible incoming webhook URL (Slack, Rocket.Chat, Mattermost, etc.). Get
+                Slack webhook urls from{" "}
                 <a href="https://api.slack.com/messaging/webhooks" target="_blank" style={{ color: "blue" }}>
                   here
                 </a>
@@ -532,7 +533,7 @@ const Settings: React.FC<SettingsPageProps> = ({ accessToken, userRole, userID,
                   <TableRow>
                     <TableHead></TableHead>
                     <TableHead></TableHead>
-                    <TableHead>Slack Webhook URL</TableHead>
+                    <TableHead>Webhook URL (Slack-compatible)</TableHead>
                   </TableRow>
                 </TableHeader>
 
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 137c67e837c..8a564a07489 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -25182,7 +25182,7 @@ export interface components {
             alert_types?: components["schemas"]["AlertType"][] | null;
             /**
              * Alerting
-             * @description List of alerting integrations. Today, just slack - `alerting: ['slack']`
+             * @description List of alerting integrations - e.g. `alerting: ['slack', 'webhook', 'email']`. 'slack' posts Slack-format messages to any Slack-compatible webhook (Slack, Rocket.Chat, Mattermost); 'webhook' posts structured JSON budget alerts to WEBHOOK_URL
              */
             alerting?: unknown[] | null;
             /**

From 30f32285103cea768c55f2f380a6422792363c91 Mon Sep 17 00:00:00 2001
From: yucheng-berri <yucheng@berri.ai>
Date: Mon, 31 Aug 2026 09:58:35 -0700
Subject: [PATCH 17/19] test(newrelic): cover static default_team_settings
 per-team routing (#38857)

* test(newrelic): cover static default_team_settings per-team routing

The dynamic POST /team/{team_id}/callback path for New Relic is tested, but
the static default_team_settings twin had no regression coverage. Add a test
that drives default_team_settings -> add_team_based_callbacks_from_config and
asserts the resolved trusted vars dispatch to BOTH the per-team metrics logger
(cost/usage) and the trace logger (LLM/agent spans), so a config-file customer
gets the same per-team routing as the API customer.

Also correct the /team/callback docstring: callback_name is a str validated
against the credential-capable callbacks, not a fixed langfuse/langsmith/gcs
Literal, and document the newrelic_api_key / newrelic_region vars.

* chore(ui): sync schema.d.ts with the /team/callback docstring

Regenerate the dashboard OpenAPI types for the add_team_callbacks description
change: callback_name is a validated str (not a langfuse/langsmith/gcs
Literal) and the newrelic_api_key / newrelic_region vars are documented.

* docs(newrelic): note LITELLM_OTEL_V2 prerequisite, trim test comments

Address review: team-scoped New Relic config is rejected with a 400 unless the
proxy runs with LITELLM_OTEL_V2=true, so document that in the /team/callback
endpoint and sync schema.d.ts. Drop the narrative setup comments in the new
test per the repo comment convention; the test name and docstring already say why.
---
 .../team_callback_endpoints.py                |  4 +-
 tests/proxy_unit_tests/test_proxy_utils.py    | 56 +++++++++++++++++++
 ui/litellm-dashboard/src/lib/http/schema.d.ts |  4 +-
 3 files changed, 62 insertions(+), 2 deletions(-)

diff --git a/litellm/proxy/management_endpoints/team_callback_endpoints.py b/litellm/proxy/management_endpoints/team_callback_endpoints.py
index 08346983f32..c2f5dbb4032 100644
--- a/litellm/proxy/management_endpoints/team_callback_endpoints.py
+++ b/litellm/proxy/management_endpoints/team_callback_endpoints.py
@@ -252,7 +252,7 @@ async def add_team_callbacks(
     Use this if if you want different teams to have different success/failure callbacks
 
     Parameters:
-    - callback_name (Literal["langfuse", "langsmith", "gcs"], required): The name of the callback to add
+    - callback_name (str, required): The name of the callback to add, e.g. "langfuse", "langsmith", "gcs", "newrelic". The value is validated against the callbacks that support team-scoped credentials
     - callback_type (Literal["success", "failure", "success_and_failure"], required): The type of callback to add. One of:
         - "success": Callback for successful LLM calls
         - "failure": Callback for failed LLM calls
@@ -268,6 +268,8 @@ async def add_team_callbacks(
         - langsmith_api_key: The API key for the Langsmith callback
         - langsmith_project: The project for the Langsmith callback
         - langsmith_base_url: The base URL for the Langsmith callback
+        - newrelic_api_key: The ingest license key for the team's New Relic account; routes both LLM/agent traces and cost metrics to that account. Requires the proxy to run with LITELLM_OTEL_V2=true, otherwise this callback is rejected with a 400
+        - newrelic_region: The New Relic region for the team's account ("us" or "eu"), riding the team's own key
 
     Example curl:
     ```
diff --git a/tests/proxy_unit_tests/test_proxy_utils.py b/tests/proxy_unit_tests/test_proxy_utils.py
index 3bde72ccd49..35de9961054 100644
--- a/tests/proxy_unit_tests/test_proxy_utils.py
+++ b/tests/proxy_unit_tests/test_proxy_utils.py
@@ -1317,6 +1317,62 @@ def test_proxy_config_state_post_init_callback_call(monkeypatch):
     assert config["litellm_settings"]["default_team_settings"][0]["team_id"] == "test"
 
 
+@pytest.mark.asyncio
+async def test_default_team_settings_newrelic_resolves_traces_and_metrics():
+    """Static `default_team_settings` is the config-file twin of POST /team/callback.
+
+    A team pinned to New Relic through `default_team_settings` must reach the
+    same two loggers the dynamic path does: the per-team metrics logger (cost
+    and usage) and the trace logger (LLM/agent spans). This proves the static
+    path resolves both, not just one, so the config-file customer gets the
+    same per-team routing as the API customer.
+    """
+    from litellm.litellm_core_utils.litellm_logging import Logging
+    from litellm.proxy.litellm_pre_call_utils import LiteLLMProxyRequestSetup
+    from litellm.proxy.proxy_server import ProxyConfig
+
+    pc = ProxyConfig()
+    pc.config = {
+        "litellm_settings": {
+            "default_team_settings": [
+                {
+                    "team_id": "team-a",
+                    "success_callback": ["newrelic"],
+                    "newrelic_api_key": "team-a-ingest-key",
+                    "newrelic_region": "eu",
+                }
+            ]
+        }
+    }
+
+    callback_metadata = LiteLLMProxyRequestSetup.add_team_based_callbacks_from_config(
+        team_id="team-a",
+        proxy_config=pc,
+    )
+
+    assert callback_metadata is not None
+    assert callback_metadata.success_callback == ["newrelic"]
+    assert callback_metadata.callback_vars == {
+        "newrelic_api_key": "team-a-ingest-key",
+        "newrelic_region": "eu",
+    }
+
+    logging_obj = Logging(
+        model="gpt-3.5-turbo",
+        messages=[{"role": "user", "content": "hi"}],
+        stream=False,
+        call_type="completion",
+        start_time=None,
+        litellm_call_id="static-nr-1",
+        function_id="static-nr-1",
+    )
+    logging_obj._trusted_callback_vars = tuple(callback_metadata.callback_vars.items())
+
+    resolved = logging_obj._resolve_dynamic_callback_string("newrelic")
+    resolved_names = {type(logger).__name__ for logger in resolved}
+    assert resolved_names == {"NewRelicMetricsLogger", "NewRelicLogger"}
+
+
 def test_proxy_config_state_get_config_state_error():
     """
     Ensures that get_config_state does not raise an error when the config is not a valid dictionary
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 8a564a07489..81a2fefe32e 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -15505,7 +15505,7 @@ export interface paths {
          *     Use this if if you want different teams to have different success/failure callbacks
          *
          *     Parameters:
-         *     - callback_name (Literal["langfuse", "langsmith", "gcs"], required): The name of the callback to add
+         *     - callback_name (str, required): The name of the callback to add, e.g. "langfuse", "langsmith", "gcs", "newrelic". The value is validated against the callbacks that support team-scoped credentials
          *     - callback_type (Literal["success", "failure", "success_and_failure"], required): The type of callback to add. One of:
          *         - "success": Callback for successful LLM calls
          *         - "failure": Callback for failed LLM calls
@@ -15521,6 +15521,8 @@ export interface paths {
          *         - langsmith_api_key: The API key for the Langsmith callback
          *         - langsmith_project: The project for the Langsmith callback
          *         - langsmith_base_url: The base URL for the Langsmith callback
+         *         - newrelic_api_key: The ingest license key for the team's New Relic account; routes both LLM/agent traces and cost metrics to that account. Requires the proxy to run with LITELLM_OTEL_V2=true, otherwise this callback is rejected with a 400
+         *         - newrelic_region: The New Relic region for the team's account ("us" or "eu"), riding the team's own key
          *
          *     Example curl:
          *     ```

From e938e89d138eec2ef20a998e09d200dfb44b71a1 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 31 Aug 2026 11:19:11 -0700
Subject: [PATCH 18/19] docs(proxy): account for budget rollover and daily
 upserts in spend wording

---
 .../internal_user_endpoints.py                | 25 +++++++++++--------
 ui/litellm-dashboard/src/lib/http/schema.d.ts | 25 +++++++++++--------
 2 files changed, 28 insertions(+), 22 deletions(-)

diff --git a/litellm/proxy/management_endpoints/internal_user_endpoints.py b/litellm/proxy/management_endpoints/internal_user_endpoints.py
index 73e993b37a1..6edb75eacb4 100644
--- a/litellm/proxy/management_endpoints/internal_user_endpoints.py
+++ b/litellm/proxy/management_endpoints/internal_user_endpoints.py
@@ -997,12 +997,13 @@ async def user_info_v2(
     This is the v2 replacement for /user/info, designed to avoid the "god endpoint" problem
     where the old endpoint loaded all keys and teams into memory.
 
-    Note on `spend`: this is the user's running budget counter, which is zeroed by the
-    budget reset job whenever `budget_reset_at` elapses (see `budget_duration`). It is NOT
+    Note on `spend`: this is the user's running budget counter, which the budget reset job
+    resets whenever `budget_reset_at` elapses (see `budget_duration`): to zero by default,
+    or to the overage above `max_budget` when `budget_rollover` is enabled. It is NOT
     lifetime or per-period historical spend. For historical spend over a date range, use
-    `/user/daily/activity` or `/user/daily/activity/aggregated`, which read immutable daily
-    spend records that are never reset. The two values are expected to diverge once a
-    budget reset has occurred within the queried period.
+    `/user/daily/activity` or `/user/daily/activity/aggregated`, which read daily spend
+    records that only ever accumulate and are never reset. The two values are expected to
+    diverge once a budget reset has occurred within the queried period.
 
     Access control:
     - Proxy admins can query any user
@@ -2694,9 +2695,10 @@ async def get_user_daily_activity(
 
     Meant to optimize querying spend data for analytics for a user.
 
-    Reads immutable daily spend records, which are never affected by budget resets.
-    This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
-    is a running budget counter zeroed on every budget reset.
+    Reads daily spend records that only ever accumulate and are never affected by budget
+    resets. Their total can legitimately exceed the `spend` field returned by
+    `/v2/user/info`, which is a running budget counter that every budget reset sets back
+    to zero (or to the overage above `max_budget` when `budget_rollover` is enabled).
 
     Returns:
     (by date)
@@ -2812,9 +2814,10 @@ async def get_user_daily_activity_aggregated(
     Aggregated analytics for a user's daily activity without pagination.
     Returns the same response shape as the paginated endpoint with page metadata set to single-page.
 
-    Reads immutable daily spend records, which are never affected by budget resets.
-    This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
-    is a running budget counter zeroed on every budget reset.
+    Reads daily spend records that only ever accumulate and are never affected by budget
+    resets. Their total can legitimately exceed the `spend` field returned by
+    `/v2/user/info`, which is a running budget counter that every budget reset sets back
+    to zero (or to the overage above `max_budget` when `budget_rollover` is enabled).
     """
     from litellm.proxy.proxy_server import prisma_client
 
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index b4526834b98..1599db10c7c 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -16210,9 +16210,10 @@ export interface paths {
          *
          *     Meant to optimize querying spend data for analytics for a user.
          *
-         *     Reads immutable daily spend records, which are never affected by budget resets.
-         *     This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
-         *     is a running budget counter zeroed on every budget reset.
+         *     Reads daily spend records that only ever accumulate and are never affected by budget
+         *     resets. Their total can legitimately exceed the `spend` field returned by
+         *     `/v2/user/info`, which is a running budget counter that every budget reset sets back
+         *     to zero (or to the overage above `max_budget` when `budget_rollover` is enabled).
          *
          *     Returns:
          *     (by date)
@@ -16246,9 +16247,10 @@ export interface paths {
          * @description Aggregated analytics for a user's daily activity without pagination.
          *     Returns the same response shape as the paginated endpoint with page metadata set to single-page.
          *
-         *     Reads immutable daily spend records, which are never affected by budget resets.
-         *     This can legitimately exceed the `spend` field returned by `/v2/user/info`, which
-         *     is a running budget counter zeroed on every budget reset.
+         *     Reads daily spend records that only ever accumulate and are never affected by budget
+         *     resets. Their total can legitimately exceed the `spend` field returned by
+         *     `/v2/user/info`, which is a running budget counter that every budget reset sets back
+         *     to zero (or to the overage above `max_budget` when `budget_rollover` is enabled).
          */
         get: operations["get_user_daily_activity_aggregated_user_daily_activity_aggregated_get"];
         put?: never;
@@ -21012,12 +21014,13 @@ export interface paths {
          *     This is the v2 replacement for /user/info, designed to avoid the "god endpoint" problem
          *     where the old endpoint loaded all keys and teams into memory.
          *
-         *     Note on `spend`: this is the user's running budget counter, which is zeroed by the
-         *     budget reset job whenever `budget_reset_at` elapses (see `budget_duration`). It is NOT
+         *     Note on `spend`: this is the user's running budget counter, which the budget reset job
+         *     resets whenever `budget_reset_at` elapses (see `budget_duration`): to zero by default,
+         *     or to the overage above `max_budget` when `budget_rollover` is enabled. It is NOT
          *     lifetime or per-period historical spend. For historical spend over a date range, use
-         *     `/user/daily/activity` or `/user/daily/activity/aggregated`, which read immutable daily
-         *     spend records that are never reset. The two values are expected to diverge once a
-         *     budget reset has occurred within the queried period.
+         *     `/user/daily/activity` or `/user/daily/activity/aggregated`, which read daily spend
+         *     records that only ever accumulate and are never reset. The two values are expected to
+         *     diverge once a budget reset has occurred within the queried period.
          *
          *     Access control:
          *     - Proxy admins can query any user

From 9f67a58198ec2b2d992a88812e2fae3c304bfd2e Mon Sep 17 00:00:00 2001
From: davida-ps <david.a@prompt.security>
Date: Mon, 31 Aug 2026 22:05:57 +0300
Subject: [PATCH 19/19] fix(guardrails): configure Prompt Security file timeout
 policy (#38083)

* fix(guardrails): fail open on Prompt Security file timeouts

* fix(guardrails): configure Prompt Security timeout policy
---
 .../prompt_security/__init__.py               |   1 +
 .../prompt_security/prompt_security.py        |  49 +++++++++
 .../guardrail_hooks/prompt_security.py        |   4 +
 .../test_prompt_security_guardrails.py        | 103 ++++++++++++++++--
 4 files changed, 148 insertions(+), 9 deletions(-)

diff --git a/litellm/proxy/guardrails/guardrail_hooks/prompt_security/__init__.py b/litellm/proxy/guardrails/guardrail_hooks/prompt_security/__init__.py
index fa1f9f3d36d..0aaba4016cd 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/prompt_security/__init__.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/prompt_security/__init__.py
@@ -20,6 +20,7 @@ def initialize_guardrail(litellm_params: "LitellmParams", guardrail: "Guardrail"
         guardrail_name=guardrail.get("guardrail_name", ""),
         event_hook=litellm_params.mode,
         default_on=litellm_params.default_on,
+        file_sanitization_fail_open=getattr(litellm_params, "file_sanitization_fail_open", None),
     )
     litellm.logging_callback_manager.add_litellm_callback(_prompt_security_callback)
 
diff --git a/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py b/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py
index 809d5e0fb31..84c4f118b00 100644
--- a/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py
+++ b/litellm/proxy/guardrails/guardrail_hooks/prompt_security/prompt_security.py
@@ -4,10 +4,12 @@ import os
 from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Final, Literal, Optional
 
+import httpx
 from fastapi import HTTPException
 from typing_extensions import ReadOnly, TypedDict
 
 from litellm._logging import verbose_proxy_logger
+from litellm.exceptions import Timeout as LiteLLMTimeout
 from litellm.integrations.custom_guardrail import (
     CustomGuardrail,
     log_guardrail_information,
@@ -24,6 +26,9 @@ if TYPE_CHECKING:
     from litellm.types.proxy.guardrails.guardrail_hooks.base import GuardrailConfigModel
 
 
+_SANITIZE_FILE_FAIL_OPEN_TIMEOUT_SECONDS: Final = 30.0
+
+
 class PromptSecurityGuardrailMissingSecrets(Exception):
     pass
 
@@ -63,6 +68,13 @@ class _SanitizeStatusResponse(TypedDict, total=False):
     metadata: ReadOnly[_SanitizeMetadata]
 
 
+class _SanitizeResult(TypedDict):
+    action: ReadOnly[str]
+    content: ReadOnly[str | None]
+    metadata: ReadOnly[_SanitizeMetadata]
+    violations: ReadOnly[Sequence[str]]
+
+
 class PromptSecurityGuardrail(CustomGuardrail):
     @classmethod
     def get_supported_event_hooks(cls) -> list[GuardrailEventHooks]:
@@ -79,6 +91,8 @@ class PromptSecurityGuardrail(CustomGuardrail):
         user: str | None = None,
         system_prompt: str | None = None,
         check_tool_results: bool | None = None,
+        file_sanitization_timeout: float = _SANITIZE_FILE_FAIL_OPEN_TIMEOUT_SECONDS,
+        file_sanitization_fail_open: bool | None = None,
         **kwargs,
     ):
         kwargs.setdefault("supported_event_hooks", list(self.get_supported_event_hooks()))
@@ -108,6 +122,8 @@ class PromptSecurityGuardrail(CustomGuardrail):
         # Configuration for file sanitization
         self.max_poll_attempts = 30  # Maximum number of polling attempts
         self.poll_interval = 2  # Seconds between polling attempts
+        self.file_sanitization_timeout = file_sanitization_timeout
+        self.file_sanitization_fail_open = file_sanitization_fail_open is not False
 
         super().__init__(**kwargs)
 
@@ -397,6 +413,39 @@ class PromptSecurityGuardrail(CustomGuardrail):
         Sanitize file content using Prompt Security API.
         Returns: dict with keys 'action', 'content', 'metadata'
         """
+        try:
+            return await asyncio.wait_for(
+                self._sanitize_file_content(file_data, filename, user_api_key_alias),
+                timeout=self.file_sanitization_timeout,
+            )
+        except (asyncio.TimeoutError, httpx.TimeoutException, LiteLLMTimeout) as exc:
+            if not self.file_sanitization_fail_open:
+                verbose_proxy_logger.error(
+                    "Prompt Security Guardrail: file sanitization for %s timed out with %s; failing closed",
+                    filename,
+                    type(exc).__name__,
+                )
+                raise HTTPException(status_code=408, detail="File sanitization timeout") from exc
+
+            verbose_proxy_logger.error(
+                "Prompt Security Guardrail: file sanitization for %s timed out with %s; failing open",
+                filename,
+                type(exc).__name__,
+            )
+            fail_open_result: Final[_SanitizeResult] = {
+                "action": "allow",
+                "content": None,
+                "metadata": {},
+                "violations": (),
+            }
+            return fail_open_result
+
+    async def _sanitize_file_content(
+        self,
+        file_data: bytes,
+        filename: str,
+        user_api_key_alias: str | None,
+    ) -> _SanitizeResult:
         headers: Final = {"APP-ID": self.api_key}
         if user_api_key_alias:
             headers["X-LiteLLM-Key-Alias"] = user_api_key_alias
diff --git a/litellm/types/proxy/guardrails/guardrail_hooks/prompt_security.py b/litellm/types/proxy/guardrails/guardrail_hooks/prompt_security.py
index 6e64f0f47a5..94f8161f44e 100644
--- a/litellm/types/proxy/guardrails/guardrail_hooks/prompt_security.py
+++ b/litellm/types/proxy/guardrails/guardrail_hooks/prompt_security.py
@@ -12,6 +12,10 @@ class PromptSecurityGuardrailConfigModel(GuardrailConfigModel):
         default=None,
         description="The API base for the Prompt Security guardrail. If not provided, the `PROMPT_SECURITY_API_BASE` environment variable is used.",
     )
+    file_sanitization_fail_open: bool = Field(
+        default=True,
+        description="Whether file sanitization timeouts allow the original file through instead of blocking the request.",
+    )
 
     @staticmethod
     def ui_friendly_name() -> str:
diff --git a/tests/test_litellm/proxy/guardrails/test_prompt_security_guardrails.py b/tests/test_litellm/proxy/guardrails/test_prompt_security_guardrails.py
index 26beaa78a46..ab4e15ff423 100644
--- a/tests/test_litellm/proxy/guardrails/test_prompt_security_guardrails.py
+++ b/tests/test_litellm/proxy/guardrails/test_prompt_security_guardrails.py
@@ -1,16 +1,16 @@
-from fastapi.exceptions import HTTPException
-from unittest.mock import patch, AsyncMock
-from httpx import Response, Request
+import asyncio
 import base64
+from unittest.mock import AsyncMock, patch
 
 import pytest
-
-from litellm.proxy.guardrails.guardrail_hooks.prompt_security.prompt_security import (
-    PromptSecurityGuardrailMissingSecrets,
-    PromptSecurityGuardrail,
-)
+from fastapi.exceptions import HTTPException
+from httpx import ReadTimeout, Request, Response
 
 import litellm
+from litellm.proxy.guardrails.guardrail_hooks.prompt_security.prompt_security import (
+    PromptSecurityGuardrail,
+    PromptSecurityGuardrailMissingSecrets,
+)
 from litellm.proxy.guardrails.init_guardrails import init_guardrails_v2
 
 
@@ -30,6 +30,7 @@ def test_prompt_security_guard_config(monkeypatch: pytest.MonkeyPatch):
                     "guardrail": "prompt_security",
                     "mode": "during_call",
                     "default_on": True,
+                    "file_sanitization_fail_open": False,
                 },
             }
         ],
@@ -41,6 +42,10 @@ def test_prompt_security_guard_config(monkeypatch: pytest.MonkeyPatch):
     assert registered[0].guardrail_name == "prompt_security"
     assert registered[0].default_on is True
     assert registered[0].event_hook == "during_call"
+    assert registered[0].file_sanitization_fail_open is False
+    config_model = registered[0].get_config_model()
+    assert config_model is not None
+    assert config_model().file_sanitization_fail_open is True
 
 
 def test_prompt_security_guard_config_no_api_key(monkeypatch: pytest.MonkeyPatch):
@@ -374,6 +379,86 @@ async def test_file_sanitization(monkeypatch: pytest.MonkeyPatch):
     assert result is not None
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "timeout",
+    (
+        litellm.Timeout(
+            message="Prompt Security upload timed out",
+            model="default-model-name",
+            llm_provider="litellm-httpx-handler",
+        ),
+        ReadTimeout(
+            "Prompt Security poll timed out",
+            request=Request(method="GET", url="https://test.prompt.security/api/sanitizeFile"),
+        ),
+    ),
+    ids=("litellm", "httpx"),
+)
+@pytest.mark.parametrize("fail_open", (True, False), ids=("fail-open", "fail-closed"))
+async def test_file_sanitization_request_timeout_policy(
+    monkeypatch: pytest.MonkeyPatch, timeout: Exception, fail_open: bool
+):
+    monkeypatch.setenv("PROMPT_SECURITY_API_KEY", "test-key")
+    monkeypatch.setenv("PROMPT_SECURITY_API_BASE", "https://test.prompt.security")
+
+    guardrail = PromptSecurityGuardrail(
+        guardrail_name="test-guard",
+        event_hook="pre_call",
+        default_on=True,
+        file_sanitization_fail_open=fail_open,
+    )
+
+    with patch.object(guardrail.async_handler, "post", AsyncMock(side_effect=timeout)):
+        if not fail_open:
+            with pytest.raises(HTTPException) as exc_info:
+                await guardrail.sanitize_file_content(b"file-content", "document.pdf")
+            assert exc_info.value.status_code == 408
+            assert exc_info.value.detail == "File sanitization timeout"
+            return
+
+        result = await guardrail.sanitize_file_content(b"file-content", "document.pdf")
+
+    assert result == {
+        "action": "allow",
+        "content": None,
+        "metadata": {},
+        "violations": (),
+    }
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("fail_open", (True, False), ids=("fail-open", "fail-closed"))
+async def test_file_sanitization_overall_timeout_policy(monkeypatch: pytest.MonkeyPatch, fail_open: bool):
+    monkeypatch.setenv("PROMPT_SECURITY_API_KEY", "test-key")
+    monkeypatch.setenv("PROMPT_SECURITY_API_BASE", "https://test.prompt.security")
+
+    guardrail = PromptSecurityGuardrail(
+        guardrail_name="test-guard",
+        event_hook="pre_call",
+        default_on=True,
+        file_sanitization_timeout=0.01,
+        file_sanitization_fail_open=fail_open,
+    )
+
+    async def hanging_post(*_args: object, **_kwargs: object) -> None:
+        await asyncio.sleep(60)
+        raise AssertionError("sanitization request should have been cancelled")
+
+    with patch.object(guardrail.async_handler, "post", side_effect=hanging_post):
+        if not fail_open:
+            with pytest.raises(HTTPException) as exc_info:
+                await guardrail.sanitize_file_content(b"file-content", "document.pdf")
+            assert exc_info.value.status_code == 408
+            assert exc_info.value.detail == "File sanitization timeout"
+            return
+
+        result = await guardrail.sanitize_file_content(b"file-content", "document.pdf")
+
+    assert result["action"] == "allow"
+    assert result["content"] is None
+
+
 @pytest.mark.asyncio
 async def test_file_sanitization_block(monkeypatch: pytest.MonkeyPatch):
     """Test that file sanitization blocks malicious files"""
@@ -544,7 +629,7 @@ async def test_role_filtering(monkeypatch: pytest.MonkeyPatch):
         return mock_response
 
     with patch.object(guardrail.async_handler, "post", side_effect=mock_post):
-        result = await guardrail.apply_guardrail(
+        await guardrail.apply_guardrail(
             inputs=inputs,
             request_data=request_data,
             input_type="request",