From 74ebc0606953402da44ff1cefb558d90c5695f55 Mon Sep 17 00:00:00 2001
From: mateo <mateo@berri.ai>
Date: Tue, 27 Oct 2026 20:05:45 +0000
Subject: [PATCH] fix(gemini): keep the placeholder resend inside the cache
 rebuild so either rejection order ends in an accepted body

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../rejected_thought_signature_retry.py       | 14 ++++++----
 .../vertex_and_google_ai_studio_gemini.py     | 28 ++++++-------------
 2 files changed, 17 insertions(+), 25 deletions(-)

diff --git a/litellm/llms/vertex_ai/gemini/rejected_thought_signature_retry.py b/litellm/llms/vertex_ai/gemini/rejected_thought_signature_retry.py
index 96aead87f61..ae51efb2b34 100644
--- a/litellm/llms/vertex_ai/gemini/rejected_thought_signature_retry.py
+++ b/litellm/llms/vertex_ai/gemini/rejected_thought_signature_retry.py
@@ -64,27 +64,29 @@ async def post_with_accepted_thought_signatures_async(
     body: RequestBody,
     post: Callable[[RequestBody], Awaitable[Sent]],
     log_request: Callable[[RequestBody], None],
-) -> Sent:
+) -> tuple[RequestBody, Sent]:
+    """Innermost retry around the raw post, so a cache rebuild (which re-transforms the messages and brings
+    the rejected signatures back) is itself wrapped by it. Returns the body the backend accepted."""
     try:
-        return await post(body)
+        return body, await post(body)
     except (httpx.HTTPStatusError, BaseLLMException) as error:
         resend_body: Final = _resend_body(body, error)
         if resend_body is None:
             raise
         log_request(resend_body)
-    return await post(resend_body)
+    return resend_body, await post(resend_body)
 
 
 def post_with_accepted_thought_signatures_sync(
     body: RequestBody,
     post: Callable[[RequestBody], Sent],
     log_request: Callable[[RequestBody], None],
-) -> Sent:
+) -> tuple[RequestBody, Sent]:
     try:
-        return post(body)
+        return body, post(body)
     except (httpx.HTTPStatusError, BaseLLMException) as error:
         resend_body: Final = _resend_body(body, error)
         if resend_body is None:
             raise
         log_request(resend_body)
-    return post(resend_body)
+    return resend_body, post(resend_body)
diff --git a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
index 4ad9b1b973b..53d412c7300 100644
--- a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
+++ b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py
@@ -3476,12 +3476,8 @@ class VertexLLM(VertexBase):
             async def post(body: RequestBody) -> ModelResponseIterator:
                 return await open_stream(client=client, data=json.dumps(body), request_body=body)
 
-            post_with_fresh_cache: Final = partial(
-                post_with_fresh_cache_async, rebuild=rebuild, post=post, log_request=log_request
-            )
-            return (
-                await post_with_accepted_thought_signatures_async(request_body, post_with_fresh_cache, log_request)
-            )[1]
+            accepted: Final = partial(post_with_accepted_thought_signatures_async, post=post, log_request=log_request)
+            return (await post_with_fresh_cache_async(request_body, rebuild, accepted, log_request))[1][1]
 
         streaming_response: Final = CustomStreamWrapper(
             completion_stream=None,
@@ -3605,12 +3601,8 @@ class VertexLLM(VertexBase):
                 )
             return response
 
-        post_with_fresh_cache: Final = partial(
-            post_with_fresh_cache_async, rebuild=rebuild, post=post, log_request=log_request
-        )
-        sent_body, first_response = await post_with_accepted_thought_signatures_async(
-            request_body, post_with_fresh_cache, log_request
-        )
+        accepted: Final = partial(post_with_accepted_thought_signatures_async, post=post, log_request=log_request)
+        _, (sent_body, first_response) = await post_with_fresh_cache_async(request_body, rebuild, accepted, log_request)
         return _transform_long_decode_hops(
             model=model,
             hops=await collect_hops_async(parse_body(first_response), sent_body, post),
@@ -3786,10 +3778,10 @@ class VertexLLM(VertexBase):
                 def post(body: RequestBody) -> ModelResponseIterator:
                     return open_stream(client=client, data=json.dumps(body), request_body=body)
 
-                post_with_fresh_cache: Final = partial(
-                    post_with_fresh_cache_sync, rebuild=rebuild, post=post, log_request=log_request
+                accepted: Final = partial(
+                    post_with_accepted_thought_signatures_sync, post=post, log_request=log_request
                 )
-                return post_with_accepted_thought_signatures_sync(data, post_with_fresh_cache, log_request)[1]
+                return post_with_fresh_cache_sync(data, rebuild, accepted, log_request)[1][1]
 
             streaming_response: Final = CustomStreamWrapper(
                 completion_stream=None,
@@ -3833,10 +3825,8 @@ class VertexLLM(VertexBase):
                 )
             return response
 
-        post_with_fresh_cache: Final = partial(
-            post_with_fresh_cache_sync, rebuild=rebuild, post=post, log_request=log_request
-        )
-        sent_body, first_response = post_with_accepted_thought_signatures_sync(data, post_with_fresh_cache, log_request)
+        accepted: Final = partial(post_with_accepted_thought_signatures_sync, post=post, log_request=log_request)
+        _, (sent_body, first_response) = post_with_fresh_cache_sync(data, rebuild, accepted, log_request)
         return _transform_long_decode_hops(
             model=model,
             hops=collect_hops_sync(parse_body(first_response), sent_body, post),