fix(bedrock): apply the 16 maxTokens floor to Kimi K3 on Converse (#43875)

* fix(bedrock): apply the 16 maxTokens floor to Kimi K3 on Converse

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

* test(e2e): use lane AWS credentials for the Kimi K3 messages test

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

* test(e2e): bypass the proxy cache and declare test metadata for the Kimi K3 probe

---------

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>

litellm/llms/bedrock/chat/converse_transformation.py

@@ -385,7 +385,7 @@ class AmazonConverseConfig(BaseConfig):
 
     @staticmethod
     def _requires_min_max_tokens(model: str) -> bool:
-        return re.search(r"openai\.gpt-\d|xai\.grok-", model) is not None
+        return re.search(r"openai\.gpt-\d|xai\.grok-|moonshotai\.kimi-k3", model) is not None
 
     def _is_nova_2_model(self, model: str) -> bool:
         """

tests/e2e/coverage_registry/llm_conversational.yaml

@@ -68,6 +68,7 @@
 - {id: llm.messages.anthropic.vision.nonstream.works, module: llm, tier: P0, subject_endpoint: messages, route: anthropic, capability: vision, streaming: nonstream, assertions: [works], source: "model_prices json", rationale: "Vision via Messages API"}
 - {id: llm.messages.anthropic.prompt_cache_5m.nonstream.works, module: llm, tier: P0, subject_endpoint: messages, route: anthropic, capability: prompt_cache_5m, streaming: nonstream, assertions: [works], source: "model_prices json", rationale: "Prompt caching via Messages API"}
 - {id: llm.messages.anthropic.thinking.nonstream.works, module: llm, tier: P1, subject_endpoint: messages, route: anthropic, capability: thinking, streaming: nonstream, assertions: [works], source: "model_prices json", rationale: "Extended thinking via Messages API"}
+- {id: llm.messages.bedrock_converse.basic.nonstream.works_for_kimi_k3_max_tokens_1, module: llm, tier: P0, subject_endpoint: messages, route: bedrock_converse, capability: basic, streaming: nonstream, assertions: [works_for_kimi_k3_max_tokens_1], source: "llm_translation/test_bedrock_kimi_k3_messages_e2e.py", rationale: "Bedrock Converse accepts max_tokens=1 on the Anthropic Messages surface after LiteLLM applies the model minimum", fail_before_fix: proven}
 - {id: llm.messages.bedrock_invoke.mid_conversation_system.nonstream.cache_hit, module: llm, tier: P0, subject_endpoint: messages, route: bedrock_invoke, capability: mid_conversation_system, streaming: nonstream, assertions: [works, cache_hit], source: "llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py", rationale: "Flagged Claude 4.8+/5 must keep mid-conversation system reminders in messages; hoisting mutates the system prefix and collapses the prompt cache (#32578/#32831/#32882)", fail_before_fix: proven}
 - {id: llm.messages.bedrock_invoke.mid_conversation_system.nonstream.works, module: llm, tier: P0, subject_endpoint: messages, route: bedrock_invoke, capability: mid_conversation_system, streaming: nonstream, assertions: [works], source: "llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py", rationale: "Claude <= 4.7 rejects role system inside messages; unflagged models must convert reminders to user turns in place (hoisting collapses the prompt cache) or every Claude Code session 400s (#32831)", fail_before_fix: proven}
 - {id: llm.messages.bedrock_invoke.web_search_server_tool.nonstream.works, module: llm, tier: P0, subject_endpoint: messages, route: bedrock_invoke, capability: web_search_server_tool, streaming: nonstream, assertions: [works], source: "llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py", rationale: "Bedrock hosts no web_search server tool, so this only works because interception rewrites it before the upstream call and the agentic loop feeds the results back in native shape; a regression that short-circuits or forwards it instead yields raw text or AWS's 400", fail_before_fix: unproven}

tests/e2e/llm_translation/test_bedrock_kimi_k3_messages_e2e.py

@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+from typing import Final
+
+import pytest
+from anthropic.types import Message, MessageParam
+from e2e_metadata import Domain, Mode, Provider, Route, Subject, meta
+from lifecycle import ResourceManager
+from models import LiteLLMParamsBody
+from proxy_client import ProxyClient
+from sdk_clients import NO_PROXY_CACHE, SdkClients
+
+pytestmark = pytest.mark.e2e
+
+BEDROCK_MODEL: Final = "bedrock/us.moonshotai.kimi-k3"
+MODEL_NAME: Final = "claude-kimi-k3"
+MODEL_SWITCH_BETA: Final = (
+    "claude-code-20250219,context-1m-2025-08-07,interleaved-thinking-2025-05-14,"
+    "redact-thinking-2026-02-12,context-management-2025-06-27,"
+    "prompt-caching-scope-2026-01-05,mid-conversation-system-2026-04-07"
+)
+
+
+def _register_model(proxy: ProxyClient, resources: ResourceManager) -> tuple[str, str]:
+    model_id: Final = proxy.create_model(
+        MODEL_NAME,
+        LiteLLMParamsBody(
+            model=BEDROCK_MODEL,
+            aws_access_key_id="os.environ/AWS_ACCESS_KEY_ID",
+            aws_secret_access_key="os.environ/AWS_SECRET_ACCESS_KEY",
+            aws_region_name="os.environ/AWS_REGION",
+            drop_params=True,
+        ),
+    )
+    resources.defer(lambda: proxy.delete_model(model_id))
+    return MODEL_NAME, resources.key(models=[MODEL_NAME])
+
+
+class TestBedrockKimiK3Messages:
+    @pytest.mark.covers("llm.messages.bedrock_converse.basic.nonstream.works_for_kimi_k3_max_tokens_1")
+    @meta(
+        Subject(
+            domain=Domain.LLM_TRANSLATION,
+            route=Route.MESSAGES,
+            providers=(Provider.BEDROCK,),
+            models=(BEDROCK_MODEL,),
+            mode=Mode.NONSTREAM,
+        )
+    )
+    def test_model_switch_probe_max_tokens_1_returns_message(
+        self, proxy: ProxyClient, resources: ResourceManager, sdk: SdkClients
+    ) -> None:
+        model, key = _register_model(proxy, resources)
+        client: Final = sdk.anthropic(key)
+        message_param: Final[MessageParam] = {
+            "role": "user",
+            "content": [{"type": "text", "text": "Hi", "cache_control": {"type": "ephemeral"}}],
+        }
+        message: Final = client.messages.create(
+            model=model,
+            max_tokens=1,
+            stream=False,
+            system="You are Claude Code, an assistant for software development.",
+            messages=[message_param],
+            extra_headers={"anthropic-beta": MODEL_SWITCH_BETA},
+            extra_query={"beta": "true"},
+            extra_body=NO_PROXY_CACHE,
+        )
+
+        assert isinstance(message, Message)
+        assert message.role == "assistant"

tests/unit/llms/bedrock/chat/test_converse_transformation.py

@@ -896,6 +896,16 @@ def test_reasoning_with_forced_tool_choice_switches_to_auto():
         ("us.xai.grok-4.6", "max_tokens", 1, 16),
         ("global.xai.grok-4.6", "max_completion_tokens", 1, 16),
         ("us.xai.grok-4.6", "max_tokens", 32, 32),
+        ("us.moonshotai.kimi-k3", "max_tokens", 1, 16),
+        ("global.moonshotai.kimi-k3", "max_completion_tokens", 1, 16),
+        ("moonshotai.kimi-k3", "max_tokens", 64, 64),
+        (
+            "arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.moonshotai.kimi-k3",
+            "max_tokens",
+            1,
+            16,
+        ),
+        ("moonshotai.kimi-k2.5", "max_tokens", 1, 1),
         ("anthropic.claude-sonnet-4-5-20250929-v1:0", "max_tokens", 1, 1),
         ("arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.openai.gpt-6-astra", "max_tokens", 1, 16),
         ("arn:aws:bedrock:us-east-1:123456789012:inference-profile/global.xai.grok-4.6", "max_tokens", 1, 16),