`ChatBedrockConverse` prevents DeepSeek V3 from returning text whenever tools are bound (introduced by #688)

[amacleay]

Since #688, ChatBedrockConverse rewrites a missing tool_choice to any for deepseek.v3* models and omits auto from their capability table. If tools are in use, the model is always forced to call a tool and can never answer in text, and asking for auto explicitly is refused by lanchain-aws code before any request is made. Bedrock itself accepts auto (and no toolChoice) on both DeepSeek V3.1 and V3.2 today.

Versions: langchain-aws 1.7.5 (also reproduced on 1.6.0 and 1.7.4), langchain-core 1.6.1, boto3 1.43.82. Region us-east-1 unless noted.

The Converse API accepts tool selection as documented

With tools, no toolChoice (Bedrock's default is auto), asked for text:

import boto3
rt = boto3.client("bedrock-runtime", region_name="us-east-1")
tool = {"toolSpec": {"name": "get_weather", "description": "Get the weather for a city.",
        "inputSchema": {"json": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}}}}
r = rt.converse(modelId="deepseek.v3.2",
                messages=[{"role": "user", "content": [{"text": "Reply with only the text OK. Do not use any tools."}]}],
                inferenceConfig={"maxTokens": 100}, toolConfig={"tools": [tool]})
print(r["stopReason"], r["output"]["message"]["content"])
# end_turn [{'text': 'OK'}]

Asked for the weather instead, the same request returns a valid toolUse block.

Through langchain-aws auto is incorrectly refused

from langchain_aws import ChatBedrockConverse
from langchain_core.tools import tool

@tool
def get_weather(city: str) -> str:
    """Get the weather for a city."""
    return f"Sunny in {city}"

llm = ChatBedrockConverse(model="deepseek.v3.2", provider="deepseek", region_name="us-east-1", max_tokens=100)
print(llm.supports_tool_choice_values)                      # ('any', 'tool')  <- no 'auto'

msg = llm.bind_tools([get_weather]).invoke("Reply with only the text OK. Do not use any tools.")
print(msg.tool_calls)                                        # [{'name': 'get_weather', ...}]  <- forced tool call
# request body sent: toolConfig.toolChoice == {"any": {}}

llm.bind_tools([get_weather], tool_choice="auto")
# ValueError: Model deepseek.v3.2 does not currently support tool_choice of type auto.

root cause:

Both come from `_resolve_tool_choice` in `bedrock_converse.py`:

if not tool_choice:
    if "deepseek.v3" in self._get_base_model():
        return _format_tool_choice("any")      # None -> any
    return None
...
if tool_choice_type in supported:              # 'auto' not in ('any', 'tool') -> raise

Other providers are unaffected

Same code, model="us.anthropic.claude-haiku-4-5-20251001-v1:0":

llm.bind_tools([get_weather]).invoke("Reply with only the text OK. Do not use any tools.").content
# 'OK'   (no toolChoice sent; caps ('auto', 'any', 'tool'))

Not a problem for some other model families: Raw Converse behaves the same for Claude, Grok 4.6, GPT-5.6 and Qwen3, and when toolChoice not provided as a parameter, we get text when asked and a tool call when appropriate.

Impact

On DeepSeek v3 with tools configured, all model calls go out with tool_choice='any', preventing text responses.

Proposed fix

1. Remove the deepseek.v3 branch in _resolve_tool_choice, and we get Bedrock defaults.
2. Add "auto" to the DeepSeek entries in the supports_tool_choice_values table.

Further details: measurements and origin

#927 doesn't resolve DeepSeek. Workaround on current releases: pass supports_tool_choice_values=("auto", "any") at construction and set tool_choice="auto" explicitly (we used a custom middleware to do this).

#688 said: "DeepSeek-V3.1 does not work with the auto and named tool choice values. While these options will not be rejected on Converse API side, only any will produce a valid tool call output." and "DeepSeek-V3.2-Exp supports tool choice in addition to any, but auto is still not supported." I cannot say whether that held in October 2025; it does not hold now.

Raw Converse, prompt "What's the weather in Boston? Use the tool.", valid toolUse (correct name, non-empty city) per 3 trials:

model	region	no toolChoice	auto	tool
deepseek.v3-v1:0 (V3.1)	us-west-2	3/3	3/3	3/3
deepseek.v3.2	us-west-2	3/3	3/3	3/3
deepseek.v3.2	us-east-1	3/3	3/3	3/3

Prompt "Reply with only the text OK.": end_turn with a text block under no toolChoice and under auto; tool_use under any, as expected.

Two-turn loop through ChatBedrockConverse.bind_tools (tool call, tool result, second call): the second call is sent with toolChoice: {"any": {}} and returns another tool call rather than the sentence asked for.

Full reproducer with an outbound-request tap, showing the toolChoice value actually sent for each case:

"""Demo: langchain-aws forces `toolChoice: any` on DeepSeek V3 whenever tools are bound."""
import json
import boto3
from langchain_aws import ChatBedrockConverse
from langchain_core.messages import HumanMessage, ToolMessage
from langchain_core.tools import tool

MODEL, REGION = "deepseek.v3.2", "us-east-1"

sent: list[dict] = []
_orig_client = boto3.Session.client
def _tapped_client(self, *args, **kwargs):
    client = _orig_client(self, *args, **kwargs)
    if (args[0] if args else kwargs.get("service_name")) == "bedrock-runtime":
        client.meta.events.register("before-call.bedrock-runtime.Converse",
                                    lambda params, **_: sent.append(json.loads(params["body"])))
    return client
boto3.Session.client = _tapped_client

@tool
def get_weather(city: str) -> str:
    """Get the weather for a city."""
    return f"Sunny, 22C in {city}"

def tool_choice_sent():
    return sent[-1].get("toolConfig", {}).get("toolChoice") if sent else None

def describe(msg):
    return f"TOOL CALL: {msg.tool_calls[0]['name']}" if msg.tool_calls else f"TEXT: {msg.content!r:.40}"

llm = ChatBedrockConverse(model=MODEL, provider="deepseek", region_name=REGION, max_tokens=200)
ASK = "Reply with only the text OK. Do not use any tools."

sent.clear(); r = llm.invoke(ASK)
print("1. no tools:            ", tool_choice_sent(), describe(r))
bound = llm.bind_tools([get_weather])
sent.clear(); r = bound.invoke(ASK)
print("2. tools, no choice:    ", tool_choice_sent(), describe(r))
try:
    llm.bind_tools([get_weather], tool_choice="auto").invoke(ASK)
except ValueError as exc:
    print("3. tools, auto:          REFUSED:", str(exc)[:80])
msgs = [HumanMessage("What's the weather in Boston? Use the tool, then answer in one sentence.")]
first = bound.invoke(msgs); msgs.append(first)
for c in first.tool_calls:
    msgs.append(ToolMessage(content=get_weather.invoke(c["args"]), tool_call_id=c["id"]))
sent.clear(); final = bound.invoke(msgs)
print("4. after tool result:   ", tool_choice_sent(), describe(final))

Output:

1. no tools:             None TEXT: 'OK'
2. tools, no choice:     {'any': {}} TOOL CALL: get_weather
3. tools, auto:          REFUSED: Model deepseek.v3.2 does not currently support tool_choice of type auto. ...
4. after tool result:    {'any': {}} TOOL CALL: get_weather