Skip to content

models - litellm - start and stop reasoning #947

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

models - litellm - start and stop reasoning #947

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
de6c95f
models - litellm - stop reasoning
pgrayy Sep 29, 2025
40bd6ff
integ test - model reasoning
pgrayy Sep 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 21 additions & 8 deletions src/strands/models/litellm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -139,22 +139,22 @@ async def stream(

logger.debug("got response from model")
yield self.format_chunk({"chunk_type": "message_start"})
yield self.format_chunk({"chunk_type": "content_start", "data_type": "text"})

tool_calls: dict[int, list[Any]] = {}
started_reasoning = False
started_text = False

async for event in response:
# Defensive: skip events with empty or missing choices
if not getattr(event, "choices", None):
continue
choice = event.choices[0]

if choice.delta.content:
yield self.format_chunk(
{"chunk_type": "content_delta", "data_type": "text", "data": choice.delta.content}
)

if hasattr(choice.delta, "reasoning_content") and choice.delta.reasoning_content:
if not started_reasoning:
yield self.format_chunk({"chunk_type": "content_start", "data_type": "reasoning_content"})
started_reasoning = True

yield self.format_chunk(
{
"chunk_type": "content_delta",
Expand All @@ -163,14 +163,27 @@ async def stream(
}
)

if choice.delta.content:
if started_reasoning:
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can a model switch between streaming reasoning content and text content in a single streamed response? If so, I think we might need to change this implementation:

Reasoning: Let me think about this problem
Text: I have thought about this problem, let me think about a solution
Reasoning: Im thinking about a solution
Text: I have thought about a solution

Not sure if any models follow this pattern?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Initially I tried to account for this but the code was a bit more complex. Since I wasn't aware of any model supporting this pattern, I decided instead to take a simpler approach and just implement for the pattern where reasoning is streamed first. I'm open to this being more flexible though.

yield self.format_chunk({"chunk_type": "content_stop", "data_type": "reasoning_content"})
started_reasoning = False

if not started_text:
yield self.format_chunk({"chunk_type": "content_start", "data_type": "text"})
started_text = True

yield self.format_chunk(
{"chunk_type": "content_delta", "data_type": "text", "data": choice.delta.content}
)

for tool_call in choice.delta.tool_calls or []:
tool_calls.setdefault(tool_call.index, []).append(tool_call)

if choice.finish_reason:
if started_text:
yield self.format_chunk({"chunk_type": "content_stop", "data_type": "text"})
break

yield self.format_chunk({"chunk_type": "content_stop", "data_type": "text"})

for tool_deltas in tool_calls.values():
yield self.format_chunk({"chunk_type": "content_start", "data_type": "tool", "data": tool_deltas[0]})

Expand Down
4 changes: 2 additions & 2 deletions tests/strands/models/test_litellm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,6 +182,8 @@ async def test_stream(litellm_acompletion, api_key, model_id, model, agenerator,
{"messageStart": {"role": "assistant"}},
{"contentBlockStart": {"start": {}}},
{"contentBlockDelta": {"delta": {"reasoningContent": {"text": "\nI'm thinking"}}}},
{"contentBlockStop": {}},
{"contentBlockStart": {"start": {}}},
{"contentBlockDelta": {"delta": {"text": "I'll calculate"}}},
{"contentBlockDelta": {"delta": {"text": "that for you"}}},
{"contentBlockStop": {}},
Expand Down Expand Up @@ -251,8 +253,6 @@ async def test_stream_empty(litellm_acompletion, api_key, model_id, model, agene
tru_events = await alist(response)
exp_events = [
{"messageStart": {"role": "assistant"}},
{"contentBlockStart": {"start": {}}},
{"contentBlockStop": {}},
{"messageStop": {"stopReason": "end_turn"}},
]

Expand Down
38 changes: 34 additions & 4 deletions tests_integ/models/providers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,10 +67,26 @@ def __init__(self):
"api_key": os.getenv("ANTHROPIC_API_KEY"),
},
model_id="claude-3-7-sonnet-20250219",
max_tokens=512,
max_tokens=2048,
params={
"thinking": {
"type": "enabled",
"budget_tokens": 1024,
},
},
),
)
bedrock = ProviderInfo(
id="bedrock",
factory=lambda: BedrockModel(
additional_request_fields={
"thinking": {
"type": "enabled",
"budget_tokens": 1024,
},
},
),
)
bedrock = ProviderInfo(id="bedrock", factory=lambda: BedrockModel())
cohere = ProviderInfo(
id="cohere",
environment_variable="COHERE_API_KEY",
Expand All @@ -84,7 +100,16 @@ def __init__(self):
),
)
litellm = ProviderInfo(
id="litellm", factory=lambda: LiteLLMModel(model_id="bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0")
id="litellm",
factory=lambda: LiteLLMModel(
model_id="bedrock/us.anthropic.claude-3-7-sonnet-20250219-v1:0",
params={
"thinking": {
"budget_tokens": 1024,
"type": "enabled",
},
},
),
)
llama = ProviderInfo(
id="llama",
Expand Down Expand Up @@ -133,7 +158,12 @@ def __init__(self):
factory=lambda: GeminiModel(
api_key=os.getenv("GOOGLE_API_KEY"),
model_id="gemini-2.5-flash",
params={"temperature": 0.7},
params={
"temperature": 0.7,
"thinking_config": {
"include_thoughts": True,
},
},
),
)

Expand Down
12 changes: 11 additions & 1 deletion tests_integ/models/test_conformance.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@

from strands import Agent
from strands.models import Model
from tests_integ.models.providers import ProviderInfo, all_providers, cohere, llama, mistral
from tests_integ.models.providers import ProviderInfo, all_providers, cohere, gemini, llama, mistral, openai, writer


def get_models():
Expand Down Expand Up @@ -60,3 +60,13 @@ class Weather(BaseModel):

assert len(result.time) > 0
assert len(result.weather) > 0


def test_stream_reasoning(skip_for, model):
skip_for([cohere, gemini, llama, mistral, openai, writer], "reasoning is not supported")
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note, reasoning is supported in gemini but there is a similar bug to what we are fixing here for litellm. We need to yield a content stop block. Will address in a follow up PR.


agent = Agent(model)
result = agent("Please reason about the equation 2+2.")

assert "reasoningContent" in result.message["content"][0]
assert result.message["content"][0]["reasoningContent"]["reasoningText"]["text"]
Loading