fix(ollama_chat): set finish_reason to tool_calls when tool calls arrive before the final chunk

[ShreeBohara]

TLDR

Problem this solves:

• ollama_chat streamed tool calls ended with finish_reason "stop"
• spec strict OpenAI clients never executed the streamed tool call

How it solves it:

• iterator remembers tool_calls seen in any earlier chunk
• final done chunk then reports finish_reason "tool_calls"

Relevant issues

Fixes #35663

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Environment for both runs: local proxy started with python -m litellm.proxy.proxy_cli --model ollama_chat/llama3.2 --host 127.0.0.1 --port 4000, ollama 0.15.6 serving llama3.2 locally, real streaming LLM calls, no mocks

curl -sN http://127.0.0.1:4000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "ollama_chat/llama3.2", "stream": true, "temperature": 0,
  "messages": [{"role": "user", "content": "What is the weather in San Francisco right now? Use the get_weather tool."}],
  "tools": [{"type": "function", "function": {"name": "get_weather", "description": "Get current weather for a location", "parameters": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}}}]}'

Before, at cb8c734 (litellm_internal_staging without this fix): the stream carries the tool call delta but the final chunk reports finish_reason "stop" (SSE trimmed to the two relevant chunks)

data: {"id":"fdd457f5-68f4-421f-aca9-d09ac47c36c8","created":1785814307,"model":"ollama_chat/llama3.2","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"id":"64712615-3171-4639-ad60-515c161a8480","function":{"arguments":"{\"location\": \"San Francisco\"}","name":"get_weather"},"type":"function","index":0}]}}]}
data: {"id":"ca3be654-bdea-463e-ab0c-6fa74abf5917","object":"chat.completion.chunk","created":1785814307,"model":"ollama_chat/llama3.2","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

After, at d216383 (this PR): same request, the tool call delta is unchanged and the final chunk now reports finish_reason "tool_calls"

data: {"id":"6e73d45b-ba59-4a62-a7c3-aaf072f2850b","created":1785814351,"model":"ollama_chat/llama3.2","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"id":"b0a71062-7b37-4fd6-9670-f080976bdb77","function":{"arguments":"{\"location\": \"San Francisco\"}","name":"get_weather"},"type":"function","index":0}]}}]}
data: {"id":"b4f96644-1e33-4b15-9318-c1fe82e01e7c","object":"chat.completion.chunk","created":1785814351,"model":"ollama_chat/llama3.2","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}

Same fix observed through /v1/messages, where the finish reason surfaces as stop_reason. Request:

curl -sN http://127.0.0.1:4000/v1/messages -H 'Content-Type: application/json' -H 'anthropic-version: 2023-06-01' -d '{
  "model": "ollama_chat/llama3.2", "max_tokens": 256, "stream": true, "temperature": 0,
  "messages": [{"role": "user", "content": "What is the weather in San Francisco right now? Use the get_weather tool."}],
  "tools": [{"name": "get_weather", "description": "Get current weather for a location", "input_schema": {"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]}}]}'

Before, at cb8c734, the tool_use block streams but the message ends as end_turn

data: {"type": "message_delta", "delta": {"stop_reason": "end_turn"}, "usage": {"input_tokens": 158, "output_tokens": 18}}

After, at d216383, the same request ends with stop_reason "tool_use"

data: {"type": "message_delta", "delta": {"stop_reason": "tool_use"}, "usage": {"input_tokens": 158, "output_tokens": 18}}

/v1/responses was also checked with the equivalent streamed request at both commits; the full event sequences are identical before and after, because the Responses emulation derives its function_call output items from the tool call deltas rather than from finish_reason, so that endpoint's output is unchanged by this PR

Type

🐛 Bug Fix

Changes

OllamaChatCompletionResponseIterator.chunk_parser only overrode finish_reason to "tool_calls" when the tool calls appeared on the same chunk as done: true. Ollama emits tool calls in a chunk with done: false, then a separate final chunk with an empty message and done_reason: "stop", so streamed tool calls ended with finish_reason "stop" and clients that gate tool execution on the finish reason discarded them. The non-streaming path already sets finish_reason correctly for this case

The iterator now sets an emitted_tool_calls flag when any chunk carries tool calls, and the done branch checks that flag instead of only the current chunk. The flag is per-stream state on the iterator instance, following the existing started_reasoning_content pattern in the same class. Behavior for tool calls arriving on the done chunk itself is unchanged since the flag is set earlier in that same call, and done_reason: "length" handling is untouched when no tool calls were streamed

Added a regression test that feeds the iterator a tool call chunk followed by a bare done chunk and asserts the final chunk reports finish_reason "tool_calls"; it fails at cb8c734 and passes with this change. Related but distinct: #35711 tracks tool call streaming for the ollama/ generate endpoint, which this PR does not touch

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[CLAassistant]

All committers have signed the CLA.

[greptile-apps]

Greptile Summary

This PR tracks whether an Ollama chat stream emitted tool calls and uses that state to correct the terminal finish reason

• Adds per-iterator tool-call state across streamed chunks
• Adds a regression test for a tool-call chunk followed by a bare stop chunk

Confidence Score: 4/5

The token-limit termination case after a streamed tool call should be fixed before merging

The new unconditional override converts done_reason="length" to "tool_calls" after any tool-call delta, hiding truncation and potentially presenting incomplete arguments as executable

Files Needing Attention: litellm/llms/ollama/chat/transformation.py, tests/test_litellm/llms/ollama/test_ollama_chat_transformation.py

Important Files Changed

Filename	Overview
litellm/llms/ollama/chat/transformation.py	Corrects split-chunk tool-call termination, but also masks token-limit termination after a tool-call delta
tests/test_litellm/llms/ollama/test_ollama_chat_transformation.py	Covers the reported stop case but omits the existing length termination combined with an earlier tool call

_{Reviews (1): Last reviewed commit: "fix(ollama_chat): set finish_reason to t..." | Re-trigger Greptile}

[greptile-apps]

Tool calls mask length termination

When Ollama emits a tool-call delta and then terminates with done_reason="length", this branch replaces the truncation signal with finish_reason="tool_calls", causing clients to treat potentially incomplete tool-call arguments as executable.

Rule Used: What: avoid backwards-incompatible changes without... (source)

Knowledge Base Used: LLM Provider Adapters

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d2163837e9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve length finish reasons after tool-call chunks

When Ollama ends a stream with done_reason="length" after a prior tool-call chunk, this unconditional override changes the terminal finish_reason from length to tool_calls. That masks max-token truncation, so strict clients may execute a tool call whose arguments were cut off instead of treating the response as incomplete; the existing streaming finalizer only upgrades stop to tool_calls for this reason. Please gate this override to stop-like endings while preserving non-stop reasons such as length.

Useful? React with 👍 / 👎.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing ShreeBohara:fix-ollama-chat-stream-tool-calls-finish-reason-35663 (d216383) with litellm_internal_staging (956d517)}