Empty model turn is replayed without content, and Ollama rejects the retry with HTTP 400

[Sora-bluesky]

Environment

• SDK version: 0.1.18
• Python version: 3.14.7
• OS / platform: Windows 11 x86_64 (10.0.26200)
• Install method: uv run --with google-antigravity
• Model server: Ollama 0.34.2 through LocalOpenAIAgentConfig(base_url="http://localhost:11434/v1")

Description

When the model returns a turn with no text and no tool calls, the SDK keeps that turn in the history as {"role": "assistant"}, with no content and no tool_calls. It then adds its retry notice as a user message and sends the history again. Ollama's OpenAI-compatible API rejects the bare assistant message with HTTP 400 invalid message content type: <nil>, because content can be left out only when tool_calls is present. The retry never reaches the model, and the session ends with AntigravityExecutionError.

I first saw this 400 three times in normal runs with gemma4:12b and qwen3.6:35b, with the built-in tools enabled. Two of those runs logged this right before the 400:

Model produced invalid output. ("model output error: model output must contain either output text or tool calls, these cannot both be empty, please try again")

The third run logged only the 400. I had no capture of those requests, so I put a logging proxy between the SDK and Ollama and forced an empty turn with a 4-token output limit. The steps below reproduced the 400 in all three of my runs, one with a Japanese prompt and two with the English prompt shown here.

Steps to Reproduce

1. Create a model whose output stops after 4 tokens. Gemma 4 starts with reasoning, so the turn ends with no text and no tool calls.

   FROM gemma4:12b
   PARAMETER num_ctx 65536
   PARAMETER num_predict 4
   

   ollama create gemma4-empty -f Modelfile
   

   The SDK sends max_completion_tokens: 65535, but Ollama 0.34.2 maps only max_tokens to num_predict, so the Modelfile limit applies.

2. Run this script:

   import asyncio
   
   from google.antigravity import Agent, LocalOpenAIAgentConfig
   
   
   async def main():
       config = LocalOpenAIAgentConfig(model="gemma4-empty", base_url="http://localhost:11434/v1")
       async with Agent(config) as agent:
           response = await agent.chat("Write quicksort in Python")
           async for text in response:
               print(text, end="", flush=True)
   
   
   asyncio.run(main())

Expected Behavior

The retry reaches the model. An assistant turn with no text and no tool calls is either left out of the history or sent with "content": "".

Actual Behavior

The first request streams back only reasoning, {"role": "assistant", "content": "", "reasoning": "The"}, and ends with finish_reason: "length". The SDK logs The maximum token limit was reached and sends a second request. This is that request as the proxy recorded it, with the system prompt, the 13 tool definitions and the local time left out:

{
  "model": "gemma4-empty",
  "stream": true,
  "tool_choice": "auto",
  "max_completion_tokens": 65535,
  "messages": [
    {"role": "system", "content": "<SDK system prompt, 5148 characters>"},
    {"role": "user", "content": "<USER_REQUEST>\nWrite quicksort in Python\n</USER_REQUEST>\n<ADDITIONAL_METADATA>\nThe current local time is: <omitted>.\n</ADDITIONAL_METADATA>"},
    {"role": "assistant"},
    {"role": "user", "content": "Error: There was a problem. \nError Message: model output error: max tokens limit reached (generation or context length exceeded). \nRetries remaining: 4"}
  ]
}

Ollama answers with:

HTTP 400
{"error":{"message":"invalid message content type: \u003cnil\u003e","type":"invalid_request_error","param":null,"code":null}}

and the run ends:

WARNING:root:System step error (HTTP 0): The maximum token limit was reached. ("model output error: max tokens limit reached (generation or context length exceeded).")
WARNING:root:System step error (HTTP 2): Agent execution terminated due to error. ("agent executor error: HTTP request failed with status 400: {\"error\":{\"message\":\"invalid message content type: \\u003cnil\\u003e\",\"type\":\"invalid_request_error\",\"param\":null,\"code\":null}}")
google.antigravity.types.AntigravityExecutionError: HTTP request failed with status 400: {"error":{"message":"invalid message content type: \u003cnil\u003e","type":"invalid_request_error","param":null,"code":null}}

The same happens without the proxy.

Additional Context

• Ollama allows a missing content only when tool_calls is present: openai/openai.go L640-L644 at v0.34.2. OpenAI's spec has the same rule for assistant messages: "Required unless tool_calls or function_call is specified" (openapi.yaml L40473-L40475).
• Sending a user message, the empty assistant message and a second user message straight to Ollama: {"role": "assistant"} returns 400, and {"role": "assistant", "content": ""} returns 200.
• Assistant turns that do have tool calls are also sent without content. Ollama accepts those because tool_calls is present.
• Possibly related: Invalid UTF-8 from ReplaceFileContent permanently bricks the conversation (proto marshal fails on every subsequent request) #189, where one bad message in the history made every later request fail.