fix(ollama): persist assistant response in message_history to fix multi-turn tool calling

[nankingjing]

Problem

The OllamaClient.chat() method never persists the assistant's response (text or tool calls) in self.message_history. All other LLM clients (AnthropicClient, OpenAIClient, OpenAICompatibleClient, GoogleClient) correctly append the assistant response to their message history after each chat call, but the Ollama client was missing this step.

This breaks multi-turn conversations with tool calling because:

1. The agent loop in BaseAgent._tool_call_handler only passes new tool result messages to subsequent chat() calls
2. The LLM client is expected to maintain the full conversation history via self.message_history (with reuse_history=True as default)
3. Without the assistant's tool-call response in history, the LLM sees an incomplete conversation: user asks task -> user provides tool results (missing the assistant's decision to call tools)

Fix

Added persistence of the assistant response to self.message_history after receiving the LLM response:

• Text responses: appended as EasyInputMessageParam(role="assistant", content=..., type="message")
• Tool call responses: appended as ResponseFunctionToolCallParam(call_id=..., name=..., arguments=..., type="function_call") for each tool call

Also added EasyInputMessageParam to the imports from openai.types.responses.

This follows the same pattern used by all other LLM client implementations.

Verification

• Verified by reading all 5 LLM client implementations -- only OllamaClient was missing history persistence
• Identified via grep confirming zero self.message_history.append() calls in OllamaClient vs multiple in every other client
• Not executed (tests require a running Ollama server with model)

[nankingjing]

Reviewed — persisting assistant responses in message_history is the correct fix for multi-turn Ollama conversations. Ready for review.

[nankingjing]

Updated the fix to use the Ollama-native message shape for assistant turns instead of OpenAI records, which rejects during multi-turn tool calling. Also added regression tests covering both text-only and tool-call persistence. Tests and lint pass locally.

[nankingjing]

Updated: the fix now persists assistant responses in Ollama-native message shape (role/tool_calls) instead of OpenAI function_call items, which ollama.chat rejects during multi-turn tool calling. Regression tests added for both text-only and tool-call persistence. @chao-peng @bytedance PTAL.