Piping certain cmd outputs to shell sage causes LiteLLM `This model does not support assistant message prefill. The conversation must end with a user message.`

````diff
diff --git a/ipyai/core.py b/ipyai/core.py
index c2e7e9e..7db83c1 100644
--- a/ipyai/core.py
+++ b/ipyai/core.py
@@ -102,7 +102,7 @@ prompt_from_lines astream_to_stdout transform_dots unload_ipython_extension""".s
_prompt_template = """{context}{prompt}"""
_tool_re = re.compile(r"&`(\w+)`")
_var_re = re.compile(r"\$`(\w+)`")
-_shell_re = re.compile(r"!`([^`]+)`")
+_shell_re = re.compile(r"(?\s*([^\n]*?)\s*```json\s*(.*?)\s*```\s*", flags=re.DOTALL)
_status_attrs = "model completion_model think search code_theme log_exact".split()
@@ -788,7 +788,7 @@ class IPyAIExtension:
parts.append("")
parts.append("\nReturn ONLY the completion text to insert immediately after the prefix."
" Do not repeat the prefix or include any explanation.")
- chat = AsyncChat(model=self.completion_model, sp=_COMPLETION_SP)
+ chat = AsyncChat(model=self.completion_model, sp=_COMPLETION_SP, cache=True)
res = await chat("\n".join(parts))
return (contents(res).content or "").strip()
````

Minimal repro. Piping this into shell sage leads to:

```
litellm.exceptions.BadRequestError: litellm.BadRequestError: AnthropicException - b'{"type":"error","error":{"type":"invalid_request_error","message":"This model does not support assistant message prefill. The conversation must end with a user message."},"request_id":"req_011CZSMPZ3QszFE6X6fwGrSC"}'
```

Might be a lisette issue.