Motivation

This is a quality-of-life issue, rather than a technical error.

• Teams thread
• When copy-and-pasting web examples (e.g., vLLM recipes in https://recipes.vllm.ai/openai/gpt-oss-120b ) into the start_command field via our WebUI, it is cumbersome to manually remove the backslashes because they are literally preserved via shlex.split() and breaks up the command execution as “\n” are passed as regular execv() argument tokens.

Required Features

• Pass-through the original (multi-line) str value all the way to the kernel runner’s service launch routine.

Impact

• Existing model defintions with list-style start_command won’t be affected.
• New or existing model definitions with str-style start_command will be affected. The string value will be interpreted as a normal shell prompt, instead of being converted to list[str] via shlex.split().
• The kernel runner’s own ${{ ... }} and { ... } variable expansion still works.

Testing Scenarios

• Copy-and-pasting vLLM recipes for a new model-definition using the custom variant should work AS-IS.

JIRA Issue: BA-6053