System messages render with role "developer", breaking chat templates

[GregoryBolshakov]

Problem

The chat template renderer emits system messages with the role developer instead of system. MessageType::SystemPrompt is annotated #[serde(rename = "developer")], and HuggingFaceChatTemplate::format builds the Jinja context with context!{ role, content }, which serializes the role through serde. So any template that branches on role == 'system' never matches the system message.

Depending on the template, the result is one of:

• the system prompt is silently replaced by the model's built-in default (e.g. Qwen2.5 emits "You are Qwen, created by Alibaba Cloud..." instead of the system message you passed),
• the system message is dropped entirely,
• the render fails with "conversation roles must alternate" (Falcon 7B, Gemma 3, Mistral 7B).

Separately, trim_blocks/lstrip_blocks aren't set on the environment (transformers sets both), and the tojson filter doesn't match Python's json.dumps, so some templates drift on whitespace and on serialized tool/JSON output.

Steps To Reproduce

• Load a chat model whose template special-cases the system role (Qwen2.5, Falcon 7B, Gemma 3, Mistral 7B, ...).
• Build a chat with a system message followed by user/assistant turns.
• Apply the chat template.
• Observe the system content is missing or replaced by a default; for Falcon 7B / Gemma 3 / Mistral 7B the render errors with "conversation roles must alternate".

Expected behavior

System messages render with role system, and the output matches transformers.apply_chat_template(..., tokenize=False) byte-for-byte.

Environment:

• Crate version: kalosm main
• Rust version: stable
• OS info: Linux

Questionnaire

• I would like to fix this and I have a solution

[D4RKAR117]

I confirm that the bug happens with any other member who doesn't belong to the Llama model family, including Phi models and GPT-oss models. It would be important to check on this.