{#- ===================================================================== -#}
{#- Delphi v0 THINKING-ALWAYS-ON chat template  (Llama-3 tokenizer)        -#}
{#- ----------------------------------------------------------------------- -#}
{#- IDENTICAL to delphi_v0.jinja2 EXCEPT the generation-prompt tail: the    -#}
{#- `enable_thinking is false` empty-block branch is REMOVED, so thinking   -#}
{#- is UNCONDITIONALLY ON and does NOT depend on any chat_template_kwargs.  -#}
{#-                                                                         -#}
{#- WHY: in SkyRL `generator.batched=true` (the fast rollout path) the      -#}
{#- inference engine does the templating and `chat_template_kwargs` are     -#}
{#- refused (skyrl_gym_generator.py:111). So `enable_thinking=true` cannot  -#}
{#- be threaded in batched mode. Serving THIS template to the vLLM engine   -#}
{#- (engine_init_kwargs.custom_chat_template_chat_completion_path) gives    -#}
{#- real thinking with batched=true (fast) — no kwarg needed, and an        -#}
{#- accidental enable_thinking=false default cannot suppress it.            -#}
{#-                                                                         -#}
{#- Assistant-turn RENDERING (training data) is byte-identical to v0 (the   -#}
{#- message loop is unchanged) — only the inference generation prompt       -#}
{#- differs — so it stays consistent with how the SFT data was rendered.    -#}
{#- Use delphi_v0.jinja2 when you need the suppressible/thinking-off path.  -#}
{#- ===================================================================== -#}
{{- bos_token }}
{%- if messages[0].role == 'system' %}
    {%- set system_message = messages[0].content %}
    {%- set loop_messages = messages[1:] %}
{%- else %}
    {%- set system_message = '' %}
    {%- set loop_messages = messages %}
{%- endif %}
{{- '<|start_header_id|>system<|end_header_id|>\n\n' }}
{%- if system_message %}
    {{- system_message }}
{%- endif %}
{%- if tools %}
    {%- if system_message %}{{- '\n\n' }}{%- endif %}
    {{- '# Tools\n\nYou may call one or more of the following functions. Emit each call as a JSON object {"name": ..., "arguments": ...} between <|tool_call|> and <|tool_call_end|>:\n' }}
    {%- for tool in tools %}
        {{- '\n' }}{{- tool | tojson }}
    {%- endfor %}
{%- endif %}
{{- '<|eot_id|>' }}
{%- for message in loop_messages %}
    {%- set content = message.content if message.content is string else '' %}
    {%- if message.role == 'user' %}
        {{- '<|start_header_id|>user<|end_header_id|>\n\n' + content + '<|eot_id|>' }}
    {%- elif message.role == 'assistant' %}
        {%- set reasoning = '' %}
        {%- if message.reasoning_content is string %}
            {%- set reasoning = message.reasoning_content %}
        {%- elif '</think>' in content %}
            {%- set reasoning = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
            {%- set content = content.split('</think>')[-1].lstrip('\n') %}
        {%- endif %}
        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
        {%- if reasoning %}
            {{- '<|start_think|>\n' + reasoning.strip('\n') + '\n<|end_think|>\n\n' }}
        {%- endif %}
        {{- content }}
        {%- if message.tool_calls %}
            {%- for tc in message.tool_calls %}
                {%- set fn = tc.function if tc.function is defined else tc %}
                {{- '\n<|tool_call|>\n{"name": "' + fn.name + '", "arguments": ' }}
                {%- if fn.arguments is string %}{{- fn.arguments }}{%- else %}{{- fn.arguments | tojson }}{%- endif %}
                {{- '}\n<|tool_call_end|>' }}
            {%- endfor %}
        {%- endif %}
        {{- '<|eot_id|>' }}
    {%- elif message.role == 'tool' %}
        {{- '<|start_header_id|>tool<|end_header_id|>\n\n<|tool_result|>\n' + content + '\n<|tool_result_end|><|eot_id|>' }}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {#- FORCE the model into the think region by prefilling the opener (R1 /     -#}
    {#- Qwen3-thinking pattern). Smoke 47525016 showed that with a FREE turn     -#}
    {#- (no prefill) this SFT ckpt emits an EMPTY <|start_think|><|end_think|>   -#}
    {#- and reasons in the answer channel — so "don't suppress" is not enough;   -#}
    {#- we prefill <|start_think|>\n so the response MUST begin inside the think -#}
    {#- region (matches the SFT assistant rendering, which is                    -#}
    {#- <|start_think|>\n<reasoning>\n<|end_think|>...). The response then is     -#}
    {#- <reasoning>\n<|end_think|>\n\n<answer>; the aime reward reads the final  -#}
    {#- boxed answer from the tail, unaffected.                                  -#}
    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n<|start_think|>\n' }}
{%- endif %}