Logic behind mapping <think> tags to reasoning_content #9433

Activity

1. github-actions commented on Jan 19, 2026

   

   github-actionsBot

   on Jan 19, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be a duplicate of existing issues. Please check:

   • MiniMax-M2 reasoning content not properly handled in OpenAI-compatible mode, should we support this ? #3555: MiniMax-M2 reasoning content not properly handled in OpenAI-compatible mode - discusses how models send <think> tags in delta.content instead of delta.reasoning_content
   • GLM 4.7 thinking process not properly formatted - missing leading tags #7779: GLM 4.7 thinking process not properly formatted - discusses <think> tag formatting issues
   • [FEATURE]: support extracting reasoning XML #6337: Feature request for extracting reasoning XML tags
   • reasoning_content not sent when using @ai-sdk/openai-compatible with deepseek-reasoner #6040: reasoning_content not sent when using @ai-sdk/openai-compatible with deepseek-reasoner - discusses reasoning_content not being sent in subsequent requests
   • DeepSeek V3.2 : reasoning_content not cleared from message history on new turns, causing excess token usage and violating API spec #5577: DeepSeek V3.2 reasoning_content not cleared from message history - comprehensive discussion of reasoning_content handling

   Feel free to ignore if none of these address your specific case.
2. github-actions

   assigned

   thdxr

   on Jan 19, 2026
3. rekram1-node commented on Jan 19, 2026

   

   rekram1-node

   on Jan 19, 2026

   Contributor

   More actions

   And if a model supports interleaved reasoning

   This should only happen if the model is configured to both:
   A) be interleaved reasoning model
   B) expect reasoning to be sent back as reasoning_content | reasoning_details
4. zikajk commented on Jan 20, 2026

   

   zikajk

   on Jan 20, 2026

   Last edited by zikajk

   Author

   More actions

   @rekram1-node
   It seems OpenCode always extracts <think>...</think> tags from model output into “reasoning” parts here.
   Then, only when model.capabilities.interleaved.field === "reasoning_content", those reasoning parts are mapped into providerOptions.openaiCompatible.reasoning_content and removed from content.

   This means a model that expects <think> tags to remain in history won’t receive them back. For example, MiniMax 2.1 docs state tags should stay in history.

   Questions:

   • Do you know of any model/provider that both outputs <think>...</think> and expects preserved/interleaved reasoning via reasoning_content / reasoning_details?
   • If not, what’s the rationale for always stripping <think> tags by default — is it purely to save context/cost, or something else?
   • Should this behavior be driven by models.dev capabilities or be configurable?

   Thank you
5. rekram1-node commented on Jan 20, 2026

   

   rekram1-node

   on Jan 20, 2026

   Contributor

   More actions

   I see what u mean hmm
6. rekram1-node commented on Jan 20, 2026

   

   rekram1-node

   on Jan 20, 2026

   Contributor

   More actions

   Yeah ig we should remove that wrapping or something hm
7. rekram1-node

   assigned

   rekram1-node

   and unassigned

   thdxr

   on Jan 20, 2026
8. zikajk commented on Jan 20, 2026

   

   zikajk

   on Jan 20, 2026

   Last edited by zikajk

   Author

   More actions

   And could you tell me if removing the <think>...</think> from history was intentional or a mistake? I've heard conflicting opinions on whether the coding agent should restore it.
9. rekram1-node commented on Jan 20, 2026

   

   rekram1-node

   on Jan 20, 2026

   Contributor

   More actions

   There is some nuance but tldr:

   • we should prolly keep rendering the tags as actual reasoning blocks
   • we shouldn't transform the way they are sent back, if they are sent as part of the assistant text, then they should stay that way

   I feel like this is best of both worlds the only reason we have that parser is for that

   👍React with 👍1Reacted by zikajk
10. daryltucker commented on Jan 26, 2026

    

    daryltucker

    on Jan 26, 2026

    Last edited by daryltucker

    More actions

    response.reasoning_text.delta is this why Thinking isn't shown in opencode now?

    #4947

    ./build/bin/llama-server   -m /mnt/ollama/ggufs/Qwen3-30B-A3B-Thinking-2507-UD-Q8_K_XL.gguf   --port 8079   --host 0.0.0
    .0   -c 65536   -np 1   -ctk q8_0 -ctv q8_0   -cb   --no-warmup   --jinja   --chat-template-file UnForced-Thinking-TEMPLATE-Responses.j2   --split-mode layer  
     -ngl 99 --reasoning-format deepseek --reasoning-budget -1
    

    {# System Prompt Logic #}
    {%- set system_message = "You are a helpful assistant. IMPORTANT: You MUST always begin your response with a <think> tag to start your reasoning process." -%}
    {%- if messages[0]['role'] == 'system' -%}
        {%- set system_message = messages[0]['content'] -%}
    {%- endif -%}
    
    <|im_start|>system
    {{ system_message }}<|im_end|>
    
    {# Loop through remaining Messages #}
    {%- for message in messages -%}
        {# Skip System Prompt if it was handled above #}
        {%- if not (loop.first and message['role'] == 'system') -%}
            <|im_start|>{{ message['role'] }}
    
            {# Allow Deep Reasoning #}
            {%- set reasoning = message['reasoning_content'] or message['thinking'] -%}
            {%- if reasoning -%}
                <think>
                {{ reasoning }}
                </think>
            {%- endif -%}
    
            {# Completion vs Response #}
            {%- if message['content'] is string -%}
                {{ message['content'] }}
            {%- elif message['content'] is iterable -%}
                {%- for item in message['content'] -%}
                    {%- if item['type'] in ['text', 'input_text'] -%}
                        {{ item['text'] }}
                    {%- endif -%}
                {%- endfor -%}
            {%- endif -%}
    
            <|im_end|>
        {%- endif -%}
    {%- endfor -%}
    
    {# Provide <think> Tag for models used to Native Thinking flow #}
    {%- if add_generation_prompt -%}
    <|im_start|>assistant
    {%- endif -%}
    

    {# System Prompt Logic #}
    {%- set system_message = "You are a helpful assistant." -%}
    {%- if messages[0]['role'] == 'system' -%}
        {%- set system_message = messages[0]['content'] -%}
    {%- endif -%}
    
    <|im_start|>system
    {{ system_message }}<|im_end|>
    
    {# Loop through remaining Messages #}
    {%- for message in messages -%}
        {# Skip System Prompt if it was handled above #}
        {%- if not (loop.first and message['role'] == 'system') -%}
            <|im_start|>{{ message['role'] }}
    
            {# Allow Deep Reasoning #}
            {%- set reasoning = message['reasoning_content'] or message['thinking'] -%}
            {%- if reasoning -%}
                <think>
                {{ reasoning }}
                </think>
            {%- endif -%}
    
            {# Completion vs Response #}
            {%- if message['content'] is string -%}
                {{ message['content'] }}
            {%- elif message['content'] is iterable -%}
                {%- for item in message['content'] -%}
                    {%- if item['type'] in ['text', 'input_text'] -%}
                        {{ item['text'] }}
                    {%- endif -%}
                {%- endfor -%}
            {%- endif -%}
    
            <|im_end|>
        {%- endif -%}
    {%- endfor -%}
    
    {# Provide <think> Tag for models used to Native Thinking flow #}
    {%- if add_generation_prompt -%}
    <|im_start|>assistant
    <think>
    {%- endif -%}
    

    I can't get it to work anymore, specifically v1/responses. Any hints or confirmation at least it's just transition pains?
11. rekram1-node commented on Jan 31, 2026

    

    rekram1-node

    on Jan 31, 2026

    Contributor

    More actions

    I removed the thing, need to make it prettier later

    @daryltucker im not sure im following urs? Im closing cause I addressed issue OP raised
12. rekram1-node

    closed this as completedon Jan 31, 2026