## Turn 1

Here's a thinking process:

1.  **Analyze User Input:**
   - **Context:** Home-lab user, OpenWebUI interface, local model via llama.cpp proxy.
   - **Problem:** Model "spends all its tokens thinking and never answers."
   - **Request:** Practical explanation, settings to check, prompt design for useful final answers.
   - **Required Sections:** 
     - # Short Diagnosis
     - # Why This Happens
     - # Settings To Check
     - # Prompt Design
     - # Benchmark Design
     - # What The Flight Recorder Should Show
     - # Practical Defaults
     - # When To Increase Budgets
     - # When To Stop A Run
   - **Tone/Style:** Accessible but technical, include good/bad prompt examples.
   - **Constraints:** Target ~4,500 tokens. Do not stop early. Prioritize completeness/structure. Keep in `message.content`.

2.  **Deconstruct Required Sections & Plan Content:**
   I need to expand each section thoroughly to meet the ~4,500 token target while maintaining high quality and technical accuracy. I'll structure each section with subsections, explanations, examples, and practical advice.

   *# Short Diagnosis*
   - Briefly state the core issue: context window exhaustion, runaway generation, misconfigured stop sequences, or over-reliance on chain-of-thought/internal reasoning tokens without a clear output trigger.
   - Mention llama.cpp specific behaviors (max tokens, stop strings, temperature/repetition penalties causing loops).

   *# Why This Happens*
   - Explain the mechanics: LLMs are autoregressive. They predict the next token until a stop condition is met.
   - Discuss "thinking" vs "answering": Many modern models are trained with explicit reasoning formats (e.g., `<think>...