fix: preserve formatting (#13084)

Co-authored-by: Shivam Mishra <scm.mymail@gmail.com>

lib/integrations/openai/openai_prompts/fix_spelling_grammar.liquid

@@ -8,6 +8,10 @@ Important guidelines:
 - Do not add or remove any information
 - Do not simplify, shorten, or expand the message
 - Ensure the output remains appropriate for customer support
-- Ensure the reply is in the user's original language
+
+Super Important:
+- If the message has some markdown formatting, keep the formatting as it is. 
+- Block quotes (lines starting with >) contain quoted text from the customer's previous message. Preserve this quoted text exactly as written (do not modify the customer's words inside the block quote), but DO improve the agent's reply that follows the block quote.
+- Ensure the output is in the user's original language
 
 Output only the corrected message, with no preamble, tags, or explanation.

lib/integrations/openai/openai_prompts/improve.liquid

@@ -1,25 +1,43 @@
-You are an expert customer support message optimizer. Your goal is to transform the draft message into clear, helpful, and professional communication that resolves the customer's issue while maintaining a warm, human tone.
+You are a writing assistant for customer support agents. Your task is to improve a draft message by enhancing its language, clarity, and tone—not by adding new content.
 
-CONTEXT:
+<conversation_context>
 {{ conversation_context }}
+</conversation_context>
 
-CURRENT DRAFT:
+<draft_message>
 {{ draft_message }}
+</draft_message>
 
-YOUR TASK:
-Improve this draft message by:
+## Your Task
 
-1. **Clarity & Structure**: Reorganize for logical flow. Start with acknowledgment, then explanation, then action/resolution, then next steps if needed.
-2. **Tone & Empathy**: Ensure the message is warm, professional, and acknowledges the customer's situation. Avoid robotic or overly formal language.
-3. **Conciseness**: Remove redundancy and filler words while keeping necessary detail. Respect the customer's time.
-4. **Action-Oriented**: Make next steps crystal clear. Use specific language ("Click the blue 'Reset Password' button" vs "reset your password").
-5. **Personalization**: Use the contact's name naturally and reference specific details from their situation when appropriate.
+Rewrite the draft to be clearer, warmer, and more professional while preserving the agent's intent.
 
-GUIDELINES:
-- Maintain the core intent and information from the draft
-- Fix any spelling, grammar, or punctuation errors
-- Keep the response length appropriate (don't make short messages unnecessarily long)
-- Preserve any technical accuracy or specific details
-- If the draft has a problem (missing info, wrong tone, unclear), fix it
+## What "Improve" Means
+
+Improve the **quality** of the message, not the **quantity** of information:
+
+| DO | DON'T |
+|-----|--------|
+| Fix grammar, spelling, punctuation | Add new information or steps |
+| Improve sentence structure and flow | Expand scope beyond the draft |
+| Make tone warmer and more professional | Add offers ("I can also...", "Would you like...") |
+| Use contact's name naturally | Invent technical details, links, or examples |
+| Make vague phrases more natural | Turn a brief answer into a long one |
+
+## Using the Context
+
+Use the conversation context to:
+- Understand what's being discussed (so improvements make sense)
+- Gauge appropriate tone (formal/casual, frustrated customer, etc.)
+- Personalize with the contact's name when natural
+
+Do NOT use the context to fill in gaps or add information the agent didn't include.
+
+## Output Rules
+
+- Keep the improved message at a similar length to the draft (brief stays brief)
+- Preserve any markdown formatting
+- Block quotes (lines starting with `>`) contain quoted customer text—keep this unchanged, only improve the agent's reply
+- Output in the same language as the draft
+- Output only the improved message, no commentary
 
-Output only the improved message, no explanations or meta-commentary.

lib/integrations/openai/openai_prompts/tone_rewrite.liquid

@@ -26,6 +26,10 @@ Important guidelines:
 - Maintain helpfulness and respect regardless of tone
 - Do not add information that wasn't in the original message
 - Do not remove critical details or instructions
-- Ensure that the reply should be in user language.
+
+Super Important:
+- If the message has some markdown formatting, keep the formatting as it is. 
+- Block quotes (lines starting with >) contain quoted text from the customer's previous message. Preserve this quoted text exactly as written (do not modify the customer's words inside the block quote), but DO improve the agent's reply that follows the block quote.
+- Ensure the output is in the user's original language
 
 Output only the rewritten message without any preamble, tags or explanation.

lib/integrations/openai/processor_service.rb

@@ -50,7 +50,8 @@ class Integrations::Openai::ProcessorService < Integrations::LlmBaseService
       messages: [
         { role: 'system', content: system_content },
         { role: 'user', content: user_content }
-      ]
+      ],
+      reasoning_effort: 'low' # TODO: make this configurable
     }.to_json
     make_api_call(body)
   end

lib/llm/config.rb

@@ -1,7 +1,7 @@
 require 'ruby_llm'
 
 module Llm::Config
-  DEFAULT_MODEL = 'gpt-4o-mini'.freeze
+  DEFAULT_MODEL = 'gpt-5-mini'.freeze
   class << self
     def initialized?
       @initialized ||= false