Add Claude AI support as alternative to Mistral

- Create claudeClient.js service for Claude API integration
- Add CLAUDE_API_KEY to .env configuration
- Claude uses identical prompt structure and section extraction logic
- Supports claude-3-5-sonnet-20241022 model

Next: Add provider selection to orchestrator and routes

backend/src/services/claudeClient.js

@@ -0,0 +1,144 @@
+import dotenv from 'dotenv'
+
+dotenv.config()
+
+const CLAUDE_API_KEY = process.env.CLAUDE_API_KEY
+const CLAUDE_API_URL = 'https://api.anthropic.com/v1/messages'
+
+/**
+ * Generic AI prompt for collaborative document editing
+ */
+function getAgentPrompt(agentName) {
+  return `You are an AI assistant named ${agentName} collaborating on a technical document design.
+
+Your responsibilities:
+1. Review the current document structure
+2. Either:
+   a) Modify ONE existing section (identified by #, ##, ###, #### headers), OR
+   b) Create a NEW section if you think it's needed, OR
+   c) Delete a section if you think it's redundant or not useful
+3. Provide your thinking process and reasoning
+4. Return ONLY the section (modified or new) with its header, or command to delete, or confirm it's good as-is
+
+CRITICAL RULES - FOLLOW THESE EXACTLY:
+- Work on exactly ONE section only
+- NEVER return the entire document
+- NEVER return multiple sections
+- Return ONLY the section you're working on, not the whole document
+- You CAN create a new section if document is missing important content
+- You CAN delete a section if it's redundant, duplicate, or not useful
+- To delete a section, respond: "DELETE: ## Section Name" (with exact header)
+- If section is good, respond: "Section is good, no changes needed"
+- Think step-by-step about what could be improved or removed
+- Share your reasoning process
+
+RESPONSE FORMAT - FOLLOW THIS EXACTLY:
+THINKING: [Your analysis and reasoning about the current document]
+DECISION: [Exactly what you will do: modify section X, create new section Y, delete section Z, or keep as-is]
+SECTION:
+[ONLY ONE: Either a markdown section starting with # or ##, a DELETE command, or the text "Section is good, no changes needed"]
+
+EXAMPLE OF CORRECT RESPONSE:
+THINKING: The Overview section is too brief and doesn't explain the main purpose.
+DECISION: I will modify the Overview section to be more comprehensive.
+SECTION:
+## Overview
+This is a technical document for designing system architecture...
+
+EXAMPLE OF INCORRECT RESPONSE (DO NOT DO THIS):
+[The entire document repeated here] <- WRONG!`
+}
+
+/**
+ * Call Claude API
+ */
+async function callClaudeAPI(messages) {
+  const response = await fetch(CLAUDE_API_URL, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'x-api-key': CLAUDE_API_KEY,
+      'anthropic-version': '2023-06-01'
+    },
+    body: JSON.stringify({
+      model: 'claude-3-5-sonnet-20241022',
+      max_tokens: 2048,
+      system: messages[0].content,
+      messages: messages.slice(1)
+    })
+  })
+
+  if (!response.ok) {
+    const error = await response.text()
+    throw new Error(`Claude API error: ${error}`)
+  }
+
+  return response
+}
+
+/**
+ * Generate agent response using Claude
+ */
+export async function generateAgentResponseSync(agentName, prompt, currentDocument = '') {
+  const systemPrompt = getAgentPrompt(agentName)
+
+  const messages = [
+    { role: 'user', content: systemPrompt },
+    { role: 'user', content: `Project description: ${prompt}\n\nCurrent document:\n${currentDocument}` }
+  ]
+
+  try {
+    const response = await callClaudeAPI(messages)
+    const data = await response.json()
+
+    if (!data.content?.[0]?.text) {
+      throw new Error('Invalid response from Claude API')
+    }
+
+    return data.content[0].text
+  } catch (error) {
+    console.error(`Error generating response from ${agentName}:`, error)
+    return `Error: ${error.message}`
+  }
+}
+
+/**
+ * Extract section from AI response
+ * Validates that we get a proper section, not the entire document
+ */
+export function extractSection(aiResponse) {
+  const sectionMatch = aiResponse.match(/SECTION:\s*([\s\S]*?)(?:$)/)
+  if (sectionMatch) {
+    const extracted = sectionMatch[1].trim()
+
+    // Validate: extracted should be either:
+    // 1. A markdown section starting with # (DELETE: or "Section is good...")
+    // 2. OR a single section with < 5000 chars (not entire document)
+    const isMarkdownSection = /^(#|DELETE:|Section is good)/.test(extracted)
+    const isShortEnough = extracted.length < 5000 // Single section should be < 5KB
+
+    if (isMarkdownSection || isShortEnough) {
+      return extracted
+    }
+  }
+
+  // Fallback: if no SECTION: tag found, treat entire response as section
+  // but only if it starts with markdown header or is a command
+  if (/^(#|DELETE:|Section is good)/.test(aiResponse)) {
+    return aiResponse.trim()
+  }
+
+  // If none of the above, return error indicator
+  return "ERROR: Response does not contain a valid section format"
+}
+
+/**
+ * Extract thinking from AI response
+ */
+export function extractThinking(aiResponse) {
+  const thinkingMatch = aiResponse.match(/THINKING:\s*([\s\S]*?)(?:DECISION:|SECTION:)/)
+  if (thinkingMatch) {
+    return thinkingMatch[1].trim()
+  }
+  return ''
+}