free-coding-models 0.3.9 → 0.3.12

package/src/anthropic-translator.js

@@ -1,440 +0,0 @@
-/**
- * @file src/anthropic-translator.js
- * @description Bidirectional wire format translation between Anthropic Messages API
- * and OpenAI Chat Completions API.
- *
- * 📖 This is the key module that enables Claude Code to work natively through the
- *    FCM proxy without needing the external Claude proxy integration.
- *    Claude Code sends requests in Anthropic format (POST /v1/messages) and this
- *    module translates them to OpenAI format for the upstream providers, then
- *    translates the responses back.
- *
- * 📖 Supports both JSON and SSE streaming modes.
- *
- * @functions
- *   → translateAnthropicToOpenAI(body) — Convert Anthropic Messages request → OpenAI chat completions
- *   → translateOpenAIToAnthropic(openaiResponse, requestModel) — Convert OpenAI JSON response → Anthropic
- *   → createAnthropicSSETransformer(requestModel) — Create a Transform stream for SSE translation
- *   → estimateAnthropicTokens(body) — Fast local token estimate for `/v1/messages/count_tokens`
- *
- * @exports translateAnthropicToOpenAI, translateOpenAIToAnthropic, createAnthropicSSETransformer, estimateAnthropicTokens
- * @see src/proxy-server.js — routes /v1/messages through this translator
- */
-
-import { Transform } from 'node:stream'
-import { randomUUID } from 'node:crypto'
-
-function normalizeThinkingText(block) {
-  if (!block || typeof block !== 'object') return ''
-  if (typeof block.text === 'string' && block.text) return block.text
-  if (typeof block.thinking === 'string' && block.thinking) return block.thinking
-  if (typeof block.summary === 'string' && block.summary) return block.summary
-  return ''
-}
-
-function contentBlocksToText(blocks, { includeThinking = false } = {}) {
-  return blocks
-    .map((block) => {
-      if (block?.type === 'thinking' && includeThinking) {
-        const thinkingText = normalizeThinkingText(block)
-        return thinkingText ? `<thinking>${thinkingText}</thinking>` : ''
-      }
-      if (block?.type === 'redacted_thinking' && includeThinking) {
-        return '<thinking>[redacted]</thinking>'
-      }
-      return block?.type === 'text' ? block.text : ''
-    })
-    .filter(Boolean)
-}
-
-function extractReasoningBlocks(message = {}) {
-  if (Array.isArray(message.reasoning)) {
-    return message.reasoning
-      .map((entry) => normalizeThinkingText(entry))
-      .filter(Boolean)
-      .map((text) => ({ type: 'thinking', thinking: text }))
-  }
-  if (typeof message.reasoning_content === 'string' && message.reasoning_content.trim()) {
-    return [{ type: 'thinking', thinking: message.reasoning_content.trim() }]
-  }
-  return []
-}
-
-/**
- * 📖 Translate an Anthropic Messages API request body to OpenAI Chat Completions format.
- *
- * Anthropic format:
- *   { model, messages: [{role, content}], system, max_tokens, stream, temperature, top_p, stop_sequences }
- *
- * OpenAI format:
- *   { model, messages: [{role, content}], max_tokens, stream, temperature, top_p, stop }
- *
- * @param {object} body — Anthropic request body
- * @returns {object} — OpenAI-compatible request body
- */
-export function translateAnthropicToOpenAI(body) {
-  // 📖 Guard against null/undefined/non-object input
-  if (!body || typeof body !== 'object') return { model: '', messages: [], stream: false }
-  if (!Array.isArray(body.messages)) body = { ...body, messages: [] }
-
-  const openaiMessages = []
-
-  // 📖 Anthropic "system" field → OpenAI system message
-  if (body.system) {
-    if (typeof body.system === 'string') {
-      openaiMessages.push({ role: 'system', content: body.system })
-    } else if (Array.isArray(body.system)) {
-      // 📖 Anthropic supports system as array of content blocks
-      const text = contentBlocksToText(body.system, { includeThinking: true }).join('\n\n')
-      if (text) openaiMessages.push({ role: 'system', content: text })
-    }
-  }
-
-  // 📖 Convert Anthropic messages to OpenAI format
-  if (Array.isArray(body.messages)) {
-    for (const msg of body.messages) {
-      const role = msg.role === 'assistant' ? 'assistant' : 'user'
-
-      if (typeof msg.content === 'string') {
-        openaiMessages.push({ role, content: msg.content })
-      } else if (Array.isArray(msg.content)) {
-        // 📖 Anthropic content blocks: [{type: "text", text: "..."}, {type: "tool_result", ...}]
-        const textParts = contentBlocksToText(msg.content, { includeThinking: true })
-        const toolResults = msg.content.filter(b => b.type === 'tool_result')
-        const toolUses = msg.content.filter(b => b.type === 'tool_use')
-
-        // 📖 Tool use blocks (assistant) → OpenAI tool_calls
-        if (toolUses.length > 0 && role === 'assistant') {
-          const toolCalls = toolUses.map(tu => ({
-            id: tu.id || randomUUID(),
-            type: 'function',
-            function: {
-              name: tu.name,
-              arguments: typeof tu.input === 'string' ? tu.input : JSON.stringify(tu.input || {}),
-            }
-          }))
-          openaiMessages.push({
-            role: 'assistant',
-            content: textParts.join('\n') || null,
-            tool_calls: toolCalls,
-          })
-        }
-        // 📖 Tool result blocks (user) → OpenAI tool messages
-        else if (toolResults.length > 0) {
-          // 📖 First push any text parts as user message
-          if (textParts.length > 0) {
-            openaiMessages.push({ role: 'user', content: textParts.join('\n') })
-          }
-          for (const tr of toolResults) {
-            const content = typeof tr.content === 'string'
-              ? tr.content
-              : Array.isArray(tr.content)
-                ? tr.content.filter(b => b.type === 'text').map(b => b.text).join('\n')
-                : JSON.stringify(tr.content || '')
-            openaiMessages.push({
-              role: 'tool',
-              tool_call_id: tr.tool_use_id || tr.id || '',
-              content,
-            })
-          }
-        }
-        // 📖 Plain text content blocks → join into single message
-        else if (textParts.length > 0) {
-          openaiMessages.push({ role, content: textParts.join('\n') })
-        }
-      }
-    }
-  }
-
-  const result = {
-    model: body.model,
-    messages: openaiMessages,
-    stream: body.stream === true,
-  }
-
-  // 📖 Map Anthropic parameters to OpenAI equivalents
-  if (body.max_tokens != null) result.max_tokens = body.max_tokens
-  if (body.temperature != null) result.temperature = body.temperature
-  if (body.top_p != null) result.top_p = body.top_p
-  if (Array.isArray(body.stop_sequences) && body.stop_sequences.length > 0) {
-    result.stop = body.stop_sequences
-  }
-
-  // 📖 Map Anthropic tools to OpenAI function tools
-  if (Array.isArray(body.tools) && body.tools.length > 0) {
-    result.tools = body.tools.map(tool => ({
-      type: 'function',
-      function: {
-        name: tool.name,
-        description: tool.description || '',
-        parameters: tool.input_schema || {},
-      }
-    }))
-  }
-
-  return result
-}
-
-/**
- * 📖 Translate an OpenAI Chat Completions JSON response to Anthropic Messages format.
- *
- * @param {object} openaiResponse — parsed OpenAI response
- * @param {string} requestModel — model name from the original request
- * @returns {object} — Anthropic Messages response
- */
-export function translateOpenAIToAnthropic(openaiResponse, requestModel) {
-  const choice = openaiResponse.choices?.[0]
-  const message = choice?.message || {}
-  const content = []
-
-  for (const reasoningBlock of extractReasoningBlocks(message)) {
-    content.push(reasoningBlock)
-  }
-
-  // 📖 Text content → Anthropic text block
-  if (message.content) {
-    content.push({ type: 'text', text: message.content })
-  }
-
-  // 📖 Tool calls → Anthropic tool_use blocks
-  if (Array.isArray(message.tool_calls)) {
-    for (const tc of message.tool_calls) {
-      let input = {}
-      try { input = JSON.parse(tc.function?.arguments || '{}') } catch { /* ignore */ }
-      content.push({
-        type: 'tool_use',
-        id: tc.id || randomUUID(),
-        name: tc.function?.name || '',
-        input,
-      })
-    }
-  }
-
-  // 📖 Fallback: Anthropic requires at least one content block — provide empty text if none
-  if (content.length === 0) {
-    content.push({ type: 'text', text: '' })
-  }
-
-  // 📖 Map OpenAI finish_reason → Anthropic stop_reason
-  let stopReason = 'end_turn'
-  if (choice?.finish_reason === 'stop') stopReason = 'end_turn'
-  else if (choice?.finish_reason === 'length') stopReason = 'max_tokens'
-  else if (choice?.finish_reason === 'tool_calls') stopReason = 'tool_use'
-
-  return {
-    id: openaiResponse.id || `msg_${randomUUID().replace(/-/g, '')}`,
-    type: 'message',
-    role: 'assistant',
-    content,
-    model: requestModel || openaiResponse.model || '',
-    stop_reason: stopReason,
-    stop_sequence: null,
-    usage: {
-      input_tokens: openaiResponse.usage?.prompt_tokens || 0,
-      output_tokens: openaiResponse.usage?.completion_tokens || 0,
-    },
-  }
-}
-
-/**
- * 📖 Create a Transform stream that converts OpenAI SSE chunks to Anthropic SSE format.
- *
- * OpenAI SSE:
- *   data: {"choices":[{"delta":{"content":"Hello"}}]}
- *
- * Anthropic SSE:
- *   event: message_start
- *   data: {"type":"message_start","message":{...}}
- *
- *   event: content_block_start
- *   data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}
- *
- *   event: content_block_delta
- *   data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Hello"}}
- *
- *   event: message_stop
- *   data: {"type":"message_stop"}
- *
- * @param {string} requestModel — model from original request
- * @returns {{ transform: Transform, getUsage: () => object }}
- */
-// 📖 Max SSE buffer size to prevent memory exhaustion from malformed streams (1 MB)
-const MAX_SSE_BUFFER = 1 * 1024 * 1024
-
-export function createAnthropicSSETransformer(requestModel) {
-  let headerSent = false
-  // 📖 Track block indices for proper content_block_start/stop/delta indexing.
-  // nextBlockIndex increments for each new content block (text or tool_use).
-  // currentBlockIndex tracks the index of the most recently opened block.
-  let nextBlockIndex = 0
-  let currentBlockIndex = -1
-  let inputTokens = 0
-  let outputTokens = 0
-  let buffer = ''
-
-  const transform = new Transform({
-    transform(chunk, encoding, callback) {
-      buffer += chunk.toString()
-      // 📖 Guard against unbounded buffer growth from malformed SSE streams
-      if (buffer.length > MAX_SSE_BUFFER) {
-        buffer = ''
-        return callback(new Error('SSE buffer overflow'))
-      }
-      const lines = buffer.split('\n')
-      // 📖 Keep the last incomplete line in the buffer
-      buffer = lines.pop() || ''
-
-      for (const line of lines) {
-        if (!line.startsWith('data: ')) continue
-        const payload = line.slice(6).trim()
-        if (payload === '[DONE]') {
-          // 📖 End of stream — close any open block, then send message_delta + message_stop
-          if (currentBlockIndex >= 0) {
-            this.push(`event: content_block_stop\ndata: ${JSON.stringify({ type: 'content_block_stop', index: currentBlockIndex })}\n\n`)
-            currentBlockIndex = -1
-          }
-          this.push(`event: message_delta\ndata: ${JSON.stringify({
-            type: 'message_delta',
-            delta: { stop_reason: 'end_turn', stop_sequence: null },
-            usage: { output_tokens: outputTokens },
-          })}\n\n`)
-          this.push(`event: message_stop\ndata: ${JSON.stringify({ type: 'message_stop' })}\n\n`)
-          continue
-        }
-
-        let parsed
-        try { parsed = JSON.parse(payload) } catch { continue }
-
-        // 📖 Send message_start header on first chunk
-        if (!headerSent) {
-          headerSent = true
-          this.push(`event: message_start\ndata: ${JSON.stringify({
-            type: 'message_start',
-            message: {
-              id: parsed.id || `msg_${randomUUID().replace(/-/g, '')}`,
-              type: 'message',
-              role: 'assistant',
-              content: [],
-              model: requestModel || parsed.model || '',
-              stop_reason: null,
-              stop_sequence: null,
-              usage: { input_tokens: inputTokens, output_tokens: 0 },
-            },
-          })}\n\n`)
-        }
-
-        const choice = parsed.choices?.[0]
-        if (!choice) continue
-        const delta = choice.delta || {}
-
-        // 📖 Track usage if present
-        if (parsed.usage) {
-          inputTokens = parsed.usage.prompt_tokens || inputTokens
-          outputTokens = parsed.usage.completion_tokens || outputTokens
-        }
-
-        // 📖 Text delta
-        if (delta.content) {
-          if (currentBlockIndex < 0 || nextBlockIndex === 0) {
-            // 📖 Open first text block
-            currentBlockIndex = nextBlockIndex++
-            this.push(`event: content_block_start\ndata: ${JSON.stringify({
-              type: 'content_block_start',
-              index: currentBlockIndex,
-              content_block: { type: 'text', text: '' },
-            })}\n\n`)
-          }
-          this.push(`event: content_block_delta\ndata: ${JSON.stringify({
-            type: 'content_block_delta',
-            index: currentBlockIndex,
-            delta: { type: 'text_delta', text: delta.content },
-          })}\n\n`)
-        }
-
-        // 📖 Tool call deltas (if model supports tool use)
-        if (Array.isArray(delta.tool_calls)) {
-          for (const tc of delta.tool_calls) {
-            if (tc.function?.name) {
-              // 📖 New tool call — close previous block if open, then start new one
-              if (currentBlockIndex >= 0) {
-                this.push(`event: content_block_stop\ndata: ${JSON.stringify({ type: 'content_block_stop', index: currentBlockIndex })}\n\n`)
-              }
-              currentBlockIndex = nextBlockIndex++
-              this.push(`event: content_block_start\ndata: ${JSON.stringify({
-                type: 'content_block_start',
-                index: currentBlockIndex,
-                content_block: {
-                  type: 'tool_use',
-                  id: tc.id || randomUUID(),
-                  name: tc.function.name,
-                  input: {},
-                },
-              })}\n\n`)
-            }
-            if (tc.function?.arguments) {
-              this.push(`event: content_block_delta\ndata: ${JSON.stringify({
-                type: 'content_block_delta',
-                index: currentBlockIndex >= 0 ? currentBlockIndex : 0,
-                delta: { type: 'input_json_delta', partial_json: tc.function.arguments },
-              })}\n\n`)
-            }
-          }
-        }
-
-        // 📖 Handle finish_reason for tool_calls — close ALL open blocks
-        if (choice.finish_reason === 'tool_calls' && currentBlockIndex >= 0) {
-          this.push(`event: content_block_stop\ndata: ${JSON.stringify({ type: 'content_block_stop', index: currentBlockIndex })}\n\n`)
-          currentBlockIndex = -1
-        }
-      }
-      callback()
-    },
-
-    flush(callback) {
-      // 📖 Process any remaining buffer
-      if (buffer.trim()) {
-        // ignore incomplete data
-      }
-      callback()
-    },
-  })
-
-  return {
-    transform,
-    getUsage: () => ({ input_tokens: inputTokens, output_tokens: outputTokens }),
-  }
-}
-
-function estimateTokenCountFromText(text) {
-  const normalized = String(text || '').trim()
-  if (!normalized) return 0
-  return Math.ceil(normalized.length / 4)
-}
-
-export function estimateAnthropicTokens(body) {
-  const openaiBody = translateAnthropicToOpenAI(body)
-  const messageTokens = Array.isArray(openaiBody.messages)
-    ? openaiBody.messages.reduce((total, message) => {
-        let nextTotal = total + 4
-        if (typeof message.content === 'string') {
-          nextTotal += estimateTokenCountFromText(message.content)
-        }
-        if (Array.isArray(message.tool_calls)) {
-          for (const toolCall of message.tool_calls) {
-            nextTotal += estimateTokenCountFromText(toolCall.function?.name || '')
-            nextTotal += estimateTokenCountFromText(toolCall.function?.arguments || '')
-          }
-        }
-        if (typeof message.tool_call_id === 'string') {
-          nextTotal += estimateTokenCountFromText(message.tool_call_id)
-        }
-        return nextTotal
-      }, 2)
-    : 0
-
-  const toolTokens = Array.isArray(body?.tools)
-    ? body.tools.reduce((total, tool) => total + estimateTokenCountFromText(JSON.stringify(tool || {})), 0)
-    : 0
-
-  return messageTokens + toolTokens
-}