refactor: Modularize proxy and add tool_use support

- Split monolithic messages.js (223 lines) into 5 focused modules
- Add full tool_use/tool_result round-trip translation
- Add x-api-key header authentication (Anthropic SDK default)
- Fix SSE Content-Type via Hono streamSSE helper
- Fix streaming usage tracking with stream_options.include_usage
- Add stop_reason mapping (end_turn, max_tokens, tool_use, stop_sequence)
- Forward stop_sequences to OpenAI stop parameter
- Handle system message as string or array of content blocks
- Use timing-safe XOR comparison for auth tokens
- Cache OpenAI client and model map across requests
- Sanitize error responses to prevent upstream detail leakage
- Use crypto.randomUUID() for unique message IDs
- Remove non-existent build/dev commands from vercel.json

src/openai-client.js

@@ -0,0 +1,31 @@
+import OpenAI from 'openai';
+
+// Cache OpenAI client and parsed model map to avoid re-creation per request
+let cachedClient = null;
+let cachedApiKey = null;
+let cachedModelMap = null;
+let cachedModelMapRaw = null;
+
+export function getOpenAIClient(env) {
+  if (cachedClient && cachedApiKey === env.OPENAI_API_KEY) {
+    return cachedClient;
+  }
+  cachedApiKey = env.OPENAI_API_KEY;
+  cachedClient = new OpenAI({ apiKey: env.OPENAI_API_KEY });
+  return cachedClient;
+}
+
+export function mapModel(claudeModel, env) {
+  const raw = env.MODEL_MAP || '';
+  if (raw !== cachedModelMapRaw) {
+    cachedModelMapRaw = raw;
+    cachedModelMap = Object.fromEntries(
+      raw.split(',').filter(Boolean).map((p) => {
+        const trimmed = p.trim();
+        const idx = trimmed.indexOf(':');
+        return idx > 0 ? [trimmed.slice(0, idx), trimmed.slice(idx + 1)] : [trimmed, trimmed];
+      })
+    );
+  }
+  return cachedModelMap[claudeModel] || claudeModel;
+}