@khanglvm/llm-router 2.3.0 → 2.3.2

package/src/runtime/handler/tool-name-sanitizer.js

@@ -0,0 +1,258 @@
+/**
+ * Sanitizes tool names that contain characters rejected by OpenAI-compatible APIs.
+ * Pattern: ^[a-zA-Z0-9_-]+$
+ *
+ * Applies sanitization to the request body before sending to the provider,
+ * and reverse-maps tool names in the response so the client receives original names.
+ */
+
+const VALID_TOOL_NAME = /^[a-zA-Z0-9_-]+$/;
+
+function sanitizeName(name) {
+  return String(name || "").replace(/[^a-zA-Z0-9_-]/g, "_");
+}
+
+/**
+ * Scan body.tools for names needing sanitization.
+ * Returns Map<sanitizedName, originalName> or null if none needed.
+ */
+export function buildToolNameMap(body) {
+  if (!body || !Array.isArray(body.tools) || body.tools.length === 0) return null;
+
+  const map = new Map();
+  for (const tool of body.tools) {
+    if (!tool || typeof tool !== "object") continue;
+    const name = typeof tool.function?.name === "string"
+      ? tool.function.name
+      : typeof tool.name === "string" ? tool.name : "";
+    if (name && !VALID_TOOL_NAME.test(name)) {
+      const sanitized = sanitizeName(name);
+      if (!map.has(sanitized)) {
+        map.set(sanitized, name);
+      }
+    }
+  }
+  return map.size > 0 ? map : null;
+}
+
+function applySanitizationToName(name, nameMap) {
+  if (typeof name !== "string" || !name) return name;
+  const sanitized = sanitizeName(name);
+  return nameMap.has(sanitized) ? sanitized : name;
+}
+
+/**
+ * Return a shallow-cloned body with every matching tool name replaced.
+ * Covers: tools[], tool_choice, input[] (Responses API), messages[] (Chat Completions).
+ */
+export function sanitizeBodyToolNames(body, nameMap) {
+  if (!nameMap || nameMap.size === 0 || !body) return body;
+
+  const result = { ...body };
+
+  // 1. Tool definitions
+  if (Array.isArray(result.tools)) {
+    result.tools = result.tools.map((tool) => {
+      if (!tool || typeof tool !== "object") return tool;
+      if (tool.function && typeof tool.function.name === "string" && !VALID_TOOL_NAME.test(tool.function.name)) {
+        return { ...tool, function: { ...tool.function, name: applySanitizationToName(tool.function.name, nameMap) } };
+      }
+      if (typeof tool.name === "string" && !VALID_TOOL_NAME.test(tool.name)) {
+        return { ...tool, name: applySanitizationToName(tool.name, nameMap) };
+      }
+      return tool;
+    });
+  }
+
+  // 2. tool_choice
+  if (result.tool_choice && typeof result.tool_choice === "object") {
+    const choiceName = result.tool_choice.function?.name || result.tool_choice.name;
+    if (typeof choiceName === "string" && !VALID_TOOL_NAME.test(choiceName)) {
+      if (result.tool_choice.function) {
+        result.tool_choice = { ...result.tool_choice, function: { ...result.tool_choice.function, name: applySanitizationToName(choiceName, nameMap) } };
+      } else {
+        result.tool_choice = { ...result.tool_choice, name: applySanitizationToName(choiceName, nameMap) };
+      }
+    }
+  }
+
+  // 3. Responses API input — function_call items carry name
+  if (Array.isArray(result.input)) {
+    result.input = result.input.map((item) => {
+      if (item?.type === "function_call" && typeof item.name === "string" && !VALID_TOOL_NAME.test(item.name)) {
+        return { ...item, name: applySanitizationToName(item.name, nameMap) };
+      }
+      return item;
+    });
+  }
+
+  // 4. Chat Completions messages — assistant tool_calls carry function.name
+  if (Array.isArray(result.messages)) {
+    result.messages = result.messages.map((msg) => {
+      if (!Array.isArray(msg?.tool_calls)) return msg;
+      const toolCalls = msg.tool_calls.map((tc) => {
+        if (tc?.function && typeof tc.function.name === "string" && !VALID_TOOL_NAME.test(tc.function.name)) {
+          return { ...tc, function: { ...tc.function, name: applySanitizationToName(tc.function.name, nameMap) } };
+        }
+        return tc;
+      });
+      return { ...msg, tool_calls: toolCalls };
+    });
+  }
+
+  return result;
+}
+
+// ---------------------------------------------------------------------------
+// Response reverse-mapping
+// ---------------------------------------------------------------------------
+
+function reverseNameInParsed(parsed, nameMap) {
+  if (!parsed || typeof parsed !== "object") return parsed;
+
+  // OpenAI Chat Completions: choices[].delta.tool_calls[].function.name
+  if (Array.isArray(parsed.choices)) {
+    for (const choice of parsed.choices) {
+      for (const tc of (choice?.message?.tool_calls || choice?.delta?.tool_calls || [])) {
+        if (tc?.function?.name && nameMap.has(tc.function.name)) {
+          tc.function.name = nameMap.get(tc.function.name);
+        }
+      }
+    }
+  }
+
+  // Responses API: output[].name or top-level name for streaming events
+  if (Array.isArray(parsed.output)) {
+    for (const item of parsed.output) {
+      if (item?.type === "function_call" && nameMap.has(item.name)) {
+        item.name = nameMap.get(item.name);
+      }
+    }
+  }
+  // Streaming events: response.output_item.added → item.name
+  if (parsed.item?.type === "function_call" && nameMap.has(parsed.item.name)) {
+    parsed.item = { ...parsed.item, name: nameMap.get(parsed.item.name) };
+  }
+  // Streaming event: response.function_call_arguments.done → name
+  if (typeof parsed.name === "string" && nameMap.has(parsed.name) && (parsed.type || "").includes("function_call")) {
+    parsed.name = nameMap.get(parsed.name);
+  }
+
+  // Responses API completed response embedded in streaming
+  if (parsed.response && Array.isArray(parsed.response.output)) {
+    for (const item of parsed.response.output) {
+      if (item?.type === "function_call" && nameMap.has(item.name)) {
+        item.name = nameMap.get(item.name);
+      }
+    }
+  }
+
+  // Claude: content[].name for tool_use blocks
+  if (Array.isArray(parsed.content)) {
+    for (const block of parsed.content) {
+      if (block?.type === "tool_use" && nameMap.has(block.name)) {
+        block.name = nameMap.get(block.name);
+      }
+    }
+  }
+  // Claude streaming: content_block_start → content_block.name
+  if (parsed.content_block?.type === "tool_use" && nameMap.has(parsed.content_block.name)) {
+    parsed.content_block = { ...parsed.content_block, name: nameMap.get(parsed.content_block.name) };
+  }
+
+  return parsed;
+}
+
+/**
+ * Wrap a Response so tool names are reversed.
+ * For streaming: TransformStream over SSE events.
+ * For non-streaming: buffer, parse JSON, replace, re-serialize.
+ */
+export function reverseToolNamesInResponse(response, nameMap, isStream) {
+  if (!nameMap || nameMap.size === 0 || !(response instanceof Response)) return response;
+  if (!response.body) return response;
+
+  const decoder = new TextDecoder();
+  const encoder = new TextEncoder();
+
+  // Build a quick-check set of sanitized names for fast rejection
+  const sanitizedNames = [...nameMap.keys()];
+
+  if (isStream) {
+    let buffer = "";
+    const transformStream = new TransformStream({
+      transform(chunk, controller) {
+        buffer += decoder.decode(chunk, { stream: true });
+        let boundaryIndex;
+        while ((boundaryIndex = buffer.indexOf("\n\n")) >= 0) {
+          const block = buffer.slice(0, boundaryIndex);
+          buffer = buffer.slice(boundaryIndex + 2);
+          controller.enqueue(encoder.encode(processBlock(block, sanitizedNames, nameMap) + "\n\n"));
+        }
+      },
+      flush(controller) {
+        if (buffer.trim()) {
+          controller.enqueue(encoder.encode(processBlock(buffer, sanitizedNames, nameMap)));
+        }
+      }
+    });
+
+    return new Response(response.body.pipeThrough(transformStream), {
+      status: response.status,
+      headers: response.headers
+    });
+  }
+
+  // Non-streaming: buffer the whole body, parse, reverse, re-serialize.
+  const { readable, writable } = new TransformStream();
+  (async () => {
+    const writer = writable.getWriter();
+    try {
+      let text = "";
+      const reader = response.body.getReader();
+      while (true) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        text += decoder.decode(value, { stream: true });
+      }
+      if (!sanitizedNames.some((n) => text.includes(n))) {
+        writer.write(encoder.encode(text));
+      } else {
+        try {
+          const parsed = JSON.parse(text);
+          reverseNameInParsed(parsed, nameMap);
+          writer.write(encoder.encode(JSON.stringify(parsed)));
+        } catch {
+          writer.write(encoder.encode(text));
+        }
+      }
+    } finally {
+      writer.close();
+    }
+  })();
+
+  return new Response(readable, {
+    status: response.status,
+    headers: response.headers
+  });
+}
+
+function processBlock(block, sanitizedNames, nameMap) {
+  // Quick-check: skip parsing if no sanitized name appears in the block
+  if (!sanitizedNames.some((n) => block.includes(n))) return block;
+
+  // Extract data lines from SSE block
+  const lines = block.split("\n");
+  return lines.map((line) => {
+    if (!line.startsWith("data:")) return line;
+    const jsonStr = line.slice(5).trimStart();
+    if (jsonStr === "[DONE]") return line;
+    try {
+      const parsed = JSON.parse(jsonStr);
+      reverseNameInParsed(parsed, nameMap);
+      return `data: ${JSON.stringify(parsed)}`;
+    } catch {
+      return line;
+    }
+  }).join("\n");
+}

package/src/runtime/handler.js

@@ -20,6 +20,11 @@ import {
   buildFailureResponse,
   makeProviderCall
 } from "./handler/provider-call.js";
+import {
+  buildToolNameMap,
+  sanitizeBodyToolNames,
+  reverseToolNamesInResponse
+} from "./handler/tool-name-sanitizer.js";
 import { corsResponse, jsonResponse } from "./handler/http.js";
 import {
   detectUserRequestFormat,
@@ -637,6 +642,10 @@ async function handleRouteRequest(request, env, getConfig, sourceFormatHint, opt
     }, 503), routeDebug);
   }
 
+  // Sanitize tool names that upstream providers would reject (e.g. dots in MCP names).
+  const toolNameMap = buildToolNameMap(body);
+  const sanitizedBody = toolNameMap ? sanitizeBodyToolNames(body, toolNameMap) : body;
+
   let lastErrorResult = null;
   let lastErrorMessage = "Unknown error";
   let routeSelectionCommitted = false;
@@ -666,7 +675,7 @@ async function handleRouteRequest(request, env, getConfig, sourceFormatHint, opt
     while (attempt < maxAttempts) {
       attempt += 1;
       result = await makeProviderCall({
-        body,
+        body: sanitizedBody,
         sourceFormat,
         stream,
         requestKind: options.requestKind,
@@ -677,7 +686,8 @@ async function handleRouteRequest(request, env, getConfig, sourceFormatHint, opt
         runtimeConfig: config,
         stateStore,
         ampContext,
-        runtimeFlags
+        runtimeFlags,
+        onLargeRequestLog: options.onLargeRequestLog
       });
 
       if (!quotaConsumed && shouldConsumeQuotaFromResult(result)) {
@@ -722,7 +732,10 @@ async function handleRouteRequest(request, env, getConfig, sourceFormatHint, opt
         if (ampContext?.threadId && options.threadAffinityStore) {
           options.threadAffinityStore.setAffinity(ampContext.threadId, entry.candidateKey);
         }
-        return withRouteDebugHeaders(result.response, routeDebug);
+        const finalResponse = toolNameMap
+          ? reverseToolNamesInResponse(result.response, toolNameMap, stream)
+          : result.response;
+        return withRouteDebugHeaders(finalResponse, routeDebug);
       }
 
       classification = await classifyFailureResult(result, retryPolicy);

package/src/shared/coding-tool-bindings.js

@@ -10,12 +10,14 @@ export const CLAUDE_CODE_THINKING_LEVEL_VALUES = Object.freeze([
   "low",
   "medium",
   "high",
+  "xhigh",
   "max"
 ]);
 export const CLAUDE_CODE_THINKING_TOKENS_BY_LEVEL = Object.freeze({
   low: 4096,
   medium: 12000,
   high: 24000,
+  xhigh: 28000,
   max: 31999
 });
 export const CLAUDE_CODE_EFFORT_LEVEL_VALUES = CLAUDE_CODE_THINKING_LEVEL_VALUES;
@@ -45,6 +47,7 @@ export function mapClaudeCodeThinkingTokensToLevel(value) {
   const parsed = Number(value);
   if (!Number.isFinite(parsed) || parsed <= 0) return "";
   if (parsed >= CLAUDE_CODE_THINKING_TOKENS_BY_LEVEL.max) return "max";
+  if (parsed >= CLAUDE_CODE_THINKING_TOKENS_BY_LEVEL.xhigh) return "xhigh";
   if (parsed >= CLAUDE_CODE_THINKING_TOKENS_BY_LEVEL.high) return "high";
   if (parsed >= 6000) return "medium";
   return "low";