metrillm-mcp 0.2.4 → 0.2.5

package/dist/index.js

@@ -735,6 +735,7 @@ var defaultKeepAlive2;
 var activeAbortControllers = /* @__PURE__ */ new Set();
 var directorySizeCache = /* @__PURE__ */ new Map();
 var modelDefinitionCache = /* @__PURE__ */ new Map();
+var outputLimitModeCache = /* @__PURE__ */ new Map();
 var NON_THINKING_SYSTEM_PROMPT = [
   "You are in non-thinking mode for benchmark reproducibility.",
   "Return only the final answer.",
@@ -765,12 +766,22 @@ function buildNativeThinkingOption(think) {
 function hasSamplingOverrides2(options) {
   return options?.top_p !== void 0 || options?.seed !== void 0;
 }
+var UNSUPPORTED_SAMPLING_FIELD_PATTERN = /unrecognized|unknown|not support|unsupported|invalid|unexpected|additional|extra/;
+var UNSUPPORTED_OUTPUT_LIMIT_FIELD_PATTERN = /unrecognized|unknown|not support|unsupported|unexpected|additional|extra|invalid field/;
 function isUnsupportedSamplingMessage(status, text) {
   if (status !== 400 && status !== 422) return false;
   const lower = text.toLowerCase();
-  const mentionsSampling = /\b(seed|top_p|topp)\b/.test(lower);
-  if (!mentionsSampling) return false;
-  return /unrecognized|unknown|not support|unsupported|invalid|unexpected|additional|extra/.test(lower);
+  if (!/\b(seed|top_p|topp)\b/.test(lower)) return false;
+  return UNSUPPORTED_SAMPLING_FIELD_PATTERN.test(lower);
+}
+function isUnsupportedOutputLimitMessage(status, text, mode) {
+  if (status !== 400 && status !== 422) return false;
+  const lower = text.toLowerCase();
+  const fieldName = mode === "legacy" ? "max_tokens" : "max_output_tokens";
+  const alternateFieldName = mode === "legacy" ? "max_output_tokens" : "max_tokens";
+  const mentionsUnsupportedCurrentField = lower.includes(fieldName) && UNSUPPORTED_OUTPUT_LIMIT_FIELD_PATTERN.test(lower);
+  const mentionsRequiredAlternateField = lower.includes(alternateFieldName) && /\b(required|missing)\b/.test(lower);
+  return mentionsUnsupportedCurrentField || mentionsRequiredAlternateField;
 }
 function extractLMStudioErrorMessage(body) {
   const trimmed = body.trim();
@@ -804,20 +815,66 @@ function buildLMStudioRequestError(kind, model, status, statusText, body) {
   const suffix = backendMessage ? ` ${backendMessage}` : "";
   return new Error(`LM Studio ${kind} failed (${status} ${statusText})${suffix}`.trim());
 }
-function buildNativeChatBody(model, prompt, options, stream, includeSampling) {
+function buildUnsupportedOutputLimitNegotiationError(kind, model, body) {
+  const backendMessage = extractLMStudioErrorMessage(body);
+  return new Error(
+    [
+      `LM Studio ${kind} failed for "${model}" because this backend rejected both max_output_tokens and max_tokens.`,
+      "MetriLLM cannot safely continue without an explicit output limit because benchmarks rely on bounded generation.",
+      backendMessage ? `Backend error: ${backendMessage}` : void 0
+    ].filter(Boolean).join(" ")
+  );
+}
+function buildNativeChatBody(model, prompt, options, stream, includeSampling, outputLimitMode) {
   const reasoning = buildNativeThinkingOption(options?.think);
+  const outputLimit = options?.num_predict !== void 0 ? options.num_predict : 512;
   return {
     model,
     input: prompt,
     temperature: options?.temperature ?? 0,
     ...includeSampling && options?.top_p !== void 0 ? { top_p: options.top_p } : {},
     ...includeSampling && options?.seed !== void 0 ? { seed: options.seed } : {},
-    max_tokens: options?.num_predict ?? 512,
+    ...outputLimitMode === "preferred" ? { max_output_tokens: outputLimit } : {},
+    ...outputLimitMode === "legacy" ? { max_tokens: outputLimit } : {},
     stream,
     ...reasoning !== void 0 ? { reasoning } : {},
     ...options?.think === false ? { system_prompt: NON_THINKING_SYSTEM_PROMPT } : {}
   };
 }
+var MAX_NEGOTIATE_RETRIES = 5;
+async function negotiateRequest(kind, model, cacheKey, options, makeRequest) {
+  let includeSampling = true;
+  let outputLimitMode = outputLimitModeCache.get(cacheKey) ?? "preferred";
+  const triedOutputLimitModes = /* @__PURE__ */ new Set([outputLimitMode]);
+  let resp = await makeRequest(includeSampling, outputLimitMode);
+  let retries = 0;
+  while (!resp.ok && retries < MAX_NEGOTIATE_RETRIES) {
+    retries++;
+    const body = await resp.text().catch(() => "");
+    if (includeSampling && hasSamplingOverrides2(options) && isUnsupportedSamplingMessage(resp.status, body)) {
+      includeSampling = false;
+      resp = await makeRequest(includeSampling, outputLimitMode);
+      continue;
+    }
+    if (isUnsupportedOutputLimitMessage(resp.status, body, outputLimitMode)) {
+      const nextMode = outputLimitMode === "preferred" ? "legacy" : !triedOutputLimitModes.has("preferred") ? "preferred" : null;
+      if (!nextMode) {
+        throw buildUnsupportedOutputLimitNegotiationError(kind, model, body);
+      }
+      outputLimitMode = nextMode;
+      triedOutputLimitModes.add(outputLimitMode);
+      resp = await makeRequest(includeSampling, outputLimitMode);
+      continue;
+    }
+    throw buildLMStudioRequestError(kind, model, resp.status, resp.statusText, body);
+  }
+  if (!resp.ok) {
+    const body = await resp.text().catch(() => "");
+    throw buildLMStudioRequestError(kind, model, resp.status, resp.statusText, body);
+  }
+  outputLimitModeCache.set(cacheKey, outputLimitMode);
+  return resp;
+}
 function getNativeStatNumber(value) {
   if (typeof value !== "number" || !Number.isFinite(value) || value < 0) return void 0;
   return value;
@@ -1530,25 +1587,18 @@ async function generate2(model, prompt, options) {
   try {
     const baseUrl = getLMStudioBaseUrl();
     const url = new URL("/api/v1/chat", baseUrl);
-    const doRequest = (includeSampling) => fetch(url, {
-      method: "POST",
-      headers: getLMStudioHeaders(),
-      body: JSON.stringify(buildNativeChatBody(model, prompt, options, false, includeSampling)),
-      signal: controller.signal
-    });
-    let resp = await doRequest(true);
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      if (hasSamplingOverrides2(options) && isUnsupportedSamplingMessage(resp.status, body)) {
-        resp = await doRequest(false);
-      } else {
-        throw buildLMStudioRequestError("generate", model, resp.status, resp.statusText, body);
-      }
-    }
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      throw buildLMStudioRequestError("generate", model, resp.status, resp.statusText, body);
-    }
+    const resp = await negotiateRequest(
+      "generate",
+      model,
+      baseUrl,
+      options,
+      (sampling, limitMode) => fetch(url, {
+        method: "POST",
+        headers: getLMStudioHeaders(),
+        body: JSON.stringify(buildNativeChatBody(model, prompt, options, false, sampling, limitMode)),
+        signal: controller.signal
+      })
+    );
     const payload = await resp.json();
     const nativeResponse = extractNativeResponse(payload);
     const response = nativeResponse.response;
@@ -1606,25 +1656,18 @@ async function generateStream2(model, prompt, callbacks, options) {
   };
   try {
     resetStallTimer();
-    const doRequest = (includeSampling) => fetch(url, {
-      method: "POST",
-      headers: getLMStudioHeaders(),
-      body: JSON.stringify(buildNativeChatBody(model, prompt, options, true, includeSampling)),
-      signal: controller.signal
-    });
-    let resp = await doRequest(true);
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      if (hasSamplingOverrides2(options) && isUnsupportedSamplingMessage(resp.status, body)) {
-        resp = await doRequest(false);
-      } else {
-        throw buildLMStudioRequestError("stream", model, resp.status, resp.statusText, body);
-      }
-    }
-    if (!resp.ok) {
-      const body = await resp.text().catch(() => "");
-      throw buildLMStudioRequestError("stream", model, resp.status, resp.statusText, body);
-    }
+    const resp = await negotiateRequest(
+      "stream",
+      model,
+      baseUrl,
+      options,
+      (sampling, limitMode) => fetch(url, {
+        method: "POST",
+        headers: getLMStudioHeaders(),
+        body: JSON.stringify(buildNativeChatBody(model, prompt, options, true, sampling, limitMode)),
+        signal: controller.signal
+      })
+    );
     if (!resp.body) {
       throw new Error("LM Studio stream response body is empty");
     }