@ai-sdk-tool/eval 0.1.2 → 0.1.3

package/dist/index.js

@@ -5,14 +5,15 @@ var colors = {
   red: "\x1B[31m",
   yellow: "\x1B[33m",
   cyan: "\x1B[36m",
-  magenta: "\x1B[35m"
+  magenta: "\x1B[35m",
+  gray: "\x1B[90m"
 };
 function printResult(result) {
-  const { model, benchmark, result: benchmarkResult } = result;
+  const { model, modelKey, benchmark, result: benchmarkResult } = result;
   const status = benchmarkResult.success ? `${colors.green}\u2714 SUCCESS${colors.reset}` : `${colors.red}\u2716 FAILURE${colors.reset}`;
   console.log(
     `
- ${colors.cyan}[${model}]${colors.reset} - ${colors.magenta}${benchmark}${colors.reset}`
+ ${colors.cyan}[${model}]${colors.reset}${modelKey ? ` ${colors.gray}(${modelKey})${colors.reset}` : ""} - ${colors.magenta}${benchmark}${colors.reset}`
   );
   console.log(
     `  \u2514 ${status} | Score: ${colors.yellow}${benchmarkResult.score.toFixed(2)}${colors.reset}`
@@ -38,6 +39,186 @@ function consoleReporter(results) {
   console.log("\n---------------------------\n");
 }
 
+// src/reporters/console.debug.ts
+var colors2 = {
+  reset: "\x1B[0m",
+  green: "\x1B[32m",
+  red: "\x1B[31m",
+  yellow: "\x1B[33m",
+  cyan: "\x1B[36m",
+  magenta: "\x1B[35m",
+  gray: "\x1B[90m",
+  bold: "\x1B[1m",
+  underline: "\x1B[4m"
+};
+function colorizeDiffLine(line) {
+  if (line.startsWith("+")) return `${colors2.green}${line}${colors2.reset}`;
+  if (line.startsWith("-")) return `${colors2.red}${line}${colors2.reset}`;
+  if (line.startsWith("@"))
+    return `${colors2.cyan}${colors2.bold}${line}${colors2.reset}`;
+  return line;
+}
+function uniqueLines(lines) {
+  const seen = /* @__PURE__ */ new Set();
+  const out = [];
+  for (const l of lines) {
+    if (seen.has(l)) continue;
+    seen.add(l);
+    out.push(l);
+  }
+  return out;
+}
+function suggestFixFromDiff(parsed) {
+  const suggestions = [];
+  const { error_type, expected, actual, diff } = parsed ?? {};
+  if (diff && diff.some((d) => d.includes("function name")) || diff && diff.some((d) => d.includes("missing function:"))) {
+    const expectedName = expected?.function;
+    const actualName = actual?.function;
+    if (expectedName && actualName && expectedName !== actualName) {
+      suggestions.push(
+        `Call the function '${expectedName}' instead of '${actualName}'.`
+      );
+    }
+    if (Array.isArray(expected?.functions)) {
+      suggestions.push(
+        `Ensure tool calls include: ${expected.functions.join(", ")}.`
+      );
+    }
+  }
+  if (diff && diff.some((d) => d.startsWith("- missing required param:"))) {
+    const missing = diff.filter((d) => d.startsWith("- missing required param:")).map((d) => d.replace("- missing required param: ", ""));
+    if (missing.length) {
+      suggestions.push(`Add required parameter(s): ${missing.join(", ")}.`);
+    }
+  }
+  if (diff && diff.some((d) => d.startsWith("+ unexpected param:"))) {
+    const extras = diff.filter((d) => d.startsWith("+ unexpected param:")).map((d) => d.replace("+ unexpected param: ", ""));
+    if (extras.length) {
+      suggestions.push(`Remove unexpected parameter(s): ${extras.join(", ")}.`);
+    }
+  }
+  if (diff && diff.some((d) => d.startsWith("@@ param "))) {
+    const targets = diff.filter((d) => d.startsWith("@@ param ")).map((d) => d.replace("@@ param ", ""));
+    for (const param of targets) {
+      const allowedLine = diff.find(
+        (d) => d.startsWith("- expected one of:")
+      );
+      if (allowedLine) {
+        const allowed = allowedLine.replace("- expected one of: ", "");
+        suggestions.push(`Set '${param}' to one of: ${allowed}.`);
+      } else {
+        suggestions.push(`Adjust '${param}' to an allowed value.`);
+      }
+    }
+  }
+  if (suggestions.length === 0 && typeof error_type === "string") {
+    if (error_type.includes("missing_required")) {
+      suggestions.push(
+        "Add all required parameters defined by the tool schema."
+      );
+    } else if (error_type.includes("unexpected_param")) {
+      suggestions.push("Remove parameters not present in the tool schema.");
+    } else if (error_type.includes("wrong_count")) {
+      suggestions.push(
+        "Adjust the number of tool calls to match expected count."
+      );
+    } else if (error_type.includes("wrong_func_name")) {
+      suggestions.push("Use the exact expected function name from the schema.");
+    } else if (error_type.includes("value_error")) {
+      suggestions.push("Choose a value from the allowed options.");
+    }
+  }
+  return uniqueLines(suggestions);
+}
+function consoleDebugReporter(results) {
+  console.log("\n--- \u{1F4CA} Evaluation Report (debug) ---");
+  for (const r of results) {
+    const { model, modelKey, benchmark, result } = r;
+    const status = result.success ? `${colors2.green}\u2714 SUCCESS${colors2.reset}` : `${colors2.red}\u2716 FAILURE${colors2.reset}`;
+    console.log(
+      `
+ ${colors2.cyan}[${model}]${colors2.reset}${modelKey ? ` ${colors2.gray}(${modelKey})${colors2.reset}` : ""} - ${colors2.magenta}${benchmark}${colors2.reset}`
+    );
+    console.log(
+      `  \u2514 ${status} | Score: ${colors2.yellow}${result.score.toFixed(2)}${colors2.reset}`
+    );
+    const metrics = Object.entries(result.metrics);
+    if (metrics.length > 0) {
+      console.log("    Metrics:");
+      for (const [k, v] of metrics) console.log(`      - ${k}: ${v}`);
+    }
+    if (result.logs && result.logs.length) {
+      const failLogs = result.logs.filter(
+        (l) => l.startsWith("[FAIL]") || l.startsWith("[ERROR]") || l.startsWith("[FATAL]") || l.startsWith("[STACK]") || l.startsWith("[DEBUG-FAIL]")
+      );
+      const hasFails = failLogs.length > 0;
+      if (hasFails) {
+        console.log(`    ${colors2.bold}Failure details:${colors2.reset}`);
+        const debugIds = /* @__PURE__ */ new Set();
+        for (const l of failLogs) {
+          if (l.startsWith("[DEBUG-FAIL]")) {
+            try {
+              const parsed = JSON.parse(l.replace(/^\[DEBUG-FAIL\] /, ""));
+              if (parsed?.id) debugIds.add(String(parsed.id));
+            } catch {
+            }
+          }
+        }
+        for (const line of failLogs) {
+          if (line.startsWith("[FAIL]")) {
+            const m = line.match(/^\[FAIL\]\s+([^:]+):/);
+            const failId = m?.[1];
+            if (failId && debugIds.has(failId)) continue;
+            console.log(`      ${colors2.red}${line}${colors2.reset}`);
+          } else if (line.startsWith("[ERROR]") || line.startsWith("[FATAL]")) {
+            console.log(`      ${colors2.yellow}${line}${colors2.reset}`);
+          } else if (line.startsWith("[STACK]")) {
+            console.log(`      ${colors2.gray}${line}${colors2.reset}`);
+          } else if (line.startsWith("[DEBUG-FAIL]")) {
+            const payload = line.replace(/^\[DEBUG-FAIL\] /, "");
+            try {
+              const parsed = JSON.parse(payload);
+              const { id, expected, actual, message, diff } = parsed;
+              console.log(
+                `      ${colors2.underline}${id}${colors2.reset} ${message ? "- " + message : ""}`
+              );
+              if (diff && Array.isArray(diff)) {
+                for (const dLine of diff)
+                  console.log("        " + colorizeDiffLine(dLine));
+              } else {
+                console.log("        expected:");
+                console.log(
+                  colors2.green + "          " + JSON.stringify(expected, null, 2).split("\n").join("\n          ") + colors2.reset
+                );
+                console.log("        actual:");
+                console.log(
+                  colors2.red + "          " + JSON.stringify(actual, null, 2).split("\n").join("\n          ") + colors2.reset
+                );
+              }
+              const suggestions = suggestFixFromDiff(parsed);
+              if (suggestions.length) {
+                console.log(
+                  `        ${colors2.bold}Suggested fix:${colors2.reset}`
+                );
+                for (const s of suggestions) console.log(`          \u2022 ${s}`);
+              }
+            } catch {
+              console.log(`      ${line}`);
+            }
+          }
+        }
+      } else {
+        const info = result.logs.filter(
+          (l) => l.startsWith("[INFO]") || l.startsWith("[PASS]")
+        );
+        for (const line of info)
+          console.log(`      ${colors2.gray}${line}${colors2.reset}`);
+      }
+    }
+  }
+  console.log("\n------------------------------------\n");
+}
+
 // src/reporters/json.ts
 function jsonReporter(results) {
   const serializableResults = results.map((r) => ({
@@ -53,30 +234,35 @@ function jsonReporter(results) {
 // src/reporters/index.ts
 var reporters = {
   console: consoleReporter,
-  json: jsonReporter
+  json: jsonReporter,
+  "console.debug": consoleDebugReporter
 };
 
 // src/evaluate.ts
-async function runSingleBenchmark(model, benchmark) {
+async function runSingleBenchmark(model, benchmark, modelKey) {
   const modelId = typeof model === "object" && model !== null && "modelId" in model && typeof model.modelId === "string" ? model.modelId : "unknown-model";
   try {
-    console.log(`[${modelId}] Running benchmark: ${benchmark.name}...`);
+    console.log(
+      `[${modelId}]${modelKey ? ` (${modelKey})` : ""} Running benchmark: ${benchmark.name}...`
+    );
     const result = await benchmark.run(model);
     console.log(
-      `[${modelId}] Finished benchmark: ${benchmark.name}. Score: ${result.score}`
+      `[${modelId}]${modelKey ? ` (${modelKey})` : ""} Finished benchmark: ${benchmark.name}. Score: ${result.score}`
     );
     return {
       model: modelId,
+      modelKey,
       benchmark: benchmark.name,
       result
     };
   } catch (error) {
     console.error(
-      `[${modelId}] Error running benchmark: ${benchmark.name}`,
+      `[${modelId}]${modelKey ? ` (${modelKey})` : ""} Error running benchmark: ${benchmark.name}`,
       error
     );
     return {
       model: modelId,
+      modelKey,
       benchmark: benchmark.name,
       result: {
         score: 0,
@@ -89,11 +275,26 @@ async function runSingleBenchmark(model, benchmark) {
 }
 async function evaluate(options) {
   const { models, benchmarks, reporter = "console" } = options;
-  const modelsArray = Array.isArray(models) ? models : [models];
+  const modelEntries = [];
+  if (Array.isArray(models)) {
+    for (const m of models) modelEntries.push([void 0, m]);
+  } else if (typeof models === "object" && models !== null && "modelId" in models) {
+    modelEntries.push([void 0, models]);
+  } else {
+    for (const [key, m] of Object.entries(
+      models
+    )) {
+      modelEntries.push([key, m]);
+    }
+  }
   const allResults = [];
-  for (const model of modelsArray) {
+  for (const [modelKey, model] of modelEntries) {
     for (const benchmark of benchmarks) {
-      const evaluationResult = await runSingleBenchmark(model, benchmark);
+      const evaluationResult = await runSingleBenchmark(
+        model,
+        benchmark,
+        modelKey
+      );
       allResults.push(evaluationResult);
     }
   }
@@ -436,7 +637,9 @@ function checkStringValue(param, modelValue, possibleAnswers) {
   if (!standardizedPossibleAnswers.includes(standardizedModelValue)) {
     return {
       valid: false,
-      error: `Invalid value for parameter '${param}': '${modelValue}'. Expected one of ${possibleAnswers.join(", ")}.`,
+      error: `Invalid value for parameter '${param}': ${JSON.stringify(
+        modelValue
+      )}. Expected one of ${JSON.stringify(possibleAnswers)}.`,
       error_type: "value_error:string"
     };
   }
@@ -490,15 +693,55 @@ function simpleFunctionChecker(funcDescription, modelToolCall, possibleAnswer) {
       if (!hasMatch) {
         return {
           valid: false,
-          error: `Invalid value for list parameter '${paramName}'.`,
+          error: `Invalid value for list parameter '${paramName}'. Got ${JSON.stringify(
+            modelValue
+          )}. Expected one of ${JSON.stringify(possibleValues)}.`,
           error_type: "value_error:list"
         };
       }
     } else {
-      if (!possibleValues.includes(modelValue)) {
+      const hasMatch = possibleValues.some((possibleValue) => {
+        if (modelValue === possibleValue) return true;
+        if (typeof modelValue === "object" && modelValue !== null && typeof possibleValue === "object" && possibleValue !== null) {
+          try {
+            const normalizeObject = (obj) => {
+              if (Array.isArray(obj)) {
+                return obj.map(normalizeObject);
+              }
+              if (obj && typeof obj === "object") {
+                const normalized = {};
+                for (const [key, value] of Object.entries(obj)) {
+                  if (Array.isArray(value) && value.length === 1 && (typeof value[0] !== "object" || value[0] === null)) {
+                    normalized[key] = value[0];
+                  } else {
+                    normalized[key] = normalizeObject(value);
+                  }
+                }
+                return normalized;
+              }
+              return obj;
+            };
+            const normalizedModel = normalizeObject(modelValue);
+            const normalizedPossible = normalizeObject(possibleValue);
+            return JSON.stringify(normalizedModel) === JSON.stringify(normalizedPossible);
+          } catch {
+            return false;
+          }
+        }
+        if (typeof modelValue === "number" && typeof possibleValue === "string") {
+          return modelValue.toString() === possibleValue;
+        }
+        if (typeof modelValue === "string" && typeof possibleValue === "number") {
+          return modelValue === possibleValue.toString();
+        }
+        return false;
+      });
+      if (!hasMatch) {
         return {
           valid: false,
-          error: `Invalid value for parameter '${paramName}': got '${modelValue}', expected one of '${possibleValues}'.`,
+          error: `Invalid value for parameter '${paramName}'. Got ${JSON.stringify(
+            modelValue
+          )}. Expected one of ${JSON.stringify(possibleValues)}.`,
           error_type: "value_error:other"
         };
       }
@@ -594,7 +837,8 @@ function check(testCase, modelOutput, possibleAnswer) {
       if (!modelOutput || modelOutput.length !== 1) {
         return {
           valid: false,
-          error: `Expected 1 function call, but got ${modelOutput?.length ?? 0}.`
+          error: `Expected 1 function call, but got ${modelOutput?.length ?? 0}.`,
+          error_type: "simple:wrong_count"
         };
       }
       return simpleFunctionChecker(
@@ -623,7 +867,11 @@ function check(testCase, modelOutput, possibleAnswer) {
     }
     return { valid: true };
   } catch (e) {
-    return { valid: false, error: `Checker Error: ${e.message}` };
+    return {
+      valid: false,
+      error: `Checker Error: ${e.message}`,
+      error_type: "checker_error"
+    };
   }
 }
 function createBfclBenchmark(name, description, testDataFile, answerDataFile) {
@@ -675,7 +923,13 @@ function createBfclBenchmark(name, description, testDataFile, answerDataFile) {
           if (copy.items) copy.items = fixSchema(copy.items);
           return copy;
         };
-        for (const testCase of testCases) {
+        const concurrencyEnv = process.env.BFCL_CONCURRENCY;
+        const concurrency = concurrencyEnv && Number.isFinite(Number(concurrencyEnv)) ? Math.max(1, Number(concurrencyEnv)) : 4;
+        logs.push(
+          `[INFO] Running ${testCases.length} test cases with concurrency=${concurrency}`
+        );
+        const runSingleCase = async (testCase) => {
+          const caseLogs = [];
           const { function: tools, question: messages } = testCase;
           try {
             const flatMessages = Array.isArray(messages) && messages.some((m) => Array.isArray(m)) ? messages.flat(1) : messages;
@@ -708,11 +962,11 @@ function createBfclBenchmark(name, description, testDataFile, answerDataFile) {
             try {
               const firstTool = transformedTools[0];
               const schemaType = firstTool?.inputSchema?.type ?? firstTool?.inputSchema?.jsonSchema?.type;
-              logs.push(
+              caseLogs.push(
                 `[DEBUG] ${testCase.id}: firstTool=${JSON.stringify(firstTool)}, schemaType=${schemaType}`
               );
             } catch (e) {
-              logs.push(
+              caseLogs.push(
                 `[DEBUG] ${testCase.id}: failed to introspect tools: ${e.message}`
               );
             }
@@ -720,14 +974,22 @@ function createBfclBenchmark(name, description, testDataFile, answerDataFile) {
               model,
               messages: flatMessages,
               tools: toolsMap,
-              toolChoice: "auto"
+              toolChoice: "auto",
+              // Pass original schema information to middleware
+              providerOptions: {
+                toolCallMiddleware: {
+                  originalToolSchemas: Object.fromEntries(
+                    transformedTools.map((t) => [t.name, t.inputSchema])
+                  )
+                }
+              }
             });
             try {
-              logs.push(
+              caseLogs.push(
                 `[DEBUG] ${testCase.id}: rawToolCalls=${JSON.stringify(toolCalls)}, finishReason=${finishReason}, text=${JSON.stringify(text)}`
               );
             } catch {
-              logs.push(
+              caseLogs.push(
                 `[DEBUG] ${testCase.id}: failed to serialize toolCalls`
               );
             }
@@ -760,20 +1022,221 @@ function createBfclBenchmark(name, description, testDataFile, answerDataFile) {
               possibleAnswer
             );
             if (checkerResult.valid) {
-              correctCount++;
-              logs.push(`[PASS] ${testCase.id}`);
+              caseLogs.push(`[PASS] ${testCase.id}`);
+              return { valid: true, logs: caseLogs };
             } else {
-              logs.push(`[FAIL] ${testCase.id}: ${checkerResult.error}`);
+              caseLogs.push(`[FAIL] ${testCase.id}: ${checkerResult.error}`);
+              try {
+                const category = testCase.id.split("_")[0];
+                const diff = [];
+                const summarizeArgs = (args) => {
+                  if (args == null) return args;
+                  if (typeof args !== "object") return args;
+                  return Object.keys(args).sort().reduce((acc, k) => {
+                    acc[k] = args[k];
+                    return acc;
+                  }, {});
+                };
+                const expected = {};
+                const actual = {};
+                if (category === "simple") {
+                  const funcDesc = tools[0];
+                  const gt = possibleAnswer.ground_truth?.[0];
+                  const expectedFuncName = funcDesc?.name;
+                  const expectedParams = gt ? gt[Object.keys(gt)[0]] : void 0;
+                  const received = restoredCalls[0];
+                  const receivedName = received?.toolName ?? received?.name;
+                  const receivedArgs = summarizeArgs(received?.args);
+                  expected.function = expectedFuncName;
+                  expected.params = expectedParams;
+                  actual.function = receivedName;
+                  actual.args = receivedArgs;
+                  if (expectedFuncName !== receivedName) {
+                    diff.push(`@@ function name`);
+                    diff.push(`- ${expectedFuncName}`);
+                    diff.push(`+ ${receivedName}`);
+                  }
+                  if (expectedParams && receivedArgs) {
+                    const required = funcDesc?.parameters?.required ?? [];
+                    for (const req of required) {
+                      if (!(req in receivedArgs)) {
+                        diff.push(`- missing required param: ${req}`);
+                      }
+                    }
+                    for (const k of Object.keys(receivedArgs)) {
+                      if (!Object.prototype.hasOwnProperty.call(expectedParams, k)) {
+                        diff.push(`+ unexpected param: ${k}`);
+                      }
+                    }
+                    for (const k of Object.keys(receivedArgs)) {
+                      if (Object.prototype.hasOwnProperty.call(expectedParams, k)) {
+                        const allowed = expectedParams[k];
+                        const got = receivedArgs[k];
+                        const includes = Array.isArray(allowed) && allowed.some((v) => {
+                          try {
+                            if (Array.isArray(got)) {
+                              return JSON.stringify(
+                                got.map((x) => String(x)).sort()
+                              ) === JSON.stringify(
+                                v.map((x) => String(x)).sort()
+                              );
+                            }
+                          } catch {
+                          }
+                          return String(v).toLowerCase().replace(/\s+/g, "") === String(got).toLowerCase().replace(/\s+/g, "");
+                        });
+                        if (!includes) {
+                          diff.push(`@@ param ${k}`);
+                          diff.push(
+                            `- expected one of: ${JSON.stringify(allowed)}`
+                          );
+                          diff.push(`+ got: ${JSON.stringify(got)}`);
+                        }
+                      }
+                    }
+                  }
+                } else {
+                  const gtArr = possibleAnswer.ground_truth ?? [];
+                  const expectedNames = gtArr.map((g) => Object.keys(g)[0]);
+                  const actualNames = restoredCalls.map(
+                    (c) => c.toolName ?? c.name
+                  );
+                  expected.functions = expectedNames;
+                  actual.functions = actualNames;
+                  if (expectedNames.length !== actualNames.length) {
+                    diff.push(`@@ call count`);
+                    diff.push(`- expected ${expectedNames.length}`);
+                    diff.push(`+ got ${actualNames.length}`);
+                  }
+                  const missing = expectedNames.filter(
+                    (n) => !actualNames.includes(n)
+                  );
+                  const extra = actualNames.filter(
+                    (n) => !expectedNames.includes(n)
+                  );
+                  for (const m of missing)
+                    diff.push(`- missing function: ${m}`);
+                  for (const e of extra)
+                    diff.push(`+ unexpected function: ${e}`);
+                  const usedActual = /* @__PURE__ */ new Set();
+                  for (const expectedObj of gtArr) {
+                    const fname = Object.keys(expectedObj)[0];
+                    let matchedIndex = -1;
+                    for (let i = 0; i < restoredCalls.length; i++) {
+                      if (usedActual.has(i)) continue;
+                      const rc = restoredCalls[i];
+                      const rcName = rc?.toolName ?? rc?.name;
+                      if (rcName === fname) {
+                        matchedIndex = i;
+                        break;
+                      }
+                    }
+                    if (matchedIndex === -1) continue;
+                    usedActual.add(matchedIndex);
+                    const received = restoredCalls[matchedIndex];
+                    const receivedArgs = summarizeArgs(received?.args);
+                    const expectedParamsAllowed = expectedObj[fname];
+                    const funcDesc = tools.find(
+                      (t) => t.name === fname
+                    );
+                    const requiredParams = funcDesc?.parameters?.required ?? [];
+                    diff.push(`@@ function ${fname}`);
+                    if (expectedParamsAllowed && receivedArgs) {
+                      for (const req of requiredParams) {
+                        if (!(req in receivedArgs)) {
+                          diff.push(`- missing required param: ${req}`);
+                        }
+                      }
+                      for (const k of Object.keys(receivedArgs)) {
+                        if (!Object.prototype.hasOwnProperty.call(
+                          expectedParamsAllowed,
+                          k
+                        )) {
+                          diff.push(`+ unexpected param: ${k}`);
+                        }
+                      }
+                      for (const k of Object.keys(receivedArgs)) {
+                        if (Object.prototype.hasOwnProperty.call(
+                          expectedParamsAllowed,
+                          k
+                        )) {
+                          const allowed = expectedParamsAllowed[k];
+                          const got = receivedArgs[k];
+                          const includes = Array.isArray(allowed) && allowed.some((v) => {
+                            try {
+                              if (Array.isArray(got)) {
+                                return JSON.stringify(
+                                  got.map((x) => String(x)).sort()
+                                ) === JSON.stringify(
+                                  v.map((x) => String(x)).sort()
+                                );
+                              }
+                            } catch {
+                            }
+                            return String(v).toLowerCase().replace(/\s+/g, "") === String(got).toLowerCase().replace(/\s+/g, "");
+                          });
+                          if (!includes) {
+                            diff.push(`@@ param ${k}`);
+                            diff.push(
+                              `- expected one of: ${JSON.stringify(allowed)}`
+                            );
+                            diff.push(`+ got: ${JSON.stringify(got)}`);
+                          }
+                        }
+                      }
+                    }
+                  }
+                }
+                caseLogs.push(
+                  `[DEBUG-FAIL] ${JSON.stringify({
+                    id: testCase.id,
+                    message: checkerResult.error,
+                    error_type: checkerResult.error_type,
+                    expected,
+                    actual,
+                    diff
+                  })}`
+                );
+              } catch {
+                caseLogs.push(
+                  `[DEBUG] ${testCase.id}: failed to build debug diff`
+                );
+              }
+              return { valid: false, logs: caseLogs };
             }
           } catch (e) {
-            logs.push(
+            caseLogs.push(
               `[ERROR] ${testCase.id}: Model generation failed: ${e?.message}`
             );
             if (e?.stack) {
-              logs.push(`[STACK] ${testCase.id}: ${e.stack}`);
+              caseLogs.push(`[STACK] ${testCase.id}: ${e.stack}`);
             }
+            return { valid: false, logs: caseLogs };
           }
-        }
+        };
+        const mapWithConcurrency = async (items, limit2, mapper) => {
+          const results = new Array(items.length);
+          let idx = 0;
+          const workers = new Array(Math.min(limit2, items.length)).fill(0).map(async () => {
+            while (true) {
+              const current = idx++;
+              if (current >= items.length) break;
+              results[current] = await mapper(items[current], current);
+            }
+          });
+          await Promise.all(workers);
+          return results;
+        };
+        const resultsPerCase = await mapWithConcurrency(
+          testCases,
+          concurrency,
+          async (tc) => runSingleCase(tc)
+        );
+        correctCount = resultsPerCase.reduce(
+          (acc, r) => acc + (r.valid ? 1 : 0),
+          0
+        );
+        for (const r of resultsPerCase) logs.push(...r.logs);
         if (testCases.length === 0) {
           return {
             score: 0,