@apo-ai/sdk 0.13.0 → 0.14.0

package/dist/agent-task/public.d.ts

@@ -1307,6 +1307,13 @@ type JudgeConfig = {
      * dual behavior (second judge as parallel, evidence-only shadow).
      */
     mode?: "cascade";
+    /**
+     * Sampling temperature for `t.judge` requests (not `t.agent`, whose session
+     * stays at 0). Unset falls back to `APO_JUDGE_TEMPERATURE`, then `0`. Some
+     * reasoning models recommend a non-zero value (DeepSeek: 0.6) to avoid
+     * repetition at greedy decoding. Recorded on the check's judge metadata.
+     */
+    temperature?: number;
 };
 /**
  * What a check knows about its scope when it invokes the judge — the

package/dist/agent-task/public.js

@@ -1262,6 +1262,16 @@ function runWithSharedPrefix(key, task2) {
   };
   return lane.warmed.then(acquire).then(() => task2().finally(release));
 }
+function judgeTemperature(configured) {
+  const raw = process.env.APO_JUDGE_TEMPERATURE?.trim();
+  const fromEnv = raw ? Number(raw) : void 0;
+  const value = configured ?? fromEnv ?? 0;
+  if (!Number.isFinite(value) || value < 0 || value > 2) {
+    const source = configured !== void 0 ? "judge.temperature" : "APO_JUDGE_TEMPERATURE";
+    throw new Error(`${source} must be a number between 0 and 2, got ${configured ?? raw}`);
+  }
+  return value;
+}
 function defaultJudgeBaseURL(baseURL) {
   return baseURL ?? process.env.OPENROUTER_BASE_URL ?? "https://openrouter.ai/api/v1";
 }
@@ -1271,6 +1281,7 @@ function defaultJudgeAPIKey(apiKey) {
 async function callJudge(args) {
   const baseURL = defaultJudgeBaseURL(args.baseURL);
   const apiKey = defaultJudgeAPIKey(args.apiKey);
+  const temperature = judgeTemperature(args.temperature);
   const { briefingText, instructionText, systemPromptText, deliverableText, secondJudgeState, secondJudgeProjected } = buildJudgePromptParts({
     values: args.values,
     instruction: args.instruction,
@@ -1304,7 +1315,7 @@ async function callJudge(args) {
       },
       { role: "user", content: instructionText }
     ],
-    temperature: 0,
+    temperature,
     response_format: schema ? judgeResponseFormat() : JSON_OBJECT_FORMAT,
     ...stream ? { stream: true, stream_options: { include_usage: true } } : {}
   });
@@ -1427,6 +1438,7 @@ async function callJudge(args) {
     }
     const judgeMetadata = async (completion) => ({
       model: args.model,
+      temperature,
       contract: judgeContractInUse(),
       prompt: { system: systemPromptText, user: instructionText },
       ...completion ? { response: completion.text, tokens: parseJudgeUsage(completion.usage) } : {},
@@ -2541,7 +2553,8 @@ function resolveJudgeConfig(judgeConfig, override) {
     baseURL: override?.baseURL ?? judgeConfig?.baseURL,
     apiKey: override?.apiKey ?? judgeConfig?.apiKey,
     prompt: override?.prompt ?? judgeConfig?.prompt,
-    mode: override?.mode ?? judgeConfig?.mode
+    mode: override?.mode ?? judgeConfig?.mode,
+    temperature: override?.temperature ?? judgeConfig?.temperature
   };
 }
 var TEST_METHOD_NAMES = [
@@ -2650,6 +2663,18 @@ function createJudgeMethod(rec, judgeConfig, judgeScope, judgeTracer) {
       );
       return;
     }
+    try {
+      judgeTemperature(effective.temperature);
+    } catch (error) {
+      rec.record(label, false, `judge failed: ${error instanceof Error ? error.message : String(error)}`, {
+        evaluator_type: "llm",
+        expected: instruction,
+        received: valueArray.length === 1 ? valueArray[0] : valueArray,
+        location,
+        outcome: "error"
+      });
+      return;
+    }
     const context2 = judgeScopeToContext(judgeScope);
     const judgeServing = servingHostFromBaseURL(effective.baseURL);
     const secondJudgeValue = opts?.secondJudgeValue;
@@ -2706,6 +2731,7 @@ function createJudgeMethod(rec, judgeConfig, judgeScope, judgeTracer) {
       baseURL: effective.baseURL,
       apiKey: effective.apiKey,
       prompt: effective.prompt,
+      ...effective.temperature !== void 0 ? { temperature: effective.temperature } : {},
       ...context2 ? { context: context2 } : {},
       ...prefetchedSecondJudge ? { prefetchedSecondJudge } : {},
       ...secondJudgeValue !== void 0 ? {