@apo-ai/sdk 0.9.2 → 0.9.4

package/dist/agent-task/public.js

@@ -377,6 +377,9 @@ function levenshtein(a, b) {
   return prev[n];
 }
 
+// src/agent-task/checks/judge.ts
+import { createHash } from "crypto";
+
 // src/agent-task/checks/second-judge.ts
 var SECOND_JUDGE_TIMEOUT_MS = 3e4;
 function resolveSecondJudgeModel() {
@@ -557,8 +560,19 @@ function parseJudgeJson(raw) {
     reasoning: "Judge response could not be parsed as JSON \u2014 the verdict is unknown, so this check is treated as a failure. The model's raw response is available in the judge metadata."
   };
 }
-var JUDGE_TIMEOUT_MS = 3e5;
 var JUDGE_IDLE_TIMEOUT_MS = 9e4;
+function envMs(name, fallback) {
+  const raw = process.env[name]?.trim();
+  if (!raw) return fallback;
+  const ms = Number(raw);
+  return Number.isFinite(ms) && ms > 0 ? ms : fallback;
+}
+function judgeFirstDataTimeoutMs() {
+  return envMs("APO_JUDGE_TIMEOUT_MS", 3e5);
+}
+function judgeMaxDurationMs() {
+  return envMs("APO_JUDGE_MAX_DURATION_MS", 12e5);
+}
 var JUDGE_RETRY_DELAY_MS = 1e3;
 var JUDGE_RETRY_AFTER_CAP_MS = 2e4;
 var JUDGE_MIN_RETRY_BUDGET_MS = 6e4;
@@ -644,6 +658,36 @@ async function readCompletion(response, onData) {
   return { text, usage, finishReason };
 }
 var prefixQueues = /* @__PURE__ */ new Map();
+function promptCacheKey(prefix) {
+  return `apo-${createHash("sha256").update(prefix).digest("hex").slice(0, 32)}`;
+}
+var cacheKeyRejectedBy = /* @__PURE__ */ new Set();
+var CACHE_KEY_REJECTED = /(unknown|unrecogni[sz]ed|unsupported|unexpected|extra|not permitted|not allowed|cannot find|invalid)[^\n]{0,80}prompt_cache_key|prompt_cache_key[^\n]{0,80}(unknown|unrecogni[sz]ed|unsupported|unexpected|not permitted|not allowed|not supported|extra)/i;
+function rejectsPromptCacheKey(status, body) {
+  return status === 400 && CACHE_KEY_REJECTED.test(body);
+}
+async function fetchWithPromptCacheKey(input, init) {
+  let body;
+  try {
+    body = typeof init?.body === "string" ? JSON.parse(init.body) : void 0;
+  } catch {
+    body = void 0;
+  }
+  if (!body || !("prompt_cache_key" in body)) return fetch(input, init);
+  const scope = `${input instanceof Request ? input.url : String(input)}\0${String(body.model)}`;
+  const { prompt_cache_key: _omitted, ...rest } = body;
+  const withoutKey = () => fetch(input, { ...init, body: JSON.stringify(rest) });
+  if (cacheKeyRejectedBy.has(scope)) return withoutKey();
+  const response = await fetch(input, init);
+  if (response.status !== 400) return response;
+  const text = await response.text();
+  if (!rejectsPromptCacheKey(response.status, text)) {
+    return new Response(text, { status: response.status, statusText: response.statusText, headers: response.headers });
+  }
+  const retry = await withoutKey();
+  if (retry.ok) cacheKeyRejectedBy.add(scope);
+  return retry;
+}
 function runWithSharedPrefix(key, task2) {
   const prev = prefixQueues.get(key) ?? Promise.resolve();
   const next = prev.then(task2, task2);
@@ -684,6 +728,7 @@ ${instructionText}`,
   const cacheKey = `${args.model}\0${briefingText}\0${deliverableText}`;
   const requestBody = (stream) => JSON.stringify({
     model: args.model,
+    prompt_cache_key: promptCacheKey(cacheKey),
     messages: [
       {
         role: "system",
@@ -702,23 +747,36 @@ ${instructionText}`,
     response_format: { type: "json_object" },
     ...stream ? { stream: true, stream_options: { include_usage: true } } : {}
   });
-  const attempt = async (stream, deadline) => {
+  const attempt = async (stream, maxDeadline) => {
     const controller = new AbortController();
     const expire = (reason, name) => () => controller.abort(new DOMException(reason, name));
-    const total = setTimeout(
-      expire(`no complete response within ${JUDGE_TIMEOUT_MS / 1e3}s`, "TimeoutError"),
-      Math.max(deadline - Date.now(), 0)
+    const firstDataMs = judgeFirstDataTimeoutMs();
+    const runaway = setTimeout(
+      expire(
+        `still streaming after ${Math.round(judgeMaxDurationMs() / 1e3)}s (APO_JUDGE_MAX_DURATION_MS)`,
+        "TimeoutError"
+      ),
+      Math.max(maxDeadline - Date.now(), 0)
+    );
+    let waiting = setTimeout(
+      expire(
+        `no data within ${Math.round(firstDataMs / 1e3)}s (APO_JUDGE_TIMEOUT_MS)`,
+        "FirstDataTimeoutError"
+      ),
+      firstDataMs
     );
     const stalled = expire(`no data for ${JUDGE_IDLE_TIMEOUT_MS / 1e3}s`, "IdleTimeoutError");
     let idle;
     const onData = () => {
+      clearTimeout(waiting);
+      waiting = void 0;
       clearTimeout(idle);
       idle = setTimeout(stalled, JUDGE_IDLE_TIMEOUT_MS);
     };
     try {
       let response;
       try {
-        response = await fetch(`${baseURL}/chat/completions`, {
+        response = await fetchWithPromptCacheKey(`${baseURL}/chat/completions`, {
           method: "POST",
           signal: controller.signal,
           headers: {
@@ -748,7 +806,8 @@ ${instructionText}`,
         return transportFailure("Judge response failed", controller.signal.reason ?? error);
       }
     } finally {
-      clearTimeout(total);
+      clearTimeout(runaway);
+      clearTimeout(waiting);
       clearTimeout(idle);
       controller.abort();
     }
@@ -768,7 +827,7 @@ ${instructionText}`,
   };
   return runWithSharedPrefix(cacheKey, async () => {
     const startedAt = Date.now();
-    const deadline = startedAt + JUDGE_TIMEOUT_MS;
+    const deadline = startedAt + judgeMaxDurationMs();
     let stream = true;
     let result = await attempt(stream, deadline);
     const truncated = result.kind === "completion" && isTruncated(result.completion);
@@ -860,7 +919,7 @@ ${args.instruction}`
 }
 
 // src/agent-task/checks/agent-session.ts
-import { createHash as createHash2 } from "crypto";
+import { createHash as createHash3 } from "crypto";
 
 // src/agent-task/tracing.ts
 function createNoopAgentTaskTraceContext() {
@@ -908,7 +967,7 @@ function isTraceableSpanId(spanId) {
 }
 
 // src/agent-task/checks/mcp-tools.ts
-import { createHash } from "crypto";
+import { createHash as createHash2 } from "crypto";
 import { isAbsolute, resolve } from "path";
 import { readFile } from "fs/promises";
 var DEFAULT_TOOL_TIMEOUT_MS = 3e4;
@@ -1065,7 +1124,7 @@ function validateMcpServerConfig(server, where) {
   }
 }
 function sha256(text) {
-  return createHash("sha256").update(text).digest("hex");
+  return createHash2("sha256").update(text).digest("hex");
 }
 function renderValue(value) {
   if (typeof value === "string") return value;
@@ -1259,8 +1318,17 @@ var RECORD_TEXT_LIMIT = 2 * 1024;
 var RECORD_BRIEFING_LIMIT = 16 * 1024;
 var RECORD_MAX_STEPS = 64;
 var RECORD_KEEP_HEAD = 56;
+function sessionPrefix(body) {
+  const messages = body.messages ?? [];
+  const firstUser = messages.find((m) => m.role === "user");
+  return [
+    String(body.model),
+    ...messages.filter((m) => m.role === "system").map((m) => m.content),
+    firstUser?.content
+  ].map((part) => typeof part === "string" ? part : JSON.stringify(part ?? null)).join("\0");
+}
 function sha2562(text) {
-  return createHash2("sha256").update(text).digest("hex");
+  return createHash3("sha256").update(text).digest("hex");
 }
 function renderValue2(value) {
   if (typeof value === "string") return value;
@@ -1543,11 +1611,16 @@ async function runAgentSession(spec) {
   const provider = createOpenAICompatible({
     name: "openrouter",
     baseURL: spec.baseURL ?? "https://openrouter.ai/api/v1",
+    fetch: fetchWithPromptCacheKey,
     apiKey: spec.apiKey ?? process.env.OPENROUTER_API_KEY ?? process.env.OPENAI_API_KEY,
     // Mark the briefing prefix cacheable — repeated samples (rejudge) re-bill
-    // nothing for the shared turn-0 context.
+    // nothing for the shared turn-0 context — and keep every step of one
+    // session on the replica holding its growing prefix. The key is per
+    // session (briefing + criterion): sessions run concurrently, unlike
+    // t.judge calls, and one key per task would pile them onto one replica.
     transformRequestBody: (body) => ({
       ...body,
+      prompt_cache_key: promptCacheKey(sessionPrefix(body)),
       messages: body.messages?.map(
         (m) => m.role === "system" && typeof m.content === "string" ? { ...m, content: [{ type: "text", text: m.content, cache_control: { type: "ephemeral" } }] } : m
       )
@@ -2290,7 +2363,7 @@ var Recorder = class {
 };
 
 // src/agent-task/checks/compact.ts
-import { createHash as createHash3 } from "crypto";
+import { createHash as createHash4 } from "crypto";
 var RECEIVED_VALUE_LIMIT = 4 * 1024;
 var JUDGE_SEGMENT_LIMIT = 16 * 1024;
 var PREVIEW_CHARS = 256;
@@ -2398,7 +2471,7 @@ function marker(bytes) {
     kind: "truncated",
     preview,
     size_bytes: bytes.length,
-    sha256: createHash3("sha256").update(bytes).digest("hex")
+    sha256: createHash4("sha256").update(bytes).digest("hex")
   };
 }
 function truncateSessionText(session) {
@@ -3043,7 +3116,7 @@ function isFileArtifact(value) {
 }
 
 // src/agent-task/deliverables/upload.ts
-import { createHash as createHash4 } from "crypto";
+import { createHash as createHash5 } from "crypto";
 import { createReadStream } from "fs";
 import { Readable } from "stream";
 var DEFAULT_FETCH = typeof fetch === "function" ? fetch : void 0;
@@ -3098,7 +3171,7 @@ async function uploadOne(name, artifact, config) {
 }
 async function hashFile(path) {
   return new Promise((resolve5, reject) => {
-    const hash = createHash4("sha256");
+    const hash = createHash5("sha256");
     let size = 0;
     const stream = createReadStream(path);
     stream.on("data", (chunk) => {
@@ -4338,7 +4411,7 @@ function usageOf2(raw) {
 }
 
 // src/agent-task/transcript-replay/otlp.ts
-import { createHash as createHash5 } from "crypto";
+import { createHash as createHash6 } from "crypto";
 var SPAN_KIND_INTERNAL = 1;
 var STATUS_CODE_UNSET = 0;
 function transcriptSessionToOtlp(session, options = {}) {
@@ -4485,7 +4558,7 @@ function messageAttrValue(role, content) {
   return JSON.stringify([{ role, content }]);
 }
 function sha256Hex(input) {
-  return createHash5("sha256").update(input).digest("hex");
+  return createHash6("sha256").update(input).digest("hex");
 }
 function deriveTraceId(source, sessionId) {
   return sha256Hex(`apo-transcript-replay|${source}|${sessionId}`).slice(0, 32);
@@ -6148,12 +6221,12 @@ function printOverallSummary(summaries, hasErrors) {
 }
 
 // src/agent-task/task-revision-manifest.ts
-import { createHash as createHash6 } from "crypto";
+import { createHash as createHash7 } from "crypto";
 function normalizeManifestPath(path) {
   return path.split("\\").join("/").normalize("NFC");
 }
 function sha256Hex2(bytes) {
-  return createHash6("sha256").update(bytes).digest("hex");
+  return createHash7("sha256").update(bytes).digest("hex");
 }
 function compareBytewiseUtf8(a, b) {
   const ba = Buffer.from(a, "utf8");