@semiont/jobs 0.6.6 → 0.6.8

package/dist/worker-main.js

@@ -1,7 +1,7 @@
-import { replyChannelsFor, createTomlConfigLoader, didToAgent, baseUrl, STARTUP_FETCH_RETRY, retryWithBackoff, serviceAccountToken, isTransientFetchError, isObject, isString, isNumber, RETRY_RULES, resourceId, getPrimaryMediaType, isGenerationJobParams, capabilitiesOf, assembleAnnotation, findClaimSpan, busRequest, BusRequestError, isArray, textSourceOf, yieldsGeometryOf, decodeRepresentation, GENERATABLE_MEDIA_TYPES, locate, createFragmentSelector, deriveViews, annotationId, estimateTokens, reconcileSelector, getLocaleEnglishName, cutChunk, chunkText } from '@semiont/core';
+import { JOB_TYPES, replyChannelsFor, createTomlConfigLoader, didToAgent, baseUrl, STARTUP_FETCH_RETRY, retryWithBackoff, serviceAccountToken, isTransientFetchError, isObject, isString, isNumber, RETRY_RULES, resourceId, getPrimaryMediaType, isGenerationJobParams, capabilitiesOf, assembleAnnotation, findClaimSpan, busRequest, BusRequestError, isArray, textSourceOf, yieldsGeometryOf, decodeRepresentation, GENERATABLE_MEDIA_TYPES, locate, createFragmentSelector, deriveViews, annotationId, estimateTokens, reconcileSelector, getLocaleEnglishName, cutChunk, chunkText } from '@semiont/core';
 import { archivistContentReads, extractPdfTextLayer, withinByteBudget, MAX_PDF_BYTES } from '@semiont/content';
 import { withSpan, SpanKind, recordJobOutcome, recordAnchorOutcome, recordDetectionCall } from '@semiont/observability';
-import { createInferenceClient, StructuredReadError } from '@semiont/inference';
+import { createInferenceClient, answerLimitsRequests, StructuredReadError } from '@semiont/inference';
 import { execFileSync } from 'child_process';
 import { existsSync, readFileSync, mkdtempSync, writeFileSync, rmSync } from 'fs';
 import { homedir, hostname, tmpdir } from 'os';
@@ -9384,18 +9384,9 @@ function createJobClaimAdapter(options) {
 function willRetryAfter(metadata, failureClass) {
   return failureClass !== "deterministic" && metadata.retryCount < metadata.maxRetries;
 }
-
-// src/types.ts
-var JOB_TYPES = /* @__PURE__ */ new Set([
-  "reference-annotation",
-  "generation",
-  "highlight-annotation",
-  "assessment-annotation",
-  "comment-annotation",
-  "tag-annotation"
-]);
+var KNOWN_JOB_TYPES = new Set(JOB_TYPES);
 function isJobType(value) {
-  return JOB_TYPES.has(value);
+  return KNOWN_JOB_TYPES.has(value);
 }
 function asJobParams(params) {
   if (typeof params.resourceId !== "string") {
@@ -12016,6 +12007,9 @@ var WORKER_CONSUMED_BROADCASTS = [
   // Cooperative cancellation of the ACTIVE job (JOB-RESTART-SAFETY P4).
   "job:cancel-requested"
 ];
+var WORKER_ANSWERED_OPERATIONS = [
+  "job:limits-requested"
+];
 var WORKER_CHANNELS = [
   ...replyChannelsFor(WORKER_AWAITED_OPERATIONS),
   ...WORKER_CONSUMED_BROADCASTS
@@ -12059,7 +12053,7 @@ async function authenticateAgent(opts) {
   );
 }
 async function startAgentWorker(opts) {
-  const { group, gatewayBaseUrl: gatewayBaseUrl2, credential: credential2, contentReads: contentReads2, logger: logger2 } = opts;
+  const { group, gatewayBaseUrl: gatewayBaseUrl2, credential: credential2, contentReads: contentReads2, reportsLimitsOf, logger: logger2 } = opts;
   const { inference } = group;
   const { protocol, host, port } = parseGatewayUrl(gatewayBaseUrl2);
   const { token: initialToken, did } = await authenticateAgent({
@@ -12091,8 +12085,9 @@ async function startAgentWorker(opts) {
     token$,
     tokenRefresher: () => session.refresh().then((t) => t ?? null),
     // Only the reply channels this process awaits — not the full bridged
-    // set. See WORKER_AWAITED_OPERATIONS.
-    channels: WORKER_CHANNELS
+    // set. See WORKER_AWAITED_OPERATIONS. The agent that reports its pool's
+    // limits also subscribes the requests it answers.
+    channels: reportsLimitsOf.length > 0 ? [...WORKER_CHANNELS, ...WORKER_ANSWERED_OPERATIONS] : WORKER_CHANNELS
   });
   const content = new HttpContentTransport(transport);
   const client = new SemiontClient(transport, content, transport);
@@ -12136,6 +12131,7 @@ async function startAgentWorker(opts) {
     contentReads: contentReads2,
     logger: logger2
   });
+  const limitsResponder = reportsLimitsOf.length > 0 ? answerLimitsRequests(transport, "job:limits-requested", reportsLimitsOf, logger2) : void 0;
   logger2.info("Agent ready", {
     did,
     provider: inference.type,
@@ -12152,19 +12148,12 @@ async function startAgentWorker(opts) {
       ...adapter.vitals()
     }),
     dispose: async () => {
+      limitsResponder?.unsubscribe();
       adapter.dispose();
       await session.dispose();
     }
   };
 }
-var ALL_JOB_TYPES = [
-  "reference-annotation",
-  "generation",
-  "highlight-annotation",
-  "assessment-annotation",
-  "comment-annotation",
-  "tag-annotation"
-];
 var configPath = join(homedir(), ".semiontconfig");
 var tomlReader = {
   readIfExists: (p) => existsSync(p) ? readFileSync(p, "utf-8") : null
@@ -12172,7 +12161,8 @@ var tomlReader = {
 var envConfig = createTomlConfigLoader(
   tomlReader,
   configPath,
-  process.env
+  process.env,
+  "worker"
 )(null);
 var workerInferenceMap = envConfig._metadata?.workers;
 if (!workerInferenceMap || Object.keys(workerInferenceMap).length === 0) {
@@ -12220,7 +12210,7 @@ function toClientConfig(w) {
   };
 }
 var groups = /* @__PURE__ */ new Map();
-for (const jobType of ALL_JOB_TYPES) {
+for (const jobType of JOB_TYPES) {
   const inference = resolveWorker(jobType);
   const key = clientKey(inference);
   let group = groups.get(key);
@@ -12247,8 +12237,17 @@ async function main() {
     }))
   });
   const workers = await Promise.all(
+    // The first agent reports every group's limits: the gateway delivers only
+    // the first reply to a request, so one agent answers for the pool.
     Array.from(groups.values()).map(
-      (group) => startAgentWorker({ group, gatewayBaseUrl, credential, contentReads, logger })
+      (group, i, all) => startAgentWorker({
+        group,
+        gatewayBaseUrl,
+        credential,
+        contentReads,
+        logger,
+        reportsLimitsOf: i === 0 ? all.map((g) => g.client) : []
+      })
     )
   );
   const health = createServer((req, res) => {