@opengeni/worker-bundle 2.0.3-canary.36079201993001 → 2.0.3

package/dist/index.js

@@ -790,7 +790,7 @@ function temporalWorkflowExecutionNotFound(error) {
   return isGrpcServiceError(error) && error.code === 5;
 }
 async function createDefaultWorkerActivities(role, dependencies = {}) {
-  return role === "control" ? (await import("./activities-control-IVTL723D.js")).createControlActivities(dependencies) : (await import("./activities-turn-AJHDYNZU.js")).createTurnActivities(dependencies);
+  return role === "control" ? (await import("./activities-control-IVTL723D.js")).createControlActivities(dependencies) : (await import("./activities-turn-EZEDZTEZ.js")).createTurnActivities(dependencies);
 }
 function resolveOpenGeniWorkflowDefinition(moduleUrl = import.meta.url) {
   const modulePath = fileURLToPath(moduleUrl).replaceAll("\\", "/");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@opengeni/worker-bundle",
-  "version": "2.0.3-canary.36079201993001",
+  "version": "2.0.3",
   "description": "OpenGeni worker entry and reusable embedded lifecycle, shipped with a release-coherent pre-bundled Temporal workflow artifact.",
   "license": "Apache-2.0",
   "repository": {
@@ -54,24 +54,24 @@
   },
   "dependencies": {
     "@llamaindex/liteparse": "2.14.2",
-    "@opengeni/agent-proto": "^0.6.0-canary.36079201993001",
-    "@opengeni/capabilities": "^0.3.4-canary.36079201993001",
-    "@opengeni/codemode": "^0.6.2-canary.36079201993001",
-    "@opengeni/codex": "^0.2.26-canary.36079201993001",
-    "@opengeni/config": "^2.1.1-canary.36079201993001",
-    "@opengeni/contracts": "^5.1.1-canary.36079201993001",
-    "@opengeni/core": "^4.0.3-canary.36079201993001",
-    "@opengeni/db": "^6.0.3-canary.36079201993001",
-    "@opengeni/documents": "^0.8.34-canary.36079201993001",
-    "@opengeni/events": "^0.4.32-canary.36079201993001",
-    "@opengeni/github": "^0.7.16-canary.36079201993001",
-    "@opengeni/network": "^0.3.1-canary.36079201993001",
-    "@opengeni/observability": "^0.8.32-canary.36079201993001",
-    "@opengeni/runtime": "^4.0.3-canary.36079201993001",
-    "@opengeni/sdk": "^7.1.1-canary.36079201993001",
-    "@opengeni/storage": "^0.2.133-canary.36079201993001",
-    "@opengeni/tool-gateway": "^0.1.13-canary.36079201993001",
-    "@opengeni/xai-subscription": "^0.1.4-canary.36079201993001",
+    "@opengeni/agent-proto": "^0.6.0",
+    "@opengeni/capabilities": "^0.3.4",
+    "@opengeni/codemode": "^0.6.2",
+    "@opengeni/codex": "^0.2.26",
+    "@opengeni/config": "^2.1.1",
+    "@opengeni/contracts": "^5.1.1",
+    "@opengeni/core": "^4.0.3",
+    "@opengeni/db": "^6.0.3",
+    "@opengeni/documents": "^0.8.34",
+    "@opengeni/events": "^0.4.32",
+    "@opengeni/github": "^0.7.16",
+    "@opengeni/network": "^0.3.1",
+    "@opengeni/observability": "^0.8.32",
+    "@opengeni/runtime": "^4.0.3",
+    "@opengeni/sdk": "^7.1.1",
+    "@opengeni/storage": "^0.2.133",
+    "@opengeni/tool-gateway": "^0.1.13",
+    "@opengeni/xai-subscription": "^0.1.4",
     "@temporalio/activity": "^1.17.0",
     "@temporalio/client": "^1.17.0",
     "@temporalio/worker": "^1.17.0",

package/src/activities/agent-turn/compaction-prep.ts

@@ -11,8 +11,7 @@ import {
   compactionThresholdTokens,
   CompactionNeededError,
   compactionProviderRejection,
-  estimateSerializedValueTokens,
-  compactionSummaryOutputTokens,
+  SUMMARY_BUFFER_TOKENS,
   type ModelResponseUsage,
 } from "@opengeni/runtime";
 import { type Settings } from "@opengeni/config";
@@ -169,13 +168,6 @@ export async function prepareCompaction(deps: CompactionPrepDeps): Promise<Compa
   const remotePrefix: RemoteCompactionPrefix = {
     agent: null,
   };
-  const portableResponsesNeedsAgentPrefix =
-    resolvedModel?.provider.api === "responses" &&
-    !(billingState.isCodexTurn && session.codexCompactionMode === "remote_v2");
-  const preparedPortableRequest = () => {
-    if (!remotePrefix.agent) throw new Error("Compaction agent is unavailable");
-    return preparedCompactionRequest(remotePrefix.agent);
-  };
 
   const promptCacheKey = acceptsPromptCacheKeyForTurn(resolvedModel) ? input.sessionId : undefined;
   const compactionUsageState = createCompactionModelUsageEventState(claimedModelUsageSourceKeys);
@@ -208,8 +200,8 @@ export async function prepareCompaction(deps: CompactionPrepDeps): Promise<Compa
       contextContributions: eventing.companyBrainContextContributions,
     });
   };
-  const compactionSummarizerFor = (systemInstructions?: string): CompactionSummarizer => {
-    const summarize: CompactionSummarizer = resolvedModel
+  const compactionSummarizerFor = (systemInstructions?: string) =>
+    resolvedModel
       ? (s: Settings, m: Array<Record<string, unknown>>) =>
           withProviderRequestContext(() =>
             summarizeContextForCompaction(s, m, {
@@ -217,37 +209,20 @@ export async function prepareCompaction(deps: CompactionPrepDeps): Promise<Compa
               provider: resolvedModel.provider,
               api: resolvedModel.provider.api,
               model: turnExecutionPolicy.upstreamModelId,
-              maxOutputTokens: compactionSummaryOutputTokens(s.contextWindowTokens),
-              ...(cancellationSignal ? { signal: cancellationSignal } : {}),
+              maxOutputTokens: SUMMARY_BUFFER_TOKENS,
               onUsage: recordCompactionUsage,
               ...(systemInstructions ? { systemInstructions } : {}),
               ...(promptCacheKey ? { promptCacheKey } : {}),
-              ...(portableResponsesNeedsAgentPrefix
-                ? { preparedRequest: preparedPortableRequest() }
-                : {}),
             }),
           )
       : (s: Settings, m: Array<Record<string, unknown>>) =>
           summarizeContextForCompaction(s, m, {
             model: turnExecutionPolicy.upstreamModelId,
-            maxOutputTokens: compactionSummaryOutputTokens(s.contextWindowTokens),
-            ...(cancellationSignal ? { signal: cancellationSignal } : {}),
+            maxOutputTokens: SUMMARY_BUFFER_TOKENS,
             onUsage: recordCompactionUsage,
             ...(systemInstructions ? { systemInstructions } : {}),
             ...(promptCacheKey ? { promptCacheKey } : {}),
           });
-    summarize.estimatePrefixTokens = () => {
-      if (resolvedModel?.provider.api === "chat") {
-        return estimateSerializedValueTokens(systemInstructions ?? "");
-      }
-      const prepared = portableResponsesNeedsAgentPrefix ? preparedPortableRequest() : null;
-      return (
-        estimateSerializedValueTokens(prepared?.systemInstructions ?? systemInstructions ?? "") +
-        (prepared ? estimateSerializedValueTokens(prepared.tools) : 0)
-      );
-    };
-    return summarize;
-  };
   // Prompt-cache prefix for remote_v2 MUST match ordinary turns:
   // tools → instructions → history. Filled after buildAgent for every
   // compact path (including operator /compact, which now builds the agent
@@ -290,9 +265,13 @@ export async function prepareCompaction(deps: CompactionPrepDeps): Promise<Compa
     ...(remoteCompactionRequester ? { requestRemoteCompactionV2: remoteCompactionRequester } : {}),
   } as const;
 
-  // Responses compaction, portable or remote, prepares the ordinary agent
-  // request first so tool schemas and instructions match the warm cache prefix.
-  // Chat providers keep the standalone portable maintenance path.
+  // Operator /compact:
+  // - portable (incl. Codex portable): early maintenance path — no
+  //   prepareTools/sandbox; summarizer only needs composed instructions.
+  // - remote_v2: fall through to prepareTools/buildAgent so the compact
+  //   request reuses the ordinary tools→instructions cache prefix, then
+  //   settle without inference. (Requester is also wired for Codex portable
+  //   turns but unused there — gate on the frozen session mode.)
   const compactionOnlyTurn = turn.source === "compaction";
   const remoteV2CompactionNeedsAgentPrefix =
     Boolean(remoteCompactionRequester) && session.codexCompactionMode === "remote_v2";
@@ -329,11 +308,7 @@ export async function prepareCompaction(deps: CompactionPrepDeps): Promise<Compa
     control.activityStatus = "idle";
     return claimedResult({ status: "idle" });
   };
-  if (
-    compactionOnlyTurn &&
-    !remoteV2CompactionNeedsAgentPrefix &&
-    !portableResponsesNeedsAgentPrefix
-  ) {
+  if (compactionOnlyTurn && !remoteV2CompactionNeedsAgentPrefix) {
     const compactionInstructions = appendWorkspaceMemory(
       appendSessionInstructions(
         appendWorkspaceGovernance(
@@ -491,7 +466,6 @@ export async function runPostAgentCompaction(
     claimedResult,
     turn,
     session,
-    resolvedModel,
     remotePrefix,
     remoteCompactionRequester,
     publishCompactionLiveEvents,
@@ -506,13 +480,10 @@ export async function runPostAgentCompaction(
   } = deps;
 
   const agentInstructions = typeof agent.instructions === "string" ? agent.instructions : "";
-  const preparedPortable =
-    resolvedModel?.provider.api === "responses" &&
-    !(remoteCompactionRequester && session.codexCompactionMode === "remote_v2");
   const compactSummarizer = compactionSummarizerFor(
     agentInstructions.trim() ? agentInstructions : undefined,
   );
-  if (remoteCompactionRequester || preparedPortable) {
+  if (remoteCompactionRequester) {
     // The prefix is captured only after this agent passes normal SDK preparation.
     remotePrefix.agent = agent;
   }
@@ -520,11 +491,7 @@ export async function runPostAgentCompaction(
   if (compactionOnlyTurn) {
     const requested = await isSessionCompactionRequested(db, input.workspaceId, input.sessionId);
     let outcome: Awaited<ReturnType<typeof maybeCompactContext>> | null = null;
-    if (
-      requested &&
-      (preparedPortable ||
-        (remoteCompactionRequester && session.codexCompactionMode === "remote_v2"))
-    ) {
+    if (requested && remoteCompactionRequester && session.codexCompactionMode === "remote_v2") {
       queuePreparedCompaction(
         agent,
         new CompactionNeededError({
@@ -643,10 +610,7 @@ export async function runPostAgentCompaction(
     return { exit: claimedResult({ status: "idle" }) };
   }
 
-  if (
-    preparedPortable ||
-    (remoteCompactionRequester && session.codexCompactionMode === "remote_v2")
-  ) {
+  if (remoteCompactionRequester && session.codexCompactionMode === "remote_v2") {
     const forced = await isSessionCompactionRequested(db, input.workspaceId, input.sessionId);
     const thresholdTokens = compactionThresholdTokens(eventing.modelRunSettings);
     if (

package/src/activities/context-compaction.ts

@@ -9,8 +9,7 @@ import {
 import {
   EmptyCompactionSummaryError,
   REMOTE_COMPACTION_V2_IMPLEMENTATION,
-  compactionSummaryOutputTokens,
-  buildSummaryItem,
+  SUMMARY_BUFFER_TOKENS,
   buildCompactionReplacementHistory,
   buildRemoteV2ReplacementHistory,
   compactionThresholdTokens,
@@ -59,13 +58,7 @@ export type MaybeCompactResult =
  * Codex, call Codex `/codex/responses` with `compaction_trigger` and persist the
  * opaque compaction item. Fail closed — never silently fall back to portable.
  */
-export type CompactionSummarizer = ((
-  settings: Settings,
-  input: CompactionItem[],
-) => Promise<string>) & {
-  /** Model-visible instructions and tool schemas outside the history estimate. */
-  estimatePrefixTokens?: () => number;
-};
+export type CompactionSummarizer = (settings: Settings, input: CompactionItem[]) => Promise<string>;
 
 /** Returns the opaque Codex remote compaction v2 item. */
 export type RemoteCompactionV2Requester = (
@@ -88,7 +81,7 @@ export async function maybeCompactContext(
   // Injectable for tests; defaults to the real provider-aware model call.
   summarize: CompactionSummarizer = (s, m) =>
     summarizeForCompaction(s, m, {
-      maxOutputTokens: compactionSummaryOutputTokens(s.contextWindowTokens),
+      maxOutputTokens: SUMMARY_BUFFER_TOKENS,
     }),
   // Operator-forced (the /compact command): bypass the budget trigger and
   // compact now if there is anything to summarize. Structural guards still hold.
@@ -307,7 +300,6 @@ async function settleSkippedAfterStart(
   reason:
     | "no_history"
     | "replacement_not_smaller"
-    | "replacement_exceeds_model_budget"
     | "replacement_unchanged"
     | "summarization_failed",
 ): Promise<Extract<MaybeCompactResult, { compacted: false }>> {
@@ -463,23 +455,12 @@ async function compactContextPortable(
   const summarized = await summarizeWithCodexOverflowTrimming(summarize, settings, items);
   const summaryBody = summarized.summaryBody;
   const retainedTokens = await retentionTokenCounts(canonicalItems, projectForWire);
-  const outputReserve = compactionSummaryOutputTokens(settings.contextWindowTokens);
-  const structuralBudget = Math.min(
-    contextInputBudgetTokens(settings) || settings.contextWindowTokens - outputReserve,
-    settings.contextWindowTokens - outputReserve,
-  );
-  const prefixTokens = Math.max(0, Math.ceil(summarize.estimatePrefixTokens?.() ?? 0));
-  const summaryTokens = estimateTokens([buildSummaryItem(summaryBody)]);
   const replacementHistory = buildCompactionReplacementHistory(
     canonicalItems,
     summaryBody,
     (item) => retainedTokens.get(item) ?? estimateTokens([item]),
-    Math.min(outputReserve, Math.max(0, structuralBudget - prefixTokens - summaryTokens)),
   );
   const estimatedTokensAfter = estimateTokens(await projectForWire(replacementHistory));
-  if (estimatedTokensAfter + prefixTokens > structuralBudget) {
-    return await settleSkippedAfterStart(db, scope, options, "replacement_exceeds_model_budget");
-  }
   const replacementFingerprint = compactionReplacementFingerprint(replacementHistory);
   const previousReplacementFingerprint = latestCompactionReplacementFingerprint(canonicalItems);
   const summaryIndex =
@@ -539,7 +520,7 @@ async function compactContextPortable(
   };
 }
 
-export async function summarizeWithCodexOverflowTrimming(
+async function summarizeWithCodexOverflowTrimming(
   summarize: CompactionSummarizer,
   settings: Settings,
   activeHistory: CompactionItem[],
@@ -550,32 +531,17 @@ export async function summarizeWithCodexOverflowTrimming(
 }> {
   // Codex's estimator is intentionally coarse. Keep the explicit checkpoint
   // request below both the effective input window and the raw window minus the
-  // requested summary. Preserve the full portable history copy on the first
-  // call whenever it fits; only trim further after an actual provider overflow.
-  // Durable active history remains untouched until applyContextCompaction.
-  const summaryAwareBudget = Math.max(
-    0,
-    settings.contextWindowTokens - compactionSummaryOutputTokens(settings.contextWindowTokens),
-  );
+  // requested summary, then leave 15% estimator headroom. This changes only the
+  // temporary summarizer input; durable active history remains untouched until
+  // applyContextCompaction succeeds under the attempt fence.
+  const summaryAwareBudget = Math.max(0, settings.contextWindowTokens - SUMMARY_BUFFER_TOKENS);
   const configuredInputBudget = contextInputBudgetTokens(settings);
   const structuralBudget = Math.min(
     configuredInputBudget > 0 ? configuredInputBudget : summaryAwareBudget,
     summaryAwareBudget,
   );
-  const prefixTokens = Math.max(0, Math.ceil(summarize.estimatePrefixTokens?.() ?? 0));
-  const initialBudget = Math.max(0, structuralBudget - prefixTokens);
+  const initialBudget = Math.floor(structuralBudget * 0.85);
   let preparation = prepareCompactionPromptInput(activeHistory, initialBudget);
-  // A checkpoint prompt without source history cannot summarize that history.
-  // Do not let a plausible-sounding reply replace durable active context.
-  const requireHistory = () => {
-    if (activeHistory.length > 0 && preparation.input.length === 1) {
-      throw new EmptyCompactionSummaryError({
-        stage: "portable_input_budget",
-        reason: "no_history_fit",
-      });
-    }
-  };
-  requireHistory();
   try {
     return {
       summaryBody: await summarize(settings, preparation.input),
@@ -585,16 +551,15 @@ export async function summarizeWithCodexOverflowTrimming(
   } catch (error) {
     if (!isContextWindowExceeded(error)) throw error;
     // The provider is more authoritative than the byte/4 estimate. Refit once
-    // to 40% of both the available target and the actual prepared estimate;
+    // to 70% of both the configured target and the actual prepared estimate;
     // then fail terminally with prior history intact. Never issue one failing
-    // request per oldest item. The retry is bounded, not a guarantee: the
-    // provider can count slightly more than twice the byte/4 estimate. The
-    // prepared prefix has already been reserved from the available budget.
+    // request per oldest item. The production incident proved that the
+    // provider can count slightly more than twice the byte/4 estimate, so a
+    // half-size retry is the smallest honest bound for that observed skew.
     const retryBudget = Math.floor(
-      Math.min(initialBudget * 0.4, preparation.estimatedInputTokens * 0.4),
+      Math.min(initialBudget * 0.5, preparation.estimatedInputTokens * 0.5),
     );
     preparation = prepareCompactionPromptInput(activeHistory, retryBudget);
-    requireHistory();
     return {
       summaryBody: await summarize(settings, preparation.input),
       preparation,