@sema-agent/core 5.55.0 → 5.56.0

package/dist/engine/session/session.js

@@ -1,7 +1,7 @@
 import { asAgentMessage, createCompactionSummaryMessage, createCustomMessage, } from "../harness/messages.js";
 import { SessionError, isValidModelChange, normalizeAnnouncedListing, normalizeCompactionStateCarrier, normalizeGitAnnouncement, normalizeReminderMark, normalizeWorkspaceState } from "../harness/types.js";
 import { normalizePromptEpoch } from "../../prompt-assembly/epoch.js";
-import { budgetInvokedSkillsRetention, readElidedMessages, readRetainedInvokedSkills, renderInvokedSkillsRetention, } from "../compaction/utils.js";
+import { budgetInvokedSkillsRetention, readElidedMessages, readRetainedInvokedSkills, readUnsummarizedMessages, renderInvokedSkillsRetention, } from "../compaction/utils.js";
 const RETENTION_CLAMP_DEFAULT_CHARS_PER_TOKEN = 4;
 const RETENTION_CLAMP_WINDOW_FRACTION = 0.5;
 export function buildSessionContext(pathEntries, opts) {
@@ -67,7 +67,7 @@ export function buildSessionContext(pathEntries, opts) {
             retainedSkills = budgetInvokedSkillsRetention(retainedSkills, budgetChars);
         }
         const retainedSkillsBlock = renderInvokedSkillsRetention(retainedSkills);
-        messages.push(asAgentMessage(createCompactionSummaryMessage(compaction.summary + retainedSkillsBlock, compaction.tokensBefore, compaction.timestamp, readElidedMessages(compaction.details))));
+        messages.push(asAgentMessage(createCompactionSummaryMessage(compaction.summary + retainedSkillsBlock, compaction.tokensBefore, compaction.timestamp, readElidedMessages(compaction.details), readUnsummarizedMessages(compaction.details))));
         const compactionIdx = pathEntries.findIndex((e) => e.type === "compaction" && e.id === compaction.id);
         let foundFirstKept = false;
         for (let i = 0; i < compactionIdx; i++) {

package/dist/orchestration/run-spec.js

@@ -114,6 +114,9 @@ export async function runSpec(runner, contract, opts) {
             postCompact: (opts.taskSpec.hooks.postCompact ?? deployBaseline.hooks.postCompact)?.bind(opts.taskSpec.hooks.postCompact ? opts.taskSpec.hooks : deployBaseline.hooks),
             stopFailure: (opts.taskSpec.hooks.stopFailure ?? deployBaseline.hooks.stopFailure)?.bind(opts.taskSpec.hooks.stopFailure ? opts.taskSpec.hooks : deployBaseline.hooks),
             permissionDenied: (opts.taskSpec.hooks.permissionDenied ?? deployBaseline.hooks.permissionDenied)?.bind(opts.taskSpec.hooks.permissionDenied ? opts.taskSpec.hooks : deployBaseline.hooks),
+            ...((opts.taskSpec.hooks.timeoutMs ?? deployBaseline.hooks.timeoutMs) !== undefined
+                ? { timeoutMs: opts.taskSpec.hooks.timeoutMs ?? deployBaseline.hooks.timeoutMs }
+                : {}),
         }
         : (opts.taskSpec.hooks ?? deployBaseline?.hooks);
     let guardRestores = 0;
@@ -130,10 +133,14 @@ export async function runSpec(runner, contract, opts) {
         ...(userHooks?.postCompact && { postCompact: userHooks.postCompact.bind(userHooks) }),
         ...(userHooks?.stopFailure && { stopFailure: userHooks.stopFailure.bind(userHooks) }),
         ...(userHooks?.permissionDenied && { permissionDenied: userHooks.permissionDenied.bind(userHooks) }),
+        ...(userHooks?.timeoutMs !== undefined ? { timeoutMs: userHooks.timeoutMs } : {}),
         async stop(ctx) {
             guardRestores += restoreFrozenPaths(snapshot, rootDir);
             if (!gateAbandoned) {
-                const report = await runOracle(contract.oracle, rootDir, { baseline, signal: opts.taskSpec.signal });
+                const gateSignal = ctx.signal !== undefined && opts.taskSpec.signal !== undefined ? AbortSignal.any([ctx.signal, opts.taskSpec.signal]) : (ctx.signal ?? opts.taskSpec.signal);
+                const report = await runOracle(contract.oracle, rootDir, { baseline, ...(gateSignal !== undefined ? { signal: gateSignal } : {}) });
+                if (ctx.signal?.aborted === true)
+                    return undefined;
                 oracleRuns++;
                 if (report.verdict === "red") {
                     oracleHadRedRun = true;

package/dist/prompts/default.d.ts

@@ -142,8 +142,8 @@ export declare const CYBER_RISK = "IMPORTANT: Assist with authorized security te
 export declare const URL_SAFETY = "IMPORTANT: You must NEVER generate or guess URLs for the user unless you are confident that the URLs are for helping the user with programming. You may use URLs provided by the user in their messages or local files.";
 /**
  * Tool-result retention reminder (design/64 §6.2 C). Always injected: core's in-task context management
- * (`clearStaleToolResults`) unconditionally clears old tool-result content before each request, so the
- * model must persist anything load-bearing into its own response.
+ * (`clearStaleToolResults`) can clear old tool-result content out of a request, and compaction can
+ * replace it with prose, so the model must persist anything load-bearing into its own response.
  *
  * RB-328 — CITATION CORRECTED. Source = the CC **88 readable tree**
  * (`collection-claude-code-source-code/original-source-code/src/constants/prompts.ts:841`,
@@ -157,8 +157,14 @@ export declare const URL_SAFETY = "IMPORTANT: You must NEVER generate or guess U
  * context-management claim (`_My` @596605-596612, the "The system will automatically compress prior
  * messages … your conversation with the user is not limited by the context window" line — an
  * overclaim {@link harnessHeadLines} explicitly refuses to copy). So 220 clears results without ever
- * telling the model to save what mattered. sema keeps the instruction: its clear is unconditional and
- * runs before EVERY request, which is a stronger reason to warn than CC's.
+ * telling the model to save what mattered. sema keeps the instruction — and the REASON is now stated
+ * as measured rather than as remembered (anchoring re-check 2026-08-23): the earlier wording here
+ * ("its clear is unconditional and runs before EVERY request") is false about our own engine.
+ * `clearStaleToolResults` returns the original array untouched while the estimate is within
+ * `editBudget` (= `contextEditFrontier`, e.g. 167000 on a 200k window) — probed: under budget, same
+ * reference, zero clears. What is true, and is warning enough, is that a long session meets THREE
+ * removers: this pass once the frontier is crossed, the compaction boundary's summary, and the
+ * per-message aggregate tool-result budget, which sheds the largest results at any size.
  */
 export declare const SUMMARIZE_TOOL_RESULTS = "When working with tool results, write down any important information you might need later in your own response, as the original tool result may be cleared or summarized from the context later.";
 export declare const EXECUTION_ENVIRONMENT: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sema-agent/core",
-  "version": "5.55.0",
+  "version": "5.56.0",
   "description": "Stateless, task-oriented AI agent core",
   "type": "module",
   "license": "BUSL-1.1",