@hmharness/kernel 0.4.4 → 0.5.0

package/dist/loop.js

@@ -7,7 +7,7 @@
  * no gate configured means deny (safe default). Between turns the
  * transcript is compacted against the context budget.
  */
-import { compactMessages, compactWithDigest } from "./context.js";
+import { compactMessages, compactWithDigest, transcriptChars } from "./context.js";
 import { adaptiveContextChars } from "./window.js";
 import { chat } from "./provider.js";
 export async function runLoop(opts) {
@@ -19,10 +19,21 @@ export async function runLoop(opts) {
     let toolUses = 0;
     const usage = { promptTokens: 0, completionTokens: 0 };
     const tools = registry.toOpenAITools();
+    let wrapupSignaled = false;
     for (let turn = 1; turn <= maxTurns; turn++) {
         const compacted = opts.summarizeContext
             ? await compactWithDigest(working, budget, opts.summarizeContext)
             : compactMessages(working, budget);
+        // Budget wrap-up signal: when context usage crosses 80%, inject a
+        // single system nudge to wind down instead of hard-cutting mid-thought
+        // (Codex's token-budget-context pattern; DeepSeek's 80% pressure trigger).
+        if (!wrapupSignaled && transcriptChars(compacted) > budget * 0.8) {
+            wrapupSignaled = true;
+            working.push({
+                role: 'system',
+                content: '[context budget] Context is running low. Wrap up the current task: summarize what was accomplished, suggest concrete next steps, and stop starting new sub-tasks.',
+            });
+        }
         const chatRes = await modelCall(provider, compacted, tools, {
             onDelta: events?.onDelta,
         });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hmharness/kernel",
-  "version": "0.4.4",
+  "version": "0.5.0",
   "description": "hmharness kernel: tool registry, provider adapters, the agent loop, session log, config. Zero runtime dependencies (Node >=22 native fetch).",
   "type": "module",
   "main": "dist/index.js",