@springbrand/agent-runtime 0.3.0-alpha.19 → 0.3.0-alpha.20

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@springbrand/agent-runtime",
-  "version": "0.3.0-alpha.19",
+  "version": "0.3.0-alpha.20",
   "type": "module",
   "files": [
     "src",

package/src/config/harness-runtime.ts

@@ -153,10 +153,9 @@ export class HarnessRuntimeConfig {
       platform: state.snapshot.bindings.platform,
       compaction: {
         enabled: true,
-        // Pi compacts above contextWindow - reserveTokens. Keep the same
-        // 15% headroom within the available input budget as the pre-Harness runtime.
-        reserveTokens: Math.ceil(state.snapshot.pi.model.contextWindow -
-          (state.snapshot.pi.model.contextWindow - state.snapshot.pi.model.maxTokens) * 0.85),
+        // Pi compacts above contextWindow - reserveTokens.
+        reserveTokens: state.snapshot.pi.model.contextWindow -
+          compactionThreshold(state.snapshot.pi.model),
         keepRecentTokens: Math.min(
           20_000,
           Math.floor(state.snapshot.pi.model.contextWindow * 0.15),
@@ -201,3 +200,21 @@ export type {
   PreparePiRuntimeOptions,
   PreparedPiRuntime,
 } from "./pinning";
+
+/**
+ * Input tokens above which a Turn compacts its history.
+ *
+ * OpenRouter rejects some requests well below a model's advertised window
+ * (seen from about 393k tokens on DeepSeek V4.1 Flash and 520k-630k on GPT-6
+ * Luna), before any provider is chosen. No model waits past this cap, so long
+ * Chats compact instead of failing.
+ */
+export const MAX_COMPACTION_THRESHOLD_TOKENS = 300_000;
+
+/** 15% headroom within the input budget, capped at {@link MAX_COMPACTION_THRESHOLD_TOKENS}. */
+export function compactionThreshold(model: { contextWindow: number; maxTokens: number }): number {
+  return Math.min(
+    Math.floor((model.contextWindow - model.maxTokens) * 0.85),
+    MAX_COMPACTION_THRESHOLD_TOKENS,
+  );
+}