@kolisachint/hoocode-agent 0.4.71 → 0.4.73

package/dist/core/agent-session.js

@@ -21,12 +21,13 @@ import { sleep } from "../utils/sleep.js";
 import { loadAgentRegistry } from "./agent-registry.js";
 import { formatNoApiKeyFoundMessage, formatNoModelSelectedMessage } from "./auth-guidance.js";
 import { executeBashWithOperations } from "./bash-executor.js";
-import { calculateContextTokens, collectEntriesForBranchSummary, compact, estimateContextTokens, generateBranchSummary, prepareCompaction, shouldCompact, } from "./compaction/index.js";
+import { calculateContextTokens, collectEntriesForBranchSummary, compact, estimateContextTokens, generateBranchSummary, prepareCompaction, serializeConversation, shouldCompact, } from "./compaction/index.js";
 import { DEFAULT_THINKING_LEVEL } from "./defaults.js";
 import { exportSessionToHtml } from "./export-html/index.js";
 import { createToolHtmlRenderer } from "./export-html/tool-renderer.js";
 import { ExtensionRunner, wrapRegisteredTools, } from "./extensions/index.js";
 import { emitSessionShutdownEvent } from "./extensions/runner.js";
+import { convertToLlm } from "./messages.js";
 import { expandPromptTemplate, tryExpandPromptTemplate } from "./prompt-templates.js";
 import { clearProviderExhaustion, isProviderQuotaError, markProviderExhausted } from "./provider-health.js";
 import { LocalInferenceRouter, resolveRoutingMode } from "./routing/local-inference.js";
@@ -283,7 +284,7 @@ export class AgentSession {
                     changed = true;
                 }
             }
-            // Optionally compress large read/bash output before it enters context.
+            // Optionally compress large bash output before it enters context.
             // Only when local-inference tool-result routing is active; on any
             // failure the original (uncompressed) content is kept (fallback to raw).
             if (!resolvedIsError) {
@@ -301,7 +302,8 @@ export class AgentSession {
     /**
      * Compress a tool result via the local executor when routing is active and the
      * tool/size qualify. Returns the compressed content blocks, or undefined to
-     * keep the original (no routing, not compressible, too small, or any failure).
+     * keep the original (no routing, not compressible, outside the size band, or
+     * any failure).
      */
     async _maybeCompressToolResult(toolName, content) {
         const router = this.localRouter;
@@ -1494,21 +1496,29 @@ export class AgentSession {
     async _compactWithRouting(preparation, primaryModel, apiKey, headers, customInstructions, signal) {
         const router = this.localRouter;
         const executor = router?.selectModel("summarization", primaryModel);
+        // Size band guards local inference globally: only route conversations within
+        // the configured byte band to the executor. Oversized conversations are slow
+        // locally and can OOM small machines, so they fall back to the primary model.
+        // Only serialize to measure size when routing is actually engaged (avoids
+        // needless work on the common primary-only path).
         if (router && executor && executor !== primaryModel) {
-            try {
-                if (!(await this._ensureExecutorServer(signal))) {
-                    throw new Error("executor server unavailable");
+            const conversationBytes = Buffer.byteLength(serializeConversation(convertToLlm(preparation.messagesToSummarize)), "utf8");
+            if (router.withinSizeBand(conversationBytes)) {
+                try {
+                    if (!(await this._ensureExecutorServer(signal))) {
+                        throw new Error("executor server unavailable");
+                    }
+                    const { apiKey: exKey, headers: exHeaders } = await this._getRequiredRequestAuth(executor);
+                    return await compact(preparation, executor, exKey, exHeaders, customInstructions, signal, 
+                    // Executor runs without thinking (validated config); summary quality
+                    // relies on /no_think via the model's promptSuffix compat setting.
+                    "off");
+                }
+                catch (error) {
+                    if (signal.aborted)
+                        throw error;
+                    logLocalInferenceFallback("summarization", error);
                 }
-                const { apiKey: exKey, headers: exHeaders } = await this._getRequiredRequestAuth(executor);
-                return await compact(preparation, executor, exKey, exHeaders, customInstructions, signal, 
-                // Executor runs without thinking (validated config); summary quality
-                // relies on /no_think via the model's promptSuffix compat setting.
-                "off");
-            }
-            catch (error) {
-                if (signal.aborted)
-                    throw error;
-                logLocalInferenceFallback("summarization", error);
             }
         }
         return compact(preparation, primaryModel, apiKey, headers, customInstructions, signal, this.thinkingLevel);