@ai-agent-forge/agent-forge 0.86.0 → 0.86.2

package/dist/core/agent-session.js

@@ -101,6 +101,16 @@ export class AgentSession {
     _autoCompactionAbortController = undefined;
     // biome-ignore lint/correctness/noUnusedPrivateClassMembers: read and written via the agent-session-* submodule facade
     _overflowRecoveryAttempted = false;
+    // biome-ignore lint/correctness/noUnusedPrivateClassMembers: read and written via the agent-session-* submodule facade
+    _overflowTrimStage = 0;
+    /**
+     * microcompact 冷却（按模型请求计数，0 = 不冷却）：relief 落地时置
+     * MICROCOMPACT_COOLDOWN_ASSISTANT_RESPONSES，此后每次 assistant message_end
+     * （无论 stopReason）减 1、下限 0；user message_start 不复位。冷却期内阈值
+     * decide 以 microcompactEnabled:false 关闭便宜缓解。微观经济动机见
+     * agent-session-compaction.ts 的常量注释（中部历史改写 × 不稳定前缀缓存）。
+     */
+    _microcompactCooldown = 0;
     // Branch summarization state
     _branchSummaryAbortController = undefined;
     _branchSummaryPromise = undefined;
@@ -440,6 +450,7 @@ export class AgentSession {
         // This ensures the UI sees the updated queue state
         if (event.type === "message_start" && event.message.role === "user") {
             this._overflowRecoveryAttempted = false;
+            this._overflowTrimStage = 0;
             const messageText = contentText(event.message.content, "");
             if (messageText) {
                 // Check steering queue first
@@ -501,6 +512,12 @@ export class AgentSession {
                 const assistantMsg = event.message;
                 if (assistantMsg.stopReason !== "error" && assistantMsg.stopReason !== "length") {
                     this._overflowRecoveryAttempted = false;
+                    this._overflowTrimStage = 0;
+                }
+                // microcompact 冷却按模型请求计数：每个完成的 assistant 响应（无论
+                // stopReason）释放一档，下限 0；冷却期内的阈值 decide 不做便宜缓解。
+                if (this._microcompactCooldown > 0) {
+                    this._microcompactCooldown -= 1;
                 }
                 // Reset retry counter immediately on successful assistant response
                 // This prevents accumulation across multiple LLM calls within a turn