fix: keep session token totals in sync (#1440) (thanks @robbyczgw-cla)

CHANGELOG.md

@@ -7,6 +7,7 @@ Docs: https://docs.clawd.bot
 ### Fixes
 - BlueBubbles: stop typing indicator on idle/no-reply. (#1439) Thanks @Nicell.
 - Auto-reply: only report a model switch when session state is available. (#1465) Thanks @robbyczgw-cla.
+- Auto-reply: keep cached context token count in sync after compaction. (#1440) Thanks @robbyczgw-cla.
 - Control UI: resolve local avatar URLs with basePath across injection + identity RPC. (#1457) Thanks @dlauer.
 - Agents: surface concrete API error details instead of generic AI service errors.
 - Docs: fix gog auth services example to include docs scope. (#1454) Thanks @zerone0x.

src/agents/pi-embedded-runner/compact.ts

@@ -375,19 +375,24 @@ export async function compactEmbeddedPiSession(params: {
               session.agent.replaceMessages(limited);
             }
             const result = await session.compact(params.customInstructions);
-            // Estimate tokens after compaction by summing token estimates for remaining messages
+            // Estimate tokens after compaction with the same context-usage heuristics.
             let tokensAfter: number | undefined;
             try {
-              tokensAfter = 0;
+              const usage =
-              for (const message of session.messages) {
+                typeof session.getContextUsage === "function"
-                tokensAfter += estimateTokens(message);
+                  ? session.getContextUsage()
+                  : undefined;
+              let estimate = usage?.tokens;
+              if (!Number.isFinite(estimate) || !estimate || estimate <= 0) {
+                estimate = 0;
+                for (const message of session.messages) {
+                  estimate += estimateTokens(message);
+                }
               }
-              // Sanity check: tokensAfter should be less than tokensBefore
+              if (Number.isFinite(estimate) && estimate > 0 && estimate <= result.tokensBefore) {
-              if (tokensAfter > result.tokensBefore) {
+                tokensAfter = estimate;
-                tokensAfter = undefined; // Don't trust the estimate
               }
             } catch {
-              // If estimation fails, leave tokensAfter undefined
               tokensAfter = undefined;
             }
             return {

src/agents/subagent-registry.ts

@@ -31,7 +31,8 @@ const subagentRuns = new Map<string, SubagentRunRecord>();
 let sweeper: NodeJS.Timeout | null = null;
 let listenerStarted = false;
 let listenerStop: (() => void) | null = null;
-let restoreAttempted = false;
+// Use var to avoid TDZ on circular init paths that can call restoreSubagentRunsOnce early.
+var restoreAttempted = false;
 
 function persistSubagentRuns() {
   try {

src/auto-reply/reply.block-streaming.test.ts

@@ -42,7 +42,7 @@ describe("block streaming", () => {
   });
 
   async function waitForCalls(fn: () => number, calls: number) {
-    const deadline = Date.now() + 1500;
+    const deadline = Date.now() + 15000;
     while (fn() < calls) {
       if (Date.now() > deadline) {
         throw new Error(`Expected ${calls} call(s), got ${fn()}`);

src/auto-reply/reply/session-updates.test.ts

@@ -2,7 +2,7 @@ import { describe, expect, it, vi } from "vitest";
 
 import type { ClawdbotConfig } from "../../config/config.js";
 import { enqueueSystemEvent, resetSystemEventsForTest } from "../../infra/system-events.js";
-import { prependSystemEvents } from "./session-updates.js";
+import { incrementCompactionCount, prependSystemEvents } from "./session-updates.js";
 
 describe("prependSystemEvents", () => {
   it("adds a local timestamp to queued system events by default", async () => {
@@ -29,3 +29,37 @@ describe("prependSystemEvents", () => {
     vi.useRealTimers();
   });
 });
+
+describe("incrementCompactionCount", () => {
+  it("updates cached total tokens after compaction without clearing input/output", async () => {
+    const sessionKey = "agent:main:main";
+    const sessionStore = {
+      [sessionKey]: {
+        sessionId: "s1",
+        updatedAt: 10,
+        compactionCount: 1,
+        totalTokens: 9_000,
+        inputTokens: 111,
+        outputTokens: 222,
+      },
+    };
+    const now = 1234;
+
+    const nextCount = await incrementCompactionCount({
+      sessionEntry: sessionStore[sessionKey],
+      sessionStore,
+      sessionKey,
+      now,
+      tokensAfter: 2_000,
+    });
+
+    expect(nextCount).toBe(2);
+    expect(sessionStore[sessionKey]).toMatchObject({
+      compactionCount: 2,
+      totalTokens: 2_000,
+      inputTokens: 111,
+      outputTokens: 222,
+      updatedAt: now,
+    });
+  });
+});

src/auto-reply/reply/session-updates.ts

@@ -237,7 +237,7 @@ export async function incrementCompactionCount(params: {
   sessionKey?: string;
   storePath?: string;
   now?: number;
-  /** Token count after compaction - if provided, updates session token counts */
+  /** Token count after compaction - if provided, updates cached context usage */
   tokensAfter?: number;
 }): Promise<number | undefined> {
   const {
@@ -252,17 +252,14 @@ export async function incrementCompactionCount(params: {
   const entry = sessionStore[sessionKey] ?? sessionEntry;
   if (!entry) return undefined;
   const nextCount = (entry.compactionCount ?? 0) + 1;
-  // Build update payload with compaction count and optionally updated token counts
+  // Build update payload with compaction count and optionally updated context usage.
   const updates: Partial<SessionEntry> = {
     compactionCount: nextCount,
     updatedAt: now,
   };
-  // If tokensAfter is provided, update the cached token counts to reflect post-compaction state
+  // If tokensAfter is provided, update the cached total to reflect post-compaction context size.
   if (tokensAfter != null && tokensAfter > 0) {
     updates.totalTokens = tokensAfter;
-    // Clear input/output breakdown since we only have the total estimate after compaction
-    updates.inputTokens = undefined;
-    updates.outputTokens = undefined;
   }
   sessionStore[sessionKey] = {
     ...entry,