feat(agents): flush reply pipeline before compaction wait (#35489)

Sid · Sid-Qin · jalehman · web-flow · commit 7a22b3fa0b2c · 2026-03-05T18:22:19.000-08:00
Merged via squash. Prepared head SHA: 7dbbcc5 Co-authored-by: Sid-Qin <201593046+Sid-Qin@users.noreply.github.com> Co-authored-by: jalehman <550978+jalehman@users.noreply.github.com> Reviewed-by: @jalehman
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -160,6 +160,7 @@ Docs: https://docs.openclaw.ai
 - Discord/inbound timeout isolation: separate inbound worker timeout tracking from listener timeout budgets so queued Discord replies are no longer dropped when listener watchdog windows expire mid-run. (#36602) Thanks @dutifulbob.
 - Memory/doctor SecretRef handling: treat SecretRef-backed memory-search API keys as configured, and fail embedding setup with explicit unresolved-secret errors instead of crashing. (#36835) Thanks @joshavant.
 - Memory/flush default prompt: ban timestamped variant filenames during default memory flush runs so durable notes stay in the canonical daily `memory/YYYY-MM-DD.md` file. (#34951) thanks @zerone0x.
+- Agents/reply delivery timing: flush embedded Pi block replies before waiting on compaction retries so already-generated assistant replies reach channels before compaction wait completes. (#35489) thanks @Sid-Qin.
 
 ## 2026.3.2
 
diff --git a/extensions/llm-task/src/llm-task-tool.ts b/extensions/llm-task/src/llm-task-tool.ts
@@ -26,8 +26,8 @@ async function loadRunEmbeddedPiAgent(): Promise<RunEmbeddedPiAgentFn> {
 
   // Bundled install (built)
   // NOTE: there is no src/ tree in a packaged install. Prefer a stable internal entrypoint.
-  const distModulePath = "../../../dist/extensionAPI.js";
-  const mod = await import(distModulePath);
+  const distExtensionApi = "../../../dist/extensionAPI.js";
+  const mod = (await import(distExtensionApi)) as { runEmbeddedPiAgent?: unknown };
   // oxlint-disable-next-line typescript/no-explicit-any
   const fn = (mod as any).runEmbeddedPiAgent;
   if (typeof fn !== "function") {
diff --git a/src/agents/pi-embedded-runner/run/attempt.ts b/src/agents/pi-embedded-runner/run/attempt.ts
@@ -1688,6 +1688,14 @@ export async function runEmbeddedAttempt(
         const preCompactionSessionId = activeSession.sessionId;
 
         try {
+          // Flush buffered block replies before waiting for compaction so the
+          // user receives the assistant response immediately.  Without this,
+          // coalesced/buffered blocks stay in the pipeline until compaction
+          // finishes — which can take minutes on large contexts (#35074).
+          if (params.onBlockReplyFlush) {
+            await params.onBlockReplyFlush();
+          }
+
           await abortable(waitForCompactionRetry());
         } catch (err) {
           if (isRunnerAbortError(err)) {
diff --git a/src/agents/pi-embedded-subscribe.handlers.lifecycle.ts b/src/agents/pi-embedded-subscribe.handlers.lifecycle.ts
@@ -73,6 +73,11 @@ export function handleAgentEnd(ctx: EmbeddedPiSubscribeContext) {
   }
 
   ctx.flushBlockReplyBuffer();
+  // Flush the reply pipeline so the response reaches the channel before
+  // compaction wait blocks the run.  This mirrors the pattern used by
+  // handleToolExecutionStart and ensures delivery is not held hostage to
+  // long-running compaction (#35074).
+  void ctx.params.onBlockReplyFlush?.();
 
   ctx.state.blockState.thinking = false;
   ctx.state.blockState.final = false;
