From 182c09b80bdd71e2dd21d8421a158ffced8f3bc6 Mon Sep 17 00:00:00 2001
From: webtecnica <75556242+webtecnica@users.noreply.github.com>
Date: Wed, 22 Jul 2026 16:59:29 -0300
Subject: [PATCH] fix(agent): prevent agent hang after tool call completes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

After a tool call completes, the conversation loop does  to
start the next API call. Between the last  (from tool
completion in tool_executor.py) and the next one (at the start of the
next API call), there is a gap. If anything during this gap takes time
— context compression, slow provider prefill, or other post-tool
processing — the combined inactivity window can exceed the gateway's
inactivity_timeout (default 120s). The gateway kills the session and the
user sees 'agent never returns a final response', even though the tool
call itself succeeded in 0.1-0.2s.

Fix: call  right before  so the gateway
sees a fresh timestamp immediately after tool results are posted,
regardless of how long the follow-up API call takes.

Fixes #69559
---
 agent/conversation_loop.py | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py
index a234213e32ca..2f452cadb49b 100644
--- a/agent/conversation_loop.py
+++ b/agent/conversation_loop.py
@@ -5865,6 +5865,17 @@ def run_conversation(
                 # Save session log incrementally (so progress is visible even if interrupted)
                 agent._session_messages = messages
                 
+                # Touch activity before continuing so the gateway's
+                # inactivity monitor never sees a stale timestamp
+                # between tool completion and the start of the next
+                # API call.  Without this, a tool-call result (which
+                # takes ~0s to process) followed by slow post-tool
+                # processing (compression, persist) and a slow
+                # follow-up API call can exceed the gateway inactivity
+                # timeout (HERMES_AGENT_TIMEOUT, default 1800s) and the
+                # gateway kills the session before the next activity
+                # touch fires (#69559, #69131).
+                agent._touch_activity(f"tool results posted, continuing iteration #{api_call_count}")
                 # Continue loop for next response
                 continue