From 2c3c92c3b7c85042b448ba1a1b73e8ca0c3f0092 Mon Sep 17 00:00:00 2001
From: S'Bussiso Dube <80188685+Sbussiso@users.noreply.github.com>
Date: Wed, 9 Sep 2026 13:38:26 -0700
Subject: [PATCH] =?UTF-8?q?Keep=20the=20agent=20warm=20=E2=80=94=20Fly's?=
 =?UTF-8?q?=20proxy=20won't=20wait=20for=20it=20to=20boot?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The agent scaled to zero, and every wakeup against a stopped machine
failed with RemoteDisconnected. Fly's proxy waits ~8s for an
auto-started machine to bind its port; this process takes longer:

  before LiteLLM   ~7s   (fit, barely)
  after LiteLLM    ~16s  (broke)
  after deferring litellm's import  ~10s  (still broke)

So deferring the import was necessary but not sufficient, and the 7s
figure says this was always marginal — LiteLLM only exposed it. The
remaining time is Python plus the MCP SDK plus Sentry, and trimming that
further means tuning against a proxy timeout we do not control and can
only observe in production.

Keeping one 512MB machine warm costs about $2/month — noise against the
LLM spend of a single run — and removes cold starts from the wakeup path
instead of racing them. Scale-to-zero was the right shape when the agent
was its own app answering a public webhook; it is not worth this.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---
 fly.toml | 27 ++++++++++++++++++++-------
 1 file changed, 20 insertions(+), 7 deletions(-)

diff --git a/fly.toml b/fly.toml
index 595269e..e9f410d 100644
--- a/fly.toml
+++ b/fly.toml
@@ -95,18 +95,31 @@ primary_region = "sjc"
 # publicly: it is reachable at http://sentinel-command.flycast:8080 over
 # 6PN, which is where SENTINEL_AGENT_WEBHOOK_URL points.
 #
-# This service is what makes scale-to-zero work at all. Fly auto-starts a
-# stopped machine on a connection to one of its SERVICES; a bare 6PN
-# connection to a stopped machine just fails. So the wakeup webhook both
-# starts the machine and delivers the work.
-#
 # Requires a private IPv6 on the app: `fly ips allocate-v6 --private`.
+#
+# NOT scaled to zero, despite that being this worker's obvious shape, and
+# despite it being how the agent ran as its own app. Fly's proxy waits
+# only ~8s for a machine it auto-started to bind its port, and this
+# process cannot reliably beat that: Python plus the MCP SDK plus Sentry
+# took ~7s before LiteLLM and ~10s after — so an auto-started machine got
+# declared unreachable and Command Center's wakeup came back
+# RemoteDisconnected. Measured, twice, on 2026-09-09.
+#
+# The 7s figure means this was always marginal and happened to fit; the
+# LiteLLM move only exposed it. Deferring litellm's import cut boot from
+# 16s to 10s, which was necessary but not sufficient, and trimming
+# further imports would be tuning against a proxy limit we do not
+# control and cannot test except in production.
+#
+# So the agent stays warm. One shared-cpu-1x/512MB machine is ~$2/month,
+# which is noise against the LLM spend of a single run, and it removes
+# cold starts from the wakeup path entirely rather than racing them.
 [[services]]
   internal_port = 8080
   protocol = "tcp"
-  auto_stop_machines = "stop"
+  auto_stop_machines = "off"
   auto_start_machines = true
-  min_machines_running = 0
+  min_machines_running = 1
   processes = ["agent"]
 
   [[services.ports]]