#Agent stuck "initializing" on gpt-5.4

1 messages · Page 1 of 1 (latest)

potent tendon
#

does anyone know what to do when an agent using gpt 5.4, regardless of trying /new or restarting gateway, just doesn't respond but hangs at "initializing agent..."? swapping models fixes it, works fine with Mimo for example. swap back to codex, back to just being stuck here. tried everything i could think of including just rebooting my machine... hermes update, hermes doctor... out of ideas now 🥲

was working fine last night, been like this since i tried using it again about 3 hours ago on telegram first.

Report https://paste.rs/zJRan
agent.log https://paste.rs/48WeY
gateway.log https://paste.rs/dWmMG

worth adding: i was working with another agent profile using Mimo on telegram earlier today, THAT agent was initially unresponsive after i granted it a permission. i /reset that one and /restart this particular profile (main), that one works now, but this one doesn't. so idk if this is because of something i did today or something else.

image.png
vital cloak
#

What your screenshot shows is not Hermes getting stuck in local startup. It is getting past startup and then sitting in the model call waiting for gpt-5.4 to produce its first response chunk.

The current logs do not look like a dead gateway or a permanently broken session. They do show a restart and one drain timeout earlier, but after that the gateway comes back cleanly and Codex-side auxiliary calls are still happening. That is why Mimo working and Codex failing points more to the Codex request path itself than to /new, /reset, or the gateway being wedged.

The first thing I would try is refreshing Codex auth completely, because Hermes has explicit recovery logic for stale Codex tokens and other clients can rotate those tokens underneath it:

  1. run codex once in your terminal
  2. then run hermes auth and re-authenticate OpenAI Codex
  3. retry gpt-5.4

If it still hangs after that, please reproduce it once and immediately send a fresh /debug or fresh agent.log snippet from that exact minute. The dump you attached is useful context, but it does not include a decisive failing main-request error for the stuck attempt itself, so the missing piece is the exact 401, 429, timeout, or transport error from the live repro.

Also include your hermes --version output when you resend it. There are newer timeout and Codex-related fixes on main than the build shown in your dump, so I want to make sure the update actually moved you onto the newer code and not just restarted the same older build.

potent tendon
#

yoo that worked, thanks! didn't know about re-authenticating like that... noted! ✍️ prayge