Agent mode connected to Github repo almost always fails (after 10 mins of analysis)...why?

Codex
,
1

I am getting this message very often when I ask a Agent mode question in the web interface ChatGPT.

About 75% of the time, the Agent mode prompt eventually fails with this: getNodeByIdOrMessageId - no node found by id: 68b6e69e-eb94-492e-84c8-d3f1240a07e4
Or “sorry something went wrong”.

My Github is connected fine, the Agent can see it, I can see it working and analyzing files in there. The failure is after 10 minutes or so.

I lost about 10+ Agent credits (I have a Plus a/c). I’ve lost so many Agent and Deep Research credits through the session crashing!

Win11, MS Edge, no VPN or weird browser extensions.

2

That failure shape feels like the run is dying after the long analysis phase, not during repo access itself.

The fix I’d want from the product side is pretty boring:

  1. separate analysis budget from execution budget
  2. checkpoint the plan before long repo scans
  3. stop retrying automatically when the error is unchanged
  4. give a typed receipt that says “analysis crashed” instead of just burning another credit

A lot of trust gets lost when the agent can clearly see the repo, spends 10 minutes thinking, and then fails with no durable state. The cost isn’t just credits, it’s losing the thread of the work.