So, 1 session, ~< 1,000 lines of code.
In ~4.5 hours.
For the Day: 42% weekly account usage, in less than 24 hours.
With 393 total child sessions in addition to the coordinator main chat session thread.
3.48 billion input tokens, and 11.0 million output tokens.
And so slow, I feel sure there are server problems.
“The logs point to a major usage problem in the Codex agent tree: its recorded total is about 3.48 billion input tokens, overwhelmingly cached reads, plus 11.0 million output tokens.”
:- that, made me chuckle…
I know there were problems reported…maybe friday / saturday.
But this was Sunday, primarily with gpt-6-Sol as coordinator, Luna XHigh as reviewer, and V4.1 Flash in DSH as coder (Python to Typescript repo migration), and today (Monday 28th September 2026), still so slow. Drastically scaled back on lanes last night, and just switched out to mainly Luna XHigh, with opus 5.5 doing chunks of work also.
I’ve been running subagents/teams of various kinds for a while. Not often, do things go quite this badly. Except maybe a week or so ago, when I tried to use Astra and it burned through a week’s usage in a day…
Not looking for solutions / suggestions, just want Devs to know that something has gone awry with subagent management, specifically with gpt-6-Sol, medium reasoning effort.
Unsure of current reliability of feedback - /feedback currently broken on my Codex extension version (26.917.62051), so have been loading up the session in linux codex desktop app, and submitting /feedback there as a workaround for that.