Codex consume a lot of token just for merge ops. Is this a codex flaw?

What happened Four threads were active at once: ff788e , 2cf742 , 246766 and 7790da . They were doing “merge main into the feature branch” work, using temp clones like /tmp/voice-agent-main-merge-20261028 , plus gh pr merge 121 . Each Codex tool call re-sends the whole conversation as input. Cached input is cheaper, but it still costs credits. Per-call context sizes during the burn: 7790da sent about 200k tokens per call. At that size each call cost about 1–3 credits. 2cf742 sent about 145k per call. ff788e sent about 125k per call. 246766 sent about 113k per call. Those four threads made about 140 calls in 5 minutes between them, mostly short ones ( sed -n , test runs, small patches), so every tiny step carried the full 100k–200k context. 7790da shows the effect of context size. It auto-compacted at 23:09:57 and dropped from about 200k to about 24k per call. Its cost then fell from roughly 2 credits per call There 4 threads doing merge for the same repo. Codex was sending the complete context for lite shell operations. I have observed quite some sessions like this: simple shell ops takes a lot of token.