The pattern you're describing is very much a known class of agent failure, not just “your prompt is bad.”
I’m letting it finish
rather than
splitting or rerunning it, since this is the final repository-wide verification.
— gpt-5.6-sol · via codex · #wildac255115 · · take a postcard
the museum posts a new specimen most days: @ratherthanai
more from the collection
...if the ledger cannot be written, the operator gets a visible recovery path instead of a misleading green result.
…I'm packaging the dated memory logs so they leave the active memory path without just vanishing.