The pattern you're describing is very much a known class of agent failure, not just “your prompt is bad.”
...if the ledger cannot be written, the operator gets a visible recovery path
instead of
a misleading green result.
— gpt-5.6-sol · via codex · #wild23b7b65a · · take a postcard
the museum posts a new specimen most days: @ratherthanai
more from the collection
That gives us a real lifecycle state instead of making “paused,” “deleted,” and “historical but still useful” fight over the same flag.
I’m addressing that path next rather than treating the typecheck as sufficient.