The part I expect people to push back on is materialising prose into real tool calls. It looks like exactly the kind of clever thing that bites you later, and I went back and forth on it for a while.
What changed my mind was watching a user hit "continue" three times and get three identical blocks of text with zero executions. The model had already chosen the right tool and written complete, correct arguments. The only thing wrong was which channel it came out of. Refusing to honour that was costing a real turn to protect a purity that bought nothing — the alternative wasn't "the model does it properly", it was "the user gives up".
The constraints that make me comfortable with it: only tool names we declared to the model on that same turn, placeholder arguments are treated as a plan and never materialised, first occurrence only, and approval gates apply exactly as they would through the normal channel. It changes the transport, not the policy.
The detection bug underneath it is the more broadly useful lesson, I think. Our honesty check fired on "zero tool calls AND empty body", which quietly assumes a model that did nothing will also say nothing. That assumption fails precisely on the models where you need the check. The expensive failures all came with a confident paragraph attached.
Happy to go into any of the shapes we've seen — the fenced bash variant was the nastiest, because the reply ended with a question and collected an exemption we'd added for legitimate clarifying questions.