Opus 5.5 keeps editing files with bash, Armin Ronacher says
Two engineers agree on the behavior and disagree on whether to push back against it.
Armin Ronacher says his current experience with Opus 5.5 is the model reaching for bash to edit files rather than using the dedicated edit tools. He was relaying a suggestion from Mario Zechner, who proposed writing an extension that inserts the instruction "mofo, don't use bash to edit files" into the context. The sad face in Ronacher's post does the rest of the work.
Zechner's follow-up is the more interesting half. He also said "don't fight the model, give in to its RL, let if flow over you", and he means it as a working position rather than a shrug. He says he has not cared about edit tool calls in a long while, because he reads diffs in VS Code either while the agent is running or after it finishes, which also gets him code navigation and the rest of the editor.
don't fight the model, give in to its RL, let if flow over you
He adds a behavioral difference between the models. GPT will use tools instead of bash if you tell it to. Claude, he says, just ignores you now when it has access to bash.
Why the shared tool set mattered
Zechner's explanation is that a year back, models across labs were apparently trained on a similar tool set, and that is no longer true. By his account GPT was one of the first to go all in on bash for everything, and Claude has now followed. He says this "give in to the RL" reasoning was one of the reasons behind pi's minimal tool set in the first place.
For anyone building a harness, that is the load bearing claim. A custom edit tool only works if the model was rewarded for calling custom edit tools during training. If the reinforcement learning pushed it toward shell commands instead, the elaborate tool surface becomes something the model routes around, and a prompt telling it to behave is a patch on a training signal.
Ronacher takes the disagreement one step further in a separate post, asking whether anyone still cares about the tool calls in the transcript at all. If they are getting worse and worse, he suggests, maybe they should just be compressed. He says he wonders what people expect at this point, which is not the same as knowing.