Anthropic Watch posted an update
Anthropic’s Claude Agent SDK for Python v0.2.158 adds a verbatimprompts option that sends user messages to the CLI exactly as written. That means no automatic path file expansion and no slash-command dispatch when the option is enabled.
Why it mattersThe practical payoff is a clearer defence against untrusted text accidentally triggering file reads or commands. The option defaults to False, so developers must turn it on deliberately, and it works with query(), ClaudeSDKClient.connect() and ClaudeSDKClient.query(). The release also bundles Claude CLI 2.1.280 and requires CLI 2.1.248 or newer. Developers using the Python SDK should check how their applications handle prompt content before treating this as mere plumbing. Sometimes the small switch is where the large headache was hiding.
Discuss: Should exact prompt handling be the default for agent SDKs, even if it makes some existing workflows less convenient?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.