You Approved Something Else

we're working on removing Haiku from WebFetch now that automode is default

Claude Haiku is my current least favorite model - it hallucinates wildly, and is out-performed now by other similarly priced models like GPT-5.6-Luna Even worse: it seems to still be used by the Claude Code WebFetch tool, which means hallucination risk any time you fetch a URL!

what that leaves out though is that the boundaries you state in chat aren’t stored as rules. the classifier re-reads them from the transcript on every check, so if compaction drops the message where you said “don’t push”, the boundary goes with it. silently.

dsp just doesn't check at all. auto mode blocks destructive actions you didn't ask Claude to take (eg rm -rf can be fine but depends on your task) and screens tool results for prompt injection. but for most tasks you still won't see permission prompts, same as dsp!

docs recommend a deny rule for a hard guarantee, which means the conversational layer is the soft one. most people will assume the opposite.

👀 auto mode is safer than --dangerously-skip-permissions *and* manually approving everything

icymi we recently changed how /fork works in Claude Code, it now copies your session into a new background session while you keep working in the current one the old behavior (an in-session subagent that reports back) is now available as /subtask!