[Feature Request] Auto mode classifier too aggressive, blocks legitimate SSH operations and forces suboptimal fallback plans

Bug Description
auto mode classifier is too aggressive... and it forces the AI into bad decision making paths. for example, if ssh'ing in and diagnosing a machine is the right move, it might kill that plan and not ssh in. the result will be a suboptimal/worse plan of hallucination (guessing) the results and then, acting on that hallucination.

it should be tuned to "still stop and ask" when there's clear ambiguity between correctness and security implication. and to simply have "Broader approval categories". (ok to ssh in for diagnosis)

Environment Info

• Platform: linux
• Terminal: wsl-Ubuntu-22.04
• Version: 2.1.114
• Feedback ID: 3e9657bb-4798-4b47-a2bc-3b6dfaa8ef89

Errors

[{"error":"Error: Request was aborted.\n    at makeRequest (/$bunfs/root/src/entrypoints/cli.js:50:3448)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-04-18T10:03:47.236Z"}]