Feedback: YOLO mode is more practical than "Ask When Needed" for long-running agentic sessions

Summary

After using both permission modes extensively in daily research-computing work (SLURM cluster, long agentic sessions with many sub-agent dispatches), I find YOLO mode significantly more usable than "Ask When Needed"(new yolo) mode, to the point that I downgraded from 0.40.1 back to 0.39.1 to keep my preferred workflow. I would like to share the reasoning and suggest improvements.

Why YOLO works better in practice

1. Predictable autonomy: In YOLO I know exactly what happens — regular tool calls proceed without interruption, and I stay focused on reviewing outcomes instead of approving individual calls. In "Ask When Needed" mode the interruption points are less predictable, which breaks flow during long multi-step tasks.
2. Sub-agent heavy workflows: My setup delegates most implementation work to sub-agents (via Agent/AgentSwarm). Approval prompts multiplied across parallel sub-agent runs make "Ask When Needed" considerably noisier than YOLO.
3. Session continuity: I often leave sessions running (tmux + cron-triggered checks). YOLO lets these complete unattended; approval-gated modes stall on prompts nobody is there to answer.

Suggestions

• Keep both modes first-class: please don't deprecate or gate YOLO behind extra friction in future releases.
• Consider making "Ask When Needed" interruption criteria more transparent/configurable (e.g. an allowlist-first behavior where approved-once tool+pattern combos stop prompting — this already partially exists, but its scope could be broader and more visible in the UI).
• A per-session summary of what would have been asked in YOLO mode (a dry-run approval log) would give the best of both worlds: full autonomy with an audit trail.

Environment

• kimi-code 0.40.1 (linux-x64, install-script installation)
• Use case: computational materials research on an HPC cluster; long-running sessions with heavy sub-agent delegation

No files, tests, or implementation entry points are named. Start by tracing how YOLO and Ask When Needed handle approval prompts, including Agent/AgentSwarm sessions; done would require a maintainer-defined change covering interruption behavior, configuration visibility, or audit logging.