Description

OpenHands v1.6.0 (March 30, 2026) added a Planning Mode where the agent generates a structured `PLAN.md` file outlining implementation steps before writing any code, and waits for explicit user approval before execution.

Zeph has `/plan ` with `confirm_before_execute = true` for multi-agent orchestration plans. The gap is at the single-agent level: when a user asks Zeph to implement something, it immediately starts writing code. There is no single-agent "propose a plan, wait for approval, then execute" mode.

Reference

• OpenHands v1.6.0: https://github.com/OpenHands/OpenHands/releases
• OpenHands blog (March 2026): https://openhands.dev/blog/openhands-product-update---march-2026

OpenHands Behavior

1. User sends a high-level request ("implement feature X")
2. Agent enters Plan Mode — generates a `PLAN.md` with enumerated steps
3. User reviews and approves/edits the plan
4. Agent switches to Code Mode and executes the plan step by step
5. A visible Task List panel shows current step progress

Zeph Gap

• Single-agent mode: no pre-execution planning gate
• Orchestration (`/plan`) has `confirm_before_execute`, but is multi-agent only
• There's no Task List panel in the TUI showing current plan step

Implementation Sketch

A new autonomy-level setting (e.g., `autonomy_level = "plan-first"`) that intercepts complex multi-tool requests and proposes a step list before executing any tools. Requires:

1. Plan generation prompt (classify whether request is "simple" or "plan-worthy")
2. Confirmation gate (show plan in CLI/TUI, wait for user Y/n)
3. Optional task list view in TUI showing step progress

Priority: P4. Addresses agentic UX quality, not functional correctness.