[Feature Request] Plan Mode

[engels74]

Feature request: Plan Mode

Add a read-only "plan first, execute after approval" mode, like the one in Claude Code, Codex, and Gemini CLI.

What it is

Agent investigates and writes a plan. Can read, search, and run exploration commands but can't edit source. User reviews the plan, then approves and drops into normal execution. Standard hotkeys across all three tools: /plan and Shift+Tab.

Prior art

• Claude Code: one entry in the permission-mode system. Minimal. docs
• Codex: /plan toggle, recommended for ambiguous/complex tasks. docs
• Gemini CLI: most fleshed out. Read-only tool allowlist, plan written to .md on disk, Ctrl+X to edit it in $EDITOR, auto model routing (Pro for planning, Flash for execution), hooks, headless auto-approve. docs

Idea / MVP

1. /plan [goal] and Shift+Tab to enter
2. Tool allowlist: read-only fs, grep, web search, read-only MCP, memory read, ask_user. Writes only to the plan file.
3. Plan written to ~/.jcode/tmp/<project>/<session>/plans/*.md
4. Approve / iterate / cancel UX at the end
5. Ctrl+E opens the plan in $EDITOR, agent picks up edits on save

Why this fits jcode specifically

• Side panel: render the plan there with mermaid diagrams, instead of jamming it into chat scrollback like the others do
• Swarm: plan becomes the dispatch spec; each worker gets its slice
• Self-dev: natural default for self-modification given the README's own warning about weaker models breaking things
• Skills: planning-flavored skills auto-activate via the existing semantic loader

Open questions

• Default on or off? (Gemini: on. Others: off.)
• Hard-block all shell, or allowlist git status, cargo check, etc.?
• Auto-checkpoint on plan approval?
• Adding a "clear context, then execute plan" feature, like codex and claude have?