@aprimediet/permission-modes

permission-modes

A Claude-Code-style Shift+Tab mode cycle for the pi coding agent.

Three modes, cycled with Shift+Tab (or via commands):

Mode	Edit/Write	Mutating bash	Read-only bash	Reads	Behavior
default	prompt each call	prompt	allow	allow	Standard "ask before mutating" workflow
plan	disabled	blocked	allow	allow	Read-only exploration; produce a Plan: and execute
auto	auto-approve	auto-approve	auto-approve	allow	Auto-approve everything; auto-follow-up until done

Cycle order: default → plan → auto → default.

Install

pi install npm:permission-modes

Or load directly for a one-off run:

pi -e ./index.ts

Usage

Keyboard

• Shift+Tab — cycle to the next mode (default → plan → auto → default)

Commands

• /default — switch to default mode
• /plan — switch to plan mode (read-only)
• /auto — switch to auto mode
• /mode [name] — with no arg, opens a selector; with default/plan/auto, switches directly
• /auto-depth <n> — set the auto-follow-up cap (0 = unlimited, default 20)
• /done — stop auto-follow-up immediately. No more "Continue" will be sent in this session until you manually switch back to auto mode. Useful when you're satisfied with the result and want pi to stop.

Flag

• --permission-mode <mode> — start in default, plan, or auto

The flag is named --permission-mode (not --mode) because pi has a built-in --mode flag for output mode (text/json/rpc).

Plan mode flow

1. Run /plan (or use Shift+Tab to cycle into plan mode).
2. The agent reads your code, asks clarifying questions, and produces a numbered Plan: block.
3. pi prompts you with three options:
   • Execute the plan — switches out of plan mode, restores edit/write tools, posts the steps to the agent, and tracks progress with a ☐/☑ widget. The agent emits [DONE:n] tags as it finishes each step; pi marks them and posts Plan Complete ✓ when all are done. If the plan has more than 2 steps, pi also exposes the todo tool and tells the agent to use it for step tracking.
   • Stay in plan mode — keep exploring.
   • Refine the plan — open an editor, type feedback, and the agent iterates.

Auto mode

In auto mode, every tool call is auto-approved. After each turn, if the assistant's last message made tool calls and does not look like a completion signal, pi auto-feeds a "Continue. Auto mode is active — proceed without asking." follow-up message. This loops until:

• the agent stops making tool calls (and emits a completion signal like "all done" / "task complete" / "finished" / "all tests pass" / "fixed the issue"), or
• the cap (/auto-depth <n>, default 20) is reached, or
• the user runs /done, or
• the user manually changes mode.

Completion detection supports Indonesian phrases like "selesai", "sudah selesai", "beres". If pi keeps going when you're done, run /done to stop it.

Auto mode is intentionally permissive: it auto-approves everything, including outside-cwd writes. That's the design.

Headless / no-UI safety

In any mode where a UI prompt would normally appear, if ctx.hasUI === false (e.g. pi -p or RPC mode without a terminal), the tool call is blocked with a clear reason — never silently allowed.

State persistence

The current mode and autoFollowUpDepth are persisted to the session via pi.appendEntry("modes", {...}). On /reload, /resume, or tree navigation, the last entry is restored (with legacy value mapping: "ask"/"normal" → default, "accept-edits" → auto).

If you were mid-plan-execution, the latest modes-execute marker is detected and the plan is rebuilt from assistant messages after it, with [DONE:n] tags re-applied so the widget reflects the actual state.

Files

permission-modes/
├── package.json
├── README.md
├── index.ts          # main factory
├── utils.ts          # isSafeCommand, plan extraction, completion signal, formatCount
├── vitest.config.ts
├── utils.test.ts
└── index.test.ts

Tests

npm test

utils.test.ts covers the bash safety classifier, plan extraction, [DONE:n] marking, and completion-signal detection.