Edit Modes & Autonomy
Two Independent Axes
Composez has two orthogonal settings that control its behavior:
- Edit mode — what the AI does (query content, edit files, replace a selection)
- Autonomy level — how the AI is orchestrated (single turn, two-phase, multi-step agent)
Any edit mode works with any autonomy level. They’re completely independent.
Edit Modes
| Mode | What it does | Command |
|---|---|---|
| Query | Answers questions without changing files | /query |
| Edit | Makes changes to manuscript files | /edit or just chat |
| Selection | Replaces a specific text range | /select to set, then chat |
Switch modes with /chat-mode:
> /chat-mode query # Switch to query mode
> /chat-mode edit # Switch to edit mode
Autonomy Levels
| Level | How it works | Command |
|---|---|---|
| Direct | Single turn — one prompt, one response (the default) | /chat-mode direct |
| Compose | Two-phase — a planning model designs changes, an editor model applies them | /compose |
| Agent | Multi-step — plans and executes a sequence of slash commands | /agent |
Direct (default)
You send a message, the AI responds once, and any edits are applied. This is the standard mode for most writing and editing work.
Compose
Uses two models in sequence:
- A planning model (the main model) analyzes your request and writes a detailed plan
- An editor model turns the plan into specific file edits
This is useful when working with models that are strong at reasoning but less precise at producing file edits. Composez picks a suitable editor model by default, or you can choose one with --editor-model.
Enter compose mode with /compose or use it one-shot:
> /compose Restructure the dialogue in scene 3 to reveal the secret gradually
Agent
The AI creates a structured YAML plan of slash commands and executes them step by step. This is ideal for complex, multi-step tasks that span multiple scenes or files.
> /agent Rewrite act 1 chapter 1 in first person POV
See Agent Mode for the full guide on plans, parallel execution, and review loops.
Combining Modes
Every edit mode works with every autonomy level. The table below shows what happens for each combination:
| Direct | Compose | Agent | |
|---|---|---|---|
| Query | AI answers directly in one turn. | Planning model drafts an answer, then a second pass refines it. No files edited. | AI creates a multi-step plan to research and answer the question. |
| Edit | AI edits files in one step. | Planning model designs changes, then an editor model implements them. | AI creates a multi-step plan of editing commands executed automatically. |
| Selection | AI replaces the selected text in one turn. | Planning model designs the replacement, then a second pass applies it. | AI creates a multi-step plan that may include research before replacing the selection. |
Legacy names
The old mode names still work:
/architectis equivalent to/compose/codeis equivalent to/edit
Switching
Use /chat-mode to switch either axis independently:
> /chat-mode query # Change edit mode to query (keeps current autonomy)
> /chat-mode compose # Change autonomy to compose (keeps current edit mode)
> /chat-mode edit # Change edit mode to edit
> /chat-mode direct # Change autonomy back to direct
Or use the shorthand commands:
> /query # Switch to query mode
> /compose # Switch to compose autonomy
> /agent # Switch to agent autonomy
The prompt indicator shows your current mode:
> normal prompt (edit + direct)
compose> compose mode active
query> query mode active