Tips & Tricks

Input editing

• Multi-line input: press Shift+Enter, Alt+Enter or Ctrl+J to insert a newline without submitting. Useful for writing code snippets or structured prompts.
• Undo/redo: Ctrl+Z or Ctrl+_ undoes the last edit. Ctrl+Shift+Z redoes. Rapid typing coalesces into one undo entry; movement and kill/yank commands separate edits.
• History recall: Ctrl+P / Ctrl+N browse previous / next prompts, restoring your unsent draft when you return past the newest entry.
• History search: Ctrl+R searches backward through input history. Ctrl+S searches forward within that search. Type to filter, Enter to insert.
• Word navigation: Alt+B / Alt+F or Alt+Left / Alt+Right jump word boundaries. Alt+Backspace / Alt+D kill a word backward / forward. Ctrl+W kills backward using whitespace boundaries, useful for paths.
• Line editing: Ctrl+A / Ctrl+E jump to start / end. Ctrl+K kills to this line’s end. Ctrl+U kills to this line’s beginning, preserving other lines. Ctrl+T transposes the characters around the cursor.
• Kill/yank: Ctrl+Y inserts the most recently killed text. Alt+Y immediately afterward replaces it with an older kill. The ring is shared across text fields, while undo remains local to each field.

Chat interaction

• Copy text: Ctrl+Shift+C copies selected text to the clipboard. Click and drag to select (when mouse capture is on, toggle with /mouse). Copied text omits terminal padding and wrap breaks, preserving paragraph breaks and code indentation.
• Scroll: mouse wheel, PageUp/PageDown (10 lines), or click the scrollbar. Ctrl+Home / Ctrl+End jump to top / bottom. Ctrl+L re-attaches auto-scroll to the bottom.
• Cancel streaming: press Esc twice. The first Esc shows a hint in the status bar (“esc again to stop agent”). The second cancels immediately.
• Toggle reasoning: Ctrl+Alt+T expands/collapses reasoning/thinking blocks. Ctrl+O expands/collapses bash output.

Session management

• Open here: Ctrl+Alt+O lists conversations open in this terminal window. Tab opens global history. Attach another conversation to keep both open; Ctrl+Alt+Left / Ctrl+Alt+Right switch between them. Ctrl+Alt+Enter starts one in the current project.
• Close a view: Ctrl+Alt+W closes the current view, or the highlighted view in a conversation picker. Its draft and queued follow-ups remain available during this TUI process, and daemon work keeps running. Closing the last view exits the TUI.
• Notifications: the bell only includes conversations open here. Questions, approvals and failures need attention; completed results stay unread until visible in a focused terminal. Without terminal focus reporting, interacting with the visible conversation acknowledges its result.
• History: /sessions lists previous sessions with timestamps. /resume <id> resumes one. See Sessions.
• Rewind: /rewind <n> truncates to the first N messages. Use /rewind with no args to see a list with snippets.
• Clear: /clear wipes the conversation context (persists to disk). /compact forces context compaction on the next turn.

Clearing preserves permission grants. If you approved a tool for the session, that approval survives /clear.

Model switching

• Quick switch: /model deepseek-v4-flash switches without leaving the keyboard. /model alone opens the picker.
• Thinking variants: /thinking high sets the reasoning effort. /thinking off disables it. Or use Ctrl+Space and select “Thinking Variant”, then press Ctrl+Alt+P repeatedly to cycle through options.
• Cost tracking: /cost shows accumulated token counts and estimated cost for the session.

@-mentions

• Type @ in the input to open the reference picker. It lists files, models, skills, and subagents in one list; typing after the @ narrows the same list. Selecting a file inserts an @path mention that gets resolved and added to the agent’s context before the prompt is sent. Selecting a model, skill, or subagent inserts the matching @model:, @skill:, or @subagent: reference.

Terminal title

The terminal title names the current conversation and project, for example mew · running · Native browser lifecycle · ~/code/mew. It shows approval, question, failure, unread-result and connection states, with counts for background conversations open in this window. Logs are redirected to /tmp/mew-<pid>.log so they don’t corrupt the TUI display.

Prompting patterns

Be specific about scope

The agent has filesystem access and tools. Tell it what to focus on:

Fix the off-by-one in the pagination logic in src/api/list.rs.
Don't touch the database layer.

Ask for investigation first

For unfamiliar code, ask the agent to investigate before making changes:

Find where the session timeout is configured and explain how it works.
Don't change anything yet.

Use subagents for heavy research

If a task involves scanning many files, delegate it so the results don’t fill your main context:

Use the researcher subagent to find all places we construct SQL queries
and report back which ones use parameterized queries.

See Subagents for the built-in options.

Break work into phases

Use the planner/builder workflow for complex tasks:

1. /persona planner to investigate and write a plan
2. Review the plan in PLAN.md
3. /persona builder to execute it

See Personas for details.

Permission workflow

If you’re doing repetitive work that needs many tool approvals:

1. Start in standard mode to review each tool call.
2. Once you trust the pattern, switch to permissive with /permissions permissive to auto-allow writes and edits.
3. For full automation, use /permissions auto to let the classifier decide each call.

Switch back to standard when you’re done:

/permissions standard

See Permissions for the full mode reference.