the-tui.md

The terminal UI

Run deepcode in your project folder to work from the terminal. Type a task in plain language, then use the commands below to choose a model, attach files, and control the conversation. New to DeepCode? Follow Your first coding task for setup.

Keys

Key	What it does
Esc	With an empty prompt, interrupt the running Turn: the Turn stops and the conversation survives. With a completion menu open, close the menu. While you are typing a message, Esc does nothing, so your draft is kept.
Ctrl+C	Clear the message you are typing. With an empty prompt, interrupt the running Turn, like Esc. While a command such as /compact or /goal wait is running, stop waiting for that command.
Ctrl+O	Cycle transcript detail: normal → verbose → summary. normal shows each tool's result in one line: a command's first and last lines, another tool's first line. verbose also shows the start and end of what came back, and the model's full reasoning. summary keeps only the answers. Applies to what comes next; while a Turn runs or an approval waits, the status line confirms the chosen mode for a few seconds.
Ctrl+D	Close the client (same as /exit); running work continues in the background.

Slash commands

Type / to list commands. /model and /resume open pickers when called without arguments: use arrow keys, type to filter, and press Enter to choose. Other commands may show their current setting or require an argument; /queue and /rename, for example, require text.

Command	Does
/help	This table, in the terminal
/new [title]	Start a fresh conversation
/resume [id|all]	Pick a Session from this directory — or all directories
/rename <title>	Rename this Session
/delete <id>	Permanently delete a stored Session
/model [connection] [id]	Show or switch connection/model — the picker shows every configured connection's full catalog
/effort [auto|off|level]	Reasoning effort for future Turns
/context [auto|tokens]	Context-window cap for future Turns — 64k, 1m, or auto to follow the model's published window
/permissions [preset]	Tool access: ask · read-only · full-access · inherit
/transcript [mode]	Same three modes Ctrl+O cycles
/raw [on|off]	Print answers exactly as the model wrote them, without Markdown styling or wrapping, for copying — see How answers are shown
/preset [id|clear]	Agent presets — selectable only while the conversation is still blank
/skills	List every discovered skill
/skill <id|name>	Choose a Skill for your next message
/plugins	List installed plugins
/mcp [action]	List, add, test, authorize MCP servers
/goal …	A durable Goal for this Session — see Goals
/queue <instruction>	Save an instruction to run after the current task
/stop	Interrupt the active Turn (same as Esc)
/retry	Re-run the last finished Turn with the current model
/reconnect	Reconnect and return to your current conversation
/clear	Clear the conversation context
/compact	Summarize older Turns to free context — see Sessions
/exit	Quit

How answers are shown

Answers are rendered as Markdown while they stream. Each line appears once its last word has arrived, and lines already printed never change:

• Headings, lists, quotes and paragraphs are styled as soon as their line is complete. Links show their address, as text (https://…).
• Paragraphs wrap at spaces, and Chinese or Japanese text also wraps between characters, so lines stay full and English words inside stay whole. A wrapped list item or quote continues under its own text.
• Code blocks appear line by line with syntax highlighting. They have no indent and DeepCode never wraps them, so a copied block is exactly what the model wrote. Your terminal wraps lines that are too long for the window.
• A table appears once its last row has arrived, because a later row can widen the columns above it.
• A line that contains a URL is wrapped by your terminal instead of DeepCode, so the address is never split.

The line the model is still writing is shown under the transcript, above the status line, until it is complete.

To copy an answer verbatim, run /raw: answers are then printed exactly as the model wrote them, Markdown syntax included, and your terminal does all the wrapping. /raw again (or /raw off) switches back. The setting lasts until you close the TUI.

Inline references

@path — include a project file in your message. Type @ and use Tab-completion to find the file:

› the parser in @src/config/loader.py rejects valid YAML — why?

$skill-name — invoke a skill inline, with completion over active skills:

› $SKILL_NAME review the changes on this branch

Replace SKILL_NAME with a name shown by /skills.

Talking while it works

You don't have to wait for a Turn to finish:

• Type a message mid-Turn and it steers the running work — the agent sees it at the next boundary: Steered the active Turn.
• /queue lines up a full instruction to run after the current Turn.
• Esc (from an empty prompt) stops the Turn; what completed stays in history.

Approvals

Under the default ask preset, sensitive tools pause for you. The card shows the whole command and where it runs, or the change an edit would make:

◆ Approval needed · Run a command
  Running a command needs your approval.
    $ rm -rf build && echo removed
    in ~/projects/app
  y allow once · a always allow `rm …` and `echo …` this Session · n deny

While it waits, the status line stops and says ◆ Waiting for your approval, and the prompt reads allow? y/a/n ›. Anything else you type is kept in the prompt, not sent, until you answer.

y/yes allows this once. n/no denies — the agent is told, and works around it. a/always allows the same kind of action for the rest of the Session, and the card says which:

• Commands — the program and its subcommand: allowing git status allows git status -s but not git push; allowing ls does not allow rm. A command joined with &&, ; or | runs without asking only when every part was allowed. A command that writes a file with >, runs through sudo, env or xargs, or runs inline code (python -c) is allowed only exactly as written.
• File edits — edits anywhere in this workspace. An edit outside it is allowed for that file only.
• Other tools — that tool.

Switching to /permissions full-access asks you to confirm explicitly before it takes effect; read-only denies mutating actions under the tool policy. These are per-Session — your other Sessions keep their own settings.

What the status line tells you

The bottom line is alive during work — a spinner plus what the agent is on:

⠸ Run · 4s · pytest -q
⠼ Thinking · 8s · High · comparing the two configs
⠋ Working · 2s
⠙ Queued · 6s · waiting for a free slot (2 running)
◆ Waiting for your approval · y allow once · a always allow `ls …` this Session · n deny

Run shows the current tool or command. Thinking appears when the provider shares reasoning progress; Working means DeepCode is waiting for its response to begin; Responding means the answer is arriving, with its unfinished line shown just above. Queued says what your message waits for: the current Turn in this Session, another Session working in the same folder, or a free slot (the service runs two Turns at once). A message queued behind a running Turn shows as 1 queued next. While an approval waits, nothing runs and the line stands still; if it stays there unexpectedly long, see Troubleshooting. When idle, the line shows your display mode and keyboard hints.

A tool call you denied shows as Denied by you, and one a permission rule or a hook refused as ⊘ … declined with its reason.

Leaving with Ctrl+D while a task runs does not stop it: the service keeps running it, and the TUI prints the command to come back to it (deepcode --resume <session id>).