Your first prompt

This page tells how to send a prompt, what you see while the agent works, and how to answer it. It starts after Install: the daemon runs, you logged in, and zsh loaded the plugin.

Send a prompt

Type a comma, a space and your prompt:

~/p/app $ , why does cargo test fail?

The prompt goes to the agent with the context of your terminal:

• the current directory and the terminal,
• the last command and its exit status,
• the recent commands of this terminal (8 by default).

context.recent_commands in config.toml sets how many recent commands the model sees, and 0 sends none. A command line that starts with a space never goes to the model.

zsh never reads a prompt as shell syntax. Quotes, * and $ stay as you typed them, and history keeps the line as you typed it. A prompt of several lines stays one prompt. A paste of several lines that all start with , sends one prompt for each line. The first runs, and the others wait behind it in their order. A paste with any other line stays one prompt.

What you see

The reply goes into your scrollback, like the output of any command. When the turn ends, your shell prompt comes back at once.

The first line names the scope of the prompt. In a registered project it is the name of the project, such as app. Outside projects it is ○ no project · ~/notes. When the conversation changes, the first line tells you that in place of the scope (see Conversations).

While the turn runs, the last row shows a spinner, what the turn does, its time and how full the model's context is:

⠴ writing  12s  ctx 43%

A command of the agent shows on one line with its time, and its last three output lines below it. When the command ends, one line stays:

$ cargo test -p app  exit 101  6.2s

The turn ends with one dim line:

done in 42s, ctx 43% (89k/206k), 1.1k out, cache 91%

cache 91% is the part of the input of the turn that came from the provider's prompt cache.

Answer a question

The default permission mode is cautious. In this mode, the agent can read files outside your secrets and run read-only commands. It writes freely in $SCRATCH, its own scratch directory, and in a registered project. Other actions ask you first.

When the agent asks for approval, press y to allow or n to deny.

When a command of the agent asks for a password or a [Y/n] answer, type it and press Enter. A password does not show. The agent sees it only if the program that asked prints it. Read which command asks before you answer. The hidden shell tells more.

Register a project

A registered project is a directory where the agent can write freely in cautious, and where the auto sandbox can write. efr never registers a directory by itself. In a repository, run:

efr project add

This registers the git work tree that holds the current directory, or the directory itself. efr project add PATH registers PATH, efr project list shows the projects and efr project remove PATH removes one.

A prompt in a git repository that is not registered shows a hint, at most one time each day for each repository. conversation.repo_hint = false turns it off:

○ ~/p/tool is a git repo but not a project: writes there ask you. efr project add ~/p/tool registers it.

Permission modes tells what each mode allows.

Work with a running turn

While a turn runs, your keys go to an input row below it (› ):

When you send a prompt while a turn runs, the prompt waits for that turn. The turn view and its keys lists all the keys.

Sticky agent mode

Type , alone, or press Ctrl+Space, to switch sticky agent mode on. A robot then stands before the text that you type, and every line goes to the agent. In a registered project, the project name stands after the robot. Outside projects, ○ stands there:

🤖 app › fix the failing test
🤖 ○ › check the disk space

In sticky mode:

• !<command> runs one shell command.
• A line that starts with , runs that , command.
• , alone switches sticky mode off.

When you press Enter, the robot becomes the first word of the line. History keeps 🤖 <prompt>, and the line goes to the agent again when you recall it.

The scope in your own prompt

efr never changes your prompt. To show the scope of the next prompt in your own prompt, use the function efr_prompt_info. It prints app or ○:

setopt prompt_subst
RPROMPT='$(efr_prompt_info)'

The variable $efr_scope holds the same text, for a prompt with no subshell: RPROMPT='${efr_scope}'. Both write each % of a project's name as %%, and each ! as !! when you set prompt_bang, so the prompt shows the name as you typed it. EFR_SCOPE_STYLE sets the style of the scope in sticky mode (fg=8 by default). Before efrd runs for the first time, there is no scope.

Commands of the plugin

A prompt of one command word alone runs that command. So , resume runs ,resume. , resume the build stays a prompt. mode, model and effort can also have one value that efr accepts: , mode auto runs ,mode auto.

A ,word that is not a plugin command, and that zsh cannot run, is a prompt that starts with the word. ,run make sends run make. Your own ,word command, function or alias still runs. A typo of a plugin command, such as ,moed auto, stays on the line with a hint. It does not go to the model.

Conversations

A terminal keeps one conversation for each registered project that you worked in, and one conversation for all places outside registered projects. A prompt goes to the conversation of the scope where you send it. The conversation changes when you send a prompt, not when you cd. The first line of the reply tells you when it changes:

↪ back to "fix the tests" in app (2 h ago)

Conversations, resume and search tells more.

Send a prompt from a script

The plugin runs efr for you. You can also run it yourself:

efr send "why does cargo test fail?"
efr new "start a new conversation with this prompt"

Both take --mode, --model and --effort. efr send --steer <text> steers the running turn. efr --help lists all the commands.

When the first prompt fails

• unauthorized. The provider of [model] has no login. efr names the login command, such as efr login openai.
• Exit code 3. No daemon runs. Start it with systemctl --user start efrd.
• The provider refuses the model. efr says so. Choose another model with ,model <model> for this terminal, or with efr config set model.name <model>.

Troubleshooting tells more.