//! `railway ca` — the cloud agent front door.
//!
//! Bare `railway ca` on a terminal opens the TUI; everything else is a
//! subcommand. `railway code` is the same launcher without the TUI, and both
//! read the same preferences file, so the choice between them is only whether
//! you want to browse first.
pub mod access;
pub mod lifecycle;
pub mod prefs;
pub mod setup;
pub mod skills_sync;
pub mod telemetry;
pub mod tui;
use std::future::Future;
use anyhow::{Context, Result};
use clap::Parser;
use colored::Colorize;
use crate::client::GQLClient;
use crate::commands::code::LaunchArgs;
use crate::config::Configs;
use crate::errors::RailwayError;
use crate::macros::is_stdout_terminal;
use crate::util::progress::create_spinner;
use prefs::AgentPrefs;
use tui::{App, Outcome};
/// Manage Railway cloud agents
#[derive(Parser)]
#[clap(
args_conflicts_with_subcommands = true,
after_help = "Examples:\n\n railway ca # browse and launch agents (TUI)\n railway ca manage # jump straight into the manage screen\n railway ca setup # choose your default agent and skills\n railway ca setup --show # print current preferences\n railway ca start --claude # skip the TUI and launch\n\n railway ca list # every agent you own, everywhere\n railway ca list -e production # just this environment\n railway ca create my-agent # a VM, without connecting to it\n railway ca ssh my-agent # connect to it (starts a session if none)\n railway ca ssh my-agent -- bash # a plain shell instead of the agent\n railway ca sleep my-agent # stop the compute bill, keep the disk\n railway ca sleep --all # every running agent you own\n railway ca delete my-agent # the agent and its disk\n\nAgents are addressed by name or id. With neither, commands use this\ndirectory's agent, or your only one, and otherwise list the candidates.\n\n`railway code` is the launcher on its own — same flags, same preferences, no\nTUI. Preferences live in ~/.railway/agent-prefs.json; a flag always wins over\nthem, and RAILWAY_CA_AGENT overrides the saved default for one run. A\ndirectory linked with `railway link` wins over the saved default project too\n— new agents land there instead.\n\nNote: requires the CLOUD_AGENTS feature to be enabled."
)]
pub struct Args {
#[clap(subcommand)]
command: Option<Command>,
/// Launch flags. Passing any of these skips the TUI and launches directly,
/// so `railway ca --claude` behaves like `railway code --claude`.
#[clap(flatten)]
launch: LaunchArgs,
}
#[derive(Parser)]
enum Command {
/// Configure how cloud agents are launched (default agent, skills)
Setup(setup::Args),
/// Open the TUI directly on the manage screen, skipping the menu
Manage,
/// Launch a coding agent on a cloud agent VM, without the TUI
Start(LaunchArgs),
/// List your cloud agents
#[clap(visible_alias = "ls")]
List(lifecycle::ListArgs),
/// Create a cloud agent VM, without connecting to it
#[clap(visible_alias = "new")]
Create(lifecycle::CreateArgs),
/// Connect to an existing cloud agent over SSH
#[clap(visible_alias = "connect")]
Ssh(lifecycle::SshArgs),
/// Wake a sleeping agent
Wake(lifecycle::WakeArgs),
/// Put an agent to sleep, keeping its disk and stopping the compute bill
Sleep(lifecycle::SleepArgs),
/// Delete an agent and everything on its disk
#[clap(visible_alias = "rm")]
Delete(lifecycle::DeleteArgs),
}
/// Time one lifecycle verb and report its outcome, passing the result through
/// unchanged.
///
/// At the dispatch rather than inside each verb because the shape is identical
/// for all of them. `ssh` is the exception and tracks itself: it ends in
/// `std::process::exit` to propagate a remote exit status, which would skip
/// anything wrapped around it.
async fn tracked(kind: &'static str, run: impl Future<Output = Result<()>>) -> Result<()> {
let started = std::time::Instant::now();
let result = run.await;
let message = result.as_ref().err().map(|e| format!("{e:#}"));
telemetry::track_lifecycle(kind, started.elapsed(), message.as_deref()).await;
result
}
pub async fn command(args: Args) -> Result<()> {
// `railway ca` is often the first Railway command someone runs, so a logged
// out user gets the login flow inline instead of an error telling them to
// run `railway login` and type this again. The command they asked for then
// continues on the credential that flow just wrote.
let interactive = is_stdout_terminal();
if needs_credential(args.command.as_ref(), interactive) {
ensure_logged_in(interactive).await?;
}
match args.command {
Some(Command::Setup(a)) => setup::command(a).await,
Some(Command::Manage) => browse_into(Some(tui::Screen::Manage)).await,
Some(Command::Start(a)) => crate::commands::code::launch(a).await,
Some(Command::List(a)) => tracked("list", lifecycle::list(a)).await,
Some(Command::Create(a)) => tracked("create", lifecycle::create(a)).await,
Some(Command::Ssh(a)) => lifecycle::ssh(a).await,
Some(Command::Wake(a)) => tracked("wake", lifecycle::wake(a)).await,
Some(Command::Sleep(a)) => tracked("sleep", lifecycle::sleep(a)).await,
Some(Command::Delete(a)) => tracked("delete", lifecycle::delete(a)).await,
None if args.launch.is_bare() && is_stdout_terminal() => browse().await,
// Flags given, or no terminal to draw on: behave like `railway code`.
// A TUI in a pipe would be gibberish, and erroring instead would break
// scripted callers that reasonably expect the launcher.
None => crate::commands::code::launch(args.launch).await,
}
}
/// Whether this invocation will talk to the API, and so needs a credential
/// before it starts. Only `setup` can get by without one — everything else
/// (the TUI, `start`, a bare launch flag) opens with a query.
fn needs_credential(command: Option<&Command>, interactive: bool) -> bool {
match command {
Some(Command::Setup(a)) => a.needs_credential(interactive),
_ => true,
}
}
/// Run the login flow in place when there is no credential to work with.
///
/// Any credential counts, including a project token: those callers are already
/// authenticated as far as this command is concerned, and `railway login`
/// short-circuits on `RAILWAY_TOKEN` anyway, so sending them there would be a
/// detour to nowhere. An expired login is not a case to handle here — `main`
/// refreshes and clears dead credentials before dispatch, so it reaches this
/// check as no credential at all.
async fn ensure_logged_in(interactive: bool) -> Result<()> {
if Configs::new()?.has_auth_credentials() {
return Ok(());
}
// Piped or scripted: the login flow would sit on a device code nobody is
// watching. Fail the way every other command does instead.
if !interactive {
return Err(RailwayError::Unauthorized.into());
}
println!("{}", "Log in to Railway to continue.".bold());
let result = crate::commands::login::prompt_login().await;
telemetry::track_login_forwarded(result.as_ref().err().map(|e| format!("{e:#}")).as_deref())
.await;
result?;
println!();
Ok(())
}
/// The TUI loop. `run` gives the terminal back whenever something needs the
/// whole screen; we do that thing and re-enter with the app state intact, which
/// is what makes connecting feel like stepping into a session and back out
/// rather than restarting the command.
async fn browse() -> Result<()> {
browse_into(None).await
}
/// Like [`browse`], but opens straight on `initial_screen` instead of the menu
/// when one is given. `railway ca manage` uses this to land on the manage
/// screen without a detour through the menu — an explicit ask, so it also
/// skips the first-run "set up cloud agents?" nudge that a bare `railway ca`
/// would show.
async fn browse_into(initial_screen: Option<tui::Screen>) -> Result<()> {
let mut configs = Configs::new()?;
let client = GQLClient::new_authorized(&configs)?;
let backboard = configs.get_backboard();
// Both under one spinner: the flag check is a small query against the same
// client, and running it beside the tree load rather than before it keeps
// the preflight off the clock for everyone who does have the flag.
let spinner = create_spinner("Loading your projects".to_string());
let loaded = tokio::try_join!(
access::ensure_enabled(&client, &configs),
tui::load_tree(&client, &configs),
);
spinner.finish_and_clear();
let (_, tree) = loaded?;
if tree.is_empty() {
println!(
"No projects with environments you can use. Create one with {} first.",
"railway init".cyan()
);
return Ok(());
}
let home = dirs::home_dir().context("Unable to get home directory")?;
let stored = AgentPrefs::load_in(&home);
let first_run = stored.is_none();
let saved = stored.unwrap_or_default();
// A linked directory wins: `railway link` (or a linked service checkout) is
// an explicit, per-directory statement of "this is the project I'm working
// in", which outranks a person-wide preference chosen once from wherever
// the terminal happened to be. The configured default is still the
// fallback when this directory has no link.
let linked = linked_target(&mut configs, &client, &tree).await;
let target = linked.clone().or_else(|| {
saved.default_project.as_ref().map(|project| tui::Target {
project_id: project.project_id.clone(),
project_name: project.project_name.clone(),
environment_id: project.environment_id.clone(),
environment_name: project.environment_name.clone(),
})
});
// Same order for the "(default)" label in the tree: a linked project shows
// as the default over whatever is in the preferences file.
let default_project_id = linked
.as_ref()
.map(|t| t.project_id.clone())
.or_else(|| saved.default_project.as_ref().map(|p| p.project_id.clone()));
let mut app = App::new(
tree,
target,
saved.agent.as_deref(),
saved.theme.as_deref(),
default_project_id,
!first_run,
);
// A launch that needs a Claude token minted comes back out here, gets one
// with the real terminal, and goes straight back in — H1's step-out-and-
// return, now only for the rare case that actually needs it.
// No preferences yet: offer to set them up rather than dropping someone in
// front of a prompt whose target, agent and skills are all unanswered.
// Environments this machine has launched an agent in, so they load without
// the user going looking. `railway ca` no longer scans every project.
app.known_environments = configs.code_agent_environments();
app.skills_source = skills_sync::populated_sources(&home)
.first()
.map(|(source, _)| source.slug.to_string());
// Mirrored so the ⌥s settings card opens showing the saved answer.
app.skills_enabled = saved.skills.enabled;
// No preferences yet: ask whether to set them up, rather than dropping
// someone in front of a prompt whose target, agent and skills are all
// unanswered. Choosing Setup from the menu skips that question — they have
// already answered it by choosing it. An explicit initial screen is its
// own answer to "what should I see first" and skips the nudge too.
if let Some(screen) = initial_screen {
app.screen = screen;
} else if first_run {
app.start_wizard(true);
}
let mut pending: Option<tui::LaunchRequest> = None;
loop {
match tui::run(&mut app, client.clone(), backboard.clone(), pending.take()).await? {
Outcome::Quit => {
// A theme picked with ⌥t is a preference, not a session
// setting — persist it on the way out rather than making the
// user set it again next time. Best-effort: failing to save it
// is not worth an error on exit.
persist_theme(&home, app.theme.slug);
return Ok(());
}
Outcome::FullScreen(req) => {
println!(
"\n{}",
format!("Attaching to {} · {}", req.agent_name, req.session_name).dimmed()
);
// The relay resumes a durable session by name, so this is the
// same session the pane had — the full terminal, none of the
// TUI's chrome, and no second copy of the work.
let result = crate::commands::ssh::native::run_native_ssh_with_opts(
&req.ssh_target,
None,
req.identity.as_deref(),
Some(crate::commands::ssh::native::DurableResume {
session_name: &req.session_name,
resume_from_last_read: false,
}),
&req.relay_opts,
);
if let Err(err) = result {
eprintln!("{} {err:#}", "Session ended with an error:".red().bold());
}
crate::commands::ssh::native::clear_mouse_tracking();
pause_for_reentry();
}
Outcome::NeedsCredential(req) => {
println!(
"\n{}",
"Claude needs a one-time token for the agent — this opens your browser."
.dimmed()
);
match crate::commands::code::ensure_claude_credential_cached(&req.harness) {
Ok(()) => pending = Some(req),
Err(err) => {
eprintln!("{} {err:#}", "Couldn't mint a credential:".red().bold());
pause_for_reentry();
}
}
}
}
}
}
fn persist_theme(home: &std::path::Path, slug: &str) {
let mut prefs = AgentPrefs::load_in(home).unwrap_or_default();
if prefs.theme.as_deref() == Some(slug) {
return;
}
prefs.theme = Some(slug.to_string());
let _ = prefs.save_in(home);
}
/// Hold the restored terminal until the user is ready, so whatever the launcher
/// printed — an error, the sleep confirmation, the reconnect hint — isn't wiped
/// by the alternate screen a frame later.
fn pause_for_reentry() {
use std::io::{BufRead, Write};
print!("\n{}", "Press enter to return to railway ca…".dimmed());
let _ = std::io::stdout().flush();
let mut line = String::new();
let _ = std::io::stdin().lock().read_line(&mut line);
}
/// Seed the prompt target from the linked project, when this directory has one
/// and it appears in the tree. Best-effort: an unlinked directory just opens
/// with no target, and the first launch asks for one.
async fn linked_target(
configs: &mut Configs,
client: &reqwest::Client,
tree: &[tui::app::WorkspaceNode],
) -> Option<tui::Target> {
let linked = configs.get_linked_project().await.ok()?;
let env_id = linked.environment.clone()?;
let _ = client;
for ws in tree {
for project in &ws.projects {
if project.id != linked.project {
continue;
}
let env = project.envs.iter().find(|e| e.id == env_id)?;
return Some(tui::Target {
project_id: project.id.clone(),
project_name: project.name.clone(),
environment_id: env.id.clone(),
environment_name: env.name.clone(),
});
}
}
None
}
#[cfg(test)]
mod tests {
use super::*;
use clap::Parser;
fn setup_args(argv: &[&str]) -> Command {
Command::Setup(setup::Args::parse_from(
std::iter::once("setup").chain(argv.iter().copied()),
))
}
#[test]
fn launch_paths_need_a_credential() {
assert!(needs_credential(None, true));
assert!(needs_credential(None, false));
assert!(needs_credential(
Some(&Command::Start(LaunchArgs::parse_from(["start"]))),
true
));
}
#[test]
fn local_setup_runs_logged_out() {
assert!(!needs_credential(Some(&setup_args(&["--show"])), true));
assert!(!needs_credential(Some(&setup_args(&["-y"])), true));
// Piped setup takes the same non-interactive path as `-y`.
assert!(!needs_credential(Some(&setup_args(&[])), false));
}
#[test]
fn interactive_setup_needs_a_credential_for_the_project_picker() {
assert!(needs_credential(Some(&setup_args(&[])), true));
}
}