codecraft/server/src/tools/bash.rs

//! Shell command execution with timeout, cancellation and live output.

use super::{Tool, ToolCtx, ToolOutput, arg_str, truncate};
use anyhow::Result;
use async_trait::async_trait;
use serde_json::{Value, json};
use std::process::Stdio;
use std::time::{Duration, Instant};
use tokio::io::{AsyncBufReadExt, BufReader};
use tokio::process::Command;

const DEFAULT_TIMEOUT_MS: u64 = 120_000;
const MAX_TIMEOUT_MS: u64 = 600_000;
const MAX_OUTPUT: usize = 30_000;

pub struct Bash;

#[async_trait]
impl Tool for Bash {
    fn name(&self) -> &str {
        "bash"
    }
    fn description(&self) -> String {
        "Run a shell command (bash) in the session working directory. stdout and stderr are \
         combined. Default timeout 2 minutes, max 10 minutes (timeout_ms). The working directory \
         does not persist between calls – use absolute paths or `cd dir && ...`. Prefer the \
         dedicated read/edit/glob/grep tools over cat/sed/find/grep. For long-running processes \
         (dev servers, watchers, long builds) set run_in_background=true: you get a shell id, \
         read new output with bash_output and stop it with kill_shell."
            .into()
    }
    fn schema(&self) -> Value {
        json!({
            "type": "object",
            "properties": {
                "command": {"type": "string"},
                "description": {"type": "string", "description": "5-10 word description of what the command does"},
                "timeout_ms": {"type": "integer"},
                "run_in_background": {"type": "boolean", "description": "Keep running; read output with bash_output"}
            },
            "required": ["command"]
        })
    }
    fn title(&self, input: &Value) -> String {
        input["description"]
            .as_str()
            .map(String::from)
            .unwrap_or_else(|| format!("Run: {}", input["command"].as_str().unwrap_or("?")))
    }
    async fn run(&self, input: Value, ctx: &ToolCtx) -> Result<ToolOutput> {
        let command = arg_str(&input, "command")?;
        let timeout = Duration::from_millis(
            input["timeout_ms"].as_u64().unwrap_or(DEFAULT_TIMEOUT_MS).min(MAX_TIMEOUT_MS),
        );
        let mut cmd = match shell_command(ctx, command) {
            Ok(c) => c,
            Err(e) => return Ok(e),
        };
        if input["run_in_background"].as_bool().unwrap_or(false) {
            let id = crate::shells::spawn(&ctx.app, &ctx.session_id, command, cmd)?;
            return Ok(ToolOutput::text(format!(
                "Started in the background with shell id {id}. Read its output with bash_output (shell_id \"{id}\") and stop it with kill_shell."
            ))
            .with_meta(json!({"shell_id": id})));
        }
        cmd.stdin(Stdio::null()).stdout(Stdio::piped()).stderr(Stdio::null()).kill_on_drop(true);
        // Own process group so the whole tree can be killed on abort/timeout.
        cmd.process_group(0);
        let mut child = cmd.spawn()?;
        let pid = child.id();
        let stdout = child.stdout.take().unwrap();
        // split() instead of lines(): non-UTF-8 output must not end the read loop.
        let mut lines = BufReader::new(stdout).split(b'\n');
        let mut output = String::new();
        let started = Instant::now();
        let mut last_progress = Instant::now();
        let deadline = tokio::time::sleep(timeout);
        tokio::pin!(deadline);

        let kill_group = |pid: Option<u32>| {
            if let Some(pid) = pid {
                // SAFETY: plain syscall; negative pid targets the process group.
                unsafe {
                    libc::kill(-(pid as i32), libc::SIGKILL);
                }
            }
        };

        let status = loop {
            tokio::select! {
                line = lines.next_segment() => match line {
                    Ok(Some(l)) => {
                        if output.len() < MAX_OUTPUT * 4 {
                            output.push_str(&String::from_utf8_lossy(&l));
                            output.push('\n');
                        }
                        if last_progress.elapsed() > Duration::from_millis(300) {
                            ctx.progress(tail(&output, 4000));
                            last_progress = Instant::now();
                        }
                    }
                    Ok(None) | Err(_) => break child.wait().await?,
                },
                _ = &mut deadline => {
                    kill_group(pid);
                    let _ = child.wait().await;
                    output.push_str(&format!("\n[command timed out after {}s and was killed]", timeout.as_secs()));
                    return Ok(ToolOutput::error(truncate(&output, MAX_OUTPUT)));
                }
                _ = ctx.cancel.cancelled() => {
                    kill_group(pid);
                    let _ = child.wait().await;
                    return Ok(ToolOutput::error("[command aborted by user]"));
                }
            }
        };

        let code = status.code().unwrap_or(-1);
        let mut text = truncate(output.trim_end(), MAX_OUTPUT);
        if text.is_empty() {
            text = "(no output)".into();
        }
        if code != 0 {
            text.push_str(&format!("\n[exit code {code}]"));
        }
        Ok(ToolOutput {
            content: text,
            is_error: code != 0,
            meta: Some(json!({"exit_code": code, "duration_ms": started.elapsed().as_millis() as u64})),
            ..Default::default()
        })
    }
}

/// The bash command for a script, sandboxed for non-admin users.
pub fn shell_command(ctx: &ToolCtx, command: &str) -> std::result::Result<Command, ToolOutput> {
    // stderr of every command (not just the last one) goes into the output.
    let script = format!("exec 2>&1\n{command}");
    Ok(match &ctx.sandbox {
        Some(spec) => {
            if !crate::sandbox::available() {
                return Err(ToolOutput::error(
                    "Shell commands are disabled for this account: the server's sandbox (bubblewrap) is not available. Ask an admin.",
                ));
            }
            crate::sandbox::command(spec, &ctx.cwd, "bash", &["-c", &script])
        }
        None => {
            let mut c = Command::new("bash");
            c.arg("-c").arg(&script).current_dir(&ctx.cwd).env("CODECRAFT", "1").env("GIT_TERMINAL_PROMPT", "0").env("PAGER", "cat");
            // The session owner's git credentials and identity.
            let owner = ctx.app.session(&ctx.session_id).ok().and_then(|s| s.owner_id).and_then(|id| ctx.app.store.user(&id).ok().flatten());
            if let Some(u) = owner {
                if let Some(cfg) = crate::gitcreds::global_config(&ctx.app, &u) {
                    c.env("GIT_CONFIG_GLOBAL", cfg);
                }
                c.envs(crate::gitcreds::identity_env(&ctx.app, &u.id));
            }
            c
        }
    })
}

fn tail(s: &str, max: usize) -> String {
    if s.len() <= max {
        return s.to_string();
    }
    let mut start = s.len() - max;
    while !s.is_char_boundary(start) {
        start += 1;
    }
    s[start..].to_string()
}