Files
llmx/codex-rs/core/src/exec.rs
oai-ragona d7a40195e6 [codex-rs] Reliability pass on networking (#658)
We currently see a behavior that looks like this:
```
2025-04-25T16:52:24.552789Z  WARN codex_core::codex: stream disconnected - retrying turn (1/10 in 232ms)...
codex> event: BackgroundEvent { message: "stream error: stream disconnected before completion: Transport error: error decoding response body; retrying 1/10 in 232ms…" }
2025-04-25T16:52:54.789885Z  WARN codex_core::codex: stream disconnected - retrying turn (2/10 in 418ms)...
codex> event: BackgroundEvent { message: "stream error: stream disconnected before completion: Transport error: error decoding response body; retrying 2/10 in 418ms…" }
```

This PR contains a few different fixes that attempt to resolve/improve
this:
1. **Remove overall client timeout.** I think
[this](https://github.com/openai/codex/pull/658/files#diff-c39945d3c42f29b506ff54b7fa2be0795b06d7ad97f1bf33956f60e3c6f19c19L173)
is perhaps the big fix -- it looks to me like this was actually timing
out even if events were still coming through, and that was causing a
disconnect right in the middle of a healthy stream.
2. **Cap response sizes.** We were frequently sending MUCH larger
responses than the upstream typescript `codex`, and that was definitely
not helping. [Fix
here](https://github.com/openai/codex/pull/658/files#diff-d792bef59aa3ee8cb0cbad8b176dbfefe451c227ac89919da7c3e536a9d6cdc0R21-R26)
for that one.
3. **Much higher idle timeout.** Our idle timeout value was much lower
than typescript.
4. **Sub-linear backoff.** We were much too aggressively backing off,
[this](https://github.com/openai/codex/pull/658/files#diff-5d5959b95c6239e6188516da5c6b7eb78154cd9cfedfb9f753d30a7b6d6b8b06R30-R33)
makes it sub-exponential but maintains the jitter and such.

I was seeing that `stream error: stream disconnected` behavior
constantly, and anecdotally I can no longer reproduce. It feels much
snappier.
2025-04-25 11:44:22 -07:00

316 lines
9.4 KiB
Rust

use std::io;
use std::path::PathBuf;
use std::process::ExitStatus;
use std::process::Stdio;
use std::sync::Arc;
use std::time::Duration;
use std::time::Instant;
use serde::Deserialize;
use tokio::io::AsyncRead;
use tokio::io::AsyncReadExt;
use tokio::io::BufReader;
use tokio::process::Command;
use tokio::sync::Notify;
use crate::error::CodexErr;
use crate::error::Result;
use crate::error::SandboxErr;
use crate::protocol::SandboxPolicy;
/// Maximum we send for each stream, which is either:
/// - 10KiB OR
/// - 256 lines
const MAX_STREAM_OUTPUT: usize = 10 * 1024;
const MAX_STREAM_OUTPUT_LINES: usize = 256;
const DEFAULT_TIMEOUT_MS: u64 = 10_000;
/// Hardcode this since it does not seem worth including the libc craate just
/// for this.
const SIGKILL_CODE: i32 = 9;
const MACOS_SEATBELT_READONLY_POLICY: &str = include_str!("seatbelt_readonly_policy.sbpl");
#[derive(Deserialize, Debug, Clone)]
pub struct ExecParams {
pub command: Vec<String>,
pub workdir: Option<String>,
/// This is the maximum time in seconds that the command is allowed to run.
#[serde(rename = "timeout")]
// The wire format uses `timeout`, which has ambiguous units, so we use
// `timeout_ms` as the field name so it is clear in code.
pub timeout_ms: Option<u64>,
}
#[derive(Clone, Copy, Debug, PartialEq)]
pub enum SandboxType {
None,
/// Only available on macOS.
MacosSeatbelt,
/// Only available on Linux.
LinuxSeccomp,
}
#[cfg(target_os = "linux")]
async fn exec_linux(
params: ExecParams,
writable_roots: &[PathBuf],
ctrl_c: Arc<Notify>,
sandbox_policy: SandboxPolicy,
) -> Result<RawExecToolCallOutput> {
crate::linux::exec_linux(params, writable_roots, ctrl_c, sandbox_policy).await
}
#[cfg(not(target_os = "linux"))]
async fn exec_linux(
_params: ExecParams,
_writable_roots: &[PathBuf],
_ctrl_c: Arc<Notify>,
_sandbox_policy: SandboxPolicy,
) -> Result<RawExecToolCallOutput> {
Err(CodexErr::Io(io::Error::new(
io::ErrorKind::InvalidInput,
"linux sandbox is not supported on this platform",
)))
}
pub async fn process_exec_tool_call(
params: ExecParams,
sandbox_type: SandboxType,
writable_roots: &[PathBuf],
ctrl_c: Arc<Notify>,
sandbox_policy: SandboxPolicy,
) -> Result<ExecToolCallOutput> {
let start = Instant::now();
let raw_output_result = match sandbox_type {
SandboxType::None => exec(params, ctrl_c).await,
SandboxType::MacosSeatbelt => {
let ExecParams {
command,
workdir,
timeout_ms,
} = params;
let seatbelt_command = create_seatbelt_command(command, writable_roots);
exec(
ExecParams {
command: seatbelt_command,
workdir,
timeout_ms,
},
ctrl_c,
)
.await
}
SandboxType::LinuxSeccomp => {
exec_linux(params, writable_roots, ctrl_c, sandbox_policy).await
}
};
let duration = start.elapsed();
match raw_output_result {
Ok(raw_output) => {
let exit_code = raw_output.exit_status.code().unwrap_or(-1);
let stdout = String::from_utf8_lossy(&raw_output.stdout).to_string();
let stderr = String::from_utf8_lossy(&raw_output.stderr).to_string();
// NOTE(ragona): This is much less restrictive than the previous check. If we exec
// a command, and it returns anything other than success, we assume that it may have
// been a sandboxing error and allow the user to retry. (The user of course may choose
// not to retry, or in a non-interactive mode, would automatically reject the approval.)
if exit_code != 0 && sandbox_type != SandboxType::None {
return Err(CodexErr::Sandbox(SandboxErr::Denied(
exit_code, stdout, stderr,
)));
}
Ok(ExecToolCallOutput {
exit_code,
stdout,
stderr,
duration,
})
}
Err(err) => {
tracing::error!("exec error: {err}");
Err(err)
}
}
}
pub fn create_seatbelt_command(command: Vec<String>, writable_roots: &[PathBuf]) -> Vec<String> {
let (policies, cli_args): (Vec<String>, Vec<String>) = writable_roots
.iter()
.enumerate()
.map(|(index, root)| {
let param_name = format!("WRITABLE_ROOT_{index}");
let policy: String = format!("(subpath (param \"{param_name}\"))");
let cli_arg = format!("-D{param_name}={}", root.to_string_lossy());
(policy, cli_arg)
})
.unzip();
let full_policy = if policies.is_empty() {
MACOS_SEATBELT_READONLY_POLICY.to_string()
} else {
let scoped_write_policy = format!("(allow file-write*\n{}\n)", policies.join(" "));
format!("{MACOS_SEATBELT_READONLY_POLICY}\n{scoped_write_policy}")
};
let mut seatbelt_command: Vec<String> = vec![
"sandbox-exec".to_string(),
"-p".to_string(),
full_policy.to_string(),
];
seatbelt_command.extend(cli_args);
seatbelt_command.push("--".to_string());
seatbelt_command.extend(command);
seatbelt_command
}
#[derive(Debug)]
pub struct RawExecToolCallOutput {
pub exit_status: ExitStatus,
pub stdout: Vec<u8>,
pub stderr: Vec<u8>,
}
#[derive(Debug)]
pub struct ExecToolCallOutput {
pub exit_code: i32,
pub stdout: String,
pub stderr: String,
pub duration: Duration,
}
pub async fn exec(
ExecParams {
command,
workdir,
timeout_ms,
}: ExecParams,
ctrl_c: Arc<Notify>,
) -> Result<RawExecToolCallOutput> {
let mut child = {
if command.is_empty() {
return Err(CodexErr::Io(io::Error::new(
io::ErrorKind::InvalidInput,
"command args are empty",
)));
}
let mut cmd = Command::new(&command[0]);
if command.len() > 1 {
cmd.args(&command[1..]);
}
if let Some(dir) = &workdir {
cmd.current_dir(dir);
}
// Do not create a file descriptor for stdin because otherwise some
// commands may hang forever waiting for input. For example, ripgrep has
// a heuristic where it may try to read from stdin as explained here:
// https://github.com/BurntSushi/ripgrep/blob/e2362d4d5185d02fa857bf381e7bd52e66fafc73/crates/core/flags/hiargs.rs#L1101-L1103
cmd.stdin(Stdio::null());
cmd.stdout(Stdio::piped())
.stderr(Stdio::piped())
.kill_on_drop(true)
.spawn()?
};
let stdout_handle = tokio::spawn(read_capped(
BufReader::new(child.stdout.take().expect("stdout is not piped")),
MAX_STREAM_OUTPUT,
MAX_STREAM_OUTPUT_LINES,
));
let stderr_handle = tokio::spawn(read_capped(
BufReader::new(child.stderr.take().expect("stderr is not piped")),
MAX_STREAM_OUTPUT,
MAX_STREAM_OUTPUT_LINES,
));
let interrupted = ctrl_c.notified();
let timeout = Duration::from_millis(timeout_ms.unwrap_or(DEFAULT_TIMEOUT_MS));
let exit_status = tokio::select! {
result = tokio::time::timeout(timeout, child.wait()) => {
match result {
Ok(Ok(exit_status)) => exit_status,
Ok(e) => e?,
Err(_) => {
// timeout
child.start_kill()?;
// Debatable whether `child.wait().await` should be called here.
synthetic_exit_status(128 + SIGKILL_CODE)
}
}
}
_ = interrupted => {
child.start_kill()?;
synthetic_exit_status(128 + SIGKILL_CODE)
}
};
let stdout = stdout_handle.await??;
let stderr = stderr_handle.await??;
Ok(RawExecToolCallOutput {
exit_status,
stdout,
stderr,
})
}
async fn read_capped<R: AsyncRead + Unpin>(
mut reader: R,
max_output: usize,
max_lines: usize,
) -> io::Result<Vec<u8>> {
let mut buf = Vec::with_capacity(max_output.min(8 * 1024));
let mut tmp = [0u8; 8192];
let mut remaining_bytes = max_output;
let mut remaining_lines = max_lines;
loop {
let n = reader.read(&mut tmp).await?;
if n == 0 {
break;
}
// Copy into the buffer only while we still have byte and line budget.
if remaining_bytes > 0 && remaining_lines > 0 {
let mut copy_len = 0;
for &b in &tmp[..n] {
if remaining_bytes == 0 || remaining_lines == 0 {
break;
}
copy_len += 1;
remaining_bytes -= 1;
if b == b'\n' {
remaining_lines -= 1;
}
}
buf.extend_from_slice(&tmp[..copy_len]);
}
// Continue reading to EOF to avoid back-pressure, but discard once caps are hit.
}
Ok(buf)
}
#[cfg(unix)]
fn synthetic_exit_status(code: i32) -> ExitStatus {
use std::os::unix::process::ExitStatusExt;
std::process::ExitStatus::from_raw(code)
}
#[cfg(windows)]
fn synthetic_exit_status(code: u32) -> ExitStatus {
use std::os::windows::process::ExitStatusExt;
std::process::ExitStatus::from_raw(code)
}