Compare commits

...

1 Commits

Author SHA1 Message Date
Michael Bolin
e2aa9dd338 refactor: route zsh-fork through unified exec 2026-03-13 13:50:43 -07:00
16 changed files with 1452 additions and 176 deletions

View File

@@ -11,7 +11,6 @@ use app_test_support::McpProcess;
use app_test_support::create_final_assistant_message_sse_response;
use app_test_support::create_mock_responses_server_sequence;
use app_test_support::create_mock_responses_server_sequence_unchecked;
use app_test_support::create_shell_command_sse_response;
use app_test_support::to_response;
use codex_app_server_protocol::CommandAction;
use codex_app_server_protocol::CommandExecutionApprovalDecision;
@@ -35,9 +34,11 @@ use codex_core::features::Feature;
use core_test_support::responses;
use core_test_support::skip_if_no_network;
use pretty_assertions::assert_eq;
use serde_json::json;
use std::collections::BTreeMap;
use std::path::Path;
use tempfile::TempDir;
use tokio::time::sleep;
use tokio::time::timeout;
#[cfg(windows)]
@@ -62,19 +63,14 @@ async fn turn_start_shell_zsh_fork_executes_command_v2() -> Result<()> {
};
eprintln!("using zsh path for zsh-fork test: {}", zsh_path.display());
// Keep the shell command in flight until we interrupt it. A fast command
// Keep the exec command in flight until we interrupt it. A fast command
// like `echo hi` can finish before the interrupt arrives on faster runners,
// which turns this into a test for post-command follow-up behavior instead
// of interrupting an active zsh-fork command.
let release_marker_escaped = release_marker.to_string_lossy().replace('\'', r#"'\''"#);
let wait_for_interrupt =
format!("while [ ! -f '{release_marker_escaped}' ]; do sleep 0.01; done");
let response = create_shell_command_sse_response(
vec!["/bin/sh".to_string(), "-c".to_string(), wait_for_interrupt],
None,
Some(5000),
"call-zsh-fork",
)?;
let response = create_zsh_fork_exec_command_sse_response(&wait_for_interrupt, "call-zsh-fork")?;
let no_op_response = responses::sse(vec![
responses::ev_response_created("resp-2"),
responses::ev_completed("resp-2"),
@@ -91,7 +87,7 @@ async fn turn_start_shell_zsh_fork_executes_command_v2() -> Result<()> {
"never",
&BTreeMap::from([
(Feature::ShellZshFork, true),
(Feature::UnifiedExec, false),
(Feature::UnifiedExec, true),
(Feature::ShellSnapshot, false),
]),
&zsh_path,
@@ -163,7 +159,7 @@ async fn turn_start_shell_zsh_fork_executes_command_v2() -> Result<()> {
assert_eq!(id, "call-zsh-fork");
assert_eq!(status, CommandExecutionStatus::InProgress);
assert!(command.starts_with(&zsh_path.display().to_string()));
assert!(command.contains("/bin/sh -c"));
assert!(command.contains(" -lc "));
assert!(command.contains("sleep 0.01"));
assert!(command.contains(&release_marker.display().to_string()));
assert_eq!(cwd, workspace);
@@ -191,14 +187,8 @@ async fn turn_start_shell_zsh_fork_exec_approval_decline_v2() -> Result<()> {
eprintln!("using zsh path for zsh-fork test: {}", zsh_path.display());
let responses = vec![
create_shell_command_sse_response(
vec![
"python3".to_string(),
"-c".to_string(),
"print(42)".to_string(),
],
None,
Some(5000),
create_zsh_fork_exec_command_sse_response(
"python3 -c 'print(42)'",
"call-zsh-fork-decline",
)?,
create_final_assistant_message_sse_response("done")?,
@@ -210,7 +200,7 @@ async fn turn_start_shell_zsh_fork_exec_approval_decline_v2() -> Result<()> {
"untrusted",
&BTreeMap::from([
(Feature::ShellZshFork, true),
(Feature::UnifiedExec, false),
(Feature::UnifiedExec, true),
(Feature::ShellSnapshot, false),
]),
&zsh_path,
@@ -326,14 +316,8 @@ async fn turn_start_shell_zsh_fork_exec_approval_cancel_v2() -> Result<()> {
};
eprintln!("using zsh path for zsh-fork test: {}", zsh_path.display());
let responses = vec![create_shell_command_sse_response(
vec![
"python3".to_string(),
"-c".to_string(),
"print(42)".to_string(),
],
None,
Some(5000),
let responses = vec![create_zsh_fork_exec_command_sse_response(
"python3 -c 'print(42)'",
"call-zsh-fork-cancel",
)?];
let server = create_mock_responses_server_sequence(responses).await;
@@ -343,7 +327,7 @@ async fn turn_start_shell_zsh_fork_exec_approval_cancel_v2() -> Result<()> {
"untrusted",
&BTreeMap::from([
(Feature::ShellZshFork, true),
(Feature::UnifiedExec, false),
(Feature::UnifiedExec, true),
(Feature::ShellSnapshot, false),
]),
&zsh_path,
@@ -441,6 +425,204 @@ async fn turn_start_shell_zsh_fork_exec_approval_cancel_v2() -> Result<()> {
Ok(())
}
#[tokio::test]
async fn turn_start_shell_zsh_fork_interrupt_kills_approved_subcommand_v2() -> Result<()> {
skip_if_no_network!(Ok(()));
let tmp = TempDir::new()?;
let codex_home = tmp.path().join("codex_home");
std::fs::create_dir(&codex_home)?;
let workspace = tmp.path().join("workspace");
std::fs::create_dir(&workspace)?;
let launch_marker = workspace.join("approved-subcommand.started");
let leaked_marker = workspace.join("approved-subcommand.leaked");
let launch_marker_display = launch_marker.display().to_string();
assert!(
!launch_marker_display.contains('\''),
"test workspace path should not contain single quotes: {launch_marker_display}"
);
let leaked_marker_display = leaked_marker.display().to_string();
assert!(
!leaked_marker_display.contains('\''),
"test workspace path should not contain single quotes: {leaked_marker_display}"
);
let Some(zsh_path) = find_test_zsh_path()? else {
eprintln!("skipping zsh fork interrupt cleanup test: no zsh executable found");
return Ok(());
};
if !supports_exec_wrapper_intercept(&zsh_path) {
eprintln!(
"skipping zsh fork interrupt cleanup test: zsh does not support EXEC_WRAPPER intercepts ({})",
zsh_path.display()
);
return Ok(());
}
let zsh_path_display = zsh_path.display().to_string();
eprintln!("using zsh path for zsh-fork test: {zsh_path_display}");
let shell_command = format!(
"/bin/sh -c 'echo started > \"{launch_marker_display}\" && /bin/sleep 0.5 && echo leaked > \"{leaked_marker_display}\" && exec /bin/sleep 100'"
);
let tool_call_arguments = serde_json::to_string(&json!({
"cmd": shell_command,
"yield_time_ms": 30_000,
}))?;
let response = responses::sse(vec![
responses::ev_response_created("resp-1"),
responses::ev_function_call(
"call-zsh-fork-interrupt-cleanup",
"exec_command",
&tool_call_arguments,
),
responses::ev_completed("resp-1"),
]);
let no_op_response = responses::sse(vec![
responses::ev_response_created("resp-2"),
responses::ev_completed("resp-2"),
]);
let server =
create_mock_responses_server_sequence_unchecked(vec![response, no_op_response]).await;
create_config_toml(
&codex_home,
&server.uri(),
"untrusted",
&BTreeMap::from([
(Feature::ShellZshFork, true),
(Feature::UnifiedExec, true),
(Feature::ShellSnapshot, false),
]),
&zsh_path,
)?;
let mut mcp = create_zsh_test_mcp_process(&codex_home, &workspace).await?;
timeout(DEFAULT_READ_TIMEOUT, mcp.initialize()).await??;
let start_id = mcp
.send_thread_start_request(ThreadStartParams {
model: Some("mock-model".to_string()),
cwd: Some(workspace.to_string_lossy().into_owned()),
..Default::default()
})
.await?;
let start_resp: JSONRPCResponse = timeout(
DEFAULT_READ_TIMEOUT,
mcp.read_stream_until_response_message(RequestId::Integer(start_id)),
)
.await??;
let ThreadStartResponse { thread, .. } = to_response::<ThreadStartResponse>(start_resp)?;
let turn_id = mcp
.send_turn_start_request(TurnStartParams {
thread_id: thread.id.clone(),
input: vec![V2UserInput::Text {
text: "run the long-lived command".to_string(),
text_elements: Vec::new(),
}],
cwd: Some(workspace.clone()),
approval_policy: Some(codex_app_server_protocol::AskForApproval::UnlessTrusted),
sandbox_policy: Some(codex_app_server_protocol::SandboxPolicy::WorkspaceWrite {
writable_roots: vec![workspace.clone().try_into()?],
read_only_access: codex_app_server_protocol::ReadOnlyAccess::FullAccess,
network_access: false,
exclude_tmpdir_env_var: false,
exclude_slash_tmp: false,
}),
model: Some("mock-model".to_string()),
effort: Some(codex_protocol::openai_models::ReasoningEffort::Medium),
summary: Some(codex_protocol::config_types::ReasoningSummary::Auto),
..Default::default()
})
.await?;
let turn_resp: JSONRPCResponse = timeout(
DEFAULT_READ_TIMEOUT,
mcp.read_stream_until_response_message(RequestId::Integer(turn_id)),
)
.await??;
let TurnStartResponse { turn } = to_response::<TurnStartResponse>(turn_resp)?;
let mut saw_target_approval = false;
while !saw_target_approval {
let server_req = timeout(
DEFAULT_READ_TIMEOUT,
mcp.read_stream_until_request_message(),
)
.await??;
let ServerRequest::CommandExecutionRequestApproval { request_id, params } = server_req
else {
panic!("expected CommandExecutionRequestApproval request");
};
let approval_command = params.command.clone().unwrap_or_default();
saw_target_approval = approval_command.contains("/bin/sh")
&& approval_command.contains(&launch_marker_display)
&& !approval_command.contains(&zsh_path_display);
mcp.send_response(
request_id,
serde_json::to_value(CommandExecutionRequestApprovalResponse {
decision: CommandExecutionApprovalDecision::Accept,
})?,
)
.await?;
}
let started_command = timeout(DEFAULT_READ_TIMEOUT, async {
loop {
let notif = mcp
.read_stream_until_notification_message("item/started")
.await?;
let started: ItemStartedNotification =
serde_json::from_value(notif.params.clone().expect("item/started params"))?;
if let ThreadItem::CommandExecution { .. } = started.item {
return Ok::<ThreadItem, anyhow::Error>(started.item);
}
}
})
.await??;
let ThreadItem::CommandExecution {
id,
process_id,
status,
command,
cwd,
..
} = started_command
else {
unreachable!("loop ensures we break on command execution items");
};
assert_eq!(id, "call-zsh-fork-interrupt-cleanup");
assert_eq!(status, CommandExecutionStatus::InProgress);
assert!(command.starts_with(&zsh_path.display().to_string()));
assert!(command.contains(" -lc "));
assert!(command.contains(&launch_marker_display));
assert_eq!(cwd, workspace);
assert!(process_id.is_some(), "process id should be present");
timeout(DEFAULT_READ_TIMEOUT, async {
loop {
if launch_marker.exists() {
return Ok::<(), anyhow::Error>(());
}
sleep(std::time::Duration::from_millis(20)).await;
}
})
.await??;
mcp.interrupt_turn_and_wait_for_aborted(
thread.id.clone(),
turn.id.clone(),
DEFAULT_READ_TIMEOUT,
)
.await?;
sleep(std::time::Duration::from_millis(750)).await;
assert!(
!leaked_marker.exists(),
"expected interrupt to stop approved subcommand before it wrote {leaked_marker_display}"
);
Ok(())
}
#[tokio::test]
async fn turn_start_shell_zsh_fork_subcommand_decline_marks_parent_declined_v2() -> Result<()> {
skip_if_no_network!(Ok(()));
@@ -472,16 +654,15 @@ async fn turn_start_shell_zsh_fork_subcommand_decline_marks_parent_declined_v2()
first_file.display(),
second_file.display()
);
let tool_call_arguments = serde_json::to_string(&serde_json::json!({
"command": shell_command,
"workdir": serde_json::Value::Null,
"timeout_ms": 5000
let tool_call_arguments = serde_json::to_string(&json!({
"cmd": shell_command,
"yield_time_ms": 5000,
}))?;
let response = responses::sse(vec![
responses::ev_response_created("resp-1"),
responses::ev_function_call(
"call-zsh-fork-subcommand-decline",
"shell_command",
"exec_command",
&tool_call_arguments,
),
responses::ev_completed("resp-1"),
@@ -502,7 +683,7 @@ async fn turn_start_shell_zsh_fork_subcommand_decline_marks_parent_declined_v2()
"untrusted",
&BTreeMap::from([
(Feature::ShellZshFork, true),
(Feature::UnifiedExec, false),
(Feature::UnifiedExec, true),
(Feature::ShellSnapshot, false),
]),
&zsh_path,
@@ -744,6 +925,21 @@ async fn create_zsh_test_mcp_process(codex_home: &Path, zdotdir: &Path) -> Resul
McpProcess::new_with_env(codex_home, &[("ZDOTDIR", Some(zdotdir.as_str()))]).await
}
fn create_zsh_fork_exec_command_sse_response(
command: &str,
call_id: &str,
) -> anyhow::Result<String> {
let tool_call_arguments = serde_json::to_string(&json!({
"cmd": command,
"yield_time_ms": 5000,
}))?;
Ok(responses::sse(vec![
responses::ev_response_created("resp-1"),
responses::ev_function_call(call_id, "exec_command", &tool_call_arguments),
responses::ev_completed("resp-1"),
]))
}
fn create_config_toml(
codex_home: &Path,
server_uri: &str,

View File

@@ -103,6 +103,7 @@ fn shell_command_for_invocation(invocation: &ToolInvocation) -> Option<(Vec<Stri
&params,
invocation.session.user_shell(),
invocation.turn.tools_config.allow_login_shell,
invocation.turn.tools_config.unified_exec_backend,
)
.ok()?;
Some((command, invocation.turn.resolve_path(params.workdir)))

View File

@@ -19,6 +19,7 @@ use crate::tools::handlers::parse_arguments_with_base_path;
use crate::tools::handlers::resolve_workdir_base_path;
use crate::tools::registry::ToolHandler;
use crate::tools::registry::ToolKind;
use crate::tools::spec::UnifiedExecBackendConfig;
use crate::unified_exec::ExecCommandRequest;
use crate::unified_exec::UnifiedExecContext;
use crate::unified_exec::UnifiedExecProcessManager;
@@ -108,6 +109,7 @@ impl ToolHandler for UnifiedExecHandler {
&params,
invocation.session.user_shell(),
invocation.turn.tools_config.allow_login_shell,
invocation.turn.tools_config.unified_exec_backend,
) {
Ok(command) => command,
Err(_) => return true,
@@ -155,6 +157,7 @@ impl ToolHandler for UnifiedExecHandler {
&args,
session.user_shell(),
turn.tools_config.allow_login_shell,
turn.tools_config.unified_exec_backend,
)
.map_err(FunctionCallError::RespondToModel)?;
@@ -321,12 +324,23 @@ pub(crate) fn get_command(
args: &ExecCommandArgs,
session_shell: Arc<Shell>,
allow_login_shell: bool,
unified_exec_backend: UnifiedExecBackendConfig,
) -> Result<Vec<String>, String> {
let model_shell = args.shell.as_ref().map(|shell_str| {
let mut shell = get_shell_by_model_provided_path(&PathBuf::from(shell_str));
shell.shell_snapshot = crate::shell::empty_shell_snapshot_receiver();
shell
});
if unified_exec_backend == UnifiedExecBackendConfig::ZshFork && args.shell.is_some() {
return Err(
"shell override is not supported when the zsh-fork backend is enabled.".to_string(),
);
}
let model_shell = if unified_exec_backend == UnifiedExecBackendConfig::ZshFork {
None
} else {
args.shell.as_ref().map(|shell_str| {
let mut shell = get_shell_by_model_provided_path(&PathBuf::from(shell_str));
shell.shell_snapshot = crate::shell::empty_shell_snapshot_receiver();
shell
})
};
let shell = model_shell.as_ref().unwrap_or(session_shell.as_ref());
let use_login_shell = match args.login {

View File

@@ -1,12 +1,16 @@
use super::*;
use crate::shell::ShellType;
use crate::shell::default_user_shell;
use crate::shell::empty_shell_snapshot_receiver;
use crate::tools::handlers::parse_arguments_with_base_path;
use crate::tools::handlers::resolve_workdir_base_path;
use crate::tools::spec::UnifiedExecBackendConfig;
use codex_protocol::models::FileSystemPermissions;
use codex_protocol::models::PermissionProfile;
use codex_utils_absolute_path::AbsolutePathBuf;
use pretty_assertions::assert_eq;
use std::fs;
use std::path::PathBuf;
use std::sync::Arc;
use tempfile::tempdir;
@@ -18,8 +22,13 @@ fn test_get_command_uses_default_shell_when_unspecified() -> anyhow::Result<()>
assert!(args.shell.is_none());
let command =
get_command(&args, Arc::new(default_user_shell()), true).map_err(anyhow::Error::msg)?;
let command = get_command(
&args,
Arc::new(default_user_shell()),
true,
UnifiedExecBackendConfig::Direct,
)
.map_err(anyhow::Error::msg)?;
assert_eq!(command.len(), 3);
assert_eq!(command[2], "echo hello");
@@ -34,8 +43,13 @@ fn test_get_command_respects_explicit_bash_shell() -> anyhow::Result<()> {
assert_eq!(args.shell.as_deref(), Some("/bin/bash"));
let command =
get_command(&args, Arc::new(default_user_shell()), true).map_err(anyhow::Error::msg)?;
let command = get_command(
&args,
Arc::new(default_user_shell()),
true,
UnifiedExecBackendConfig::Direct,
)
.map_err(anyhow::Error::msg)?;
assert_eq!(command.last(), Some(&"echo hello".to_string()));
if command
@@ -55,8 +69,13 @@ fn test_get_command_respects_explicit_powershell_shell() -> anyhow::Result<()> {
assert_eq!(args.shell.as_deref(), Some("powershell"));
let command =
get_command(&args, Arc::new(default_user_shell()), true).map_err(anyhow::Error::msg)?;
let command = get_command(
&args,
Arc::new(default_user_shell()),
true,
UnifiedExecBackendConfig::Direct,
)
.map_err(anyhow::Error::msg)?;
assert_eq!(command[2], "echo hello");
Ok(())
@@ -70,8 +89,13 @@ fn test_get_command_respects_explicit_cmd_shell() -> anyhow::Result<()> {
assert_eq!(args.shell.as_deref(), Some("cmd"));
let command =
get_command(&args, Arc::new(default_user_shell()), true).map_err(anyhow::Error::msg)?;
let command = get_command(
&args,
Arc::new(default_user_shell()),
true,
UnifiedExecBackendConfig::Direct,
)
.map_err(anyhow::Error::msg)?;
assert_eq!(command[2], "echo hello");
Ok(())
@@ -82,8 +106,13 @@ fn test_get_command_rejects_explicit_login_when_disallowed() -> anyhow::Result<(
let json = r#"{"cmd": "echo hello", "login": true}"#;
let args: ExecCommandArgs = parse_arguments(json)?;
let err = get_command(&args, Arc::new(default_user_shell()), false)
.expect_err("explicit login should be rejected");
let err = get_command(
&args,
Arc::new(default_user_shell()),
false,
UnifiedExecBackendConfig::Direct,
)
.expect_err("explicit login should be rejected");
assert!(
err.contains("login shell is disabled by config"),
@@ -92,6 +121,30 @@ fn test_get_command_rejects_explicit_login_when_disallowed() -> anyhow::Result<(
Ok(())
}
#[test]
fn get_command_rejects_model_shell_override_for_zsh_fork_backend() -> anyhow::Result<()> {
let json = r#"{"cmd": "echo hello", "shell": "/bin/bash"}"#;
let args: ExecCommandArgs = parse_arguments(json)?;
let session_shell = Arc::new(Shell {
shell_type: ShellType::Zsh,
shell_path: PathBuf::from("/tmp/configured-zsh-fork-shell"),
shell_snapshot: empty_shell_snapshot_receiver(),
});
let err = get_command(
&args,
session_shell,
true,
UnifiedExecBackendConfig::ZshFork,
)
.expect_err("shell override should be rejected for zsh-fork backend");
assert!(
err.contains("shell override is not supported"),
"unexpected error: {err}"
);
Ok(())
}
#[test]
fn exec_command_args_resolve_relative_additional_permissions_against_workdir() -> anyhow::Result<()>
{

View File

@@ -50,10 +50,10 @@ use codex_shell_escalation::ShellCommandExecutor;
use codex_shell_escalation::Stopwatch;
use codex_utils_absolute_path::AbsolutePathBuf;
use std::collections::HashMap;
use std::path::Path;
use std::path::PathBuf;
use std::sync::Arc;
use std::time::Duration;
use tokio::sync::RwLock;
use tokio_util::sync::CancellationToken;
use uuid::Uuid;
@@ -91,7 +91,7 @@ pub(super) async fn try_run_zsh_fork(
req: &ShellRequest,
attempt: &SandboxAttempt<'_>,
ctx: &ToolCtx,
command: &[String],
shell_command: &[String],
) -> Result<Option<ExecToolCallOutput>, ToolError> {
let Some(shell_zsh_path) = ctx.session.services.shell_zsh_path.as_ref() else {
tracing::warn!("ZshFork backend specified, but shell_zsh_path is not configured.");
@@ -106,8 +106,10 @@ pub(super) async fn try_run_zsh_fork(
return Ok(None);
}
let ParsedShellCommand { script, login, .. } = extract_shell_script(shell_command)?;
let spec = build_command_spec(
command,
shell_command,
&req.cwd,
&req.env,
req.timeout_ms.into(),
@@ -119,14 +121,14 @@ pub(super) async fn try_run_zsh_fork(
.env_for(spec, req.network.as_ref())
.map_err(|err| ToolError::Codex(err.into()))?;
let crate::sandboxing::ExecRequest {
command,
command: sandbox_command,
cwd: sandbox_cwd,
env: sandbox_env,
network: sandbox_network,
expiration: _sandbox_expiration,
sandbox,
windows_sandbox_level,
windows_sandbox_private_desktop: _windows_sandbox_private_desktop,
windows_sandbox_private_desktop,
sandbox_permissions,
sandbox_policy,
file_system_sandbox_policy,
@@ -134,16 +136,14 @@ pub(super) async fn try_run_zsh_fork(
justification,
arg0,
} = sandbox_exec_request;
let ParsedShellCommand { script, login, .. } = extract_shell_script(&command)?;
let host_zsh_path =
resolve_host_zsh_path(sandbox_env.get("PATH").map(String::as_str), &sandbox_cwd);
let effective_timeout = Duration::from_millis(
req.timeout_ms
.unwrap_or(crate::exec::DEFAULT_EXEC_COMMAND_TIMEOUT_MS),
);
let exec_policy = Arc::new(RwLock::new(
ctx.session.services.exec_policy.current().as_ref().clone(),
));
let command_executor = CoreShellCommandExecutor {
command,
command: sandbox_command,
cwd: sandbox_cwd,
sandbox_policy,
file_system_sandbox_policy,
@@ -152,6 +152,7 @@ pub(super) async fn try_run_zsh_fork(
env: sandbox_env,
network: sandbox_network,
windows_sandbox_level,
windows_sandbox_private_desktop,
sandbox_permissions,
justification,
arg0,
@@ -164,6 +165,8 @@ pub(super) async fn try_run_zsh_fork(
.clone(),
codex_linux_sandbox_exe: ctx.turn.codex_linux_sandbox_exe.clone(),
use_legacy_landlock: ctx.turn.features.use_legacy_landlock(),
shell_zsh_path: ctx.session.services.shell_zsh_path.clone(),
host_zsh_path: host_zsh_path.clone(),
};
let main_execve_wrapper_exe = ctx
.session
@@ -192,7 +195,6 @@ pub(super) async fn try_run_zsh_fork(
req.additional_permissions_preapproved,
);
let escalation_policy = CoreShellActionProvider {
policy: Arc::clone(&exec_policy),
session: Arc::clone(&ctx.session),
turn: Arc::clone(&ctx.turn),
call_id: ctx.call_id.clone(),
@@ -205,6 +207,7 @@ pub(super) async fn try_run_zsh_fork(
approval_sandbox_permissions,
prompt_permissions: req.additional_permissions.clone(),
stopwatch: stopwatch.clone(),
host_zsh_path,
};
let escalate_server = EscalateServer::new(
@@ -225,6 +228,7 @@ pub(crate) async fn prepare_unified_exec_zsh_fork(
req: &crate::tools::runtimes::unified_exec::UnifiedExecRequest,
_attempt: &SandboxAttempt<'_>,
ctx: &ToolCtx,
shell_command: &[String],
exec_request: ExecRequest,
) -> Result<Option<PreparedUnifiedExecZshFork>, ToolError> {
let Some(shell_zsh_path) = ctx.session.services.shell_zsh_path.as_ref() else {
@@ -240,7 +244,7 @@ pub(crate) async fn prepare_unified_exec_zsh_fork(
return Ok(None);
}
let parsed = match extract_shell_script(&exec_request.command) {
let parsed = match extract_shell_script(shell_command) {
Ok(parsed) => parsed,
Err(err) => {
tracing::warn!("ZshFork unified exec fallback: {err:?}");
@@ -256,22 +260,37 @@ pub(crate) async fn prepare_unified_exec_zsh_fork(
return Ok(None);
}
let exec_policy = Arc::new(RwLock::new(
ctx.session.services.exec_policy.current().as_ref().clone(),
));
let ExecRequest {
command,
cwd,
env,
network,
expiration: _expiration,
sandbox,
windows_sandbox_level,
windows_sandbox_private_desktop,
sandbox_permissions,
sandbox_policy,
file_system_sandbox_policy,
network_sandbox_policy,
justification,
arg0,
} = &exec_request;
let host_zsh_path = resolve_host_zsh_path(env.get("PATH").map(String::as_str), cwd);
let command_executor = CoreShellCommandExecutor {
command: exec_request.command.clone(),
cwd: exec_request.cwd.clone(),
sandbox_policy: exec_request.sandbox_policy.clone(),
file_system_sandbox_policy: exec_request.file_system_sandbox_policy.clone(),
network_sandbox_policy: exec_request.network_sandbox_policy,
sandbox: exec_request.sandbox,
env: exec_request.env.clone(),
network: exec_request.network.clone(),
windows_sandbox_level: exec_request.windows_sandbox_level,
sandbox_permissions: exec_request.sandbox_permissions,
justification: exec_request.justification.clone(),
arg0: exec_request.arg0.clone(),
command: command.clone(),
cwd: cwd.clone(),
sandbox_policy: sandbox_policy.clone(),
file_system_sandbox_policy: file_system_sandbox_policy.clone(),
network_sandbox_policy: *network_sandbox_policy,
sandbox: *sandbox,
env: env.clone(),
network: network.clone(),
windows_sandbox_level: *windows_sandbox_level,
windows_sandbox_private_desktop: *windows_sandbox_private_desktop,
sandbox_permissions: *sandbox_permissions,
justification: justification.clone(),
arg0: arg0.clone(),
sandbox_policy_cwd: ctx.turn.cwd.clone(),
macos_seatbelt_profile_extensions: ctx
.turn
@@ -281,6 +300,8 @@ pub(crate) async fn prepare_unified_exec_zsh_fork(
.clone(),
codex_linux_sandbox_exe: ctx.turn.codex_linux_sandbox_exe.clone(),
use_legacy_landlock: ctx.turn.features.use_legacy_landlock(),
shell_zsh_path: ctx.session.services.shell_zsh_path.clone(),
host_zsh_path: host_zsh_path.clone(),
};
let main_execve_wrapper_exe = ctx
.session
@@ -293,7 +314,6 @@ pub(crate) async fn prepare_unified_exec_zsh_fork(
)
})?;
let escalation_policy = CoreShellActionProvider {
policy: Arc::clone(&exec_policy),
session: Arc::clone(&ctx.session),
turn: Arc::clone(&ctx.turn),
call_id: ctx.call_id.clone(),
@@ -309,6 +329,7 @@ pub(crate) async fn prepare_unified_exec_zsh_fork(
),
prompt_permissions: req.additional_permissions.clone(),
stopwatch: Stopwatch::unlimited(),
host_zsh_path,
};
let escalate_server = EscalateServer::new(
@@ -328,7 +349,6 @@ pub(crate) async fn prepare_unified_exec_zsh_fork(
}
struct CoreShellActionProvider {
policy: Arc<RwLock<Policy>>,
session: Arc<crate::codex::Session>,
turn: Arc<crate::codex::TurnContext>,
call_id: String,
@@ -341,6 +361,7 @@ struct CoreShellActionProvider {
approval_sandbox_permissions: SandboxPermissions,
prompt_permissions: Option<PermissionProfile>,
stopwatch: Stopwatch,
host_zsh_path: Option<PathBuf>,
}
#[allow(clippy::large_enum_variant)]
@@ -378,6 +399,66 @@ fn execve_prompt_is_rejected_by_policy(
}
}
fn paths_match(lhs: &Path, rhs: &Path) -> bool {
lhs == rhs
|| match (lhs.canonicalize(), rhs.canonicalize()) {
(Ok(lhs), Ok(rhs)) => lhs == rhs,
_ => false,
}
}
fn resolve_host_zsh_path(_path_env: Option<&str>, _cwd: &Path) -> Option<PathBuf> {
fn canonicalize_best_effort(path: PathBuf) -> PathBuf {
path.canonicalize().unwrap_or(path)
}
fn is_executable_file(path: &Path) -> bool {
std::fs::metadata(path).is_ok_and(|metadata| {
metadata.is_file() && {
#[cfg(unix)]
{
use std::os::unix::fs::PermissionsExt;
metadata.permissions().mode() & 0o111 != 0
}
#[cfg(not(unix))]
{
true
}
}
})
}
fn find_zsh_in_dirs(dirs: impl IntoIterator<Item = PathBuf>) -> Option<PathBuf> {
dirs.into_iter().find_map(|dir| {
let candidate = dir.join("zsh");
is_executable_file(&candidate).then(|| canonicalize_best_effort(candidate))
})
}
// Keep nested-zsh rewrites limited to canonical host shell installations.
// PATH shadowing from repos, Nix environments, or tool shims should not be
// treated as the host shell.
find_zsh_in_dirs(
["/bin", "/usr/bin", "/usr/local/bin", "/opt/homebrew/bin"]
.into_iter()
.map(PathBuf::from),
)
}
fn is_unconfigured_zsh_exec(
program: &AbsolutePathBuf,
shell_zsh_path: Option<&Path>,
host_zsh_path: Option<&Path>,
) -> bool {
let Some(shell_zsh_path) = shell_zsh_path else {
return false;
};
let Some(host_zsh_path) = host_zsh_path else {
return false;
};
paths_match(program.as_path(), host_zsh_path) && !paths_match(program.as_path(), shell_zsh_path)
}
impl CoreShellActionProvider {
fn decision_driven_by_policy(matched_rules: &[RuleMatch], decision: Decision) -> bool {
matched_rules.iter().any(|rule_match| {
@@ -489,6 +570,10 @@ impl CoreShellActionProvider {
command,
workdir,
None,
// Intercepted exec prompts happen after the original tool call has
// started, so we do not attach an execpolicy amendment payload here.
// Amendments are currently surfaced only from the top-level tool
// request path.
None,
None,
additional_permissions,
@@ -713,28 +798,35 @@ impl EscalationPolicy for CoreShellActionProvider {
.await;
}
let evaluation = {
let policy = self.policy.read().await;
evaluate_intercepted_exec_policy(
&policy,
program,
argv,
InterceptedExecPolicyContext {
approval_policy: self.approval_policy,
sandbox_policy: &self.sandbox_policy,
file_system_sandbox_policy: &self.file_system_sandbox_policy,
sandbox_permissions: self.approval_sandbox_permissions,
enable_shell_wrapper_parsing:
ENABLE_INTERCEPTED_EXEC_POLICY_SHELL_WRAPPER_PARSING,
},
)
};
let policy = self.session.services.exec_policy.current();
let evaluation = evaluate_intercepted_exec_policy(
policy.as_ref(),
program,
argv,
InterceptedExecPolicyContext {
approval_policy: self.approval_policy,
sandbox_policy: &self.sandbox_policy,
file_system_sandbox_policy: &self.file_system_sandbox_policy,
sandbox_permissions: self.approval_sandbox_permissions,
enable_shell_wrapper_parsing: ENABLE_INTERCEPTED_EXEC_POLICY_SHELL_WRAPPER_PARSING,
},
);
// When true, means the Evaluation was due to *.rules, not the
// fallback function.
let decision_driven_by_policy =
Self::decision_driven_by_policy(&evaluation.matched_rules, evaluation.decision);
let needs_escalation =
self.sandbox_permissions.requires_escalated_permissions() || decision_driven_by_policy;
// Keep zsh-fork interception alive across nested shells: if an
// intercepted exec targets the known host `zsh` path instead of the
// configured zsh-fork binary, force it through escalation so the
// executor can rewrite the program path back to the configured shell.
let force_zsh_fork_reexec = is_unconfigured_zsh_exec(
program,
self.session.services.shell_zsh_path.as_deref(),
self.host_zsh_path.as_deref(),
);
let needs_escalation = self.sandbox_permissions.requires_escalated_permissions()
|| decision_driven_by_policy
|| force_zsh_fork_reexec;
let decision_source = if decision_driven_by_policy {
DecisionSource::PrefixRule
@@ -875,6 +967,7 @@ struct CoreShellCommandExecutor {
env: HashMap<String, String>,
network: Option<codex_network_proxy::NetworkProxy>,
windows_sandbox_level: WindowsSandboxLevel,
windows_sandbox_private_desktop: bool,
sandbox_permissions: SandboxPermissions,
justification: Option<String>,
arg0: Option<String>,
@@ -883,6 +976,8 @@ struct CoreShellCommandExecutor {
macos_seatbelt_profile_extensions: Option<MacOsSeatbeltProfileExtensions>,
codex_linux_sandbox_exe: Option<PathBuf>,
use_legacy_landlock: bool,
shell_zsh_path: Option<PathBuf>,
host_zsh_path: Option<PathBuf>,
}
struct PrepareSandboxedExecParams<'a> {
@@ -925,7 +1020,7 @@ impl ShellCommandExecutor for CoreShellCommandExecutor {
expiration: ExecExpiration::Cancellation(cancel_rx),
sandbox: self.sandbox,
windows_sandbox_level: self.windows_sandbox_level,
windows_sandbox_private_desktop: false,
windows_sandbox_private_desktop: self.windows_sandbox_private_desktop,
sandbox_permissions: self.sandbox_permissions,
sandbox_policy: self.sandbox_policy.clone(),
file_system_sandbox_policy: self.file_system_sandbox_policy.clone(),
@@ -956,7 +1051,8 @@ impl ShellCommandExecutor for CoreShellCommandExecutor {
env: HashMap<String, String>,
execution: EscalationExecution,
) -> anyhow::Result<PreparedExec> {
let command = join_program_and_argv(program, argv);
let program = self.rewrite_intercepted_program_for_zsh_fork(program);
let command = join_program_and_argv(&program, argv);
let Some(first_arg) = argv.first() else {
return Err(anyhow::anyhow!(
"intercepted exec request must contain argv[0]"
@@ -1027,7 +1123,33 @@ impl ShellCommandExecutor for CoreShellCommandExecutor {
}
impl CoreShellCommandExecutor {
#[allow(clippy::too_many_arguments)]
fn rewrite_intercepted_program_for_zsh_fork(
&self,
program: &AbsolutePathBuf,
) -> AbsolutePathBuf {
let Some(shell_zsh_path) = self.shell_zsh_path.as_ref() else {
return program.clone();
};
if !is_unconfigured_zsh_exec(
program,
Some(shell_zsh_path.as_path()),
self.host_zsh_path.as_deref(),
) {
return program.clone();
}
match AbsolutePathBuf::from_absolute_path(shell_zsh_path) {
Ok(rewritten) => rewritten,
Err(err) => {
tracing::warn!(
"failed to rewrite intercepted zsh path {} to configured shell {}: {err}",
program.display(),
shell_zsh_path.display(),
);
program.clone()
}
}
}
fn prepare_sandboxed_exec(
&self,
params: PrepareSandboxedExecParams<'_>,
@@ -1082,7 +1204,7 @@ impl CoreShellCommandExecutor {
codex_linux_sandbox_exe: self.codex_linux_sandbox_exe.as_ref(),
use_legacy_landlock: self.use_legacy_landlock,
windows_sandbox_level: self.windows_sandbox_level,
windows_sandbox_private_desktop: false,
windows_sandbox_private_desktop: self.windows_sandbox_private_desktop,
})?;
if let Some(network) = exec_request.network.as_ref() {
network.apply_to_env(&mut exec_request.env);
@@ -1105,23 +1227,21 @@ struct ParsedShellCommand {
}
fn extract_shell_script(command: &[String]) -> Result<ParsedShellCommand, ToolError> {
// Commands reaching zsh-fork can be wrapped by environment/sandbox helpers, so
// we search for the first `-c`/`-lc` triple anywhere in the argv rather
// than assuming it is the first positional form.
if let Some((program, script, login)) = command.windows(3).find_map(|parts| match parts {
[program, flag, script] if flag == "-c" => {
Some((program.to_owned(), script.to_owned(), false))
if let [program, flag, script, ..] = command {
if flag == "-c" {
return Ok(ParsedShellCommand {
program: program.to_owned(),
script: script.to_owned(),
login: false,
});
}
[program, flag, script] if flag == "-lc" => {
Some((program.to_owned(), script.to_owned(), true))
if flag == "-lc" {
return Ok(ParsedShellCommand {
program: program.to_owned(),
script: script.to_owned(),
login: true,
});
}
_ => None,
}) {
return Ok(ParsedShellCommand {
program,
script,
login,
});
}
Err(ToolError::Rejected(

View File

@@ -6,8 +6,10 @@ use super::ParsedShellCommand;
use super::commands_for_intercepted_exec_policy;
use super::evaluate_intercepted_exec_policy;
use super::extract_shell_script;
use super::is_unconfigured_zsh_exec;
use super::join_program_and_argv;
use super::map_exec_result;
use super::resolve_host_zsh_path;
#[cfg(target_os = "macos")]
use crate::config::Constrained;
#[cfg(target_os = "macos")]
@@ -50,6 +52,7 @@ use codex_utils_absolute_path::AbsolutePathBuf;
use pretty_assertions::assert_eq;
#[cfg(target_os = "macos")]
use std::collections::HashMap;
use std::path::Path;
use std::path::PathBuf;
use std::time::Duration;
@@ -203,39 +206,16 @@ fn extract_shell_script_preserves_login_flag() {
}
#[test]
fn extract_shell_script_supports_wrapped_command_prefixes() {
assert_eq!(
extract_shell_script(&[
"/usr/bin/env".into(),
"CODEX_EXECVE_WRAPPER=1".into(),
"/bin/zsh".into(),
"-lc".into(),
"echo hello".into()
])
.unwrap(),
ParsedShellCommand {
program: "/bin/zsh".to_string(),
script: "echo hello".to_string(),
login: true,
}
);
assert_eq!(
extract_shell_script(&[
"sandbox-exec".into(),
"-p".into(),
"sandbox_policy".into(),
"/bin/zsh".into(),
"-c".into(),
"pwd".into(),
])
.unwrap(),
ParsedShellCommand {
program: "/bin/zsh".to_string(),
script: "pwd".to_string(),
login: false,
}
);
fn extract_shell_script_rejects_wrapped_command_prefixes() {
let err = extract_shell_script(&[
"/usr/bin/env".into(),
"CODEX_EXECVE_WRAPPER=1".into(),
"/bin/zsh".into(),
"-lc".into(),
"echo hello".into(),
])
.unwrap_err();
assert!(matches!(err, super::ToolError::Rejected(_)));
}
#[test]
@@ -274,6 +254,80 @@ fn join_program_and_argv_replaces_original_argv_zero() {
);
}
#[test]
fn is_unconfigured_zsh_exec_matches_non_configured_zsh_paths() {
let program = AbsolutePathBuf::try_from(host_absolute_path(&["bin", "zsh"])).unwrap();
let host = PathBuf::from(host_absolute_path(&["bin", "zsh"]));
let configured = PathBuf::from(host_absolute_path(&["tmp", "codex-zsh"]));
assert!(is_unconfigured_zsh_exec(
&program,
Some(configured.as_path()),
Some(host.as_path()),
));
}
#[test]
fn is_unconfigured_zsh_exec_ignores_non_zsh_or_configured_paths() {
let configured = PathBuf::from(host_absolute_path(&["tmp", "codex-zsh"]));
let host = PathBuf::from(host_absolute_path(&["bin", "zsh"]));
let configured_program = AbsolutePathBuf::try_from(configured.clone()).unwrap();
assert!(!is_unconfigured_zsh_exec(
&configured_program,
Some(configured.as_path()),
Some(host.as_path()),
));
let non_zsh =
AbsolutePathBuf::try_from(host_absolute_path(&["usr", "bin", "python3"])).unwrap();
assert!(!is_unconfigured_zsh_exec(
&non_zsh,
Some(configured.as_path()),
Some(host.as_path()),
));
assert!(!is_unconfigured_zsh_exec(
&non_zsh,
None,
Some(host.as_path()),
));
}
#[test]
fn is_unconfigured_zsh_exec_does_not_match_non_host_zsh_named_binaries() {
let program = AbsolutePathBuf::try_from(host_absolute_path(&["tmp", "repo", "zsh"])).unwrap();
let configured = PathBuf::from(host_absolute_path(&["tmp", "codex-zsh"]));
let host = PathBuf::from(host_absolute_path(&["bin", "zsh"]));
assert!(!is_unconfigured_zsh_exec(
&program,
Some(configured.as_path()),
Some(host.as_path()),
));
}
#[test]
fn resolve_host_zsh_path_ignores_repo_local_path_shadowing() {
let shadow_dir = tempfile::tempdir().expect("create shadow dir");
let cwd_dir = tempfile::tempdir().expect("create cwd dir");
let fake_zsh = shadow_dir.path().join("zsh");
std::fs::write(&fake_zsh, "#!/bin/sh\nexit 0\n").expect("write fake zsh");
#[cfg(unix)]
{
use std::os::unix::fs::PermissionsExt;
let mut permissions = std::fs::metadata(&fake_zsh)
.expect("metadata for fake zsh")
.permissions();
permissions.set_mode(0o755);
std::fs::set_permissions(&fake_zsh, permissions).expect("chmod fake zsh");
}
let path_env =
std::env::join_paths([shadow_dir.path(), Path::new("/usr/bin"), Path::new("/bin")])
.expect("join PATH")
.into_string()
.expect("PATH should be UTF-8");
let resolved = resolve_host_zsh_path(Some(&path_env), cwd_dir.path());
assert_ne!(resolved.as_deref(), Some(fake_zsh.as_path()));
}
#[test]
fn commands_for_intercepted_exec_policy_parses_plain_shell_wrappers() {
let program = AbsolutePathBuf::try_from(host_absolute_path(&["bin", "bash"])).unwrap();
@@ -660,6 +714,7 @@ async fn prepare_escalated_exec_turn_default_preserves_macos_seatbelt_extensions
file_system_sandbox_policy: read_only_file_system_sandbox_policy(),
network_sandbox_policy: NetworkSandboxPolicy::Restricted,
windows_sandbox_level: WindowsSandboxLevel::Disabled,
windows_sandbox_private_desktop: false,
sandbox_permissions: SandboxPermissions::UseDefault,
justification: None,
arg0: None,
@@ -670,6 +725,8 @@ async fn prepare_escalated_exec_turn_default_preserves_macos_seatbelt_extensions
}),
codex_linux_sandbox_exe: None,
use_legacy_landlock: false,
shell_zsh_path: None,
host_zsh_path: None,
};
let prepared = executor
@@ -712,6 +769,7 @@ async fn prepare_escalated_exec_permissions_preserve_macos_seatbelt_extensions()
file_system_sandbox_policy: unrestricted_file_system_sandbox_policy(),
network_sandbox_policy: NetworkSandboxPolicy::Enabled,
windows_sandbox_level: WindowsSandboxLevel::Disabled,
windows_sandbox_private_desktop: false,
sandbox_permissions: SandboxPermissions::UseDefault,
justification: None,
arg0: None,
@@ -719,6 +777,8 @@ async fn prepare_escalated_exec_permissions_preserve_macos_seatbelt_extensions()
macos_seatbelt_profile_extensions: None,
codex_linux_sandbox_exe: None,
use_legacy_landlock: false,
shell_zsh_path: None,
host_zsh_path: None,
};
let permissions = Permissions {
@@ -787,6 +847,7 @@ async fn prepare_escalated_exec_permission_profile_unions_turn_and_requested_mac
file_system_sandbox_policy: read_only_file_system_sandbox_policy(),
network_sandbox_policy: NetworkSandboxPolicy::from(&sandbox_policy),
windows_sandbox_level: WindowsSandboxLevel::Disabled,
windows_sandbox_private_desktop: false,
sandbox_permissions: SandboxPermissions::UseDefault,
justification: None,
arg0: None,
@@ -797,6 +858,8 @@ async fn prepare_escalated_exec_permission_profile_unions_turn_and_requested_mac
}),
codex_linux_sandbox_exe: None,
use_legacy_landlock: false,
shell_zsh_path: None,
host_zsh_path: None,
};
let prepared = executor

View File

@@ -36,9 +36,10 @@ pub(crate) async fn maybe_prepare_unified_exec(
req: &UnifiedExecRequest,
attempt: &SandboxAttempt<'_>,
ctx: &ToolCtx,
shell_command: &[String],
exec_request: ExecRequest,
) -> Result<Option<PreparedUnifiedExecSpawn>, ToolError> {
imp::maybe_prepare_unified_exec(req, attempt, ctx, exec_request).await
imp::maybe_prepare_unified_exec(req, attempt, ctx, shell_command, exec_request).await
}
#[cfg(unix)]
@@ -51,21 +52,29 @@ mod imp {
#[derive(Debug)]
struct ZshForkSpawnLifecycle {
escalation_session: EscalationSession,
escalation_session: Option<EscalationSession>,
}
impl SpawnLifecycle for ZshForkSpawnLifecycle {
fn inherited_fds(&self) -> Vec<i32> {
self.escalation_session
.env()
.get(ESCALATE_SOCKET_ENV_VAR)
.as_ref()
.and_then(|escalation_session| {
escalation_session.env().get(ESCALATE_SOCKET_ENV_VAR)
})
.and_then(|fd| fd.parse().ok())
.into_iter()
.collect()
}
fn after_spawn(&mut self) {
self.escalation_session.close_client_socket();
if let Some(escalation_session) = self.escalation_session.as_ref() {
escalation_session.close_client_socket();
}
}
fn after_exit(&mut self) {
self.escalation_session = None;
}
}
@@ -82,10 +91,17 @@ mod imp {
req: &UnifiedExecRequest,
attempt: &SandboxAttempt<'_>,
ctx: &ToolCtx,
shell_command: &[String],
exec_request: ExecRequest,
) -> Result<Option<PreparedUnifiedExecSpawn>, ToolError> {
let Some(prepared) =
unix_escalation::prepare_unified_exec_zsh_fork(req, attempt, ctx, exec_request).await?
let Some(prepared) = unix_escalation::prepare_unified_exec_zsh_fork(
req,
attempt,
ctx,
shell_command,
exec_request,
)
.await?
else {
return Ok(None);
};
@@ -93,7 +109,7 @@ mod imp {
Ok(Some(PreparedUnifiedExecSpawn {
exec_request: prepared.exec_request,
spawn_lifecycle: Box::new(ZshForkSpawnLifecycle {
escalation_session: prepared.escalation_session,
escalation_session: Some(prepared.escalation_session),
}),
}))
}
@@ -117,9 +133,10 @@ mod imp {
req: &UnifiedExecRequest,
attempt: &SandboxAttempt<'_>,
ctx: &ToolCtx,
shell_command: &[String],
exec_request: ExecRequest,
) -> Result<Option<PreparedUnifiedExecSpawn>, ToolError> {
let _ = (req, attempt, ctx, exec_request);
let _ = (req, attempt, ctx, shell_command, exec_request);
Ok(None)
}
}

View File

@@ -222,7 +222,11 @@ impl<'a> ToolRuntime<UnifiedExecRequest, UnifiedExecProcess> for UnifiedExecRunt
let exec_env = attempt
.env_for(spec, req.network.as_ref())
.map_err(|err| ToolError::Codex(err.into()))?;
match zsh_fork_backend::maybe_prepare_unified_exec(req, attempt, ctx, exec_env).await? {
match zsh_fork_backend::maybe_prepare_unified_exec(
req, attempt, ctx, &command, exec_env,
)
.await?
{
Some(prepared) => {
return self
.manager

View File

@@ -314,8 +314,6 @@ impl ToolsConfig {
);
let shell_type = if !features.enabled(Feature::ShellTool) {
ConfigShellToolType::Disabled
} else if features.enabled(Feature::ShellZshFork) {
ConfigShellToolType::ShellCommand
} else if features.enabled(Feature::UnifiedExec) && unified_exec_allowed {
// If ConPTY not supported (for old Windows versions), fallback on ShellCommand.
if codex_utils_pty::conpty_supported() {
@@ -326,6 +324,8 @@ impl ToolsConfig {
} else if model_info.shell_type == ConfigShellToolType::UnifiedExec && !unified_exec_allowed
{
ConfigShellToolType::ShellCommand
} else if features.enabled(Feature::ShellZshFork) {
ConfigShellToolType::ShellCommand
} else {
model_info.shell_type
};
@@ -583,6 +583,7 @@ fn create_approval_parameters(
fn create_exec_command_tool(
allow_login_shell: bool,
exec_permission_approvals_enabled: bool,
unified_exec_backend: UnifiedExecBackendConfig,
) -> ToolSpec {
let mut properties = BTreeMap::from([
(
@@ -600,12 +601,6 @@ fn create_exec_command_tool(
),
},
),
(
"shell".to_string(),
JsonSchema::String {
description: Some("Shell binary to launch. Defaults to the user's default shell.".to_string()),
},
),
(
"tty".to_string(),
JsonSchema::Boolean {
@@ -633,6 +628,16 @@ fn create_exec_command_tool(
},
),
]);
if unified_exec_backend != UnifiedExecBackendConfig::ZshFork {
properties.insert(
"shell".to_string(),
JsonSchema::String {
description: Some(
"Shell binary to launch. Defaults to the user's default shell.".to_string(),
),
},
);
}
if allow_login_shell {
properties.insert(
"login".to_string(),
@@ -2529,6 +2534,7 @@ pub(crate) fn build_specs_with_discoverable_tools(
create_exec_command_tool(
config.allow_login_shell,
exec_permission_approvals_enabled,
config.unified_exec_backend,
),
true,
config.code_mode_enabled,

View File

@@ -446,7 +446,7 @@ fn test_full_toolset_specs_for_gpt5_codex_unified_exec_web_search() {
// Build expected from the same helpers used by the builder.
let mut expected: BTreeMap<String, ToolSpec> = BTreeMap::from([]);
for spec in [
create_exec_command_tool(true, false),
create_exec_command_tool(true, false, UnifiedExecBackendConfig::Direct),
create_write_stdin_tool(),
PLAN_TOOL.clone(),
create_request_user_input_tool(CollaborationModesConfig::default()),
@@ -1364,7 +1364,7 @@ fn test_build_specs_default_shell_present() {
}
#[test]
fn shell_zsh_fork_prefers_shell_command_over_unified_exec() {
fn shell_zsh_fork_uses_unified_exec_when_enabled() {
let config = test_config();
let model_info = ModelsManager::construct_model_info_offline_for_tests("o3", &config);
let mut features = Features::with_defaults();
@@ -1382,7 +1382,7 @@ fn shell_zsh_fork_prefers_shell_command_over_unified_exec() {
windows_sandbox_level: WindowsSandboxLevel::Disabled,
});
assert_eq!(tools_config.shell_type, ConfigShellToolType::ShellCommand);
assert_eq!(tools_config.shell_type, ConfigShellToolType::UnifiedExec);
assert_eq!(
tools_config.shell_command_backend,
ShellCommandBackendConfig::ZshFork
@@ -1391,6 +1391,19 @@ fn shell_zsh_fork_prefers_shell_command_over_unified_exec() {
tools_config.unified_exec_backend,
UnifiedExecBackendConfig::ZshFork
);
let (tools, _) = build_specs(&tools_config, Some(HashMap::new()), None, &[]).build();
let exec_spec = find_tool(&tools, "exec_command");
let ToolSpec::Function(exec_tool) = &exec_spec.spec else {
panic!("exec_command should be a function tool spec");
};
let JsonSchema::Object { properties, .. } = &exec_tool.parameters else {
panic!("exec_command parameters should be an object schema");
};
assert!(
!properties.contains_key("shell"),
"exec_command should omit `shell` when zsh-fork backend forces the configured shell",
);
}
#[test]

View File

@@ -1,6 +1,7 @@
#![allow(clippy::module_inception)]
use std::sync::Arc;
use std::sync::Mutex as StdMutex;
use std::sync::atomic::AtomicBool;
use std::sync::atomic::Ordering;
use tokio::sync::Mutex;
@@ -36,6 +37,13 @@ pub(crate) trait SpawnLifecycle: std::fmt::Debug + Send + Sync {
}
fn after_spawn(&mut self) {}
/// Releases resources that needed to stay alive until the child process was
/// fully launched or until the session was torn down.
///
/// This hook must tolerate being called during normal process exit as well
/// as early termination paths, and it is guaranteed to run at most once.
fn after_exit(&mut self) {}
}
pub(crate) type SpawnLifecycleHandle = Box<dyn SpawnLifecycle>;
@@ -66,7 +74,8 @@ pub(crate) struct UnifiedExecProcess {
output_drained: Arc<Notify>,
output_task: JoinHandle<()>,
sandbox_type: SandboxType,
_spawn_lifecycle: SpawnLifecycleHandle,
_spawn_lifecycle: Arc<StdMutex<SpawnLifecycleHandle>>,
spawn_lifecycle_released: Arc<AtomicBool>,
}
impl UnifiedExecProcess {
@@ -74,7 +83,7 @@ impl UnifiedExecProcess {
process_handle: ExecCommandSession,
initial_output_rx: tokio::sync::broadcast::Receiver<Vec<u8>>,
sandbox_type: SandboxType,
spawn_lifecycle: SpawnLifecycleHandle,
spawn_lifecycle: Arc<StdMutex<SpawnLifecycleHandle>>,
) -> Self {
let output_buffer = Arc::new(Mutex::new(HeadTailBuffer::default()));
let output_notify = Arc::new(Notify::new());
@@ -119,6 +128,19 @@ impl UnifiedExecProcess {
output_task,
sandbox_type,
_spawn_lifecycle: spawn_lifecycle,
spawn_lifecycle_released: Arc::new(AtomicBool::new(false)),
}
}
fn release_spawn_lifecycle(&self) {
if self
.spawn_lifecycle_released
.swap(true, std::sync::atomic::Ordering::AcqRel)
{
return;
}
if let Ok(mut lifecycle) = self._spawn_lifecycle.lock() {
lifecycle.after_exit();
}
}
@@ -157,6 +179,7 @@ impl UnifiedExecProcess {
}
pub(super) fn terminate(&self) {
self.release_spawn_lifecycle();
self.output_closed.store(true, Ordering::Release);
self.output_closed_notify.notify_waiters();
self.process_handle.terminate();
@@ -232,12 +255,19 @@ impl UnifiedExecProcess {
mut exit_rx,
} = spawned;
let output_rx = codex_utils_pty::combine_output_receivers(stdout_rx, stderr_rx);
let managed = Self::new(process_handle, output_rx, sandbox_type, spawn_lifecycle);
let spawn_lifecycle = Arc::new(StdMutex::new(spawn_lifecycle));
let managed = Self::new(
process_handle,
output_rx,
sandbox_type,
Arc::clone(&spawn_lifecycle),
);
let exit_ready = matches!(exit_rx.try_recv(), Ok(_) | Err(TryRecvError::Closed));
if exit_ready {
managed.signal_exit();
managed.release_spawn_lifecycle();
managed.check_for_sandbox_denial().await?;
return Ok(managed);
}
@@ -247,14 +277,22 @@ impl UnifiedExecProcess {
.is_ok()
{
managed.signal_exit();
managed.release_spawn_lifecycle();
managed.check_for_sandbox_denial().await?;
return Ok(managed);
}
tokio::spawn({
let cancellation_token = managed.cancellation_token.clone();
let spawn_lifecycle = Arc::clone(&spawn_lifecycle);
let spawn_lifecycle_released = Arc::clone(&managed.spawn_lifecycle_released);
async move {
let _ = exit_rx.await;
if !spawn_lifecycle_released.swap(true, Ordering::AcqRel)
&& let Ok(mut lifecycle) = spawn_lifecycle.lock()
{
lifecycle.after_exit();
}
cancellation_token.cancel();
}
});

View File

@@ -1,6 +1,7 @@
use anyhow::Context;
use std::fs;
use std::path::Path;
use std::process::Stdio;
use std::time::Duration;
pub async fn wait_for_pid_file(path: &Path) -> anyhow::Result<String> {
@@ -24,6 +25,7 @@ pub async fn wait_for_pid_file(path: &Path) -> anyhow::Result<String> {
pub fn process_is_alive(pid: &str) -> anyhow::Result<bool> {
let status = std::process::Command::new("kill")
.args(["-0", pid])
.stderr(Stdio::null())
.status()
.context("failed to probe process liveness with kill -0")?;
Ok(status.success())

View File

@@ -18,6 +18,10 @@ pub struct ZshForkRuntime {
}
impl ZshForkRuntime {
pub fn zsh_path(&self) -> &Path {
&self.zsh_path
}
fn apply_to_config(
&self,
config: &mut Config,
@@ -91,6 +95,29 @@ where
builder.build(server).await
}
pub async fn build_unified_exec_zsh_fork_test<F>(
server: &wiremock::MockServer,
runtime: ZshForkRuntime,
approval_policy: AskForApproval,
sandbox_policy: SandboxPolicy,
pre_build_hook: F,
) -> Result<TestCodex>
where
F: FnOnce(&Path) + Send + 'static,
{
let mut builder = test_codex()
.with_pre_build_hook(pre_build_hook)
.with_config(move |config| {
runtime.apply_to_config(config, approval_policy, sandbox_policy);
config.use_experimental_unified_exec_tool = true;
config
.features
.enable(Feature::UnifiedExec)
.expect("test config should allow feature update");
});
builder.build(server).await
}
fn find_test_zsh_path() -> Result<Option<PathBuf>> {
let repo_root = codex_utils_cargo_bin::repo_root()?;
let dotslash_zsh = repo_root.join("codex-rs/app-server/tests/suite/zsh");

View File

@@ -35,6 +35,7 @@ use core_test_support::test_codex::TestCodex;
use core_test_support::test_codex::test_codex;
use core_test_support::wait_for_event;
use core_test_support::wait_for_event_with_timeout;
use core_test_support::zsh_fork::build_unified_exec_zsh_fork_test;
use core_test_support::zsh_fork::build_zsh_fork_test;
use core_test_support::zsh_fork::restrictive_workspace_write_policy;
use core_test_support::zsh_fork::zsh_fork_runtime;
@@ -1985,6 +1986,158 @@ async fn approving_execpolicy_amendment_persists_policy_and_skips_future_prompts
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[cfg(unix)]
async fn unified_exec_zsh_fork_execpolicy_amendment_skips_later_subcommands() -> Result<()> {
skip_if_no_network!(Ok(()));
let Some(runtime) = zsh_fork_runtime("unified exec zsh-fork execpolicy amendment test")? else {
return Ok(());
};
let approval_policy = AskForApproval::UnlessTrusted;
let sandbox_policy = SandboxPolicy::new_read_only_policy();
let server = start_mock_server().await;
let test = build_unified_exec_zsh_fork_test(
&server,
runtime,
approval_policy,
sandbox_policy.clone(),
|_| {},
)
.await?;
let allow_prefix_path = test.cwd.path().join("allow-prefix-zsh-fork.txt");
let _ = fs::remove_file(&allow_prefix_path);
let call_id = "allow-prefix-zsh-fork";
let command = "touch allow-prefix-zsh-fork.txt && touch allow-prefix-zsh-fork.txt";
let event = exec_command_event(
call_id,
command,
Some(1_000),
SandboxPermissions::UseDefault,
None,
)?;
let _ = mount_sse_once(
&server,
sse(vec![
ev_response_created("resp-zsh-fork-allow-prefix-1"),
event,
ev_completed("resp-zsh-fork-allow-prefix-1"),
]),
)
.await;
let results = mount_sse_once(
&server,
sse(vec![
ev_assistant_message("msg-zsh-fork-allow-prefix-1", "done"),
ev_completed("resp-zsh-fork-allow-prefix-2"),
]),
)
.await;
submit_turn(
&test,
"allow-prefix-zsh-fork",
approval_policy,
sandbox_policy,
)
.await?;
let expected_execpolicy_amendment = ExecPolicyAmendment::new(vec![
"touch".to_string(),
"allow-prefix-zsh-fork.txt".to_string(),
]);
let mut saw_parent_approval = false;
let mut saw_subcommand_approval = false;
loop {
let event = wait_for_event(&test.codex, |event| {
matches!(
event,
EventMsg::ExecApprovalRequest(_) | EventMsg::TurnComplete(_)
)
})
.await;
match event {
EventMsg::TurnComplete(_) => break,
EventMsg::ExecApprovalRequest(approval) => {
let command_parts = approval.command.clone();
let last_arg = command_parts.last().map(String::as_str).unwrap_or_default();
if last_arg == command {
assert!(
!saw_parent_approval,
"unexpected duplicate parent approval: {command_parts:?}"
);
saw_parent_approval = true;
test.codex
.submit(Op::ExecApproval {
id: approval.effective_approval_id(),
turn_id: None,
decision: ReviewDecision::Approved,
})
.await?;
continue;
}
let is_touch_subcommand = command_parts
.iter()
.any(|part| part == "allow-prefix-zsh-fork.txt")
&& command_parts
.first()
.is_some_and(|part| part.ends_with("/touch") || part == "touch");
if is_touch_subcommand {
assert!(
!saw_subcommand_approval,
"execpolicy amendment should suppress later matching subcommand approvals: {command_parts:?}"
);
saw_subcommand_approval = true;
assert_eq!(
approval.proposed_execpolicy_amendment,
Some(expected_execpolicy_amendment.clone())
);
test.codex
.submit(Op::ExecApproval {
id: approval.effective_approval_id(),
turn_id: None,
decision: ReviewDecision::ApprovedExecpolicyAmendment {
proposed_execpolicy_amendment: expected_execpolicy_amendment
.clone(),
},
})
.await?;
continue;
}
test.codex
.submit(Op::ExecApproval {
id: approval.effective_approval_id(),
turn_id: None,
decision: ReviewDecision::Approved,
})
.await?;
}
other => panic!("unexpected event: {other:?}"),
}
}
assert!(saw_parent_approval, "expected parent unified-exec approval");
assert!(
saw_subcommand_approval,
"expected at least one intercepted touch approval"
);
let result = parse_result(&results.single_request().function_call_output(call_id));
assert_eq!(result.exit_code.unwrap_or(0), 0);
assert!(
allow_prefix_path.exists(),
"expected touch command to complete after approving the first intercepted subcommand; output: {}",
result.stdout
);
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[cfg(unix)]
async fn matched_prefix_rule_runs_unsandboxed_under_zsh_fork() -> Result<()> {

View File

@@ -20,6 +20,7 @@ use core_test_support::skip_if_no_network;
use core_test_support::test_codex::TestCodex;
use core_test_support::wait_for_event;
use core_test_support::wait_for_event_match;
use core_test_support::zsh_fork::build_unified_exec_zsh_fork_test;
use core_test_support::zsh_fork::build_zsh_fork_test;
use core_test_support::zsh_fork::restrictive_workspace_write_policy;
use core_test_support::zsh_fork::zsh_fork_runtime;
@@ -50,6 +51,13 @@ fn shell_command_arguments(command: &str) -> Result<String> {
}))?)
}
fn exec_command_arguments(command: &str) -> Result<String> {
Ok(serde_json::to_string(&json!({
"cmd": command,
"yield_time_ms": 500,
}))?)
}
async fn submit_turn_with_policies(
test: &TestCodex,
prompt: &str,
@@ -89,6 +97,38 @@ echo 'zsh-fork-stderr' >&2
)
}
#[cfg(unix)]
fn write_repo_skill_with_shell_script_contents(
repo_root: &Path,
name: &str,
script_name: &str,
script_contents: &str,
) -> Result<PathBuf> {
use std::os::unix::fs::PermissionsExt;
let skill_dir = repo_root.join(".agents").join("skills").join(name);
let scripts_dir = skill_dir.join("scripts");
fs::create_dir_all(&scripts_dir)?;
fs::write(repo_root.join(".git"), "gitdir: here")?;
fs::write(
skill_dir.join("SKILL.md"),
format!(
r#"---
name: {name}
description: {name} skill
---
"#
),
)?;
let script_path = scripts_dir.join(script_name);
fs::write(&script_path, script_contents)?;
let mut permissions = fs::metadata(&script_path)?.permissions();
permissions.set_mode(0o755);
fs::set_permissions(&script_path, permissions)?;
Ok(script_path)
}
#[cfg(unix)]
fn write_skill_with_shell_script_contents(
home: &Path,
@@ -541,6 +581,168 @@ permissions:
Ok(())
}
#[cfg(unix)]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn unified_exec_zsh_fork_prompts_for_skill_script_execution() -> Result<()> {
skip_if_no_network!(Ok(()));
let Some(runtime) = zsh_fork_runtime("unified exec zsh-fork skill prompt test")? else {
return Ok(());
};
let server = start_mock_server().await;
let tool_call_id = "uexec-zsh-fork-skill-call";
let test = build_unified_exec_zsh_fork_test(
&server,
runtime,
AskForApproval::OnRequest,
SandboxPolicy::new_workspace_write_policy(),
|home| {
write_skill_with_shell_script(home, "mbolin-test-skill", "hello-mbolin.sh").unwrap();
write_skill_metadata(
home,
"mbolin-test-skill",
r#"
permissions:
file_system:
read:
- "./data"
write:
- "./output"
"#,
)
.unwrap();
},
)
.await?;
let (script_path_str, command) = skill_script_command(&test, "hello-mbolin.sh")?;
let arguments = exec_command_arguments(&command)?;
let mocks =
mount_function_call_agent_response(&server, tool_call_id, &arguments, "exec_command").await;
submit_turn_with_policies(
&test,
"use $mbolin-test-skill",
AskForApproval::OnRequest,
SandboxPolicy::new_workspace_write_policy(),
)
.await?;
let approval = wait_for_exec_approval_request(&test)
.await
.expect("expected exec approval request before completion");
assert_eq!(approval.call_id, tool_call_id);
assert_eq!(approval.command, vec![script_path_str.clone()]);
assert_eq!(
approval.available_decisions,
Some(vec![
ReviewDecision::Approved,
ReviewDecision::ApprovedForSession,
ReviewDecision::Abort,
])
);
assert_eq!(
approval.additional_permissions,
Some(PermissionProfile {
file_system: Some(FileSystemPermissions {
read: Some(vec![absolute_path(
&test.codex_home_path().join("skills/mbolin-test-skill/data"),
)]),
write: Some(vec![absolute_path(
&test
.codex_home_path()
.join("skills/mbolin-test-skill/output"),
)]),
}),
..Default::default()
})
);
test.codex
.submit(Op::ExecApproval {
id: approval.effective_approval_id(),
turn_id: None,
decision: ReviewDecision::Denied,
})
.await?;
wait_for_turn_complete(&test).await;
let call_output = mocks
.completion
.single_request()
.function_call_output(tool_call_id);
let output = call_output["output"].as_str().unwrap_or_default();
assert!(
output.contains("Execution denied: User denied execution"),
"expected rejection marker in function_call_output: {output:?}"
);
Ok(())
}
#[cfg(unix)]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn unified_exec_zsh_fork_keeps_skill_loading_pinned_to_turn_cwd() -> Result<()> {
skip_if_no_network!(Ok(()));
let Some(runtime) = zsh_fork_runtime("unified exec zsh-fork turn cwd skill test")? else {
return Ok(());
};
let server = start_mock_server().await;
let tool_call_id = "uexec-zsh-fork-repo-skill-call";
let test = build_unified_exec_zsh_fork_test(
&server,
runtime,
AskForApproval::OnRequest,
SandboxPolicy::new_workspace_write_policy(),
|_| {},
)
.await?;
let repo_root = test.cwd_path().join("repo");
let script_path = write_repo_skill_with_shell_script_contents(
&repo_root,
"repo-skill",
"repo-skill.sh",
"#!/bin/sh\necho 'repo-skill-output'\n",
)?;
let script_path_quoted = shlex::try_join([script_path.to_string_lossy().as_ref()])?;
let repo_root_quoted = shlex::try_join([repo_root.to_string_lossy().as_ref()])?;
let command = format!("cd {repo_root_quoted} && {script_path_quoted}");
let arguments = exec_command_arguments(&command)?;
let mocks =
mount_function_call_agent_response(&server, tool_call_id, &arguments, "exec_command").await;
submit_turn_with_policies(
&test,
"run the repo skill after changing directories",
AskForApproval::OnRequest,
SandboxPolicy::new_workspace_write_policy(),
)
.await?;
let approval = wait_for_exec_approval_request(&test).await;
assert!(
approval.is_none(),
"changing directories inside unified exec should not load repo-local skills from the shell cwd",
);
let call_output = mocks
.completion
.single_request()
.function_call_output(tool_call_id);
let output = call_output["output"].as_str().unwrap_or_default();
assert!(
output.contains("repo-skill-output"),
"expected repo skill script to run without skill-specific approval when only the shell cwd changes: {output:?}"
);
Ok(())
}
/// Permissionless skills should inherit the turn sandbox without prompting.
#[cfg(unix)]
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]

View File

@@ -1,6 +1,8 @@
use std::collections::HashMap;
use std::ffi::OsStr;
use std::fs;
#[cfg(unix)]
use std::path::PathBuf;
use std::sync::OnceLock;
use anyhow::Context;
@@ -32,6 +34,10 @@ use core_test_support::test_codex::test_codex;
use core_test_support::wait_for_event;
use core_test_support::wait_for_event_match;
use core_test_support::wait_for_event_with_timeout;
#[cfg(unix)]
use core_test_support::zsh_fork::build_unified_exec_zsh_fork_test;
#[cfg(unix)]
use core_test_support::zsh_fork::zsh_fork_runtime;
use pretty_assertions::assert_eq;
use regex_lite::Regex;
use serde_json::Value;
@@ -155,6 +161,27 @@ fn collect_tool_outputs(bodies: &[Value]) -> Result<HashMap<String, ParsedUnifie
Ok(outputs)
}
#[cfg(unix)]
fn process_text_binary_path(pid: &str) -> Result<PathBuf> {
let output = std::process::Command::new("lsof")
.args(["-Fn", "-a", "-p", pid, "-d", "txt"])
.output()
.with_context(|| format!("failed to inspect process {pid} executable mapping with lsof"))?;
if !output.status.success() {
return Err(anyhow::anyhow!(
"lsof failed for pid {pid} with status {:?}",
output.status.code()
));
}
let stdout = String::from_utf8(output.stdout).context("lsof output was not UTF-8")?;
let path = stdout
.lines()
.find_map(|line| line.strip_prefix('n'))
.ok_or_else(|| anyhow::anyhow!("lsof did not report a text binary path for pid {pid}"))?;
Ok(PathBuf::from(path))
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn unified_exec_intercepts_apply_patch_exec_command() -> Result<()> {
skip_if_no_network!(Ok(()));
@@ -1323,7 +1350,6 @@ async fn exec_command_reports_chunk_and_exit_metadata() -> Result<()> {
.into_iter()
.map(|request| request.body_json())
.collect::<Vec<_>>();
let outputs = collect_tool_outputs(&bodies)?;
let metadata = outputs
.get(call_id)
@@ -1445,7 +1471,6 @@ async fn unified_exec_defaults_to_pipe() -> Result<()> {
.into_iter()
.map(|request| request.body_json())
.collect::<Vec<_>>();
let outputs = collect_tool_outputs(&bodies)?;
let output = outputs
.get(call_id)
@@ -1539,7 +1564,6 @@ async fn unified_exec_can_enable_tty() -> Result<()> {
.into_iter()
.map(|request| request.body_json())
.collect::<Vec<_>>();
let outputs = collect_tool_outputs(&bodies)?;
let output = outputs
.get(call_id)
@@ -1624,7 +1648,6 @@ async fn unified_exec_respects_early_exit_notifications() -> Result<()> {
.into_iter()
.map(|request| request.body_json())
.collect::<Vec<_>>();
let outputs = collect_tool_outputs(&bodies)?;
let output = outputs
.get(call_id)
@@ -1823,6 +1846,350 @@ async fn write_stdin_returns_exit_metadata_and_clears_session() -> Result<()> {
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[cfg(unix)]
async fn unified_exec_zsh_fork_keeps_python_repl_attached_to_zsh_session() -> Result<()> {
skip_if_no_network!(Ok(()));
let Some(runtime) = zsh_fork_runtime("unified exec zsh-fork tty session test")? else {
return Ok(());
};
let configured_zsh_path =
fs::canonicalize(runtime.zsh_path()).unwrap_or_else(|_| runtime.zsh_path().to_path_buf());
let python = match which("python3") {
Ok(path) => path,
Err(_) => {
eprintln!("python3 not found in PATH, skipping zsh-fork python repl test.");
return Ok(());
}
};
let server = start_mock_server().await;
let test = build_unified_exec_zsh_fork_test(
&server,
runtime,
AskForApproval::Never,
SandboxPolicy::new_workspace_write_policy(),
|_| {},
)
.await?;
let start_call_id = "uexec-zsh-fork-python-start";
let send_call_id = "uexec-zsh-fork-python-pid";
let exit_call_id = "uexec-zsh-fork-python-exit";
let start_command = format!("{}; :", python.display());
let start_args = serde_json::json!({
"cmd": start_command,
"yield_time_ms": 500,
"tty": true,
});
let send_args = serde_json::json!({
"chars": "import os; print('CODEX_PY_PID=' + str(os.getpid()))\r\n",
"session_id": 1000,
"yield_time_ms": 500,
});
let exit_args = serde_json::json!({
"chars": "import sys; sys.exit(0)\r\n",
"session_id": 1000,
"yield_time_ms": 500,
});
let responses = vec![
sse(vec![
ev_response_created("resp-1"),
ev_function_call(
start_call_id,
"exec_command",
&serde_json::to_string(&start_args)?,
),
ev_completed("resp-1"),
]),
sse(vec![
ev_response_created("resp-2"),
ev_function_call(
send_call_id,
"write_stdin",
&serde_json::to_string(&send_args)?,
),
ev_completed("resp-2"),
]),
sse(vec![
ev_assistant_message("msg-1", "python is running"),
ev_completed("resp-3"),
]),
sse(vec![
ev_response_created("resp-4"),
ev_function_call(
exit_call_id,
"write_stdin",
&serde_json::to_string(&exit_args)?,
),
ev_completed("resp-4"),
]),
sse(vec![
ev_assistant_message("msg-2", "all done"),
ev_completed("resp-5"),
]),
];
let request_log = mount_sse_sequence(&server, responses).await;
test.codex
.submit(Op::UserTurn {
items: vec![UserInput::Text {
text: "test unified exec zsh-fork tty behavior".into(),
text_elements: Vec::new(),
}],
final_output_json_schema: None,
cwd: test.cwd_path().to_path_buf(),
approval_policy: AskForApproval::Never,
sandbox_policy: SandboxPolicy::new_workspace_write_policy(),
model: test.session_configured.model.clone(),
effort: None,
summary: None,
service_tier: None,
collaboration_mode: None,
personality: None,
})
.await?;
wait_for_event(&test.codex, |event| {
matches!(event, EventMsg::TurnComplete(_))
})
.await;
let requests = request_log.requests();
assert!(!requests.is_empty(), "expected at least one POST request");
let bodies = requests
.into_iter()
.map(|request| request.body_json())
.collect::<Vec<_>>();
let outputs = collect_tool_outputs(&bodies)?;
let start_output = outputs
.get(start_call_id)
.expect("missing start output for exec_command");
let process_id = start_output
.process_id
.clone()
.expect("expected process id from exec_command");
assert!(
start_output.exit_code.is_none(),
"initial exec_command should leave the PTY session running"
);
let send_output = outputs
.get(send_call_id)
.expect("missing write_stdin output");
let normalized = send_output.output.replace("\r\n", "\n");
let python_pid = Regex::new(r"CODEX_PY_PID=(\d+)")
.expect("valid python pid marker regex")
.captures(&normalized)
.and_then(|captures| captures.get(1))
.map(|value| value.as_str().to_string())
.with_context(|| format!("missing python pid in output {normalized:?}"))?;
assert!(
process_is_alive(&python_pid)?,
"python process should still be alive after printing its pid, got output {normalized:?}"
);
assert_eq!(send_output.process_id.as_deref(), Some(process_id.as_str()));
assert!(
send_output.exit_code.is_none(),
"write_stdin should not report an exit code while the process is still running"
);
let zsh_pid = std::process::Command::new("ps")
.args(["-o", "ppid=", "-p", &python_pid])
.output()
.context("failed to look up python parent pid")?;
let zsh_pid = String::from_utf8(zsh_pid.stdout)
.context("python parent pid output is not UTF-8")?
.trim()
.to_string();
assert!(
!zsh_pid.is_empty(),
"expected python parent pid to identify the zsh session"
);
assert!(
process_is_alive(&zsh_pid)?,
"expected zsh parent process {zsh_pid} to still be alive"
);
let zsh_command = std::process::Command::new("ps")
.args(["-o", "command=", "-p", &zsh_pid])
.output()
.context("failed to look up zsh parent command")?;
let zsh_command =
String::from_utf8(zsh_command.stdout).context("zsh parent command output is not UTF-8")?;
assert!(
zsh_command.contains("zsh"),
"expected python parent command to be zsh, got {zsh_command:?}"
);
let zsh_text_binary = process_text_binary_path(&zsh_pid)?;
let zsh_text_binary = fs::canonicalize(&zsh_text_binary).unwrap_or(zsh_text_binary);
assert_eq!(
zsh_text_binary, configured_zsh_path,
"python parent shell should run with configured zsh-fork binary, got {:?} ({zsh_command:?})",
zsh_text_binary,
);
test.codex
.submit(Op::UserTurn {
items: vec![UserInput::Text {
text: "shut down the python repl".into(),
text_elements: Vec::new(),
}],
final_output_json_schema: None,
cwd: test.cwd_path().to_path_buf(),
approval_policy: AskForApproval::Never,
sandbox_policy: SandboxPolicy::new_workspace_write_policy(),
model: test.session_configured.model.clone(),
effort: None,
summary: None,
service_tier: None,
collaboration_mode: None,
personality: None,
})
.await?;
wait_for_event(&test.codex, |event| {
matches!(event, EventMsg::TurnComplete(_))
})
.await;
let requests = request_log.requests();
assert!(!requests.is_empty(), "expected at least one POST request");
let bodies = requests
.into_iter()
.map(|request| request.body_json())
.collect::<Vec<_>>();
let outputs = collect_tool_outputs(&bodies)?;
let exit_output = outputs
.get(exit_call_id)
.expect("missing exit output after requesting python shutdown");
assert!(
exit_output.exit_code.is_none() || exit_output.exit_code == Some(0),
"exit request should either leave cleanup to the background watcher or report success directly, got {exit_output:?}"
);
wait_for_process_exit(&python_pid).await?;
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[cfg(unix)]
async fn unified_exec_zsh_fork_rewrites_nested_zsh_exec_to_configured_binary() -> Result<()> {
skip_if_no_network!(Ok(()));
let Some(runtime) = zsh_fork_runtime("unified exec zsh-fork nested zsh rewrite test")? else {
return Ok(());
};
let configured_zsh_path =
fs::canonicalize(runtime.zsh_path()).unwrap_or_else(|_| runtime.zsh_path().to_path_buf());
let host_zsh = match which("zsh") {
Ok(path) => path,
Err(_) => {
eprintln!("zsh not found in PATH, skipping nested zsh rewrite test.");
return Ok(());
}
};
let server = start_mock_server().await;
let test = build_unified_exec_zsh_fork_test(
&server,
runtime,
AskForApproval::Never,
SandboxPolicy::new_workspace_write_policy(),
|_| {},
)
.await?;
let start_call_id = "uexec-zsh-fork-nested-start";
let nested_command = format!(
"exec {} -lc 'echo CODEX_NESTED_ZSH_PID=$$; sleep 3; :'",
host_zsh.display(),
);
let start_args = serde_json::json!({
"cmd": nested_command,
"yield_time_ms": 500,
"tty": true,
});
let responses = vec![
sse(vec![
ev_response_created("resp-nested-1"),
ev_function_call(
start_call_id,
"exec_command",
&serde_json::to_string(&start_args)?,
),
ev_completed("resp-nested-1"),
]),
sse(vec![
ev_assistant_message("msg-nested-1", "done"),
ev_completed("resp-nested-2"),
]),
];
let request_log = mount_sse_sequence(&server, responses).await;
test.codex
.submit(Op::UserTurn {
items: vec![UserInput::Text {
text: "test nested zsh rewrite behavior".into(),
text_elements: Vec::new(),
}],
final_output_json_schema: None,
cwd: test.cwd_path().to_path_buf(),
approval_policy: AskForApproval::Never,
sandbox_policy: SandboxPolicy::new_workspace_write_policy(),
model: test.session_configured.model.clone(),
effort: None,
summary: None,
service_tier: None,
collaboration_mode: None,
personality: None,
})
.await?;
wait_for_event(&test.codex, |event| {
matches!(event, EventMsg::TurnComplete(_))
})
.await;
let requests = request_log.requests();
let bodies = requests
.into_iter()
.map(|request| request.body_json())
.collect::<Vec<_>>();
let outputs = collect_tool_outputs(&bodies)?;
let start_output = outputs
.get(start_call_id)
.expect("missing start output for nested zsh exec_command");
let normalized = start_output.output.replace("\r\n", "\n");
let nested_zsh_pid = Regex::new(r"CODEX_NESTED_ZSH_PID=(\d+)")
.expect("valid nested zsh pid regex")
.captures(&normalized)
.and_then(|captures| captures.get(1))
.map(|value| value.as_str().to_string())
.with_context(|| format!("missing nested zsh pid marker in output {normalized:?}"))?;
assert!(
process_is_alive(&nested_zsh_pid)?,
"nested zsh process should be running before release, got output {normalized:?}"
);
let nested_text_binary = process_text_binary_path(&nested_zsh_pid)?;
let nested_text_binary = fs::canonicalize(&nested_text_binary).unwrap_or(nested_text_binary);
assert_eq!(
nested_text_binary, configured_zsh_path,
"nested zsh exec should be rewritten to configured zsh-fork binary, got {:?}",
nested_text_binary,
);
Ok(())
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn unified_exec_emits_end_event_when_session_dies_via_stdin() -> Result<()> {
skip_if_no_network!(Ok(()));