Phase 1: Repository & Infrastructure Setup

- Renamed directories: codex-rs -> llmx-rs, codex-cli -> llmx-cli
- Updated package.json files:
  - Root: llmx-monorepo
  - CLI: @llmx/llmx
  - SDK: @llmx/llmx-sdk
- Updated pnpm workspace configuration
- Renamed binary: codex.js -> llmx.js
- Updated environment variables: CODEX_* -> LLMX_*
- Changed repository URLs to valknar/llmx

🤖 Generated with Claude Code
This commit is contained in:
Sebastian Krüger
2025-11-11 14:01:52 +01:00
parent 052b052832
commit f237fe560d
1151 changed files with 41 additions and 35 deletions

109
llmx-rs/exec/src/cli.rs Normal file
View File

@@ -0,0 +1,109 @@
use clap::Parser;
use clap::ValueEnum;
use codex_common::CliConfigOverrides;
use std::path::PathBuf;
#[derive(Parser, Debug)]
#[command(version)]
pub struct Cli {
/// Action to perform. If omitted, runs a new non-interactive session.
#[command(subcommand)]
pub command: Option<Command>,
/// Optional image(s) to attach to the initial prompt.
#[arg(long = "image", short = 'i', value_name = "FILE", value_delimiter = ',', num_args = 1..)]
pub images: Vec<PathBuf>,
/// Model the agent should use.
#[arg(long, short = 'm')]
pub model: Option<String>,
#[arg(long = "oss", default_value_t = false)]
pub oss: bool,
/// Select the sandbox policy to use when executing model-generated shell
/// commands.
#[arg(long = "sandbox", short = 's', value_enum)]
pub sandbox_mode: Option<codex_common::SandboxModeCliArg>,
/// Configuration profile from config.toml to specify default options.
#[arg(long = "profile", short = 'p')]
pub config_profile: Option<String>,
/// Convenience alias for low-friction sandboxed automatic execution (-a on-failure, --sandbox workspace-write).
#[arg(long = "full-auto", default_value_t = false)]
pub full_auto: bool,
/// Skip all confirmation prompts and execute commands without sandboxing.
/// EXTREMELY DANGEROUS. Intended solely for running in environments that are externally sandboxed.
#[arg(
long = "dangerously-bypass-approvals-and-sandbox",
alias = "yolo",
default_value_t = false,
conflicts_with = "full_auto"
)]
pub dangerously_bypass_approvals_and_sandbox: bool,
/// Tell the agent to use the specified directory as its working root.
#[clap(long = "cd", short = 'C', value_name = "DIR")]
pub cwd: Option<PathBuf>,
/// Allow running Codex outside a Git repository.
#[arg(long = "skip-git-repo-check", default_value_t = false)]
pub skip_git_repo_check: bool,
/// Path to a JSON Schema file describing the model's final response shape.
#[arg(long = "output-schema", value_name = "FILE")]
pub output_schema: Option<PathBuf>,
#[clap(skip)]
pub config_overrides: CliConfigOverrides,
/// Specifies color settings for use in the output.
#[arg(long = "color", value_enum, default_value_t = Color::Auto)]
pub color: Color,
/// Print events to stdout as JSONL.
#[arg(long = "json", alias = "experimental-json", default_value_t = false)]
pub json: bool,
/// Specifies file where the last message from the agent should be written.
#[arg(long = "output-last-message", short = 'o', value_name = "FILE")]
pub last_message_file: Option<PathBuf>,
/// Initial instructions for the agent. If not provided as an argument (or
/// if `-` is used), instructions are read from stdin.
#[arg(value_name = "PROMPT", value_hint = clap::ValueHint::Other)]
pub prompt: Option<String>,
}
#[derive(Debug, clap::Subcommand)]
pub enum Command {
/// Resume a previous session by id or pick the most recent with --last.
Resume(ResumeArgs),
}
#[derive(Parser, Debug)]
pub struct ResumeArgs {
/// Conversation/session id (UUID). When provided, resumes this session.
/// If omitted, use --last to pick the most recent recorded session.
#[arg(value_name = "SESSION_ID")]
pub session_id: Option<String>,
/// Resume the most recent recorded session (newest) without specifying an id.
#[arg(long = "last", default_value_t = false, conflicts_with = "session_id")]
pub last: bool,
/// Prompt to send after resuming the session. If `-` is used, read from stdin.
#[arg(value_name = "PROMPT", value_hint = clap::ValueHint::Other)]
pub prompt: Option<String>,
}
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, ValueEnum)]
#[value(rename_all = "kebab-case")]
pub enum Color {
Always,
Never,
#[default]
Auto,
}

View File

@@ -0,0 +1,45 @@
use std::path::Path;
use codex_core::config::Config;
use codex_core::protocol::Event;
use codex_core::protocol::SessionConfiguredEvent;
pub(crate) enum CodexStatus {
Running,
InitiateShutdown,
Shutdown,
}
pub(crate) trait EventProcessor {
/// Print summary of effective configuration and user prompt.
fn print_config_summary(
&mut self,
config: &Config,
prompt: &str,
session_configured: &SessionConfiguredEvent,
);
/// Handle a single event emitted by the agent.
fn process_event(&mut self, event: Event) -> CodexStatus;
fn print_final_output(&mut self) {}
}
pub(crate) fn handle_last_message(last_agent_message: Option<&str>, output_file: &Path) {
let message = last_agent_message.unwrap_or_default();
write_last_message_file(message, Some(output_file));
if last_agent_message.is_none() {
eprintln!(
"Warning: no last agent message; wrote empty content to {}",
output_file.display()
);
}
}
fn write_last_message_file(contents: &str, last_message_path: Option<&Path>) {
if let Some(path) = last_message_path
&& let Err(e) = std::fs::write(path, contents)
{
eprintln!("Failed to write last message file {path:?}: {e}");
}
}

View File

@@ -0,0 +1,599 @@
use codex_common::elapsed::format_duration;
use codex_common::elapsed::format_elapsed;
use codex_core::config::Config;
use codex_core::protocol::AgentMessageEvent;
use codex_core::protocol::AgentReasoningRawContentEvent;
use codex_core::protocol::BackgroundEventEvent;
use codex_core::protocol::DeprecationNoticeEvent;
use codex_core::protocol::ErrorEvent;
use codex_core::protocol::Event;
use codex_core::protocol::EventMsg;
use codex_core::protocol::ExecCommandBeginEvent;
use codex_core::protocol::ExecCommandEndEvent;
use codex_core::protocol::FileChange;
use codex_core::protocol::McpInvocation;
use codex_core::protocol::McpToolCallBeginEvent;
use codex_core::protocol::McpToolCallEndEvent;
use codex_core::protocol::PatchApplyBeginEvent;
use codex_core::protocol::PatchApplyEndEvent;
use codex_core::protocol::SessionConfiguredEvent;
use codex_core::protocol::StreamErrorEvent;
use codex_core::protocol::TaskCompleteEvent;
use codex_core::protocol::TurnAbortReason;
use codex_core::protocol::TurnDiffEvent;
use codex_core::protocol::WarningEvent;
use codex_core::protocol::WebSearchEndEvent;
use codex_protocol::num_format::format_with_separators;
use owo_colors::OwoColorize;
use owo_colors::Style;
use shlex::try_join;
use std::collections::HashMap;
use std::path::PathBuf;
use std::time::Instant;
use crate::event_processor::CodexStatus;
use crate::event_processor::EventProcessor;
use crate::event_processor::handle_last_message;
use codex_common::create_config_summary_entries;
use codex_protocol::plan_tool::StepStatus;
use codex_protocol::plan_tool::UpdatePlanArgs;
/// This should be configurable. When used in CI, users may not want to impose
/// a limit so they can see the full transcript.
const MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL: usize = 20;
pub(crate) struct EventProcessorWithHumanOutput {
call_id_to_patch: HashMap<String, PatchApplyBegin>,
// To ensure that --color=never is respected, ANSI escapes _must_ be added
// using .style() with one of these fields. If you need a new style, add a
// new field here.
bold: Style,
italic: Style,
dimmed: Style,
magenta: Style,
red: Style,
green: Style,
cyan: Style,
yellow: Style,
/// Whether to include `AgentReasoning` events in the output.
show_agent_reasoning: bool,
show_raw_agent_reasoning: bool,
last_message_path: Option<PathBuf>,
last_total_token_usage: Option<codex_core::protocol::TokenUsageInfo>,
final_message: Option<String>,
}
impl EventProcessorWithHumanOutput {
pub(crate) fn create_with_ansi(
with_ansi: bool,
config: &Config,
last_message_path: Option<PathBuf>,
) -> Self {
let call_id_to_patch = HashMap::new();
if with_ansi {
Self {
call_id_to_patch,
bold: Style::new().bold(),
italic: Style::new().italic(),
dimmed: Style::new().dimmed(),
magenta: Style::new().magenta(),
red: Style::new().red(),
green: Style::new().green(),
cyan: Style::new().cyan(),
yellow: Style::new().yellow(),
show_agent_reasoning: !config.hide_agent_reasoning,
show_raw_agent_reasoning: config.show_raw_agent_reasoning,
last_message_path,
last_total_token_usage: None,
final_message: None,
}
} else {
Self {
call_id_to_patch,
bold: Style::new(),
italic: Style::new(),
dimmed: Style::new(),
magenta: Style::new(),
red: Style::new(),
green: Style::new(),
cyan: Style::new(),
yellow: Style::new(),
show_agent_reasoning: !config.hide_agent_reasoning,
show_raw_agent_reasoning: config.show_raw_agent_reasoning,
last_message_path,
last_total_token_usage: None,
final_message: None,
}
}
}
}
struct PatchApplyBegin {
start_time: Instant,
auto_approved: bool,
}
/// Timestamped helper. The timestamp is styled with self.dimmed.
macro_rules! ts_msg {
($self:ident, $($arg:tt)*) => {{
eprintln!($($arg)*);
}};
}
impl EventProcessor for EventProcessorWithHumanOutput {
/// Print a concise summary of the effective configuration that will be used
/// for the session. This mirrors the information shown in the TUI welcome
/// screen.
fn print_config_summary(
&mut self,
config: &Config,
prompt: &str,
session_configured_event: &SessionConfiguredEvent,
) {
const VERSION: &str = env!("CARGO_PKG_VERSION");
ts_msg!(
self,
"OpenAI Codex v{} (research preview)\n--------",
VERSION
);
let mut entries = create_config_summary_entries(config);
entries.push((
"session id",
session_configured_event.session_id.to_string(),
));
for (key, value) in entries {
eprintln!("{} {}", format!("{key}:").style(self.bold), value);
}
eprintln!("--------");
// Echo the prompt that will be sent to the agent so it is visible in the
// transcript/logs before any events come in. Note the prompt may have been
// read from stdin, so it may not be visible in the terminal otherwise.
ts_msg!(self, "{}\n{}", "user".style(self.cyan), prompt);
}
fn process_event(&mut self, event: Event) -> CodexStatus {
let Event { id: _, msg } = event;
match msg {
EventMsg::Error(ErrorEvent { message }) => {
let prefix = "ERROR:".style(self.red);
ts_msg!(self, "{prefix} {message}");
}
EventMsg::Warning(WarningEvent { message }) => {
ts_msg!(
self,
"{} {message}",
"warning:".style(self.yellow).style(self.bold)
);
}
EventMsg::DeprecationNotice(DeprecationNoticeEvent { summary, details }) => {
ts_msg!(
self,
"{} {summary}",
"deprecated:".style(self.magenta).style(self.bold)
);
if let Some(details) = details {
ts_msg!(self, " {}", details.style(self.dimmed));
}
}
EventMsg::BackgroundEvent(BackgroundEventEvent { message }) => {
ts_msg!(self, "{}", message.style(self.dimmed));
}
EventMsg::StreamError(StreamErrorEvent { message }) => {
ts_msg!(self, "{}", message.style(self.dimmed));
}
EventMsg::TaskStarted(_) => {
// Ignore.
}
EventMsg::TaskComplete(TaskCompleteEvent { last_agent_message }) => {
let last_message = last_agent_message.as_deref();
if let Some(output_file) = self.last_message_path.as_deref() {
handle_last_message(last_message, output_file);
}
self.final_message = last_agent_message;
return CodexStatus::InitiateShutdown;
}
EventMsg::TokenCount(ev) => {
self.last_total_token_usage = ev.info;
}
EventMsg::AgentReasoningSectionBreak(_) => {
if !self.show_agent_reasoning {
return CodexStatus::Running;
}
eprintln!();
}
EventMsg::AgentReasoningRawContent(AgentReasoningRawContentEvent { text }) => {
if self.show_raw_agent_reasoning {
ts_msg!(
self,
"{}\n{}",
"thinking".style(self.italic).style(self.magenta),
text,
);
}
}
EventMsg::AgentMessage(AgentMessageEvent { message }) => {
ts_msg!(
self,
"{}\n{}",
"codex".style(self.italic).style(self.magenta),
message,
);
}
EventMsg::ExecCommandBegin(ExecCommandBeginEvent { command, cwd, .. }) => {
eprint!(
"{}\n{} in {}",
"exec".style(self.italic).style(self.magenta),
escape_command(&command).style(self.bold),
cwd.to_string_lossy(),
);
}
EventMsg::ExecCommandEnd(ExecCommandEndEvent {
aggregated_output,
duration,
exit_code,
..
}) => {
let duration = format!(" in {}", format_duration(duration));
let truncated_output = aggregated_output
.lines()
.take(MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL)
.collect::<Vec<_>>()
.join("\n");
match exit_code {
0 => {
let title = format!(" succeeded{duration}:");
ts_msg!(self, "{}", title.style(self.green));
}
_ => {
let title = format!(" exited {exit_code}{duration}:");
ts_msg!(self, "{}", title.style(self.red));
}
}
eprintln!("{}", truncated_output.style(self.dimmed));
}
EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
call_id: _,
invocation,
}) => {
ts_msg!(
self,
"{} {}",
"tool".style(self.magenta),
format_mcp_invocation(&invocation).style(self.bold),
);
}
EventMsg::McpToolCallEnd(tool_call_end_event) => {
let is_success = tool_call_end_event.is_success();
let McpToolCallEndEvent {
call_id: _,
result,
invocation,
duration,
} = tool_call_end_event;
let duration = format!(" in {}", format_duration(duration));
let status_str = if is_success { "success" } else { "failed" };
let title_style = if is_success { self.green } else { self.red };
let title = format!(
"{} {status_str}{duration}:",
format_mcp_invocation(&invocation)
);
ts_msg!(self, "{}", title.style(title_style));
if let Ok(res) = result {
let val: serde_json::Value = res.into();
let pretty =
serde_json::to_string_pretty(&val).unwrap_or_else(|_| val.to_string());
for line in pretty.lines().take(MAX_OUTPUT_LINES_FOR_EXEC_TOOL_CALL) {
eprintln!("{}", line.style(self.dimmed));
}
}
}
EventMsg::WebSearchEnd(WebSearchEndEvent { call_id: _, query }) => {
ts_msg!(self, "🌐 Searched: {query}");
}
EventMsg::PatchApplyBegin(PatchApplyBeginEvent {
call_id,
auto_approved,
changes,
}) => {
// Store metadata so we can calculate duration later when we
// receive the corresponding PatchApplyEnd event.
self.call_id_to_patch.insert(
call_id,
PatchApplyBegin {
start_time: Instant::now(),
auto_approved,
},
);
ts_msg!(
self,
"{}",
"file update".style(self.magenta).style(self.italic),
);
// Pretty-print the patch summary with colored diff markers so
// it's easy to scan in the terminal output.
for (path, change) in changes.iter() {
match change {
FileChange::Add { content } => {
let header = format!(
"{} {}",
format_file_change(change),
path.to_string_lossy()
);
eprintln!("{}", header.style(self.magenta));
for line in content.lines() {
eprintln!("{}", line.style(self.green));
}
}
FileChange::Delete { content } => {
let header = format!(
"{} {}",
format_file_change(change),
path.to_string_lossy()
);
eprintln!("{}", header.style(self.magenta));
for line in content.lines() {
eprintln!("{}", line.style(self.red));
}
}
FileChange::Update {
unified_diff,
move_path,
} => {
let header = if let Some(dest) = move_path {
format!(
"{} {} -> {}",
format_file_change(change),
path.to_string_lossy(),
dest.to_string_lossy()
)
} else {
format!("{} {}", format_file_change(change), path.to_string_lossy())
};
eprintln!("{}", header.style(self.magenta));
// Colorize diff lines. We keep file header lines
// (--- / +++) without extra coloring so they are
// still readable.
for diff_line in unified_diff.lines() {
if diff_line.starts_with('+') && !diff_line.starts_with("+++") {
eprintln!("{}", diff_line.style(self.green));
} else if diff_line.starts_with('-')
&& !diff_line.starts_with("---")
{
eprintln!("{}", diff_line.style(self.red));
} else {
eprintln!("{diff_line}");
}
}
}
}
}
}
EventMsg::PatchApplyEnd(PatchApplyEndEvent {
call_id,
stdout,
stderr,
success,
..
}) => {
let patch_begin = self.call_id_to_patch.remove(&call_id);
// Compute duration and summary label similar to exec commands.
let (duration, label) = if let Some(PatchApplyBegin {
start_time,
auto_approved,
}) = patch_begin
{
(
format!(" in {}", format_elapsed(start_time)),
format!("apply_patch(auto_approved={auto_approved})"),
)
} else {
(String::new(), format!("apply_patch('{call_id}')"))
};
let (exit_code, output, title_style) = if success {
(0, stdout, self.green)
} else {
(1, stderr, self.red)
};
let title = format!("{label} exited {exit_code}{duration}:");
ts_msg!(self, "{}", title.style(title_style));
for line in output.lines() {
eprintln!("{}", line.style(self.dimmed));
}
}
EventMsg::TurnDiff(TurnDiffEvent { unified_diff }) => {
ts_msg!(
self,
"{}",
"file update:".style(self.magenta).style(self.italic)
);
eprintln!("{unified_diff}");
}
EventMsg::AgentReasoning(agent_reasoning_event) => {
if self.show_agent_reasoning {
ts_msg!(
self,
"{}\n{}",
"thinking".style(self.italic).style(self.magenta),
agent_reasoning_event.text,
);
}
}
EventMsg::SessionConfigured(session_configured_event) => {
let SessionConfiguredEvent {
session_id: conversation_id,
model,
reasoning_effort: _,
history_log_id: _,
history_entry_count: _,
initial_messages: _,
rollout_path: _,
} = session_configured_event;
ts_msg!(
self,
"{} {}",
"codex session".style(self.magenta).style(self.bold),
conversation_id.to_string().style(self.dimmed)
);
ts_msg!(self, "model: {}", model);
eprintln!();
}
EventMsg::PlanUpdate(plan_update_event) => {
let UpdatePlanArgs { explanation, plan } = plan_update_event;
// Header
ts_msg!(self, "{}", "Plan update".style(self.magenta));
// Optional explanation
if let Some(explanation) = explanation
&& !explanation.trim().is_empty()
{
ts_msg!(self, "{}", explanation.style(self.italic));
}
// Pretty-print the plan items with simple status markers.
for item in plan {
match item.status {
StepStatus::Completed => {
ts_msg!(self, " {} {}", "".style(self.green), item.step);
}
StepStatus::InProgress => {
ts_msg!(self, " {} {}", "".style(self.cyan), item.step);
}
StepStatus::Pending => {
ts_msg!(
self,
" {} {}",
"".style(self.dimmed),
item.step.style(self.dimmed)
);
}
}
}
}
EventMsg::ViewImageToolCall(view) => {
ts_msg!(
self,
"{} {}",
"viewed image".style(self.magenta),
view.path.display()
);
}
EventMsg::TurnAborted(abort_reason) => match abort_reason.reason {
TurnAbortReason::Interrupted => {
ts_msg!(self, "task interrupted");
}
TurnAbortReason::Replaced => {
ts_msg!(self, "task aborted: replaced by a new task");
}
TurnAbortReason::ReviewEnded => {
ts_msg!(self, "task aborted: review ended");
}
},
EventMsg::ShutdownComplete => return CodexStatus::Shutdown,
EventMsg::WebSearchBegin(_)
| EventMsg::ExecApprovalRequest(_)
| EventMsg::ApplyPatchApprovalRequest(_)
| EventMsg::ExecCommandOutputDelta(_)
| EventMsg::GetHistoryEntryResponse(_)
| EventMsg::McpListToolsResponse(_)
| EventMsg::ListCustomPromptsResponse(_)
| EventMsg::RawResponseItem(_)
| EventMsg::UserMessage(_)
| EventMsg::EnteredReviewMode(_)
| EventMsg::ExitedReviewMode(_)
| EventMsg::AgentMessageDelta(_)
| EventMsg::AgentReasoningDelta(_)
| EventMsg::AgentReasoningRawContentDelta(_)
| EventMsg::ItemStarted(_)
| EventMsg::ItemCompleted(_)
| EventMsg::AgentMessageContentDelta(_)
| EventMsg::ReasoningContentDelta(_)
| EventMsg::ReasoningRawContentDelta(_)
| EventMsg::UndoCompleted(_)
| EventMsg::UndoStarted(_) => {}
}
CodexStatus::Running
}
fn print_final_output(&mut self) {
if let Some(usage_info) = &self.last_total_token_usage {
eprintln!(
"{}\n{}",
"tokens used".style(self.magenta).style(self.italic),
format_with_separators(usage_info.total_token_usage.blended_total())
);
}
// If the user has not piped the final message to a file, they will see
// it twice: once written to stderr as part of the normal event
// processing, and once here on stdout. We print the token summary above
// to help break up the output visually in that case.
#[allow(clippy::print_stdout)]
if let Some(message) = &self.final_message {
if message.ends_with('\n') {
print!("{message}");
} else {
println!("{message}");
}
}
}
}
fn escape_command(command: &[String]) -> String {
try_join(command.iter().map(String::as_str)).unwrap_or_else(|_| command.join(" "))
}
fn format_file_change(change: &FileChange) -> &'static str {
match change {
FileChange::Add { .. } => "A",
FileChange::Delete { .. } => "D",
FileChange::Update {
move_path: Some(_), ..
} => "R",
FileChange::Update {
move_path: None, ..
} => "M",
}
}
fn format_mcp_invocation(invocation: &McpInvocation) -> String {
// Build fully-qualified tool name: server.tool
let fq_tool_name = format!("{}.{}", invocation.server, invocation.tool);
// Format arguments as compact JSON so they fit on one line.
let args_str = invocation
.arguments
.as_ref()
.map(|v: &serde_json::Value| serde_json::to_string(v).unwrap_or_else(|_| v.to_string()))
.unwrap_or_default();
if args_str.is_empty() {
format!("{fq_tool_name}()")
} else {
format!("{fq_tool_name}({args_str})")
}
}

View File

@@ -0,0 +1,501 @@
use std::collections::HashMap;
use std::path::PathBuf;
use std::sync::atomic::AtomicU64;
use crate::event_processor::CodexStatus;
use crate::event_processor::EventProcessor;
use crate::event_processor::handle_last_message;
use crate::exec_events::AgentMessageItem;
use crate::exec_events::CommandExecutionItem;
use crate::exec_events::CommandExecutionStatus;
use crate::exec_events::ErrorItem;
use crate::exec_events::FileChangeItem;
use crate::exec_events::FileUpdateChange;
use crate::exec_events::ItemCompletedEvent;
use crate::exec_events::ItemStartedEvent;
use crate::exec_events::ItemUpdatedEvent;
use crate::exec_events::McpToolCallItem;
use crate::exec_events::McpToolCallItemError;
use crate::exec_events::McpToolCallItemResult;
use crate::exec_events::McpToolCallStatus;
use crate::exec_events::PatchApplyStatus;
use crate::exec_events::PatchChangeKind;
use crate::exec_events::ReasoningItem;
use crate::exec_events::ThreadErrorEvent;
use crate::exec_events::ThreadEvent;
use crate::exec_events::ThreadItem;
use crate::exec_events::ThreadItemDetails;
use crate::exec_events::ThreadStartedEvent;
use crate::exec_events::TodoItem;
use crate::exec_events::TodoListItem;
use crate::exec_events::TurnCompletedEvent;
use crate::exec_events::TurnFailedEvent;
use crate::exec_events::TurnStartedEvent;
use crate::exec_events::Usage;
use crate::exec_events::WebSearchItem;
use codex_core::config::Config;
use codex_core::protocol::AgentMessageEvent;
use codex_core::protocol::AgentReasoningEvent;
use codex_core::protocol::Event;
use codex_core::protocol::EventMsg;
use codex_core::protocol::ExecCommandBeginEvent;
use codex_core::protocol::ExecCommandEndEvent;
use codex_core::protocol::FileChange;
use codex_core::protocol::McpToolCallBeginEvent;
use codex_core::protocol::McpToolCallEndEvent;
use codex_core::protocol::PatchApplyBeginEvent;
use codex_core::protocol::PatchApplyEndEvent;
use codex_core::protocol::SessionConfiguredEvent;
use codex_core::protocol::TaskCompleteEvent;
use codex_core::protocol::TaskStartedEvent;
use codex_core::protocol::WebSearchEndEvent;
use codex_protocol::plan_tool::StepStatus;
use codex_protocol::plan_tool::UpdatePlanArgs;
use serde_json::Value as JsonValue;
use tracing::error;
use tracing::warn;
pub struct EventProcessorWithJsonOutput {
last_message_path: Option<PathBuf>,
next_event_id: AtomicU64,
// Tracks running commands by call_id, including the associated item id.
running_commands: HashMap<String, RunningCommand>,
running_patch_applies: HashMap<String, PatchApplyBeginEvent>,
// Tracks the todo list for the current turn (at most one per turn).
running_todo_list: Option<RunningTodoList>,
last_total_token_usage: Option<codex_core::protocol::TokenUsage>,
running_mcp_tool_calls: HashMap<String, RunningMcpToolCall>,
last_critical_error: Option<ThreadErrorEvent>,
}
#[derive(Debug, Clone)]
struct RunningCommand {
command: String,
item_id: String,
}
#[derive(Debug, Clone)]
struct RunningTodoList {
item_id: String,
items: Vec<TodoItem>,
}
#[derive(Debug, Clone)]
struct RunningMcpToolCall {
server: String,
tool: String,
item_id: String,
arguments: JsonValue,
}
impl EventProcessorWithJsonOutput {
pub fn new(last_message_path: Option<PathBuf>) -> Self {
Self {
last_message_path,
next_event_id: AtomicU64::new(0),
running_commands: HashMap::new(),
running_patch_applies: HashMap::new(),
running_todo_list: None,
last_total_token_usage: None,
running_mcp_tool_calls: HashMap::new(),
last_critical_error: None,
}
}
pub fn collect_thread_events(&mut self, event: &Event) -> Vec<ThreadEvent> {
match &event.msg {
EventMsg::SessionConfigured(ev) => self.handle_session_configured(ev),
EventMsg::AgentMessage(ev) => self.handle_agent_message(ev),
EventMsg::AgentReasoning(ev) => self.handle_reasoning_event(ev),
EventMsg::ExecCommandBegin(ev) => self.handle_exec_command_begin(ev),
EventMsg::ExecCommandEnd(ev) => self.handle_exec_command_end(ev),
EventMsg::McpToolCallBegin(ev) => self.handle_mcp_tool_call_begin(ev),
EventMsg::McpToolCallEnd(ev) => self.handle_mcp_tool_call_end(ev),
EventMsg::PatchApplyBegin(ev) => self.handle_patch_apply_begin(ev),
EventMsg::PatchApplyEnd(ev) => self.handle_patch_apply_end(ev),
EventMsg::WebSearchBegin(_) => Vec::new(),
EventMsg::WebSearchEnd(ev) => self.handle_web_search_end(ev),
EventMsg::TokenCount(ev) => {
if let Some(info) = &ev.info {
self.last_total_token_usage = Some(info.total_token_usage.clone());
}
Vec::new()
}
EventMsg::TaskStarted(ev) => self.handle_task_started(ev),
EventMsg::TaskComplete(_) => self.handle_task_complete(),
EventMsg::Error(ev) => {
let error = ThreadErrorEvent {
message: ev.message.clone(),
};
self.last_critical_error = Some(error.clone());
vec![ThreadEvent::Error(error)]
}
EventMsg::Warning(ev) => {
let item = ThreadItem {
id: self.get_next_item_id(),
details: ThreadItemDetails::Error(ErrorItem {
message: ev.message.clone(),
}),
};
vec![ThreadEvent::ItemCompleted(ItemCompletedEvent { item })]
}
EventMsg::StreamError(ev) => vec![ThreadEvent::Error(ThreadErrorEvent {
message: ev.message.clone(),
})],
EventMsg::PlanUpdate(ev) => self.handle_plan_update(ev),
_ => Vec::new(),
}
}
fn get_next_item_id(&self) -> String {
format!(
"item_{}",
self.next_event_id
.fetch_add(1, std::sync::atomic::Ordering::SeqCst)
)
}
fn handle_session_configured(&self, payload: &SessionConfiguredEvent) -> Vec<ThreadEvent> {
vec![ThreadEvent::ThreadStarted(ThreadStartedEvent {
thread_id: payload.session_id.to_string(),
})]
}
fn handle_web_search_end(&self, ev: &WebSearchEndEvent) -> Vec<ThreadEvent> {
let item = ThreadItem {
id: self.get_next_item_id(),
details: ThreadItemDetails::WebSearch(WebSearchItem {
query: ev.query.clone(),
}),
};
vec![ThreadEvent::ItemCompleted(ItemCompletedEvent { item })]
}
fn handle_agent_message(&self, payload: &AgentMessageEvent) -> Vec<ThreadEvent> {
let item = ThreadItem {
id: self.get_next_item_id(),
details: ThreadItemDetails::AgentMessage(AgentMessageItem {
text: payload.message.clone(),
}),
};
vec![ThreadEvent::ItemCompleted(ItemCompletedEvent { item })]
}
fn handle_reasoning_event(&self, ev: &AgentReasoningEvent) -> Vec<ThreadEvent> {
let item = ThreadItem {
id: self.get_next_item_id(),
details: ThreadItemDetails::Reasoning(ReasoningItem {
text: ev.text.clone(),
}),
};
vec![ThreadEvent::ItemCompleted(ItemCompletedEvent { item })]
}
fn handle_exec_command_begin(&mut self, ev: &ExecCommandBeginEvent) -> Vec<ThreadEvent> {
let item_id = self.get_next_item_id();
let command_string = match shlex::try_join(ev.command.iter().map(String::as_str)) {
Ok(command_string) => command_string,
Err(e) => {
warn!(
call_id = ev.call_id,
"Failed to stringify command: {e:?}; skipping item.started"
);
ev.command.join(" ")
}
};
self.running_commands.insert(
ev.call_id.clone(),
RunningCommand {
command: command_string.clone(),
item_id: item_id.clone(),
},
);
let item = ThreadItem {
id: item_id,
details: ThreadItemDetails::CommandExecution(CommandExecutionItem {
command: command_string,
aggregated_output: String::new(),
exit_code: None,
status: CommandExecutionStatus::InProgress,
}),
};
vec![ThreadEvent::ItemStarted(ItemStartedEvent { item })]
}
fn handle_mcp_tool_call_begin(&mut self, ev: &McpToolCallBeginEvent) -> Vec<ThreadEvent> {
let item_id = self.get_next_item_id();
let server = ev.invocation.server.clone();
let tool = ev.invocation.tool.clone();
let arguments = ev.invocation.arguments.clone().unwrap_or(JsonValue::Null);
self.running_mcp_tool_calls.insert(
ev.call_id.clone(),
RunningMcpToolCall {
server: server.clone(),
tool: tool.clone(),
item_id: item_id.clone(),
arguments: arguments.clone(),
},
);
let item = ThreadItem {
id: item_id,
details: ThreadItemDetails::McpToolCall(McpToolCallItem {
server,
tool,
arguments,
result: None,
error: None,
status: McpToolCallStatus::InProgress,
}),
};
vec![ThreadEvent::ItemStarted(ItemStartedEvent { item })]
}
fn handle_mcp_tool_call_end(&mut self, ev: &McpToolCallEndEvent) -> Vec<ThreadEvent> {
let status = if ev.is_success() {
McpToolCallStatus::Completed
} else {
McpToolCallStatus::Failed
};
let (server, tool, item_id, arguments) =
match self.running_mcp_tool_calls.remove(&ev.call_id) {
Some(running) => (
running.server,
running.tool,
running.item_id,
running.arguments,
),
None => {
warn!(
call_id = ev.call_id,
"Received McpToolCallEnd without begin; synthesizing new item"
);
(
ev.invocation.server.clone(),
ev.invocation.tool.clone(),
self.get_next_item_id(),
ev.invocation.arguments.clone().unwrap_or(JsonValue::Null),
)
}
};
let (result, error) = match &ev.result {
Ok(value) => {
let result = McpToolCallItemResult {
content: value.content.clone(),
structured_content: value.structured_content.clone(),
};
(Some(result), None)
}
Err(message) => (
None,
Some(McpToolCallItemError {
message: message.clone(),
}),
),
};
let item = ThreadItem {
id: item_id,
details: ThreadItemDetails::McpToolCall(McpToolCallItem {
server,
tool,
arguments,
result,
error,
status,
}),
};
vec![ThreadEvent::ItemCompleted(ItemCompletedEvent { item })]
}
fn handle_patch_apply_begin(&mut self, ev: &PatchApplyBeginEvent) -> Vec<ThreadEvent> {
self.running_patch_applies
.insert(ev.call_id.clone(), ev.clone());
Vec::new()
}
fn map_change_kind(&self, kind: &FileChange) -> PatchChangeKind {
match kind {
FileChange::Add { .. } => PatchChangeKind::Add,
FileChange::Delete { .. } => PatchChangeKind::Delete,
FileChange::Update { .. } => PatchChangeKind::Update,
}
}
fn handle_patch_apply_end(&mut self, ev: &PatchApplyEndEvent) -> Vec<ThreadEvent> {
if let Some(running_patch_apply) = self.running_patch_applies.remove(&ev.call_id) {
let status = if ev.success {
PatchApplyStatus::Completed
} else {
PatchApplyStatus::Failed
};
let item = ThreadItem {
id: self.get_next_item_id(),
details: ThreadItemDetails::FileChange(FileChangeItem {
changes: running_patch_apply
.changes
.iter()
.map(|(path, change)| FileUpdateChange {
path: path.to_str().unwrap_or("").to_string(),
kind: self.map_change_kind(change),
})
.collect(),
status,
}),
};
return vec![ThreadEvent::ItemCompleted(ItemCompletedEvent { item })];
}
Vec::new()
}
fn handle_exec_command_end(&mut self, ev: &ExecCommandEndEvent) -> Vec<ThreadEvent> {
let Some(RunningCommand { command, item_id }) = self.running_commands.remove(&ev.call_id)
else {
warn!(
call_id = ev.call_id,
"ExecCommandEnd without matching ExecCommandBegin; skipping item.completed"
);
return Vec::new();
};
let status = if ev.exit_code == 0 {
CommandExecutionStatus::Completed
} else {
CommandExecutionStatus::Failed
};
let item = ThreadItem {
id: item_id,
details: ThreadItemDetails::CommandExecution(CommandExecutionItem {
command,
aggregated_output: ev.aggregated_output.clone(),
exit_code: Some(ev.exit_code),
status,
}),
};
vec![ThreadEvent::ItemCompleted(ItemCompletedEvent { item })]
}
fn todo_items_from_plan(&self, args: &UpdatePlanArgs) -> Vec<TodoItem> {
args.plan
.iter()
.map(|p| TodoItem {
text: p.step.clone(),
completed: matches!(p.status, StepStatus::Completed),
})
.collect()
}
fn handle_plan_update(&mut self, args: &UpdatePlanArgs) -> Vec<ThreadEvent> {
let items = self.todo_items_from_plan(args);
if let Some(running) = &mut self.running_todo_list {
running.items = items.clone();
let item = ThreadItem {
id: running.item_id.clone(),
details: ThreadItemDetails::TodoList(TodoListItem { items }),
};
return vec![ThreadEvent::ItemUpdated(ItemUpdatedEvent { item })];
}
let item_id = self.get_next_item_id();
self.running_todo_list = Some(RunningTodoList {
item_id: item_id.clone(),
items: items.clone(),
});
let item = ThreadItem {
id: item_id,
details: ThreadItemDetails::TodoList(TodoListItem { items }),
};
vec![ThreadEvent::ItemStarted(ItemStartedEvent { item })]
}
fn handle_task_started(&mut self, _: &TaskStartedEvent) -> Vec<ThreadEvent> {
self.last_critical_error = None;
vec![ThreadEvent::TurnStarted(TurnStartedEvent {})]
}
fn handle_task_complete(&mut self) -> Vec<ThreadEvent> {
let usage = if let Some(u) = &self.last_total_token_usage {
Usage {
input_tokens: u.input_tokens,
cached_input_tokens: u.cached_input_tokens,
output_tokens: u.output_tokens,
}
} else {
Usage::default()
};
let mut items = Vec::new();
if let Some(running) = self.running_todo_list.take() {
let item = ThreadItem {
id: running.item_id,
details: ThreadItemDetails::TodoList(TodoListItem {
items: running.items,
}),
};
items.push(ThreadEvent::ItemCompleted(ItemCompletedEvent { item }));
}
if let Some(error) = self.last_critical_error.take() {
items.push(ThreadEvent::TurnFailed(TurnFailedEvent { error }));
} else {
items.push(ThreadEvent::TurnCompleted(TurnCompletedEvent { usage }));
}
items
}
}
impl EventProcessor for EventProcessorWithJsonOutput {
fn print_config_summary(&mut self, _: &Config, _: &str, ev: &SessionConfiguredEvent) {
self.process_event(Event {
id: "".to_string(),
msg: EventMsg::SessionConfigured(ev.clone()),
});
}
#[allow(clippy::print_stdout)]
fn process_event(&mut self, event: Event) -> CodexStatus {
let aggregated = self.collect_thread_events(&event);
for conv_event in aggregated {
match serde_json::to_string(&conv_event) {
Ok(line) => {
println!("{line}");
}
Err(e) => {
error!("Failed to serialize event: {e:?}");
}
}
}
let Event { msg, .. } = event;
if let EventMsg::TaskComplete(TaskCompleteEvent { last_agent_message }) = msg {
if let Some(output_file) = self.last_message_path.as_deref() {
handle_last_message(last_agent_message.as_deref(), output_file);
}
CodexStatus::InitiateShutdown
} else {
CodexStatus::Running
}
}
}

View File

@@ -0,0 +1,246 @@
use mcp_types::ContentBlock as McpContentBlock;
use serde::Deserialize;
use serde::Serialize;
use serde_json::Value as JsonValue;
use ts_rs::TS;
/// Top-level JSONL events emitted by codex exec
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
#[serde(tag = "type")]
pub enum ThreadEvent {
/// Emitted when a new thread is started as the first event.
#[serde(rename = "thread.started")]
ThreadStarted(ThreadStartedEvent),
/// Emitted when a turn is started by sending a new prompt to the model.
/// A turn encompasses all events that happen while agent is processing the prompt.
#[serde(rename = "turn.started")]
TurnStarted(TurnStartedEvent),
/// Emitted when a turn is completed. Typically right after the assistant's response.
#[serde(rename = "turn.completed")]
TurnCompleted(TurnCompletedEvent),
/// Indicates that a turn failed with an error.
#[serde(rename = "turn.failed")]
TurnFailed(TurnFailedEvent),
/// Emitted when a new item is added to the thread. Typically the item will be in an "in progress" state.
#[serde(rename = "item.started")]
ItemStarted(ItemStartedEvent),
/// Emitted when an item is updated.
#[serde(rename = "item.updated")]
ItemUpdated(ItemUpdatedEvent),
/// Signals that an item has reached a terminal state—either success or failure.
#[serde(rename = "item.completed")]
ItemCompleted(ItemCompletedEvent),
/// Represents an unrecoverable error emitted directly by the event stream.
#[serde(rename = "error")]
Error(ThreadErrorEvent),
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ThreadStartedEvent {
/// The identified of the new thread. Can be used to resume the thread later.
pub thread_id: String,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS, Default)]
pub struct TurnStartedEvent {}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct TurnCompletedEvent {
pub usage: Usage,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct TurnFailedEvent {
pub error: ThreadErrorEvent,
}
/// Describes the usage of tokens during a turn.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS, Default)]
pub struct Usage {
/// The number of input tokens used during the turn.
pub input_tokens: i64,
/// The number of cached input tokens used during the turn.
pub cached_input_tokens: i64,
/// The number of output tokens used during the turn.
pub output_tokens: i64,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ItemStartedEvent {
pub item: ThreadItem,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ItemCompletedEvent {
pub item: ThreadItem,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ItemUpdatedEvent {
pub item: ThreadItem,
}
/// Fatal error emitted by the stream.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ThreadErrorEvent {
pub message: String,
}
/// Canonical representation of a thread item and its domain-specific payload.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ThreadItem {
pub id: String,
#[serde(flatten)]
pub details: ThreadItemDetails,
}
/// Typed payloads for each supported thread item type.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
#[serde(tag = "type", rename_all = "snake_case")]
pub enum ThreadItemDetails {
/// Response from the agent.
/// Either a natural-language response or a JSON string when structured output is requested.
AgentMessage(AgentMessageItem),
/// Agent's reasoning summary.
Reasoning(ReasoningItem),
/// Tracks a command executed by the agent. The item starts when the command is
/// spawned, and completes when the process exits with an exit code.
CommandExecution(CommandExecutionItem),
/// Represents a set of file changes by the agent. The item is emitted only as a
/// completed event once the patch succeeds or fails.
FileChange(FileChangeItem),
/// Represents a call to an MCP tool. The item starts when the invocation is
/// dispatched and completes when the MCP server reports success or failure.
McpToolCall(McpToolCallItem),
/// Captures a web search request. It starts when the search is kicked off
/// and completes when results are returned to the agent.
WebSearch(WebSearchItem),
/// Tracks the agent's running to-do list. It starts when the plan is first
/// issued, updates as steps change state, and completes when the turn ends.
TodoList(TodoListItem),
/// Describes a non-fatal error surfaced as an item.
Error(ErrorItem),
}
/// Response from the agent.
/// Either a natural-language response or a JSON string when structured output is requested.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct AgentMessageItem {
pub text: String,
}
/// Agent's reasoning summary.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ReasoningItem {
pub text: String,
}
/// The status of a command execution.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq, Default, TS)]
#[serde(rename_all = "snake_case")]
pub enum CommandExecutionStatus {
#[default]
InProgress,
Completed,
Failed,
}
/// A command executed by the agent.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct CommandExecutionItem {
pub command: String,
pub aggregated_output: String,
pub exit_code: Option<i32>,
pub status: CommandExecutionStatus,
}
/// A set of file changes by the agent.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct FileUpdateChange {
pub path: String,
pub kind: PatchChangeKind,
}
/// The status of a file change.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
#[serde(rename_all = "snake_case")]
pub enum PatchApplyStatus {
Completed,
Failed,
}
/// A set of file changes by the agent.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct FileChangeItem {
pub changes: Vec<FileUpdateChange>,
pub status: PatchApplyStatus,
}
/// Indicates the type of the file change.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
#[serde(rename_all = "snake_case")]
pub enum PatchChangeKind {
Add,
Delete,
Update,
}
/// The status of an MCP tool call.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq, Default, TS)]
#[serde(rename_all = "snake_case")]
pub enum McpToolCallStatus {
#[default]
InProgress,
Completed,
Failed,
}
/// Result payload produced by an MCP tool invocation.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct McpToolCallItemResult {
pub content: Vec<McpContentBlock>,
pub structured_content: Option<JsonValue>,
}
/// Error details reported by a failed MCP tool invocation.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct McpToolCallItemError {
pub message: String,
}
/// A call to an MCP tool.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct McpToolCallItem {
pub server: String,
pub tool: String,
#[serde(default)]
pub arguments: JsonValue,
pub result: Option<McpToolCallItemResult>,
pub error: Option<McpToolCallItemError>,
pub status: McpToolCallStatus,
}
/// A web search request.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct WebSearchItem {
pub query: String,
}
/// An error notification.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct ErrorItem {
pub message: String,
}
/// An item in agent's to-do list.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct TodoItem {
pub text: String,
pub completed: bool,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, TS)]
pub struct TodoListItem {
pub items: Vec<TodoItem>,
}

427
llmx-rs/exec/src/lib.rs Normal file
View File

@@ -0,0 +1,427 @@
// - In the default output mode, it is paramount that the only thing written to
// stdout is the final message (if any).
// - In --json mode, stdout must be valid JSONL, one event per line.
// For both modes, any other output must be written to stderr.
#![deny(clippy::print_stdout)]
mod cli;
mod event_processor;
mod event_processor_with_human_output;
pub mod event_processor_with_jsonl_output;
pub mod exec_events;
pub use cli::Cli;
use codex_core::AuthManager;
use codex_core::BUILT_IN_OSS_MODEL_PROVIDER_ID;
use codex_core::ConversationManager;
use codex_core::NewConversation;
use codex_core::auth::enforce_login_restrictions;
use codex_core::config::Config;
use codex_core::config::ConfigOverrides;
use codex_core::git_info::get_git_repo_root;
use codex_core::protocol::AskForApproval;
use codex_core::protocol::Event;
use codex_core::protocol::EventMsg;
use codex_core::protocol::Op;
use codex_core::protocol::SessionSource;
use codex_ollama::DEFAULT_OSS_MODEL;
use codex_protocol::config_types::SandboxMode;
use codex_protocol::user_input::UserInput;
use event_processor_with_human_output::EventProcessorWithHumanOutput;
use event_processor_with_jsonl_output::EventProcessorWithJsonOutput;
use opentelemetry_appender_tracing::layer::OpenTelemetryTracingBridge;
use serde_json::Value;
use std::io::IsTerminal;
use std::io::Read;
use std::path::PathBuf;
use supports_color::Stream;
use tracing::debug;
use tracing::error;
use tracing::info;
use tracing_subscriber::EnvFilter;
use tracing_subscriber::prelude::*;
use crate::cli::Command as ExecCommand;
use crate::event_processor::CodexStatus;
use crate::event_processor::EventProcessor;
use codex_core::default_client::set_default_originator;
use codex_core::find_conversation_path_by_id_str;
pub async fn run_main(cli: Cli, codex_linux_sandbox_exe: Option<PathBuf>) -> anyhow::Result<()> {
if let Err(err) = set_default_originator("codex_exec".to_string()) {
tracing::warn!(?err, "Failed to set codex exec originator override {err:?}");
}
let Cli {
command,
images,
model: model_cli_arg,
oss,
config_profile,
full_auto,
dangerously_bypass_approvals_and_sandbox,
cwd,
skip_git_repo_check,
color,
last_message_file,
json: json_mode,
sandbox_mode: sandbox_mode_cli_arg,
prompt,
output_schema: output_schema_path,
config_overrides,
} = cli;
// Determine the prompt source (parent or subcommand) and read from stdin if needed.
let prompt_arg = match &command {
// Allow prompt before the subcommand by falling back to the parent-level prompt
// when the Resume subcommand did not provide its own prompt.
Some(ExecCommand::Resume(args)) => args.prompt.clone().or(prompt),
None => prompt,
};
let prompt = match prompt_arg {
Some(p) if p != "-" => p,
// Either `-` was passed or no positional arg.
maybe_dash => {
// When no arg (None) **and** stdin is a TTY, bail out early unless the
// user explicitly forced reading via `-`.
let force_stdin = matches!(maybe_dash.as_deref(), Some("-"));
if std::io::stdin().is_terminal() && !force_stdin {
eprintln!(
"No prompt provided. Either specify one as an argument or pipe the prompt into stdin."
);
std::process::exit(1);
}
// Ensure the user knows we are waiting on stdin, as they may
// have gotten into this state by mistake. If so, and they are not
// writing to stdin, Codex will hang indefinitely, so this should
// help them debug in that case.
if !force_stdin {
eprintln!("Reading prompt from stdin...");
}
let mut buffer = String::new();
if let Err(e) = std::io::stdin().read_to_string(&mut buffer) {
eprintln!("Failed to read prompt from stdin: {e}");
std::process::exit(1);
} else if buffer.trim().is_empty() {
eprintln!("No prompt provided via stdin.");
std::process::exit(1);
}
buffer
}
};
let output_schema = load_output_schema(output_schema_path);
let (stdout_with_ansi, stderr_with_ansi) = match color {
cli::Color::Always => (true, true),
cli::Color::Never => (false, false),
cli::Color::Auto => (
supports_color::on_cached(Stream::Stdout).is_some(),
supports_color::on_cached(Stream::Stderr).is_some(),
),
};
// Build fmt layer (existing logging) to compose with OTEL layer.
let default_level = "error";
// Build env_filter separately and attach via with_filter.
let env_filter = EnvFilter::try_from_default_env()
.or_else(|_| EnvFilter::try_new(default_level))
.unwrap_or_else(|_| EnvFilter::new(default_level));
let fmt_layer = tracing_subscriber::fmt::layer()
.with_ansi(stderr_with_ansi)
.with_writer(std::io::stderr)
.with_filter(env_filter);
let sandbox_mode = if full_auto {
Some(SandboxMode::WorkspaceWrite)
} else if dangerously_bypass_approvals_and_sandbox {
Some(SandboxMode::DangerFullAccess)
} else {
sandbox_mode_cli_arg.map(Into::<SandboxMode>::into)
};
// When using `--oss`, let the bootstrapper pick the model (defaulting to
// gpt-oss:20b) and ensure it is present locally. Also, force the builtin
// `oss` model provider.
let model = if let Some(model) = model_cli_arg {
Some(model)
} else if oss {
Some(DEFAULT_OSS_MODEL.to_owned())
} else {
None // No model specified, will use the default.
};
let model_provider = if oss {
Some(BUILT_IN_OSS_MODEL_PROVIDER_ID.to_string())
} else {
None // No specific model provider override.
};
// Load configuration and determine approval policy
let overrides = ConfigOverrides {
model,
review_model: None,
config_profile,
// Default to never ask for approvals in headless mode. Feature flags can override.
approval_policy: Some(AskForApproval::Never),
sandbox_mode,
cwd: cwd.map(|p| p.canonicalize().unwrap_or(p)),
model_provider,
codex_linux_sandbox_exe,
base_instructions: None,
developer_instructions: None,
compact_prompt: None,
include_apply_patch_tool: None,
show_raw_agent_reasoning: oss.then_some(true),
tools_web_search_request: None,
experimental_sandbox_command_assessment: None,
additional_writable_roots: Vec::new(),
};
// Parse `-c` overrides.
let cli_kv_overrides = match config_overrides.parse_overrides() {
Ok(v) => v,
Err(e) => {
eprintln!("Error parsing -c overrides: {e}");
std::process::exit(1);
}
};
let config = Config::load_with_cli_overrides(cli_kv_overrides, overrides).await?;
if let Err(err) = enforce_login_restrictions(&config).await {
eprintln!("{err}");
std::process::exit(1);
}
let otel = codex_core::otel_init::build_provider(&config, env!("CARGO_PKG_VERSION"));
#[allow(clippy::print_stderr)]
let otel = match otel {
Ok(otel) => otel,
Err(e) => {
eprintln!("Could not create otel exporter: {e}");
std::process::exit(1);
}
};
if let Some(provider) = otel.as_ref() {
let otel_layer = OpenTelemetryTracingBridge::new(&provider.logger).with_filter(
tracing_subscriber::filter::filter_fn(codex_core::otel_init::codex_export_filter),
);
let _ = tracing_subscriber::registry()
.with(fmt_layer)
.with(otel_layer)
.try_init();
} else {
let _ = tracing_subscriber::registry().with(fmt_layer).try_init();
}
let mut event_processor: Box<dyn EventProcessor> = match json_mode {
true => Box::new(EventProcessorWithJsonOutput::new(last_message_file.clone())),
_ => Box::new(EventProcessorWithHumanOutput::create_with_ansi(
stdout_with_ansi,
&config,
last_message_file.clone(),
)),
};
if oss {
codex_ollama::ensure_oss_ready(&config)
.await
.map_err(|e| anyhow::anyhow!("OSS setup failed: {e}"))?;
}
let default_cwd = config.cwd.to_path_buf();
let default_approval_policy = config.approval_policy;
let default_sandbox_policy = config.sandbox_policy.clone();
let default_model = config.model.clone();
let default_effort = config.model_reasoning_effort;
let default_summary = config.model_reasoning_summary;
if !skip_git_repo_check && get_git_repo_root(&default_cwd).is_none() {
eprintln!("Not inside a trusted directory and --skip-git-repo-check was not specified.");
std::process::exit(1);
}
let auth_manager = AuthManager::shared(
config.codex_home.clone(),
true,
config.cli_auth_credentials_store_mode,
);
let conversation_manager = ConversationManager::new(auth_manager.clone(), SessionSource::Exec);
// Handle resume subcommand by resolving a rollout path and using explicit resume API.
let NewConversation {
conversation_id: _,
conversation,
session_configured,
} = if let Some(ExecCommand::Resume(args)) = command {
let resume_path = resolve_resume_path(&config, &args).await?;
if let Some(path) = resume_path {
conversation_manager
.resume_conversation_from_rollout(config.clone(), path, auth_manager.clone())
.await?
} else {
conversation_manager
.new_conversation(config.clone())
.await?
}
} else {
conversation_manager
.new_conversation(config.clone())
.await?
};
// Print the effective configuration and prompt so users can see what Codex
// is using.
event_processor.print_config_summary(&config, &prompt, &session_configured);
info!("Codex initialized with event: {session_configured:?}");
let (tx, mut rx) = tokio::sync::mpsc::unbounded_channel::<Event>();
{
let conversation = conversation.clone();
tokio::spawn(async move {
loop {
tokio::select! {
_ = tokio::signal::ctrl_c() => {
tracing::debug!("Keyboard interrupt");
// Immediately notify Codex to abort any inflight task.
conversation.submit(Op::Interrupt).await.ok();
// Exit the inner loop and return to the main input prompt. The codex
// will emit a `TurnInterrupted` (Error) event which is drained later.
break;
}
res = conversation.next_event() => match res {
Ok(event) => {
debug!("Received event: {event:?}");
let is_shutdown_complete = matches!(event.msg, EventMsg::ShutdownComplete);
if let Err(e) = tx.send(event) {
error!("Error sending event: {e:?}");
break;
}
if is_shutdown_complete {
info!("Received shutdown event, exiting event loop.");
break;
}
},
Err(e) => {
error!("Error receiving event: {e:?}");
break;
}
}
}
}
});
}
// Package images and prompt into a single user input turn.
let mut items: Vec<UserInput> = images
.into_iter()
.map(|path| UserInput::LocalImage { path })
.collect();
items.push(UserInput::Text { text: prompt });
let initial_prompt_task_id = conversation
.submit(Op::UserTurn {
items,
cwd: default_cwd,
approval_policy: default_approval_policy,
sandbox_policy: default_sandbox_policy,
model: default_model,
effort: default_effort,
summary: default_summary,
final_output_json_schema: output_schema,
})
.await?;
info!("Sent prompt with event ID: {initial_prompt_task_id}");
// Run the loop until the task is complete.
// Track whether a fatal error was reported by the server so we can
// exit with a non-zero status for automation-friendly signaling.
let mut error_seen = false;
while let Some(event) = rx.recv().await {
if matches!(event.msg, EventMsg::Error(_)) {
error_seen = true;
}
let shutdown: CodexStatus = event_processor.process_event(event);
match shutdown {
CodexStatus::Running => continue,
CodexStatus::InitiateShutdown => {
conversation.submit(Op::Shutdown).await?;
}
CodexStatus::Shutdown => {
break;
}
}
}
event_processor.print_final_output();
if error_seen {
std::process::exit(1);
}
Ok(())
}
async fn resolve_resume_path(
config: &Config,
args: &crate::cli::ResumeArgs,
) -> anyhow::Result<Option<PathBuf>> {
if args.last {
let default_provider_filter = vec![config.model_provider_id.clone()];
match codex_core::RolloutRecorder::list_conversations(
&config.codex_home,
1,
None,
&[],
Some(default_provider_filter.as_slice()),
&config.model_provider_id,
)
.await
{
Ok(page) => Ok(page.items.first().map(|it| it.path.clone())),
Err(e) => {
error!("Error listing conversations: {e}");
Ok(None)
}
}
} else if let Some(id_str) = args.session_id.as_deref() {
let path = find_conversation_path_by_id_str(&config.codex_home, id_str).await?;
Ok(path)
} else {
Ok(None)
}
}
fn load_output_schema(path: Option<PathBuf>) -> Option<Value> {
let path = path?;
let schema_str = match std::fs::read_to_string(&path) {
Ok(contents) => contents,
Err(err) => {
eprintln!(
"Failed to read output schema file {}: {err}",
path.display()
);
std::process::exit(1);
}
};
match serde_json::from_str::<Value>(&schema_str) {
Ok(value) => Some(value),
Err(err) => {
eprintln!(
"Output schema file {} is not valid JSON: {err}",
path.display()
);
std::process::exit(1);
}
}
}

40
llmx-rs/exec/src/main.rs Normal file
View File

@@ -0,0 +1,40 @@
//! Entry-point for the `codex-exec` binary.
//!
//! When this CLI is invoked normally, it parses the standard `codex-exec` CLI
//! options and launches the non-interactive Codex agent. However, if it is
//! invoked with arg0 as `codex-linux-sandbox`, we instead treat the invocation
//! as a request to run the logic for the standalone `codex-linux-sandbox`
//! executable (i.e., parse any -s args and then run a *sandboxed* command under
//! Landlock + seccomp.
//!
//! This allows us to ship a completely separate set of functionality as part
//! of the `codex-exec` binary.
use clap::Parser;
use codex_arg0::arg0_dispatch_or_else;
use codex_common::CliConfigOverrides;
use codex_exec::Cli;
use codex_exec::run_main;
#[derive(Parser, Debug)]
struct TopCli {
#[clap(flatten)]
config_overrides: CliConfigOverrides,
#[clap(flatten)]
inner: Cli,
}
fn main() -> anyhow::Result<()> {
arg0_dispatch_or_else(|codex_linux_sandbox_exe| async move {
let top_cli = TopCli::parse();
// Merge root-level overrides into inner CLI struct so downstream logic remains unchanged.
let mut inner = top_cli.inner;
inner
.config_overrides
.raw_overrides
.splice(0..0, top_cli.config_overrides.raw_overrides);
run_main(inner, codex_linux_sandbox_exe).await?;
Ok(())
})
}