codex/codex-cli/src/cli.tsx at main · tomasgrozman/codex

458 lines (418 loc) · 14.3 KB
#!/usr/bin/env node
import "dotenv/config";
// Hack to suppress deprecation warnings (punycode)
// eslint-disable-next-line @typescript-eslint/no-explicit-any
(process as any).noDeprecation = true;
import type { AppRollout } from "./app";
import type { ApprovalPolicy } from "./approvals";
import type { CommandConfirmation } from "./utils/agent/agent-loop";
import type { AppConfig } from "./utils/config";
import type { ResponseItem } from "openai/resources/responses/responses";
import App from "./app";
import { runSinglePass } from "./cli_singlepass";
import { AgentLoop } from "./utils/agent/agent-loop";
import { initLogger } from "./utils/agent/log";
import { ReviewDecision } from "./utils/agent/review";
import { AutoApprovalMode } from "./utils/auto-approval-mode";
  loadConfig,
  PRETTY_PRINT,
  INSTRUCTIONS_FILEPATH,
} from "./utils/config";
import { createInputItem } from "./utils/input-utils";
  isModelSupportedForResponses,
  preloadModels,
} from "./utils/model-utils.js";
import { parseToolCall } from "./utils/parsers";
import { onExit, setInkRenderer } from "./utils/terminal";
import chalk from "chalk";
import { spawnSync } from "child_process";
import fs from "fs";
import { render } from "ink";
import meow from "meow";
import path from "path";
import React from "react";
// Call this early so `tail -F "$TMPDIR/oai-codex/codex-cli-latest.log"` works
// immediately. This must be run with DEBUG=1 for logging to work.
initLogger();
// TODO: migrate to new versions of quiet mode
//     -q, --quiet    Non-interactive quiet mode that only prints final message
//     -j, --json     Non-interactive JSON output mode that prints JSON messages
const cli = meow(
    $ codex [options] <prompt>
    $ codex completion <bash|zsh|fish>
    -h, --help                 Show usage and exit
    -m, --model <model>        Model to use for completions (default: o4-mini)
    -i, --image <path>         Path(s) to image files to include as input
    -v, --view <rollout>       Inspect a previously saved rollout instead of starting a session
    -q, --quiet                Non-interactive mode that only prints the assistant's final output
    -c, --config               Open the instructions file in your editor
    -a, --approval-mode <mode> Override the approval policy: 'suggest', 'auto-edit', or 'full-auto'
    --auto-edit                Automatically approve file edits; still prompt for commands
    --full-auto                Automatically approve edits and commands when executed in the sandbox
    --no-project-doc           Do not automatically include the repository's 'codex.md'
    --project-doc <file>       Include an additional markdown file at <file> as context
    --full-stdout              Do not truncate stdout/stderr from command outputs
  Dangerous options
    --dangerously-auto-approve-everything
                               Skip all confirmation prompts and execute commands without
                               sandboxing. Intended solely for ephemeral local testing.
  Experimental options
    -f, --full-context         Launch in "full-context" mode which loads the entire repository
                               into context and applies a batch of edits in one go. Incompatible
                               with all other flags, except for --model.
    $ codex "Write and run a python program that prints ASCII art"
    $ codex -q "fix build issues"
    $ codex completion bash
    importMeta: import.meta,
    autoHelp: true,
    flags: {
      // misc
      help: { type: "boolean", aliases: ["h"] },
      view: { type: "string" },
      model: { type: "string", aliases: ["m"] },
      image: { type: "string", isMultiple: true, aliases: ["i"] },
      quiet: {
        type: "boolean",
        aliases: ["q"],
        description: "Non-interactive quiet mode",
      config: {
        type: "boolean",
        aliases: ["c"],
        description: "Open the instructions file in your editor",
      dangerouslyAutoApproveEverything: {
        type: "boolean",
        description:
          "Automatically approve all commands without prompting. This is EXTREMELY DANGEROUS and should only be used in trusted environments.",
      autoEdit: {
        type: "boolean",
        description: "Automatically approve edits; prompt for commands.",
      fullAuto: {
        type: "boolean",
        description:
          "Automatically run commands in a sandbox; only prompt for failures.",
      approvalMode: {
        type: "string",
        aliases: ["a"],
        description:
          "Determine the approval mode for Codex (default: suggest) Values: suggest, auto-edit, full-auto",
      noProjectDoc: {
        type: "boolean",
        description: "Disable automatic inclusion of project‑level codex.md",
      projectDoc: {
        type: "string",
        description: "Path to a markdown file to include as project doc",
      fullStdout: {
        type: "boolean",
        description:
          "Disable truncation of command stdout/stderr messages (show everything)",
        aliases: ["no-truncate"],
      // Experimental mode where whole directory is loaded in context and model is requested
      // to make code edits in a single pass.
      fullContext: {
        type: "boolean",
        aliases: ["f"],
        description: `Run in full-context editing approach. The model is given the whole code
          directory as context and performs changes in one go without acting.`,
// Handle 'completion' subcommand before any prompting or API calls
if (cli.input[0] === "completion") {
  const shell = cli.input[1] || "bash";
  const scripts: Record<string, string> = {
    bash: `# bash completion for codex
_codex_completion() {
  local cur
  cur="\${COMP_WORDS[COMP_CWORD]}"
  COMPREPLY=( $(compgen -o default -o filenames -- "\${cur}") )
complete -F _codex_completion codex`,
    zsh: `# zsh completion for codex
#compdef codex
  _arguments '*:filename:_files'
    fish: `# fish completion for codex
complete -c codex -a '(_fish_complete_path)' -d 'file path'`,
  const script = scripts[shell];
  if (!script) {
    // eslint-disable-next-line no-console
    console.error(`Unsupported shell: ${shell}`);
    process.exit(1);
  // eslint-disable-next-line no-console
  console.log(script);
  process.exit(0);
// Show help if requested
if (cli.flags.help) {
  cli.showHelp();
// Handle config flag: open instructions file in editor and exit
if (cli.flags.config) {
  // Ensure configuration and instructions file exist
    loadConfig();
  } catch {
    // ignore errors
  const filePath = INSTRUCTIONS_FILEPATH;
  const editor =
    process.env["EDITOR"] || (process.platform === "win32" ? "notepad" : "vi");
  spawnSync(editor, [filePath], { stdio: "inherit" });
  process.exit(0);
// ---------------------------------------------------------------------------
// API key handling
// ---------------------------------------------------------------------------
const apiKey = process.env["OPENAI_API_KEY"];
if (!apiKey) {
  // eslint-disable-next-line no-console
  console.error(
    `\n${chalk.red("Missing OpenAI API key.")}\n\n` +
      `Set the environment variable ${chalk.bold("OPENAI_API_KEY")} ` +
      `and re-run this command.\n` +
      `You can create a key here: ${chalk.bold(
        chalk.underline("https://platform.openai.com/account/api-keys"),
      )}\n`,
  process.exit(1);
const fullContextMode = Boolean(cli.flags.fullContext);
let config = loadConfig(undefined, undefined, {
  cwd: process.cwd(),
  disableProjectDoc: Boolean(cli.flags.noProjectDoc),
  projectDocPath: cli.flags.projectDoc as string | undefined,
  isFullContext: fullContextMode,
const prompt = cli.input[0];
const model = cli.flags.model;
const imagePaths = cli.flags.image as Array<string> | undefined;
  ...config,
  model: model ?? config.model,
if (!(await isModelSupportedForResponses(config.model))) {
  // eslint-disable-next-line no-console
  console.error(
    `The model "${config.model}" does not appear in the list of models ` +
      `available to your account. Double‑check the spelling (use\n` +
      `  openai models list\n` +
      `to see the full list) or choose another model with the --model flag.`,
  process.exit(1);
let rollout: AppRollout | undefined;
if (cli.flags.view) {
  const viewPath = cli.flags.view;
  const absolutePath = path.isAbsolute(viewPath)
    ? viewPath
    : path.join(process.cwd(), viewPath);
    const content = fs.readFileSync(absolutePath, "utf-8");
    rollout = JSON.parse(content) as AppRollout;
  } catch (error) {
    // eslint-disable-next-line no-console
    console.error("Error reading rollout file:", error);
    process.exit(1);
// If we are running in --fullcontext mode, do that and exit.
if (fullContextMode) {
  await runSinglePass({
    originalPrompt: prompt,
    config,
    rootPath: process.cwd(),
  onExit();
  process.exit(0);
// If we are running in --quiet mode, do that and exit.
const quietMode = Boolean(cli.flags.quiet);
const autoApproveEverything = Boolean(
  cli.flags.dangerouslyAutoApproveEverything,
const fullStdout = Boolean(cli.flags.fullStdout);
if (quietMode) {
  process.env["CODEX_QUIET_MODE"] = "1";
  if (!prompt || prompt.trim() === "") {
    // eslint-disable-next-line no-console
    console.error(
      'Quiet mode requires a prompt string, e.g.,: codex -q "Fix bug #123 in the foobar project"',
    process.exit(1);
  await runQuietMode({
    prompt: prompt as string,
    imagePaths: imagePaths || [],
    approvalPolicy: autoApproveEverything
      ? AutoApprovalMode.FULL_AUTO
      : AutoApprovalMode.SUGGEST,
    config,
  onExit();
  process.exit(0);
// Default to the "suggest" policy.
// Determine the approval policy to use in interactive mode.
// Priority (highest → lowest):
// 1. --fullAuto – run everything automatically in a sandbox.
// 2. --dangerouslyAutoApproveEverything – run everything **without** a sandbox
//    or prompts.  This is intended for completely trusted environments.  Since
//    it is more dangerous than --fullAuto we deliberately give it lower
//    priority so a user specifying both flags still gets the safer behaviour.
// 3. --autoEdit – automatically approve edits, but prompt for commands.
// 4. Default – suggest mode (prompt for everything).
const approvalPolicy: ApprovalPolicy =
  cli.flags.fullAuto || cli.flags.approvalMode === "full-auto"
    ? AutoApprovalMode.FULL_AUTO
    : cli.flags.autoEdit || cli.flags.approvalMode === "auto-edit"
    ? AutoApprovalMode.AUTO_EDIT
    : AutoApprovalMode.SUGGEST;
preloadModels();
const instance = render(
    prompt={prompt}
    config={config}
    rollout={rollout}
    imagePaths={imagePaths}
    approvalPolicy={approvalPolicy}
    fullStdout={fullStdout}
    patchConsole: process.env["DEBUG"] ? false : true,
setInkRenderer(instance);
function formatResponseItemForQuietMode(item: ResponseItem): string {
  if (!PRETTY_PRINT) {
    return JSON.stringify(item);
  switch (item.type) {
    case "message": {
      const role = item.role === "assistant" ? "assistant" : item.role;
      const txt = item.content
        .map((c) => {
          if (c.type === "output_text" || c.type === "input_text") {
            return c.text;
          if (c.type === "input_image") {
            return "<Image>";
          if (c.type === "input_file") {
            return c.filename;
          if (c.type === "refusal") {
            return c.refusal;
          return "?";
        .join(" ");
      return `${role}: ${txt}`;
    case "function_call": {
      const details = parseToolCall(item);
      return `$ ${details?.cmdReadableText ?? item.name}`;
    case "function_call_output": {
      // @ts-expect-error metadata unknown on ResponseFunctionToolCallOutputItem
      const meta = item.metadata as ExecOutputMetadata;
      const parts: Array<string> = [];
      if (typeof meta?.exit_code === "number") {
        parts.push(`code: ${meta.exit_code}`);
      if (typeof meta?.duration_seconds === "number") {
        parts.push(`duration: ${meta.duration_seconds}s`);
      const header = parts.length > 0 ? ` (${parts.join(", ")})` : "";
      return `command.stdout${header}\n${item.output}`;
    default: {
      return JSON.stringify(item);
async function runQuietMode({
  imagePaths,
  approvalPolicy,
  prompt: string;
  imagePaths: Array<string>;
  approvalPolicy: ApprovalPolicy;
  config: AppConfig;
}): Promise<void> {
  const agent = new AgentLoop({
    model: config.model,
    config: config,
    instructions: config.instructions,
    approvalPolicy,
    onItem: (item: ResponseItem) => {
      // eslint-disable-next-line no-console
      console.log(formatResponseItemForQuietMode(item));
    onLoading: () => {
      /* intentionally ignored in quiet mode */
    getCommandConfirmation: (
      _command: Array<string>,
    ): Promise<CommandConfirmation> => {
      return Promise.resolve({ review: ReviewDecision.NO_CONTINUE });
    onLastResponseId: () => {
      /* intentionally ignored in quiet mode */
  const inputItem = await createInputItem(prompt, imagePaths);
  await agent.run([inputItem]);
const exit = () => {
  onExit();
  process.exit(0);
process.on("SIGINT", exit);
process.on("SIGQUIT", exit);
process.on("SIGTERM", exit);
// ---------------------------------------------------------------------------
// Fallback for Ctrl‑C when stdin is in raw‑mode
// ---------------------------------------------------------------------------
if (process.stdin.isTTY) {
  // Ensure we do not leave the terminal in raw mode if the user presses
  // Ctrl‑C while some other component has focus and Ink is intercepting
  // input. Node does *not* emit a SIGINT in raw‑mode, so we listen for the
  // corresponding byte (0x03) ourselves and trigger a graceful shutdown.
  const onRawData = (data: Buffer | string): void => {
    const str = Buffer.isBuffer(data) ? data.toString("utf8") : data;
    if (str === "\u0003") {
      exit();
  process.stdin.on("data", onRawData);
// Ensure terminal clean‑up always runs, even when other code calls
// `process.exit()` directly.
process.once("exit", onExit);
Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FilesExpand file tree

cli.tsx

Latest commit

History

cli.tsx

File metadata and controls
