import { spawn } from 'child_process';import { promisify } from 'util';import { execFile as execFileCallback } from 'child_process';import { createReadStream, writeFileSync, unlinkSync } from 'fs';import { join } from 'path';import { tmpdir } from 'os';import dotenv from 'dotenv';import type { Runner, RunnerResult, RunnerOptions, RunnerStatus, TokenUsage } from './types.js';import { debug, debugPrompt, debugResponse } from '../logger.js';import { isValidModelName } from '../config.js'; const execFile = promisify(execFileCallback); // Validate model name to prevent injection (defense in depth)// OpenAI Codex CLI binary namesconst CODEX_BINARIES = ['codex', 'openai-codex']; export class CodexRunner implements Runner { name = 'codex'; displayName = 'OpenAI Codex'; installHint = 'npm install -g @openai/codex'; private binaryPath: string = 'codex'; async isAvailable(): Promise<boolean> { for (const binary of CODEX_BINARIES) { try { await execFile('which', [binary]); this.binaryPath = binary; debug(`Found Codex CLI at: ${binary}`); return true; } catch { // Try next binary } } debug('Codex CLI not found', { tried: CODEX_BINARIES }); return false; } async checkStatus(): Promise<RunnerStatus> { const installed = await this.isAvailable(); if (!installed) { return { installed: false, ready: false, error: 'Not installed' }; } // Check version let version: string | undefined; try { const { stdout, stderr } = await execFile(this.binaryPath, ['--version']); version = (stdout || stderr).trim(); } catch { // Version check might fail } // Check for OpenAI API key (load .env from cwd if not set) if (!process.env.OPENAI_API_KEY) { dotenv.config({ path: join(process.cwd(), '.env') }); } if (!process.env.OPENAI_API_KEY) { return { installed: true, ready: false, version, error: 'OPENAI_API_KEY not set (set in .env or environment)' }; } return { installed: true, ready: true, version }; } async run(workdir: string, prompt: string, options?: RunnerOptions): Promise<RunnerResult> { // Guard: Don't run with empty prompt if (!prompt || prompt.trim().length === 0) { debug('Empty prompt - skipping codex run'); return { success: false, output: '', error: 'No prompt provided (nothing to fix)' }; } // Validate model before writing sensitive prompt to disk if (options?.model && !isValidModelName(options.model)) { return { success: false, output: '', error: `Invalid model name: ${options.model}` }; } // Guard: Reject non-OpenAI models — Codex only supports OpenAI models. // WHY: If model rotation recommends an Anthropic/Google model (e.g., after // a runner switch), sending it to the OpenAI API wastes a retry cycle and // triggers a confusing "model does not exist" auth error. if (options?.model && /^(claude|anthropic|gemini)/i.test(options.model)) { const msg = `Model "${options.model}" is not an OpenAI model — skipping for Codex`; debug(msg); return { success: false, output: '', error: msg, errorType: 'model' }; } // Write prompt to temp file for stdin piping const promptFile = join(tmpdir(), `prr-prompt.${process.pid}.${Date.now()}.txt`); writeFileSync(promptFile, prompt, { encoding: 'utf-8', mode: 0o600 }); debug('Wrote prompt to file', { promptFile, length: prompt.length }); const promptSlug = debugPrompt('codex', prompt, { workdir, model: options?.model }); const cleanupPromptFile = () => { try { unlinkSync(promptFile); } catch { // Ignore cleanup errors } }; // Build args for `codex exec` - the non-interactive mode // WHY: Interactive mode requires TTY for cursor position queries, which fails in automation. // `codex exec` is designed for non-interactive/CI use and doesn't need a TTY. const args: string[] = ['exec']; // Force API key auth so Codex uses OPENAI_API_KEY instead of cached ChatGPT/Teams login. // See https://github.com/openai/codex/issues/2733 — when a Teams subscription login is active, // Codex can ignore OPENAI_API_KEY and use the session (which may 401 for API calls). args.push('--config', 'forced_login_method=api'); // Bypass sandbox and approvals - prr controls the execution environment // WHY: Codex's landlock sandbox can fail in some environments (containers, etc.) // prr already isolates work in a cloned workdir, so external sandboxing is sufficient args.push('--dangerously-bypass-approvals-and-sandbox'); // Set working directory args.push('-C', workdir); // Add model if specified if (options?.model) { args.push('--model', options.model); } // Add extra directories if specified if (options?.codexAddDirs && options.codexAddDirs.length > 0) { for (const dir of options.codexAddDirs) { if (!dir) continue; args.push('--add-dir', dir); } } // JSONL stdout so we can parse turn.completed (token usage) and item.completed agent_message (final output) args.push('--json'); // Read prompt from stdin args.push('-'); return new Promise((resolve) => { const modelInfo = options?.model ? ` (model: ${options.model})` : ''; console.log(`\nRunning: ${this.binaryPath} exec${modelInfo} [prompt via stdin]\n`); debug('Codex exec command', { binary: this.binaryPath, args: args.filter(a => a !== '-'), // Don't log the stdin marker workdir, model: options?.model, promptLength: prompt.length, }); // Prefer key from options (config), then process.env, then .env from cwd let apiKey = options?.openaiApiKey?.trim(); if (!apiKey && !process.env.OPENAI_API_KEY) { dotenv.config({ path: join(process.cwd(), '.env') }); } if (!apiKey) { apiKey = process.env.OPENAI_API_KEY?.trim(); } const spawnEnv: NodeJS.ProcessEnv = { ...process.env, CI: '1', NO_COLOR: '1', FORCE_COLOR: '0', }; if (apiKey) { spawnEnv.OPENAI_API_KEY = apiKey; // codex exec uses CODEX_API_KEY for implicit login, not OPENAI_API_KEY (openai/codex#7323). spawnEnv.CODEX_API_KEY = apiKey; } // Force Codex to use the real OpenAI API so our key works (see openai/codex#9153). // If OPENAI_BASE_URL is set (e.g. to a proxy or api.z.ai), Codex may call that URL and get 401. delete spawnEnv.OPENAI_BASE_URL; delete spawnEnv.OPENAI_API_BASE; const child = spawn(this.binaryPath, args, { cwd: workdir, stdio: ['pipe', 'pipe', 'pipe'], env: spawnEnv, }); // Pipe prompt to stdin const promptStream = createReadStream(promptFile); if (!child.stdin) { const error = 'Codex process stdin is unavailable'; debug(error); promptStream.destroy(); cleanupPromptFile(); resolve({ success: false, output: '', error }); return; } promptStream.pipe(child.stdin); promptStream.on('error', (err) => { debug('Error reading prompt file', { error: err.message }); }); let stdout = ''; let stderr = ''; let lineBuffer = ''; const aggregatedUsage: TokenUsage = { input_tokens: 0, cached_input_tokens: 0, output_tokens: 0 }; let lastAgentMessage = ''; child.stdout?.on('data', (data) => { const str = data.toString(); stdout += str; lineBuffer += str; const lines = lineBuffer.split('\n'); lineBuffer = lines.pop() ?? ''; for (const line of lines) { const trimmed = line.trim(); if (!trimmed) continue; try { const obj = JSON.parse(trimmed) as { type?: string; usage?: TokenUsage; item?: { type?: string; text?: string } }; if (obj.type === 'turn.completed' && obj.usage) { aggregatedUsage.input_tokens += obj.usage.input_tokens ?? 0; aggregatedUsage.cached_input_tokens = (aggregatedUsage.cached_input_tokens ?? 0) + (obj.usage.cached_input_tokens ?? 0); aggregatedUsage.output_tokens += obj.usage.output_tokens ?? 0; } else if (obj.type === 'item.completed' && obj.item?.type === 'agent_message' && typeof obj.item.text === 'string') { lastAgentMessage = obj.item.text; } } catch { // Not JSON or malformed; ignore line } } }); child.stderr?.on('data', (data) => { const str = data.toString(); stderr += str; process.stderr.write(str); }); child.on('close', (code) => { cleanupPromptFile(); // Parse any remaining line (no trailing newline) if (lineBuffer.trim()) { try { const obj = JSON.parse(lineBuffer.trim()) as { type?: string; usage?: TokenUsage; item?: { type?: string; text?: string } }; if (obj.type === 'turn.completed' && obj.usage) { aggregatedUsage.input_tokens += obj.usage.input_tokens ?? 0; aggregatedUsage.cached_input_tokens = (aggregatedUsage.cached_input_tokens ?? 0) + (obj.usage.cached_input_tokens ?? 0); aggregatedUsage.output_tokens += obj.usage.output_tokens ?? 0; } else if (obj.type === 'item.completed' && obj.item?.type === 'agent_message' && typeof obj.item.text === 'string') { lastAgentMessage = obj.item.text; } } catch { // ignore } } const output = lastAgentMessage || stdout; debugResponse(promptSlug, 'codex', output, { exitCode: code, stderrLength: stderr.length, usage: aggregatedUsage }); // Safety check: detect cursor position error even in exec mode (shouldn't happen, but just in case) const hasCursorError = isCursorPositionError(stdout) || isCursorPositionError(stderr); if (hasCursorError) { debug('Codex cursor position error detected in exec mode - unexpected environment issue'); resolve({ success: false, output, error: 'Codex cursor position error: TTY/PTY environment issue. This is unexpected in exec mode.', errorType: 'environment', }); return; } const usage = (aggregatedUsage.input_tokens || aggregatedUsage.output_tokens) ? aggregatedUsage : undefined; if (code === 0) { resolve({ success: true, output, usage }); } else { // Check for common error patterns const combinedOutput = stdout + stderr; // AUTH (401 / missing key) FIRST — so we don't misclassify as quota. // 401 Unauthorized / Missing bearer = API key not set or invalid; user needs to fix env, not rotate. if (/\b401\b|unauthorized|missing bearer|basic authentication in header/i.test(combinedOutput)) { const authMsg = /missing bearer|basic authentication/i.test(combinedOutput) ? 'OpenAI API key missing or invalid (401 Unauthorized). For codex exec we set CODEX_API_KEY from your key; ensure OPENAI_API_KEY (or config) is set so we can pass it. If OPENAI_BASE_URL is set, unset it (see github.com/openai/codex/issues/9153). See also #7323 (codex exec needs CODEX_API_KEY).' : (stderr || 'Authentication error'); debug('Codex auth error (401 or missing key)'); resolve({ success: false, output, error: authMsg, errorType: 'auth', usage }); } else if (/authentication|invalid.*key|api.*key.*rejected/i.test(combinedOutput)) { resolve({ success: false, output, error: stderr || `Authentication error`, errorType: 'auth', usage }); } else if (/quota exceeded|rate.?limit|too many requests|billing|exceeded.*plan|tokens?.used/i.test(combinedOutput)) { const tokensMatch = combinedOutput.match(/tokens?\s*used\s*[\n:]*\s*([\d,]+)/i); const tokenInfo = tokensMatch ? ` (${tokensMatch[1]} tokens used)` : ''; const errorMsg = `Quota/rate limit exceeded${tokenInfo}`; debug('Quota exceeded - will rotate to next tool/model', { error: errorMsg }); resolve({ success: false, output, error: errorMsg, errorType: 'quota', usage }); } else if (/does not exist|model.*not found|you do not have access/i.test(combinedOutput)) { // Model doesn't exist or API key lacks access - bail immediately // WHY: Retrying won't help. Need different model or API access. const modelError = combinedOutput.match(/The model [`'"]?([^`'"]+)[`'"]? does not exist/i); const errorMsg = modelError ? `Model "${modelError[1]}" does not exist or is not accessible` : stderr || 'Model not found or not accessible'; debug('Model access error - bailing immediately', { error: errorMsg }); resolve({ success: false, output, error: errorMsg, errorType: 'auth', usage }); } else if (/permission denied|cannot write|read-only/i.test(combinedOutput)) { resolve({ success: false, output, error: stderr || `Permission error`, errorType: 'permission', usage }); } else { resolve({ success: false, output, error: stderr || `Process exited with code ${code}`, usage }); } } }); child.on('error', (err) => { cleanupPromptFile(); resolve({ success: false, output: lastAgentMessage || stdout, error: err.message }); }); }); }} function isCursorPositionError(output?: string): boolean { // Codex throws this when it can't query terminal cursor position // WHY: Interactive mode uses TUI elements that need cursor position // This shouldn't happen in exec mode, but we check as a safety measure if (!output) return false; const cleaned = output.replace(/\x1b\[[0-9;]*[a-zA-Z]/g, '').replace(/[\x00-\x1f]/g, ' '); return /cursor.{0,10}position.{0,10}could.{0,10}not.{0,10}be.{0,10}read/i.test(cleaned);}
import { spawn } from 'child_process';
import { promisify } from 'util';
import { execFile as execFileCallback } from 'child_process';
import { createReadStream, writeFileSync, unlinkSync } from 'fs';
import { join } from 'path';
import { tmpdir } from 'os';
import dotenv from 'dotenv';
import type { Runner, RunnerResult, RunnerOptions, RunnerStatus, TokenUsage } from './types.js';
import { debug, debugPrompt, debugResponse } from '../logger.js';
import { isValidModelName } from '../config.js';
const execFile = promisify(execFileCallback);
// Validate model name to prevent injection (defense in depth)
// OpenAI Codex CLI binary names
const CODEX_BINARIES = ['codex', 'openai-codex'];
export class CodexRunner implements Runner {
name = 'codex';
displayName = 'OpenAI Codex';
installHint = 'npm install -g @openai/codex';
private binaryPath: string = 'codex';
async isAvailable(): Promise<boolean> {
for (const binary of CODEX_BINARIES) {
try {
await execFile('which', [binary]);
this.binaryPath = binary;
debug(`Found Codex CLI at: ${binary}`);
return true;
} catch {
// Try next binary
}
}
debug('Codex CLI not found', { tried: CODEX_BINARIES });
return false;
}
async checkStatus(): Promise<RunnerStatus> {
const installed = await this.isAvailable();
if (!installed) {
return { installed: false, ready: false, error: 'Not installed' };
}
// Check version
let version: string | undefined;
try {
const { stdout, stderr } = await execFile(this.binaryPath, ['--version']);
version = (stdout || stderr).trim();
} catch {
// Version check might fail
}
// Check for OpenAI API key (load .env from cwd if not set)
if (!process.env.OPENAI_API_KEY) {
dotenv.config({ path: join(process.cwd(), '.env') });
}
if (!process.env.OPENAI_API_KEY) {
return { installed: true, ready: false, version, error: 'OPENAI_API_KEY not set (set in .env or environment)' };
}
return { installed: true, ready: true, version };
}
async run(workdir: string, prompt: string, options?: RunnerOptions): Promise<RunnerResult> {
// Guard: Don't run with empty prompt
if (!prompt || prompt.trim().length === 0) {
debug('Empty prompt - skipping codex run');
return { success: false, output: '', error: 'No prompt provided (nothing to fix)' };
}
// Validate model before writing sensitive prompt to disk
if (options?.model && !isValidModelName(options.model)) {
return { success: false, output: '', error: `Invalid model name: ${options.model}` };
}
// Guard: Reject non-OpenAI models — Codex only supports OpenAI models.
// WHY: If model rotation recommends an Anthropic/Google model (e.g., after
// a runner switch), sending it to the OpenAI API wastes a retry cycle and
// triggers a confusing "model does not exist" auth error.
if (options?.model && /^(claude|anthropic|gemini)/i.test(options.model)) {
const msg = `Model "${options.model}" is not an OpenAI model — skipping for Codex`;
debug(msg);
return { success: false, output: '', error: msg, errorType: 'model' };
}
// Write prompt to temp file for stdin piping
const promptFile = join(tmpdir(), `prr-prompt.${process.pid}.${Date.now()}.txt`);
writeFileSync(promptFile, prompt, { encoding: 'utf-8', mode: 0o600 });
debug('Wrote prompt to file', { promptFile, length: prompt.length });
const promptSlug = debugPrompt('codex', prompt, { workdir, model: options?.model });
const cleanupPromptFile = () => {
try {
unlinkSync(promptFile);
} catch {
// Ignore cleanup errors
}
};
// Build args for `codex exec` - the non-interactive mode
// WHY: Interactive mode requires TTY for cursor position queries, which fails in automation.
// `codex exec` is designed for non-interactive/CI use and doesn't need a TTY.
const args: string[] = ['exec'];
// Force API key auth so Codex uses OPENAI_API_KEY instead of cached ChatGPT/Teams login.
// See https://github.com/openai/codex/issues/2733 — when a Teams subscription login is active,
// Codex can ignore OPENAI_API_KEY and use the session (which may 401 for API calls).
args.push('--config', 'forced_login_method=api');
// Bypass sandbox and approvals - prr controls the execution environment
// WHY: Codex's landlock sandbox can fail in some environments (containers, etc.)
// prr already isolates work in a cloned workdir, so external sandboxing is sufficient
args.push('--dangerously-bypass-approvals-and-sandbox');
// Set working directory
args.push('-C', workdir);
// Add model if specified
if (options?.model) {
args.push('--model', options.model);
}
// Add extra directories if specified
if (options?.codexAddDirs && options.codexAddDirs.length > 0) {
for (const dir of options.codexAddDirs) {
if (!dir) continue;
args.push('--add-dir', dir);
}
}
// JSONL stdout so we can parse turn.completed (token usage) and item.completed agent_message (final output)
args.push('--json');
// Read prompt from stdin
args.push('-');
return new Promise((resolve) => {
const modelInfo = options?.model ? ` (model: ${options.model})` : '';
console.log(`\nRunning: ${this.binaryPath} exec${modelInfo} [prompt via stdin]\n`);
debug('Codex exec command', {
binary: this.binaryPath,
args: args.filter(a => a !== '-'), // Don't log the stdin marker
workdir,
model: options?.model,
promptLength: prompt.length,
});
// Prefer key from options (config), then process.env, then .env from cwd
let apiKey = options?.openaiApiKey?.trim();
if (!apiKey && !process.env.OPENAI_API_KEY) {
dotenv.config({ path: join(process.cwd(), '.env') });
}
if (!apiKey) {
apiKey = process.env.OPENAI_API_KEY?.trim();
}
const spawnEnv: NodeJS.ProcessEnv = {
...process.env,
CI: '1',
NO_COLOR: '1',
FORCE_COLOR: '0',
};
if (apiKey) {
spawnEnv.OPENAI_API_KEY = apiKey;
// codex exec uses CODEX_API_KEY for implicit login, not OPENAI_API_KEY (openai/codex#7323).
spawnEnv.CODEX_API_KEY = apiKey;
}
// Force Codex to use the real OpenAI API so our key works (see openai/codex#9153).
// If OPENAI_BASE_URL is set (e.g. to a proxy or api.z.ai), Codex may call that URL and get 401.
delete spawnEnv.OPENAI_BASE_URL;
delete spawnEnv.OPENAI_API_BASE;
const child = spawn(this.binaryPath, args, {
cwd: workdir,
stdio: ['pipe', 'pipe', 'pipe'],
env: spawnEnv,
});
// Pipe prompt to stdin
const promptStream = createReadStream(promptFile);
if (!child.stdin) {
const error = 'Codex process stdin is unavailable';
debug(error);
promptStream.destroy();
cleanupPromptFile();
resolve({ success: false, output: '', error });
return;
}
promptStream.pipe(child.stdin);
promptStream.on('error', (err) => {
debug('Error reading prompt file', { error: err.message });
});
let stdout = '';
let stderr = '';
let lineBuffer = '';
const aggregatedUsage: TokenUsage = { input_tokens: 0, cached_input_tokens: 0, output_tokens: 0 };
let lastAgentMessage = '';
child.stdout?.on('data', (data) => {
const str = data.toString();
stdout += str;
lineBuffer += str;
const lines = lineBuffer.split('\n');
lineBuffer = lines.pop() ?? '';
for (const line of lines) {
const trimmed = line.trim();
if (!trimmed) continue;
try {
const obj = JSON.parse(trimmed) as { type?: string; usage?: TokenUsage; item?: { type?: string; text?: string } };
if (obj.type === 'turn.completed' && obj.usage) {
aggregatedUsage.input_tokens += obj.usage.input_tokens ?? 0;
aggregatedUsage.cached_input_tokens = (aggregatedUsage.cached_input_tokens ?? 0) + (obj.usage.cached_input_tokens ?? 0);
aggregatedUsage.output_tokens += obj.usage.output_tokens ?? 0;
} else if (obj.type === 'item.completed' && obj.item?.type === 'agent_message' && typeof obj.item.text === 'string') {
lastAgentMessage = obj.item.text;
}
} catch {
// Not JSON or malformed; ignore line
}
}
});
child.stderr?.on('data', (data) => {
const str = data.toString();
stderr += str;
process.stderr.write(str);
});
child.on('close', (code) => {
cleanupPromptFile();
// Parse any remaining line (no trailing newline)
if (lineBuffer.trim()) {
try {
const obj = JSON.parse(lineBuffer.trim()) as { type?: string; usage?: TokenUsage; item?: { type?: string; text?: string } };
if (obj.type === 'turn.completed' && obj.usage) {
aggregatedUsage.input_tokens += obj.usage.input_tokens ?? 0;
aggregatedUsage.cached_input_tokens = (aggregatedUsage.cached_input_tokens ?? 0) + (obj.usage.cached_input_tokens ?? 0);
aggregatedUsage.output_tokens += obj.usage.output_tokens ?? 0;
} else if (obj.type === 'item.completed' && obj.item?.type === 'agent_message' && typeof obj.item.text === 'string') {
lastAgentMessage = obj.item.text;
}
} catch {
// ignore
}
}
const output = lastAgentMessage || stdout;
debugResponse(promptSlug, 'codex', output, { exitCode: code, stderrLength: stderr.length, usage: aggregatedUsage });
// Safety check: detect cursor position error even in exec mode (shouldn't happen, but just in case)
const hasCursorError = isCursorPositionError(stdout) || isCursorPositionError(stderr);
if (hasCursorError) {
debug('Codex cursor position error detected in exec mode - unexpected environment issue');
resolve({
success: false,
output,
error: 'Codex cursor position error: TTY/PTY environment issue. This is unexpected in exec mode.',
errorType: 'environment',
});
return;
}
const usage = (aggregatedUsage.input_tokens || aggregatedUsage.output_tokens) ? aggregatedUsage : undefined;
if (code === 0) {
resolve({ success: true, output, usage });
} else {
// Check for common error patterns
const combinedOutput = stdout + stderr;
// AUTH (401 / missing key) FIRST — so we don't misclassify as quota.
// 401 Unauthorized / Missing bearer = API key not set or invalid; user needs to fix env, not rotate.
if (/\b401\b|unauthorized|missing bearer|basic authentication in header/i.test(combinedOutput)) {
const authMsg = /missing bearer|basic authentication/i.test(combinedOutput)
? 'OpenAI API key missing or invalid (401 Unauthorized). For codex exec we set CODEX_API_KEY from your key; ensure OPENAI_API_KEY (or config) is set so we can pass it. If OPENAI_BASE_URL is set, unset it (see github.com/openai/codex/issues/9153). See also #7323 (codex exec needs CODEX_API_KEY).'
: (stderr || 'Authentication error');
debug('Codex auth error (401 or missing key)');
resolve({ success: false, output, error: authMsg, errorType: 'auth', usage });
} else if (/authentication|invalid.*key|api.*key.*rejected/i.test(combinedOutput)) {
resolve({ success: false, output, error: stderr || `Authentication error`, errorType: 'auth', usage });
} else if (/quota exceeded|rate.?limit|too many requests|billing|exceeded.*plan|tokens?.used/i.test(combinedOutput)) {
const tokensMatch = combinedOutput.match(/tokens?\s*used\s*[\n:]*\s*([\d,]+)/i);
const tokenInfo = tokensMatch ? ` (${tokensMatch[1]} tokens used)` : '';
const errorMsg = `Quota/rate limit exceeded${tokenInfo}`;
debug('Quota exceeded - will rotate to next tool/model', { error: errorMsg });
resolve({ success: false, output, error: errorMsg, errorType: 'quota', usage });
} else if (/does not exist|model.*not found|you do not have access/i.test(combinedOutput)) {
// Model doesn't exist or API key lacks access - bail immediately
// WHY: Retrying won't help. Need different model or API access.
const modelError = combinedOutput.match(/The model [`'"]?([^`'"]+)[`'"]? does not exist/i);
const errorMsg = modelError
? `Model "${modelError[1]}" does not exist or is not accessible`
: stderr || 'Model not found or not accessible';
debug('Model access error - bailing immediately', { error: errorMsg });
resolve({ success: false, output, error: errorMsg, errorType: 'auth', usage });
} else if (/permission denied|cannot write|read-only/i.test(combinedOutput)) {
resolve({ success: false, output, error: stderr || `Permission error`, errorType: 'permission', usage });
} else {
resolve({ success: false, output, error: stderr || `Process exited with code ${code}`, usage });
}
}
});
child.on('error', (err) => {
cleanupPromptFile();
resolve({ success: false, output: lastAgentMessage || stdout, error: err.message });
});
});
}
}
function isCursorPositionError(output?: string): boolean {
// Codex throws this when it can't query terminal cursor position
// WHY: Interactive mode uses TUI elements that need cursor position
// This shouldn't happen in exec mode, but we check as a safety measure
if (!output) return false;
const cleaned = output.replace(/\x1b\[[0-9;]*[a-zA-Z]/g, '').replace(/[\x00-\x1f]/g, ' ');
return /cursor.{0,10}position.{0,10}could.{0,10}not.{0,10}be.{0,10}read/i.test(cleaned);
}