Run forever as a resident worker in ONE Codex turn: every 5 minutes poll a watched directory for new task files and execute each one to completion, then keep watching โ the turn never ends. Use when the user says "queue-loop", "๋ฌดํ ๋ฃจํ๋ก ๋๋ ค", "ํ ๊ฐ์ํด์ ๊ณ์ ์คํ", "resident worker", "watch a directory and run new tasks forever", or wants a single invocation that keeps picking up dropped commands indefinitely. Because the work stays inside one already-running turn, it keeps issuing model requests across usage-limit windows (the workspace spend cap gates only NEW top-level turns, not an in-flight one). Do NOT use for a single one-shot task, or when the user wants the turn to finish and report back.
Use when the user wants Korean prose cleaned of AI tone, stiff old-fashioned (ํ ์๋ฒ์ง/ํ์ฅ/๊ณต๋ฌธ์) tone, translationese, and excessive quotation marks - or asks to "๊ธ ๋ค๋ฌ์ด", "์์ฐ์ค๋ฝ๊ฒ", "AI ๋งํฌ ์ ๊ฑฐ", "ํ ์๋ฒ์ง ๋งํฌ ๋นผ", "deslop ํ๊ตญ์ด", or "writing-fixer". Removes as many AI and unnatural expressions as possible in one thorough pass while preserving meaning, code, math, and markup. Korean prose only.
Inject a task into the queue-loop worker's queue so a running Codex resident worker picks it up on its next poll (~5 min). Use when the user says "queue-submit", "ํ์ ์์ ๋ฃ์ด", "task ์ฃผ์ ํด", "์ด๊ฑฐ ์์ปคํํ ๋๊ฒจ", "submit a task to the codex worker", "enqueue this for queue-loop", or wants to hand a job off to the always-on worker instead of doing it inline. It only writes a *.task file into the watched queue directory (default ~/codex-queue) โ it does NOT run the task in this session. Do not use when the user wants the work done right now in the current session.
ํ๊ตญ์ด ํ ์คํธ์์ AI๊ฐ ์ด ํ์ , ๋ฒ์ญ์ฒด, ๋ถ์์ฐ์ค๋ฌ์ด ํํ์ ์ ๊ฑฐํ๊ณ ์์ฐ์ค๋ฌ์ด ํ๊ตญ์ด๋ก ๋ค์ ์จ ์ฃผ๋ ์คํฌ. ๋งํฌ ํต์ผ, ๋ฌธ์ฅ๋ถํธ ๊ต์ , ํผ๋ ํํ ์ ๋ฆฌ, ์์ด ์ง์ญ์ฒด ์ ๊ฑฐ๋ฅผ ์ํํ๋ค.
Iteratively review the current repository, plan fixes from the reviews, and implement the plans. Runs three prompts in strict order (deep review โ plan from reviews โ implement with ralph) and repeats the full three-prompt cycle up to N times (default 100). Context is cleaned between iterations by spawning each full cycle in a fresh subagent.
Browser configuration: viewport sizing, device emulation, geolocation, offline mode, color scheme, and HTTP settings.
Browser debugging: execute JavaScript, view console output and page errors, record traces, and run performance profiling.
Browser interaction commands: clicking, typing, filling forms, selecting options, scrolling, dragging, file uploads, and raw mouse/keyboard input.