Codex Record & Replay
A macOS Codex feature that turns a demonstrated workflow into a reusable, inspectable, editable skill. You show Codex a recurring task once — filing an expense report, submitting a time-off request — and it captures that demo as a skill you can later replay and edit. Shipped in Codex app 26.616 on 2026-06-18. Source: X/@OpenAIDevs, 2026-06-18
What it is
Record & Replay closes the gap between "I did this once by hand" and "this is an automation." You control when recording starts and stops; Codex watches the demonstration and produces a skill — not an opaque macro, but an inspectable, editable artifact you can review and modify before reusing. The recurring example use cases (expense reports, time-off requests) are exactly the repetitive GUI/desktop chores that don't fit a clean API. Source: X/@OpenAIDevs, 2026-06-18; developers.openai.com/codex/record-and-replay
Requirements and availability
- macOS only, shipped in the Codex app build 26.616 (2026-06-18). Source: developers.openai.com/codex/changelog, 2026-06-18
- Requires Computer Use enabled (by the user or their administrator).
- Initial availability excludes the European Economic Area, the United Kingdom, and Switzerland; rolling out to select markets with broader availability to come. Source: developers.openai.com/codex/changelog, 2026-06-18; X/@OpenAIDevs thread, 2026-06-18
Same changelog also added bulk actions to automation run history and local↔remote thread handoff. Source: developers.openai.com/codex/changelog, 2026-06-18
Demo artifact
The launch video demonstrates the intended flow as a real browser/desktop task: Codex observes a YouTube upload workflow, captures the steps, asks for the relevant files, runs the workflow again, and leaves the generated skill inspectable. The useful lesson for the wiki is the boundary: Record & Replay is best for repeated GUI workflows with no stable API, while hand-authored SKILL.md files remain better for repo-local procedures that need source control, tests, or cross-harness portability. Source: X/@OpenAIDevs video artifact, 2026-06-18
Why it matters to Kevin
This is the No One-Off Work principle shipped as a product feature: do a repetitive task once, demonstrate it, and it becomes a codified, reusable, inspectable skill — the same "discovery → codify → reuse" lifecycle this wiki enforces, but for desktop/GUI workflows that have no API. It complements the agent-skills ecosystem (see Skills.sh - Agent Skills Registry, Matt Pocock's Skills (Skills for Real Engineers)): authored skills come from writing, Record & Replay skills come from demonstration. Because the output is editable rather than a brittle recording, it stays in the "skills as durable, reviewable artifacts" lineage rather than RPA-style screen scraping. It also leans on computer use — the model drives the desktop UI directly. For Kevin's stack: a candidate for turning recurring local chores (and demo flows) into Codex skills without hand-writing each one.
Where it fits
| Feature | Skill source | Surface |
|---|---|---|
| Codex Record & Replay | Demonstration (record a GUI workflow) | macOS desktop, Computer Use |
| Matt Pocock's Skills (Skills for Real Engineers) / Skills.sh - Agent Skills Registry | Hand-authored Markdown | Cross-agent, repo-based |
| Cursor Orchestrate Skill | Agent-spawned sub-skills | Cursor SDK |
Timeline
- 2026-06-21 | Captured from Kevin's X bookmark of @OpenAIDevs' Record & Replay launch (2026-06-18, ~9.6K bookmarks). Researched the OpenAI Codex changelog (build 26.616) for requirements (macOS, Computer Use, regional exclusions). Framed as the no-one-off-work lifecycle for GUI workflows. Source: X/@OpenAIDevs, 2026-06-18; Source: developers.openai.com/codex/changelog, 2026-06-18
- 2026-06-30 | One-by-one artifact review inspected the local launch thumbnail/video and reconfirmed the official Codex docs/changelog route: no local skill import is needed because this is a Codex feature that generates user-owned skills on demand. Source: X bookmark artifact audit, 2026-06-30