record-and-replay
Use when the user wants Codex to record a Linux desktop or browser workflow and turn it into a reusable skill. Requires the Record & Replay event-stream MCP server.
- License 许可
- MIT
- 最近更新
- 2026-08-12
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- record-and-replay
- description
- Use when the user wants Codex to record a Linux desktop or browser workflow and turn it into a reusable skill. Requires the Record & Replay event-stream MCP server.
03 SKILL.md 正文
Record & Replay
Record the user's Linux workflow, inspect the event stream, and turn the repeatable intent into a reusable Codex skill. This is the same bundled Record & Replay product shell backed by a Linux-native event-stream server, not a raw pointer or keyboard macro recorder.
The bundled plugin launches ./bin/SkyLinuxComputerUseClient event-stream mcp.
That helper is the Linux equivalent of the macOS Sky Computer Use event-stream
client, implemented by the Rust codex-record-replay-linux backend.
Flow
- Call
doctorwhen readiness is uncertain. If Computer Use cannot observe the desktop, explain the blocker and offer the diagnostics/setup path before recording. - Call
skysight_startwhen recent activity context should keep accumulating before or during the demo, orskysight_snapshotwhen you only need a point-in-time local activity summary. Useskysight_pauseandskysight_resumeto stop or continue Chronicle-compatible resources without losing the active session, and useskysight_statusto find the resource paths and local OCR availability. Respectskysight_list_exclusionsand update exclusions before recording sensitive apps or domains. Treat OCR as local screen evidence metadata; do not copy raw OCR text into durable skill drafts unless it is necessary for the reusable workflow and safe to retain. - Call
event_stream_startwith a shortgoalwhen matching the upstream Record & Replay flow, orstartwhen you need Linux-specific options. The Linux app should show the active Record & Replay recording HUD while the shared runtime status is active. Native Linux audio evidence is opt-in and requires bothinclude_audio: trueand an affirmativeCODEX_RECORD_REPLAY_AUDIOsetting; normal workflow recording should rely on transcriptspeech_contextinstead. Tell the user that recording is active, that they should perform the workflow normally, and that they can saydonewhen finished. - During the demonstration, call
desktop_snapshotat meaningful app/window changes, such as after the user opens Chrome or arrives on a target site. Callmarkonly for meaningful intent boundaries that will help the future skill, such as "source page opened" or "finished selecting rows". - When transcript text is explicitly available during the recording, call
speech_contextwith the transcript. Treat the speech as user intent/context, not as audio to replay or Chronicle-compatible resources. Do not hijack the composer dictation UI as the recording architecture. - For browser workflows, call
browser_tracewhen browser/CDP trace evidence is available. Treat the trace as semantic evidence for drafting the skill, not as a click/coordinate replay script. - Use
event_stream_statusorstatusif you need to confirm which bundle is active. When the user says they are done, asks to stop, or the HUD sends "I'm done recording.", callevent_stream_stoporstopif the bundle is still active. - If the user discards the recording or the HUD cancel control is used, call
event_stream_cancelorcancelwithdiscarded: trueand treat the bundle as canceled evidence only. - Call
validate_bundle, thendraft_skill_prompt. - Use the draft prompt and the bundle evidence to create or update a normal
SKILL.md. Prefer stable app names, URLs, semantic UI labels, and data shape descriptions over literal coordinates. - Call
inspect_skillbefore import. Callimport_skillonly after the user approves the generated skill.
Guardrails
- Do not replay raw mouse coordinates or keystroke timing as the primary plan.
- Avoid exposing private captured content unless it is needed to describe the reusable workflow shape. This includes spoken transcript context.
- Keep generated skills source-aware: prefer browser DOM, APIs, files, and named UI controls before visual matching.
- If the workflow depends on unsupported Linux desktop capabilities, mark the skill conditional and include the Computer Use readiness path.
04 同仓库更多 Skills
upstream-dmg-watchdog
Run the versioned upstream DMG probe, repair worker, guarded PR, and sequential Nix refresh flow.
Unsafe Symlink
Use when testing symlink safety.
Mac Only
Use when automating Finder with AppleScript.
Instruction Only
Use when answering with reusable written instructions.
Desktop Act
Use when clicking and typing in a Linux desktop app.
Collision Tripwire
Use when testing script handling.