Skip to main content

Mobile and remote workflow

The development Linux server is the stable execution environment while the owner can move between Mac, phone, and other clients.

If you forget everything

  1. Open this handbook.
  2. For normal development, tell ChatGPT “계속”.
  3. ChatGPT reloads the active Work Packet, Git/GitHub state, and authorized development-server context, then implements/tests directly.
  4. Only when you explicitly want Cursor for the current packet, say “Cursor에 넘겨줘”. ChatGPT must record IMPLEMENTER=CURSOR and revalidate the dispatch boundary before /work-resume.
  5. After an explicitly opted-in Cursor task, “Cursor 끝났어. 계속” returns control to ChatGPT for independent exact-HEAD verification.

Default remote flow

Browser login, account, credential, payment, or approval steps that cannot be completed remotely still require owner interaction.

Optional Cursor persistent sessions

Only after explicit owner opt-in and IMPLEMENTER=CURSOR, the verified target worktree may use:
ChatGPT must verify the target repository/worktree and current Work Packet before launch, then confirm the session appears in agent persist list. Without explicit packet opt-in, no Cursor session is launched. Launching a persistent session requires an interactive Linux/macOS TTY. A non-interactive shell can fail with:
The owner normally only needs the management commands:
stop terminates the running session and should only be used intentionally. ChatGPT should report the persistent session name and a ready-to-copy attach command after starting it. It must not silently replace the handoff with a non-persistent --print process unless the owner explicitly allows that fallback.

Ask ChatGPT for implementation progress

You do not have to SSH into the development server just to learn how implementation or a long test is progressing. When the development server is authorized for remote inspection, ask ChatGPT directly:
ChatGPT should inspect the actual development-server state instead of asking the owner to copy terminal output. Depending on the workstream, that inspection should include:
  • the target repository/worktree and any explicitly opted-in persistent Cursor session (agent persist list when applicable)
  • the running Cursor process/session and its current child command or test
  • branch, exact HEAD, dirty/clean state, and synchronization with the remote
  • current test position, remaining tests, and known long-running tests
  • the active [AI Work] Work Packet, PR, CI, and linked issue state when applicable
A useful status report is concise and evidence-based:

Completion-time estimates

When the owner asks when a task will finish, ChatGPT may estimate the remaining time from observable evidence such as:
  • suite start time and elapsed time
  • current test index versus scheduled test count
  • remaining long-running tests and their timeout ceilings
  • prior timings for the same exact or comparable test path, when available
The result should be a range such as “about 10–20 minutes remaining”, not a false exact promise. If the runtime cannot be inspected or there is not enough timing evidence, ChatGPT should say that the ETA cannot be determined reliably. For an explicitly opted-in Cursor packet, agent persist attach <session> remains available for raw interactive observation. In the default ChatGPT-primary path, simply ask ChatGPT for status, blockers, test progress, or an ETA.

Telegram

Telegram is only a completion-notification channel. Work state and verification records stay in GitHub and the repository. Send COMPLETE only when the workstream is actually complete. Use the repository, tests, CI, PR/Issue state, and actual runtime results to determine the current state.
Last modified on September 29, 2026