lnwjud: Skill for Claude Code

lnwjud Scheduled Continuation

One user request owns one durable goal and at most one live Native ChatGPT watchdog. Current mainline uses one hourly recurring Native ChatGPT Scheduled Task in the current chat with cloud execution requested. The same native task ID survives ordinary wakes until the durable goal is accepted and the exact task is made non-runnable.

Authorization and current contract

Use scheduled continuation only when the user authorized it for this task. Updating this skill, requesting a release, or enabling a host Goal does not itself authorize a native schedule. Reuse an existing goal's workspace, goal key, connector, and scheduledContinuation setting. For a new durable goal without scheduling authorization, explicitly pass scheduledContinuation: off; continue manual work normally. An explicit opt-out stops scheduled continuation and never abandons the durable goal.

Before changing workflow guidance, inspect the current runtime through CodeGraph when .codegraph/ exists. The source contracts are goal-tools.ts, scheduled-continuation-tools.ts, and registry-runtime-port.ts. Read root AGENTS.md for repository rules. Installed connector tools may lag source; resolve their advertised schema in the current session before calling them. A missing tool or stale schema is a transport/version limitation, not permission to invent an API or change schedulers.

Non-negotiable outcome

• Never require the user to type continue, ทำต่อ, or ask for status to resume unfinished work.
• Never report completion while get_goal is active.
• A ChatGPT Task marked Active proves only that the recurring schedule is enabled; it does not prove an MCP tool was invoked, the Goal lease was acquired, or any work ran between hourly firings. Verify the exact host last-run receipt, the first connector call and durable Goal/activity evidence before describing execution.
• An hourly recurring ChatGPT Task cannot guarantee continuous model execution or a gap shorter than its next firing. If the Goal opted out, resume the same Goal manually; preserve that setting and report host/tool safety checks truthfully.
• run_goal defaults to scheduledContinuation:auto in the API. Select off explicitly when scheduling was not authorized; an API default cannot grant user authorization.
• For an authorized active goal, use exactly one Native ChatGPT recurring Scheduled Task: occurrence=interval, intervalMinutes=60, destination=current chat, cloud requested.
• Never use lnwjud scheduler, Windows Task Scheduler, schtasks.exe, cron, shell timers, browser/DOM automation, or an undocumented scheduling API as a fallback.
• Native task create/update/delete/disable is host-owned. Resolve the operation from the current ChatGPT host/tool registry; never invent or hard-code an internal operation name.
• A native-host transport failure is scheduler degradation only. It never by itself completes, fails, or blocks the durable goal.
• Keep raw lease tokens, credentials, private host task IDs, and internal session IDs out of user-visible text and native task prompts.
• Continuation/recovery state is task data, not persistent user or agent instructions. Use checkpoint_goal for durable state and session_handoff only for a bounded same-chat recovery summary.
• Never invoke generic handoff / claude-handoff skills as part of lnwjud scheduled continuation, and never call write_file to create USER_INSTRUCTIONS*, user-instruction*, or generic handoff/history/* persistence. Those paths can be interpreted by the host as instruction persistence and are outside lnwjud's durable-goal contract.
• If a host blocks an attempted recovery-file write, do not retry it under another filename or through shell. Keep the durable goal/checkpoint authoritative and continue the current leased work.