let-my-agent-sleep

name	let-my-agent-sleep
description	Start long-running training, evaluation, preprocessing, or batch jobs with Let My Agent Sleep handoff semantics in OpenCode. Use when a task would otherwise require waiting, polling, or repeated continue loops until a job finishes.
compatibility	opencode

Let My Agent Sleep

HARD RULE: DO NOT POLL AFTER HANDOFF

AFTER LMAS_HANDOFF v1, THE JOB IS HANDED OFF.

UNTIL LMAS_COMPLETION_EVENT v1 ARRIVES, OR THE USER EXPLICITLY ASKS FOR STATUS:

DO NOT POLL. DO NOT CALL lmas_status. DO NOT TAIL LOGS. DO NOT READ STDOUT OR STDERR. DO NOT INSPECT ARTIFACTS. DO NOT CONTINUE THE LOOP JUST BECAUSE A TODO IS STILL OPEN.

Stop the current turn after handoff.

Treat long-running jobs as handoff tasks, not completion tasks.

Use this skill when a command is likely to run for minutes or hours, including model training, evaluation, dataset preprocessing, benchmark suites, large migrations, or batch inference.

Workflow

1. Prepare the command and run a short sanity check when practical.
2. Start the real long job with the lmas_start tool.
3. After the tool returns LMAS_HANDOFF v1, stop the current loop. Do not poll the log or wait for completion.
4. When LMAS_COMPLETION_EVENT v1 appears later in the session, use lmas_status if you need a quick run summary, then inspect stdout/stderr/artifacts and continue the original task.

Required Behavior

• Do not run long jobs directly with the normal bash tool when lmas_start is available.
• If you maintain a TODO list, scope the long-job TODO to "start with LMAS and record handoff". Mark it complete immediately after LMAS_HANDOFF v1; do not leave a TODO that means "wait for job completion".
• Do not call skill_mcp for Let My Agent Sleep. This is an OpenCode plugin tool, not an MCP server.
• If lmas_start is not available in the tool list, the OpenCode plugin is not loaded in the current session. Tell the user to restart OpenCode after running lmas install --agent opencode; do not invent .opencode/tools paths or start a normal shell fallback unless the user explicitly asks for manual fallback.
• Do not treat an unfinished long job as a failure after LMAS_HANDOFF v1.
• After LMAS_HANDOFF v1, do not check status, inspect artifacts, relaunch, or debug the job unless the user explicitly asks for a status check. A user asking whether enough time has passed, why there is no update, or whether the job finished counts as a status check request.
• If lmas_status reports RUNNING or FINALIZING, STOP IMMEDIATELY. DO NOT POLL. DO NOT TAIL LOGS. DO NOT INSPECT ARTIFACTS. DO NOT CALL lmas_status AGAIN until LMAS_COMPLETION_EVENT v1 arrives or the user explicitly asks for another status check.
• If the user explicitly asks to cancel, stop, or terminate an LMAS run, call lmas_cancel. Do not kill tmux sessions or job processes directly.
• If a user later asks for status and LMAS_STATUS v1 reports LOST, inspect watcher.log and stderr.log, report the run as lost, and ask before relaunching. Do not silently start a replacement job.
• Preserve the original command in metadata.
• Make the completion response concrete: cite the run id, status, exit code, and relevant log/artifact paths.
• After completion, prefer lmas_status over manually reading metadata files. Read stdout/stderr first; read metadata only when command context is unclear.

Path Rules

• Prefer workspace-local paths under .lmas/ for all LMAS run data and artifacts.
• Use .lmas/runs as the default run directory.
• Use .lmas/artifacts/<task-name-or-timestamp> as the default artifact directory for task outputs and reproducible helper scripts.
• Use /tmp only when the current workspace is not writable or the user explicitly requests a temporary system path.

OpenCode Setup

lmas_start, lmas_status, lmas_cancel, and diagnostic lmas_info are provided by the installed OpenCode plugin. They are not project-local files and should not be looked up under .opencode/tools.

The installer updates the OpenCode config, usually ~/.config/opencode/opencode.jsonc, and installs this skill under ~/.config/opencode/skills/let-my-agent-sleep/SKILL.md. OpenCode must be restarted after install so it reloads plugin tools.

OpenCode must know the server URL used for completion injection. The plugin usually receives it from the running OpenCode server. When needed, set:

export LMAS_OPENCODE_SERVER_URL=http://127.0.0.1:4096

The tool supplies LMAS_OPENCODE_SESSION_ID from the current OpenCode tool context.

The OpenCode plugin supplies its current server URL to lmas_start; do not accept a model-provided server or notification URL. For a direct CLI fallback on a non-default server, the user or host environment must set LMAS_OPENCODE_SERVER_URL to the server that owns the session id. Otherwise completion injection fails and resume_prompt.txt remains the fallback.