night-watchman

How to decide whether repeated work should become a script, an agent, or a skill. Use when the main thread catches itself doing something inline for the second time, when asked "should this be a script/agent/skill", or when a task keeps recurring without a fixed home for it.

Wiring guidance for dispatching a failure signal (an alert, a failed check) to the diagnose-and-pr agent and handling its output. Use when a project wants to turn monitoring findings into draft PRs instead of manual triage, when deciding what counts as in-scope for that agent, or when the owner asks to diagnose/debug a specific failing behaviour and wants the disciplined loop (a red command before any theory), not a one-line question.

Stress-test a plan, design, or decision by interviewing the owner in rounds until nothing is left assumed, then hand the settled decisions to to-issues. Use when the user says "grill me", "grill this plan", asks to pressure-test an idea before work starts, or a to-issues decision ticket needs an owner conversation.

Writing a session wrap-up note for whoever picks this up next — a fresh session, or a different person. Use at session end, when the user says "write up where we left off", or before closing a session that did real work.

The opening move for a fresh session in a project running this plugin's unattended-operations model — orient from docs/, verify every awaiting-deployment ticket against reality, and fan startable work out to pinned-model subagents in parallel. Load when the user says "start on the next open tickets", "pick up where we left off", "what can be worked on", or opens a session with no specific task.

Shared-lib conventions and the shipped-bug rules for scripts/ — bash 3.2 limits, subshell die, pipe_ok, secrets in argv, and structural network isolation in selftests. Load before writing or reviewing anything in scripts/.

The docs/ routing table and the ticket workflow — status-is-a-directory-move (or a tracker transition, if one is adopted), the frontmatter/field contract, body-vs-progress-note rules, and the UNVERIFIED discipline. Load for session wrap-up, ticket filing, ticket transitions, or any durable-knowledge capture.

Turn a finished planning, grilling, or design conversation into independently-workable tickets as frontmatter Markdown files on disk, so the session can be closed and the work picked up cold by parallel agents in git worktrees. Use this whenever a grilling session ends, whenever the user says "turn this into tickets/issues/a work log", "write this up so I can start a fresh session", "break this into work I can parallelise", or after any long discussion that settled a bunch of decisions and now needs to survive the context window. Also use it when a plan already exists but the tickets are too coupled or too vague for an agent to execute without asking questions.

Mine a settled planning, grilling, or design conversation into a decision-list JSON conforming to work-order's decision-list/FORMAT.md — the mining half of to-issues, split so that ticket authoring lives in work-order's emit-tickets skill instead of here. Use whenever a grilling session ends and the decisions need to leave the conversation as data, before anything decides how they become ticket files.

Keep an append-only decision trail for an unattended or dispatch-provider wave (one row per fork, pivot, revert, or verified unit), audit it against the transcript at wrap-up, and have a second-tier reviewer write an Attention section for the human. Use when a wave is dispatched through a dispatch provider, or any run the owner is not watching live.