Agent-to-agent (a2a) delegation invokes another coding-agent harness as a subagent and continues the same delegated conversation across turns. Use when an agent needs to summon Codex, Claude Code, OpenCode, or Hermes Agent through a shell or terminal tool, retain the child session, send follow-up instructions, or coordinate ongoing agent-to-agent work.
Delegate implementation across parallel subagents in isolated worktrees. Use when the operator asks to fan out, parallelize, split up, or distribute implementation work. Apply mutual-agreement to every slice. The coordinator must not implement.
Fleet topology — machines, hosts, roles, SSH aliases, Tailscale hostnames, and services running on each. Load this whenever the user mentions a host by name (bingus, grimoire, yeowoolmac, oci-ubuntu), asks about the fleet, wants to run something on a remote machine, or when SSH/remote operations are needed.
Keep documentation fresh after a change by routing each edit to its one canonical surface. Trigger whenever you change shell/ssh config, an AI harness or MCP server, backup/kernel/monitoring setup, a `setup` module, a repo-template scaffold or skill, agent rules (AGENTS/FLEET/skills), or GPU/inference infra — i.e. any config, module, or workflow that has a documented home.
When overwriting existing content, write only the intended replacement — never leave traces of the correction process. Trigger when replacing, deleting, or rewriting content that already exists.
Delegate work to a subagent with mutual review gates. Parent owns the plan and the commit. Subagent owns the implementation. Neither proceeds without the other's agreement.
Read-only adversarial repository-boundary audit. Inventories the complete physical workspace — tracked, ignored, generated, cached, backed-up — challenges every item's right to exist where it lives, and produces a ranked audit, an operator decision sheet, and a gated migration sequence. Never deletes or moves anything. Trigger on "housekeeping", "boundary audit", "repo audit", "audit the workspace", "what can we delete", or similar.
Activates when the user is prototyping, building greenfield projects, or explicitly working pre-MVP. Trigger phrases: "we're just prototyping", "greenfield", "pre-MVP", "from scratch", "new project", "spike", "throw it away", "rewrite not refactor", "stop shimming", "just break it". Deactivation: "exit prototype mode", "stop prototype mode".