delegation

Assess actual progress toward an agent-work goal from bounded evidence, choose CONTINUE, RE-PLAN, or CUT, and publish a dated, evidence-linked result. Use for standing project bearings and architecture-drift checks; do not use for ordinary implementation review.

Continue authorized ongoing project work through pauses, wave closeout, and status turns by selecting finite useful ready work and preserving evidence. Use when work needs to progress without serializing independent lanes; not for unattended scheduling or new authority.

Use when something only the owner can decide comes up, when writing or reading the owner's decisions page, or when choosing between blocking on a question and recording it and moving on. NOT for peer asks between equal sessions — that is multi.

Use when a prompt decomposes into independent research, review, audit, or analysis lanes that should run as parallel subagents — 'investigate these three', 'review this from several angles', 'audit all X', 'sweep the codebase for Y' — or when deciding whether and how to delegate any task to subagents. NOT for multi-file feature builds: that is team-build. NOT for messaging, briefing, or handing off to an EQUAL session you do not own: that is the multi skill.

Start, prepare, find, or stop a local dev server (Next.js/Vite/any `npm run dev`) — ALWAYS through `server-manager-agent`, never `npm run dev &`/nohup. Use whenever you are about to run a dev server, need a worktree's node_modules (fresh git/Orca worktree = clone deps, don't `npm install`), need to know what's running on which port/worktree, or need to stop a server. This makes the server visible, killable, and switchable in Ben's `accounts` app, keeps it out of your own process group (so it survives you and can't take Claude down when killed), and enforces the port bands (3000 = cadma only). Triggers: "start the dev server", "run the app", "spin up next dev", "which port is X on", "kill the server", "restart the server", "install deps in this worktree", "npm install", "node_modules missing", "sh: next: command not found".

Use when a build, worktree, or run has finished and something needs cleaning up, before starting a new build in a project that hasn't been swept in a while, or when asked what's stale, what's using disk, or what can be removed. Reports two classes, SAFE and JUDGMENT, and only ever acts on SAFE. NOT a replacement for a builder's own cleanup: a builder that leaves a worktree behind still has to say so, by name, in its report.

Use when talking to an EQUAL agent session you do not own — asking a peer for a review, a fact or a hand-off; acking or answering a peer's ask; reporting a result or a blocker to a peer; or escalating a decision only Ben can make. Works between Claude and Codex sessions on any machine. NOT for subordinate work you own and will review — that is delegate (fan-out) or team-build (builds). NOT for Orca orchestration dispatch, worker tasks, mailboxes or task DAGs, even if the orchestration skill matched first.

Read and write Notion pages/databases (rewrite docs, edit blocks, update titles, query DBs). Use whenever the user shares an app.notion.com / notion.so link or asks to edit, rewrite, append to, or read a Notion page. Access is already configured — do NOT ask the user to set up an integration token or MCP.

Use when building a substantial multi-file feature with an agent team — builder/reviewer pipeline, tiered execution (mid tier writes, high tier verifies), spec-driven parallel implementation. This is the DEFAULT build method for any substantial multi-file feature unless the user instructs otherwise. NOT for research/review/audit fan-out without a build: that is delegate. NOT for messaging, briefing, or handing off to an EQUAL session you do not own: that is the multi skill.