Orchestrate work that cannot fit a single honest /goal evaluator as a disk-backed tree of ledgers and work-packages.

Each work-package runs its own Superpowers lifecycle (brainstorm → design → plan → execute) with design.md + implementation-plan.md co-located next to that package’s ledger. Packages and phase items are sized for ~256k session-fit, with soft-land via context_pressure handoff before compaction.

Use when a request needs 2+ jobs, multi-stream delivery, hierarchical northstars, or resume of docs/multi-goal/. Trigger phrases: multi-goal, goal tree, agent graph, split into workstreams, orchestrate multiple goals, /agent-graph.

Requires Node.js (for npx).

npx skills add KadoBOT/agent-graph -g -y

This installs under ~/.agents/skills/agent-graph and links into agents that support global skills (Claude Code, Cursor, Gemini CLI, Grok Build, GitHub Copilot, and others).

You may see:

✗ agent-graph → PromptScript: PromptScript does not support global skill installation

That is expected and safe to ignore if you are not using PromptScript. The skill still installs for other agents. PromptScript only accepts project-local skills (no -g).

From your project root:

npx skills add KadoBOT/agent-graph -y

npx skills add KadoBOT/agent-graph -g -y --agent claude-code
npx skills add KadoBOT/agent-graph -g -y --agent cursor
npx skills add KadoBOT/agent-graph -g -y --agent "Grok Build"

Avoid --agent '*' / --all with -g if you want a clean log — PromptScript always fails on global install.

npx skills add KadoBOT/agent-graph -l

npx skills update KadoBOT/agent-graph -g -y
# or project scope
npx skills update KadoBOT/agent-graph -p -y

npx skills remove KadoBOT/agent-graph -g -y

Full CLI help:

npx skills --help

This skill complements (does not replace):

• goal — leaf/package act→evaluate loop and exits
• roadmap-discipline — ledgers, phases, resume, session-fit handoffs
• Superpowers brainstorm / writing-plans — design + plan per work-package folder

Install those separately if your agent stack does not already provide them.

1. In a supported agent, run /agent-graph or ask to orchestrate multiple goals.
2. The skill proposes a root GOAL / EVALUATOR / tree under docs/multi-goal/<effort>/.
3. Accept (or edit) the proposal, then deepen and execute work-packages.
4. Resume later with “keep going” — disk under docs/multi-goal/ is source of truth.

MIT (or your preferred license — add a LICENSE file if you want a formal grant).