Fused-Gaming

expand-ecosystem

Discover, evaluate, and install verified @h4shed skills, tools, and agents from the live npm ecosystem and the Fused-Gaming-Skill-MCP registry into this project, so the plugin's tool/skill/agent set can grow without hand-editing config. Use when the operator asks to add a capability, when the capability-scout records a gap, or when asked to sync/update the installed ecosystem.

Fused-Gaming 3 3 Updated 2mo ago
GitHub

Install

npx skillscat add fused-gaming/fused-gaming-skill-mcp/expand-ecosystem

Install via the SkillsCat registry.

SKILL.md

Expand Ecosystem

Grow this plugin's installed capability set from the live, verified @h4shed ecosystem — never from a guessed or frozen package list.

This skill is the executable counterpart to agents/capability-scout.md and follows references/capability-policy.md and the four-step foundation order in references/architecture.md. It does not replace the coordinator; it is invoked by the coordinator or the capability-scout when a gap is confirmed.

Step 1 — Discover, don't guess

Run the deterministic discovery tool. Never invent a package name.

node scripts/discover-ecosystem.mjs --out syncpulse-package-inventory.json

This queries https://registry.npmjs.org/-/v1/search?text=maintainer:h4shed&size=250 and cross-references Fused-Gaming/Fused-Gaming-Skill-MCP's registry/REGISTRY.md, then writes a normalized inventory with per-package classification (foundation/skill/tool/hub/cli/docs/session/other) and compatibilityDecision.

This step is read-only (P0 tier per capability-policy.md) and requires no gate.

Step 1b — Sync the registries

node scripts/sync-registries.mjs

This upserts Step 1's discovery results into registry/tools.registry.json (refreshing known entries, appending new ones as source: "discovered"), regenerates registry/agents.registry.json's localAgents from agents/*.md, records any newly discovered agent-like package under discoveredAgentPackages, and validates every config/routing-table.json rule against both registries — printing a warning for any agent or tool reference that doesn't resolve. Also P0/P1: it only ever writes the two registry files, never package.json, never an install.

Step 2 — Foundation check

Before installing anything else, confirm the four foundation packages are present and installed in this exact order, each with its own gate — never combined into one install command:

  1. @h4shed/mcp-core
  2. @h4shed/mcp-cli
  3. @h4shed/skill-syncpulse
  4. @h4shed/syncpulse-hub

inventory.foundationStatus from Step 1 tells you which are discoverable. If a foundation package is not yet an explicit direct dependency in package.json, install it before touching anything else, even if it is already present transitively.

Step 3 — Route the request

If the operator asked for a specific capability, map it through the routing matrix in the ecosystem bootstrap prompt (operator-request router / capability routing matrix — see references/capability-policy.md for the risk framing and the project's own SyncPulse Ecosystem Bootstrap prompt for the full table). Prefer:

  • one primary skill for domain instructions/workflow,
  • the smallest set of tool packages for execution,
  • the non-deprecated, newest-stable package when two overlap.

If the operator said "sync" or "update the ecosystem," treat every verifiedH4shed: true package not yet declared as a direct dependency as a candidate, grouped in this order:

  1. shared libraries required by already-selected packages
  2. skills matched to outstanding operator intent
  3. tool wrappers required by those skills
  4. docs/design-system/session/specialized packages
  5. everything else, only if the operator has explicitly requested a full sync

Step 4 — Risk-tier the candidate

For each candidate, classify per references/capability-policy.md:

Tier What Gate
P0 docs, schemas, inventory metadata auto-adopt
P1 read-only skill/tool sandbox, then auto-promote only if config/policy.example.yaml (or the project's copy) sets p1_auto_promote: true
P2 repository-write-capable tool coordinator approval
P3 credentials, deployment, billing, self-expanding privileges explicit human approval — always

Never install a package whose registry metadata does not list h4shed as a maintainer, regardless of how closely the name matches. compatibilityDecision: "review-unscoped-provenance" in the inventory means stop and ask, not proceed.

Step 5 — Install, register, smoke-test

Use the package manager already in use for this project (check for pnpm-lock.yaml / yarn.lock / bun.lock(b) / package-lock.json; default to npm only if none exists). Install compatible packages in small batches so a failure is attributable to one package. Then:

  • confirm the package resolves (require.resolve / import check, or the package's documented smoke test),
  • if it is a skill, confirm it is discoverable through @h4shed/mcp-core's registry interface,
  • if it is a CLI, run its --help (never an undocumented mutating command),
  • record the installed version.

Do not use --force or --legacy-peer-deps as a first response to a conflict — diagnose first.

Step 6 — Record and report

Update syncpulse-package-inventory.json (from Step 1) and summarize for the operator:

  • what was discovered vs. what was installed vs. what was skipped, with reasons
  • foundation status
  • any P2/P3 items still awaiting a human decision gate
  • how to re-run discovery later (node scripts/discover-ecosystem.mjs)

Never claim a package is "verified" on the strength of installation alone — verification requires the resolve/import/smoke-test step in Step 5.

What this skill does not do

  • It does not publish packages, deploy anything, or push commits.
  • It does not write secrets or invent credentials.
  • It does not install anything outside the @h4shed scope, or an unscoped package whose maintainer metadata doesn't confirm h4shed, without an explicit human decision gate.