RADIX WikiRADIX Wiki

Introduction

radix-skills is an installable agent skill that gives an AI coding agent a routed set of Radix reference material and instructs it to check what it reads against local copies of the Radix source. The repository publishes a single skill, radix, whose stated purpose is to let an agent "build or debug Radix code from local source checkouts rather than memory or web snippets". It is the work of xstelea, who also maintains Radix Context, Hookah and the Awesome Radix MCP Servers index, and who described himself in the Radix DLT Official Telegram on 6 August 2026 as previously "lead of the dApp engineering team", having "put all my accumulated radix knowledge into that skill".

A skill differs from a context pack in what it does at load time. Radix Context copies 23 documents into a project and leaves the agent to find them. A skill is a described capability the agent selects on its own: the skills.sh listing carries a description that tells the agent when the skill applies, and the skill body then routes to the one or two guides that fit the task instead of loading everything. The two projects are complementary rather than successive — radix-skills names radix-context in its own research strategy as an optional "candidate map for topics and vocabulary".

What It Contains

The skill ships 40 reference files totalling roughly 534 KB, of which 38 are topic guides named guide-*.md. Each guide states the source paths it is drawn from, a mental model, an example, and usage notes for one area of Radix work. The coverage is wide enough to be worth listing:

  • Transactions — manifest syntax and instructions, the V1/V2 lifecycle, Radix Engine Toolkit call sites, the Rust radix-transactions crate, subintents and pre-authorisations, receipts and events, and costing and fees.
  • Ledger primitives — resources and vaults, metadata, native pools, packages and components, access rules, role assignment, access controllers, account lockers, royalties, and SBOR.
  • Client and wallet — the Gateway API, transaction streaming with state-version checkpointing, the Radix Wallet connect flow, and ROLA challenge generation and verification.
  • Staking — validator behaviour, consensus-manager reads, liquid stake units, claim NFTs, unstaking and owner-badge operations.
  • Tooling — the rdx CLI, the Rust scrypto and resim binaries, rtmc/rtmd manifest conversion, Effect service patterns, TypeScript configuration, dependency and supply-chain hygiene, testing, and error diagnostics.
  • Payments — a guide to x402 payment middleware, facilitator validation and settlement, which the wiki covers separately at AI Agents & x402 Payments.

Alongside the guides the repository carries a 40 KB Python validation script, validate-radix-skill.py, that checks the guides against the source repositories they cite, and an openai.yaml agent descriptor.

Installation and Prerequisites

The skill installs from the skills.sh registry, an open-source skill index published by Vercel Labs, which detects the coding agent in use:

npx skills add xstelea/radix-skills --skill radix

Installation alone is not enough to use it as designed. The skill's prerequisite step requires two source checkouts at the root of the repository being worked on:

./.repos/radix-web3.js
./.repos/radixdlt-scrypto

A bundled script, prepare-radix-sources.sh, clones both, adds them to .gitignore, and installs the rdx-cli npm package globally when the rdx binary is missing. The script is written to fail rather than report success when git or npm is unavailable, when an existing checkout points at the wrong remote.origin.url, or when rdx-cli installs but rdx is not on PATH. The skill's own setup notes say a failed setup means "Radix-specific work is not source-backed yet".

One consequence is recorded in the repository's own architecture decision record, which chose local clones over submodules or vendoring: existing clones are never updated automatically, so a stale checkout will silently ground the agent in old source until someone refreshes it deliberately.

Design: Routing, Not Recall

The skill's research strategy is explicit about the order of authority, and it puts its own prose last. An agent is told to load the one matching guide, inspect similar code already in the user's repository, optionally consult radix-context as a vocabulary map, and then verify every context-derived fact against the source paths the guide names before using it in code. Public web documentation is reserved for questions local source cannot answer. Cloned repositories are read-only: the skill forbids importing application code from .repos.

That ordering is a direct response to the failure mode these packages exist to fix. A model asked to write Scrypto or a transaction manifest from memory tends to produce plausible code against an API version that no longer exists, because Radix's surface has changed substantially across Babylon and its protocol updates while the public corpus the model trained on did not. Pinning the answer to a checkout on disk makes the failure visible instead of silent — a missing or mismatched clone is an error the setup script raises, whereas a confidently wrong API call is not.

Provenance and Caveats

The 6 August 2026 announcement was a repost, not a release. The repository has been public since 21 June 2026 and its entire history is two commits made that day, with no code changes in the seven weeks between publication and the announcement that drew the community's attention. The skills.sh listing recorded six installs and the GitHub repository zero stars and one fork on the day it was announced — useful context for anyone reading the reaction as evidence of adoption.

Three points a prospective user should check for themselves:

  • No licence is declared. The repository carries no LICENSE file. By GitHub's own stated default, publishing a repository without a licence grants other users no rights to reproduce, distribute or create derivative works beyond viewing and forking it — whatever the one-line install command implies. This is unlike the author's Hookah, which is MIT-licensed.
  • One of the two source anchors is not official. radixdlt/radixdlt-scrypto is the canonical Scrypto repository, but radix-web3.js is the author's own library — itself unlicensed — rather than a radixdlt org project. Guidance about "the Radix codebase" therefore means the official engine plus one third-party TypeScript client, and the majority of the guides describe that client's package layout.
  • The registry's automated audits are not unanimous. skills.sh reports three scans of the skill: Gen Agent Trust Hub pass, Socket pass, and Snyk warn. A skill is instructions an agent will act on, and the setup script installs a global npm package and edits .gitignore, so the contents are worth reading before installing.