craft

Personal Claude Code collection of opinionated engineering skills and agents.

craft owns its Rust and Nix knowledge in full and is self-contained — it declares no plugin dependencies and carries no generic engineering-process content. Language-agnostic discipline (how to run a review, how to verify a claim, the debugging method) is intentionally out of scope; bring your own process skills. craft's skills cross-reference each other — that cross-linking is the collection's cohesion.

craft/
├── .claude-plugin/
│   ├── plugin.json        # plugin metadata
│   └── marketplace.json   # single-plugin marketplace (source: ./)
├── agents/                # diff/structure review agents
│   ├── rust-reviewer.md
│   ├── rust-security-scanner.md
│   ├── rust-miri.md
│   └── rust-architecture-reviewer.md
├── workflows/             # multi-agent orchestration scripts
│   ├── rust-review.js
│   ├── rust-audit.js
│   └── triage-findings.js
├── lib/                   # tested helpers shared (inlined) by the workflows
│   └── run-record.mjs
└── skills/                # (abbreviated)
    ├── rust-review/
    ├── rust-testing/
    └── …

/plugin marketplace add gurinderu/craft
/plugin install craft@craft

craft also runs in opencode via a contained adapter layer under opencode/: the skills are shared verbatim (opencode reads the Anthropic skill spec), the 4 review agents are translated to opencode agent files, and the rust-audit / triage-findings workflows ship as one contained TS/Bun plugin (/rust-audit, /triage-findings). Install with opencode/install.sh (project-scoped by default). See opencode/README.md for details, the model strategy, and parity caveats.

workflows/review.js is the default diff-review path and the generic engine: it auto-detects the language(s) in the diff via an inline PROFILES registry (Rust and Nix built in), and for each active language a scout sizes the diff and picks lenses + rigor, lenses fan out in parallel and loop until dry, every finding is adversarially and self-verified, and a synthesis step merges everything into one Confirmed/Suspected report with a verdict — depth scales to the diff automatically. workflows/rust-review.js and workflows/nix-review.js are thin pins that force a single language.

workflows/rust-audit.js is the full pre-release audit: a scout detects the diff base, unsafe, the workspace crates and their dependency edges, then many dimensions run in parallel — per-crate review, inter-crate contracts, architecture, crate-decomposition, security, Miri (if unsafe), semver, build-matrix, deps, unused-crate detection (with verification), and test/doc health — and a final step synthesizes one severity-ranked report.

workflows/triage-findings.js validates findings (craft agents + GitHub PR comments), dedups and conflict-checks them, and renders one ordered fix plan + a triage ledger — no edits.

Both review workflows and the standalone agents emit a structured run record after each run; see docs/observability.md.

Run a workflow against a path or fixed diff base:

Workflow({ scriptPath: "workflows/rust-audit.js", args: { base: "origin/main" } })

args.base is optional — without it the scout falls back to merge-base HEAD origin/main → main → HEAD~1.

• Skills use progressive disclosure: a lean SKILL.md entry plus sub-files loaded on demand.
• Crate versions in examples track the latest stable release; edition = "2024".
• Skills reference each other (e.g. rust-review points at rust-testing for how to write missing tests). Generic engineering-process discipline is out of scope — craft names no external process skills and declares no plugin dependencies.