fab-plugins

A versioned Claude Code marketplace for fabian's custom skills — 78 skills packaged into 6 short-namespaced plugins, git-tracked and version-pinned at https://github.com/fabiantax/claude-plugins.

Install

# 1. add the marketplace (do this once per machine)
/plugin marketplace add https://github.com/fabiantax/claude-plugins

# 2. install the plugins you want (short namespaces — fab, mesh, quant, strix, ml, shared)
/plugin install fab@fab-plugins        # portable — daily-driver skills — install anywhere
/plugin install mesh@fab-plugins       # host-bound (A2A mesh + services) — opt-in
/plugin install quant@fab-plugins      # quant value stream
/plugin install strix@fab-plugins      # Strix box only (GPU/ROCm/llama.cpp) — opt-in
/plugin install ml@fab-plugins         # ML eval/bench/training
/plugin install shared@fab-plugins     # Claude-Flow community skills (github-*/swarm-*/skill-*) — opt-in

Plugin skills are namespaced — invoked as <plugin>:<skill>, e.g. fab:loopit, mesh:gitea, strix:llama-cpp-rocm, ml:code-bench. They cannot conflict with skills at other levels (per the Claude Code plugin spec).

Plugins

Plugin (namespace)	Portability	Skills	defaultEnabled
fab	portable	loopit (+/loopit), deliberate, prioritize, ship, plan-and-decompose, visual-plan, visual-recap, quick-recap, catch-up, handover, triz, creative-thinking, creative-thinking-ml, stay-within-limits, rust-decouple, preview, mermaid-local, loop-improvement, svelte-error-handling, svelte-performance, scrapling, emd-optimization	true
mesh	host-bound	fab-agent-runtime, fab-agent-add-mesh, fab-agent-add-peer, mesh-context, graphfusion, gitea, gitea-pm, gitea-bots, woodpecker, bifrost, tensorzero-gateway, openobserve	false
quant	portable-ish	fab-swarm-trading, indicator-creator, quant-consult, efficient-frontier, fable-efficient, cosmos-gl, casbin-ecosystem, pi-coding-agent, moshi-best-practices	true
strix	Strix box	llama-cpp-rocm, llama-cpp-vulkan, vllm, vllm-internals, model-runtime, rocm-profiling, rdna35-architecture, pytorch-rocm, hipblas-internals, triton-kernels, qwen36-architecture, container-ml-stack, toolbox-ml	false
ml	portable	model-guide, model-picker, model-quantization, code-bench, niah-bench, llm-eval-overview, thinking-eval, huggingface-workflow, gpu-bench-pipeline, mlflow-experiments, model-training, gepa, skillopt-rust-bugfix	true
shared	portable	github-code-review, github-multi-repo, github-project-management, github-release-management, github-workflow-automation, swarm-advanced, swarm-orchestration, skill-builder, skill-creator	false

defaultEnabled: false means the plugin installs disabled — enable the ones you need with /plugin after install, so a public install elsewhere doesn't pull in host-specific paths.

Versioning

• Each plugin carries an explicit version in its plugin.json.
• marketplace.json entries pin to a matching version.
• Reproducible installs are anchored to a git tag (v0.2.0); ref pins are added to the marketplace entries when a stable tag is cut.

v0.2.0 — plugin rename (breaking identity change)

Plugins were renamed for a cleaner namespace prefix (the part you type):

v0.1.0	v0.2.0
fab-coding-loop	fab
fab-mesh	mesh
fab-quant	quant
strix-machine	strix
fab-ml-lab	ml

The marketplace name (fab-plugins) and skill contents are unchanged. To upgrade: uninstall the old names, then install the new — claude plugin uninstall fab-coding-loop@fab-plugins etc., then claude plugin install fab@fab-plugins etc.

Repository layout

.claude-plugin/marketplace.json        # the marketplace (lists all plugins)
plugins/
  <plugin>/                            # fab, mesh, quant, strix, ml, shared
    .claude-plugin/plugin.json         # plugin manifest (strict mode)
    skills/<name>/SKILL.md             # + bundled assets (scripts/, references/)
    commands/loopit.md                 # (fab only)

shared — repo-sourced community skills

Unlike the other 5 plugins (sourced read-only from ~/.claude/skills/), shared holds cross-repo-duplicated Claude-Flow community skills (github-*, swarm-*, skill-builder/skill-creator) that don't live in ~/.claude/skills. They are sourced from a repo checkout via REPO_SOURCES in build.py — currently neural-trader/.claude/skills/, which holds the newest ("9/9 certified") generation. This dedupes the drift where the same community skills were vendored into GraphFusion (older), neural-trader (newest), Atlas, and the Alies repos. Consumer repos install shared@fab-plugins at project scope and delete their vendored copies — one versioned source of truth. See the plan in docs/commit history for the drift table.

Coexistence with unversioned ~/.claude/skills/

The unversioned copies in ~/.claude/skills/ and ~/.claude/commands/loopit.md are intentionally kept for now. Because plugin skills are namespaced (fab:loopit) they do not shadow bare /loopit — the two coexist:

• bare /loopit, /prioritize, /deliberate, /ship → the unversioned copies
• fab:loopit, fab:deliberate, … → the versioned plugin copies

Retiring the unversioned copies later would change bare invocation to the namespaced form everywhere, so it's deferred until explicitly wanted. Don't retire without re-confirming.

The conversion + frontmatter normalization is reproducible via build.py (reads ~/.claude/skills/ read-only, writes into plugins/).

Rebuilding

python3 build.py   # re-copies + normalizes into plugins/
                    # ~/.claude/skills/ for fab/mesh/quant/strix/ml,
                    # neural-trader/.claude/skills/ for shared (via REPO_SOURCES)

Out of scope

Repo-level skills (company-loop, per-repo skills) stay versioned via their own repos' git.

speckit-* — removed (superseded by the upstream specify CLI)

A speckit plugin was added in v0.3.0 to dedupe the speckit-* skills vendored across the mesh repos. It has been removed — those skills are now managed by the upstream specify CLI, which already provides version pinning, per-file manifest hashes, and diff-aware upgrades. The plugin re-invented a mechanism upstream ships. The repos manage their speckit-* skills directly via .specify/ (GraphFusion, fab-trader, localscout, atlas) or are migrated to it (fab-agent-mesh → specify init).