vdx

A thin layer over existing tools: a unified lifecycle vocabulary + a versioned maturity rubric with drift detection + a structured manifest for the AI agent.

Working name. Subject to rename.

---

The problem

AI lifted the parallelism ceiling: instead of one primary stack (PHP), the same developer now juggles dozens of products at once across PHP, Node, Python, and Go. That creates two compounding pains:

1. Operational friction. Every project has its own way to "bring up / shut down / rebuild / run tests." The depth and vocabulary of commands varies wildly from repo to repo.
2. Quality drift. There is no objective measure of maturity. As the personal quality bar rises, older projects quietly fall behind, and there is no place to notice it.

Concept

vdx is built as a combine of ready-made tools. Custom code goes only into the things nobody else does — which is precisely why vdx exists.

The defensible core — four things

#	What	Why nothing off-the-shelf covers it
1	Normalized verb vocabulary up / down / build / test / check / fix	task runners (mise/Task/just) don't prescribe a vocabulary
2	Versioned maturity rubric + drift engine	Soundcheck/Cortex/OpsLevel mutate the standard live; they don't work locally or on third-party repos
3	Shared-infra orchestration (Global Traefik as a dependency of up)	no runner models cross-repo dependencies
4	A structured, executable manifest for the AI agent	AGENTS.md is deliberately prose; Claude Code skills (vdx-discover) react to project context

What vdx will NOT do: — ship its own task runner (mise already does this polyglot and well); — claim "discover-and-cache" as a novel loop (Claude Code auto-memory already writes findings between sessions natively; vdx's novelty is that the recorded path is structured and executable, not prose).

Install

Prerequisites:

• Node.js ≥ 20 (the CLI ships as ESM TypeScript via tsx, no build step required).
• mise — vdx delegates lifecycle execution to mise run <verb>.
• git — for vdx audit rubric resolution (github.com/.../vdx-rubric-vodmal@<tag>) and for vdx publish.

Then pick one — same binary, same behavior, different trade-offs:

# A. Daily-use install (recommended for any repeated `vdx <verb>` workflow)
npm i -g @vodmal/vdx-cli
vdx audit

# B. Zero-install, run from anywhere (recommended for CI, one-shot trial,
#    onboarding a teammate, or any "I just want to look")
#
# NOTE: `-p` is required because the binary `vdx` doesn't match the scoped
# package name `@vodmal/vdx-cli` — npx can't auto-infer the entry point.
npx -y -p @vodmal/vdx-cli vdx audit

Rule of thumb:

• Global when you run vdx many times per session — pays no per-invocation latency.
• npx when you don't want a global install or you're in a clean / ephemeral environment (CI runner, fresh VM, someone else's laptop). Each npx call adds ~200–500 ms for resolve+stat.

Either way you get two binaries: vdx (the CLI) and vdx-mcp (the MCP server consumed by the Claude Code plugin — itself launched via npx from the plugin manifest, so plugin users never need a global install).

Run vdx doctor first. It checks your environment (Node version, vdx on PATH, git, mise, npm auth, container runtime, Claude Code plugin) and shows you exactly what to install or configure before using anything else.

Quick start

# 1. Score any project against the owner baseline rubric
vdx audit

# 2. Generate a mise.toml that wires native scripts to vdx's 6 verbs
vdx init

# 3. Run lifecycle verbs (pass-through to `mise run <verb>`)
vdx build
vdx test
vdx check
vdx up    # bring services up
vdx down  # tear down

# 4. Publish a library (Node MVP; PHP/Python coming in Y.3)
vdx publish minor   # bump, npm publish (with OTP), git commit+tag

# 5. Check your local environment (Node version, mise, git, docker, Claude Code plugin)
vdx doctor

vdx init works deterministically on Node/PHP/Python/Go/Ruby projects. For unknown or meta (monorepo-with-subpackage) stacks, pass --stack <id> or — preferably — invoke vdx inside Claude Code with the plugin installed and let the vdx-discover skill bootstrap interactively.

Claude Code plugin

vdx ships a Claude Code plugin (MCP server + skill + hook) so the agent can call vdx_build, vdx_audit, etc. natively and bootstrap unknown projects.

Inline marketplace (this repo serves the plugin directly):

// ~/.claude/settings.json
{
  "extraKnownMarketplaces": [
    "github.com/VoDmAl/vdx/marketplace"
  ]
}

…or point at a local clone: "/abs/path/to/vdx/marketplace". Then /plugin install vdx@vdx inside Claude Code.

The plugin gives you:

• MCP tools — list_capabilities, vdx_up/down/build/test/check/fix, vdx_audit, vdx_record_success_path.
• vdx-discover skill — auto-triggers when a project lacks mise.toml (or audits at L0/L1 on lifecycle-interface); walks native scripts, verifies them, records the success path.
• PostToolUse hook — passive watermark of successful lifecycle runs into ~/.cache/vdx/last-success-path.log.

Architecture

Layer	Tool	Decision
Task runner	mise	D3
Capability manifest	mise.toml + vdx meta-block; projected into AGENTS.md ## Commands	D4
Fact gathering for audit	OpenSSF Scorecard · Qlty CLI · MegaLinter (planned)	D5
Scaffold-fix on own projects	copier (copier update, 3-way merge) — planned	D5
Claude Code packaging	plugin: MCP server + skill + hook	D6

Full rationale and open questions live in docs/decisions.md.

Scenarios

Familiar stack. vdx up → infra precheck → mise run up → native command. Instant.

Unfamiliar stack via Claude Code. Through MCP, the agent knows the project's 6 verbs and can invoke them without reading docker files. vdx audit scores it against your rubric even in unfamiliar tooling.

No manifest yet. vdx falls back to detection — make/composer/ npm — and calls them directly. Once the agent finds a working success path, a hook records it into mise.toml — the next run is instant.

Drift. The owner rubric moved forward → vdx audit shows the gap on an old project. On your own repos, copier update applies the template fixes; on third-party repos, it's a read-only report.

Two-repo layout

Repo	What	Release cadence
vdx (this)	dev hub: research, spec, evaluator (cli/), Claude Code plugin (plugin/)	changes often
vdx-rubric-vodmal	canonical owner-baseline rubric (vdx-rubric.yaml)	semver tags (v0.3.0+)

Manifest references in projects, e.g. baseline: github.com/VoDmAl/vdx-rubric-vodmal@v0.3.0, point at a specific semver tag of the canonical repo. Sync rules between the two repos are documented in CLAUDE.md ("External repo").

Current status: CLI v0.4.0 on npm, rubric v0.3.1, D12 (publish) MVP validated end-to-end

Spec D1–D12 ratified; core delivered in steps A–X.1.c within a single working session (see PROJECT_CHANGELOG.md for the full per-step log and HANDOFF.md for live state). Highlights:

• Step A — rubric v0.2 + draft spec (format, predicate DSL, drift algorithm).
• Step B — canonical repo vdx-rubric-vodmal published on GitHub.
• Step C — native evaluator (TypeScript, ~600 LOC under cli/src/).
• Step D — evaluator tuned via smoke tests on 3 reference projects.
• Step E — vdx init: stack autodetect → mapping native tasks to the 6 verbs → generation of mise.toml (no AGENTS.md — Claude Code users get the vdx-discover skill from the plugin instead).
• Step F — MCP server vdx-mcp on stdio with 9 tools.
• Step G — Claude Code plugin vdx/plugin/ (.claude-plugin/, .mcp.json, skill, hook).
• Step H — dogfooding: vdx now has its own root mise.toml, baseline audit.
• Step I — stack detector sees depth-1 sub-packages (monorepo / dev hub).
• Step J — applies_to filter in rubric v0.2.2: stack-irrelevant axes get drift_kind: excluded and don't count toward overall.
• Step K — [vdx].primary_subpackage + resolveSubpackageCtx (O31 closed).
• Step L — .github/workflows/ci.yml + npm test alias (ci L0→L3).
• Step M — matrix.node-version: [20, 22] (ci L3→L4, first axis at max).
• Step N — CLI published on npm as @vodmal/vdx-cli; the plugin is now marketplace-ready (npx -y -p @vodmal/vdx-cli@latest vdx-mcp).
• Step O — O33 closed (subpackage stack lift): for axes with applies_to, the evaluator now uses the subpackage's stack instead of the root's. For vdx this removes the excluded mask from 5 stack axes → an honest score.
• Steps P–U — vitest harness (tests axis L0→L3), new release-artifact axis with applies_when lib/app gate (rubric v0.3.1, O34/O35 closed), vdx init transparency + monorepo fallback (O39 closed), CLI bumped to v0.3.0 on npm.
• Step V — D12 (publish) accepted as 7th lifecycle verb after parallel landscape research; vdx publish lib-gated via the new applies_when predicate, stack-specific defaults institutionalized inside vdx (npm/composer/twine/cargo/gem) so users don't learn ecosystem-specific commands.
• Step X.1.a–X.1.c — D12 MVP for Node delivered and dogfooded: executePublish() runs the full pipeline (bump → npm publish with interactive OTP via stdio: 'inherit' → git commit + tag, no push) and was used to ship @vodmal/vdx-cli@0.4.0 end-to-end.

Research artifacts

Artifact	File	Status
Onboarding for a new session	HANDOFF.md	current
Research journal (Decided / Open / Observed)	docs/decisions.md	current
Landscape of comparable tools (does / doesn't)	docs/landscape.md	current
Maturity rubric v0.2 (calibrated)	docs/maturity-rubric.md	current
Build-vs-adopt decision	docs/decision.md	current
Change log	PROJECT_CHANGELOG.md	maintained

Internal research artifacts above are kept in Russian — they are the author's working notes. The outward-facing surface (this README, the CLI README, plugin README, and canonical-rubric repo) is English.

Core spec (v0.3.0)

Document	What it specifies
docs/specs/rubric-format.md	Format of the vdx-rubric.yaml owner baseline + predicate DSL + applies_to
docs/specs/manifest-format.md	The [vdx] block in mise.toml + AGENTS.md projection
docs/specs/overrides-format.md	.vdx-overrides.yml for per-project overrides
docs/specs/mcp-api.md	Contract of the vdx MCP server
docs/specs/drift-algorithm.md	Level computation + drift algorithm
docs/specs/vdx-rubric.example.yaml	Mirror of the canonical instance (for documentation)

Dogfooding

vdx audits itself against its own rubric. The declaration in mise.toml is stack = "meta" (documentation + a nested CLI in cli/).

Current position: overall L0. The real blocker is the ci axis (vdx lacked GitHub Actions until step L; the supporting axes still need real artifacts). Five stack-specific axes (tests, static-analysis, code-style, dependency-hygiene, mock-infra) correctly receive drift_kind: excluded under the meta stack and don't count toward overall. Lifecycle-interface is at L2 (3 of 6 verbs: build, test, check — up/down/fix are absent by design, vdx is not a service).

Next steps inside vdx: real tests (vitest), CI workflow, and sub-package-aware predicates (O31) so the nested cli/ is audited as a Node project. The full list of open questions lives in docs/decisions.md and HANDOFF.md.