claude-native-inventory

Inventory and route Claude Code, Cowork, and Claude Platform capabilities, including Messages API runtime, server/client tools, Files/PDF, context/cache/compaction, skills, commands, agents, hooks, MCP, SDK, plugins, and output. Cowork claims still require a Cowork-session probe.

SKILL.md

Quality

Evals

Security

Claude-Native Capability Inventory

Canonical inventory of Claude Code / Claude Cowork tools and capabilities and the skills that govern each one. This is the Claude-native successor to the Codex base claude-code-official-inventory: same taxonomy and source map, but with a runtime-status column that must be populated from the current session instead of doc-only assumptions. The checked-in baseline stays conservative. Treat the inventory as an implementation contract before creating, porting, or operating any capability skill.

Operating Rule

Sources of truth are surface-specific:

Live Claude Code/Cowork runtime — what the current session actually exposes (tools, agent types, installed skills, MCP servers, hooks firing). Probe before claiming a capability is present. Tag findings [CÓDIGO] / [CONFIG].
Claude Platform audit v2 — the 35 official pages, claims, hashes and coverage under ../../docs/audits/claude-platform-2026-07-14-v2/. Platform capabilities are reference until an application runtime is separately probed. Tag [DOC][CÓDIGO].
Claude Code official docs — https://code.claude.com/docs/llms.txt as the discovery index for Code field behavior and version gates. Tag [DOC].

When runtime and docs disagree, record both; do not erase the difference.

Runtime Probe

Before asserting availability, gather deterministic package/CLI evidence (read-only):

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/runtime_probe.py" --json

The probe proves package counts, CLI version when executable, and observable marketplace directories. For tools, agent types, MCP servers, and hooks, inspect the active session surface (tool list, Agent subagent types, mcp__* tools) rather than guessing. If a surface cannot be probed in the current runtime (e.g. Claude Cowork from a Code session), mark it no-verificado [coverage_gap], never disponible.

Runtime-Status Vocabulary

Status	Meaning
`disponible`	Confirmed present in the probed live runtime.
`experimental`	Present but research-preview / feature-gated / provider-limited.
`reference`	Real capability, but a CLI/config feature with no in-session tool to probe.
`no-verificado`	Could not be confirmed from the current runtime; `[coverage_gap]`.

Capability Matrix

The checked-in matrix is a package/reference baseline, not a live-session snapshot. Rows that need runtime observation remain no-verificado; an invocation may overlay current evidence but must not write disponible back without a dated trace.

Family	Capability / surface	Governing skill	Runtime-status	Evidence
Tools	Built-in tools (Read, Edit, Bash, Agent, Skill, ToolSearch…)	`tool-permission-policy`, `tool-use-design`	no-verificado	active session surface required `[coverage_gap]`
Claude Platform	Tool definitions (`name`, `description`, `input_schema`, `strict`, `tool_choice`, `allowed_callers`, `defer_loading`)	`tool-use-design`	reference	v2 official Platform audit + focused evals `[DOC][CÓDIGO]`
Claude Platform	Messages API tool execution loop (`tool_use`/`tool_result`, ordering, stop handling, streaming, retries)	`claude-api-tool-runtime`	reference	v2 official Platform audit `[DOC][CÓDIGO]`; application runtime varies `[coverage_gap]`
Claude Platform	Anthropic-executed server tools, web search/fetch, code execution, advisor, tool search, programmatic calling	`claude-api-server-tools`	reference	7 source pages + offline validator/evals `[DOC][CÓDIGO]`
Claude Platform	Application-owned memory, bash, text-editor, and computer-use handlers	`claude-api-client-tools`	reference	4 source pages + offline validator/evals `[DOC][CÓDIGO]`
Claude Platform	Files API lifecycle and PDF/document inputs	`claude-api-files-documents`	reference	2 source pages + offline validator/evals `[DOC][CÓDIGO]`
Claude Platform	Token counting, caching/diagnostics, context editing, compaction, windows, mid-conversation controls	`claude-api-context-management`	reference	10 source pages + offline validator/evals `[DOC][CÓDIGO]`
Claude Code	Custom skill/command extension patterns	`custom-tooling-extension`	reference	package + Claude Code docs `[CÓDIGO][DOC]`; activation requires probe
Agents	Custom subagents (`agents/*.md`, Explore/Plan/general-purpose)	`agent-creator`	no-verificado	active `Agent` surface required `[coverage_gap]`
Agents	Agent view / background sessions	`agent-view-ops`	reference	documented CLI feature `[DOC]`; current availability unprobed
Agents	Agent teams (teammate messaging)	`agent-team-orchestration`	no-verificado	active teammate surface required `[coverage_gap]`
Agents	Dynamic JS workflows	`dynamic-workflow-forge`	reference	Agent SDK contract + offline fixtures `[DOC][CÓDIGO]`
Isolation	Git worktrees	`worktree-isolation`	no-verificado	active worktree primitives required `[coverage_gap]`
Automation	Hooks (lifecycle events)	`hook-engineering`	no-verificado	config validates offline; firing requires live trace `[coverage_gap]`
Automation	Channels (external event push)	`channels-integration`	no-verificado	active channel surface required `[coverage_gap]`
Automation	Scheduled tasks (Cron primitives)	`scheduled-task-ops`	no-verificado	docs exist; active scheduler surface required `[coverage_gap]`
Automation	Goal mode (`/goal`)	`goal-mode-ops`	reference	slash command, no in-session tool `[DOC]`
Integrations	MCP servers / tools	`mcp-engineering`	no-verificado	active `mcp__*` surface required `[coverage_gap]`
Integrations	Headless / Agent SDK	`headless-sdk-automation`	reference	package/docs contract; current runtime unprobed `[DOC]`
Integrations	Deep links (`claude-cli://`)	`deep-link-runbook`	reference	URL scheme, no in-session tool `[DOC]`
Output	Artifacts / widgets	`artifact-publishing`	no-verificado	active artifact/widget surface required `[coverage_gap]`
Output	Output styles	`output-style-authoring`	reference	config feature `[DOC]`
Output	Terse / concise mode (preserva evidencia)	`terse-output`	reference	packaged guidance `[CÓDIGO]`; invocation unprobed
Packaging	Plugins & marketplaces	`plugin-builder`	no-verificado	`runtime_probe.py` reports observable directories; active load requires smoke `[coverage_gap]`
Tools (kata)	Built-in tool selection	`katas-builtin-tool-selection`	reference	`[DOC]`
Tools (kata)	MCP server configuration	`katas-mcp-server-configuration`	reference	`[DOC]`
Tools (kata)	Tool description quality	`katas-tool-description-quality`	reference	`[DOC]`
Tools (kata)	PreToolUse guardrails	`katas-pretooluse-guardrails`	reference	`[DOC]`
Tools (kata)	PostToolUse normalization	`katas-posttooluse-normalization`	reference	`[DOC]`

Authoring & Routing (Tranche 2)

Capability	Governing skill	Runtime-status
Authoring SKILL.md	`claude-skill-authoring`	reference
Choosing one operating primitive from a structured contract	`prompt-contract-operating-primitives`	reference
Authoring slash commands (merged into skills)	`claude-command-authoring`	reference
Choosing a parallel primitive	`agentic-parallelism-decision`	reference
Owning a Messages API client-side tool loop	`claude-api-tool-runtime`	reference
Owning Claude Platform server-tool plans	`claude-api-server-tools`	reference
Owning Claude Platform client-tool handlers	`claude-api-client-tools`	reference
Owning Files API/PDF contracts	`claude-api-files-documents`	reference
Owning context/cache/compaction policy	`claude-api-context-management`	reference
Choosing a reasoning pattern (Structured/Skeleton/Chain-of-Code/ToT/Committee)	`metacognition`	reference
Plugin lifecycle hub (build→validate→deploy→maintain→fix) + deploy orientation	`plugin-ops`	reference
Hardening a skill to the DoD (excellence gate)	`toolkit-hardener`	reference
Crear un arnes agentico (5 subsistemas + goal contract, stack-verified)	`agent-harness-creator`	reference
Auditar/puntuar un arnes existente (28 checks, anti-gaming)	`agent-harness-audit`	reference
Diagnosticar drift docs-vs-realidad de un arnes	`agent-harness-diagnose`	reference
Reparar un arnes con el patch emitido por el validador	`agent-harness-repair`	reference
Extender un arnes (features, packages monorepo con boundaries)	`agent-harness-extend`	reference
Evolucionar un arnes (reglas, playbooks, retiro de docs cumplidos)	`agent-harness-evolve`	reference
Refactorizar un arnes (deletion-only, insertions=0)	`agent-harness-refactor`	reference
Emitir y validar un bloque hooks de settings.json (scaffold_hook / validate_hook)	`hook-scaffold`	reference
Convertir una regla en prosa en una hookify rule validada	`hook-rules`	reference
Spec-first: scaffold y validacion de skill.yaml antes de crear el skill	`skill-spec`	reference
Gate de cargabilidad de un plugin antes de instalar / tras deploy	`plugin-reload`	reference
Auditoria de consistencia de un paquete multi-documento (crossrefs, orphans, naming)	`consistency-audit`	reference
Experimentos acotados fail-closed (ExperimentSpec v2, limite de ciclos)	`bounded-experiment`	reference
Simplificar estructura (downgrade graph→secuencia→prompt)	`agent-harness-simplify`	reference
Probar que los frenos del arnes muerden (prueba de vida, tri-estado)	`agent-harness-selftest`	reference
Compilar un objetivo NL en DAG validado (olas, write_sets, guardian)	`graph-spec-compiler`	reference
Disciplina de ejecucion de graphs (aprobacion, receipts, fallos tipados)	`graph-run-discipline`	reference
Disenar controles falsificables (METHOD, falsificacion, satisfacibilidad)	`graph-controls-designer`	reference

Plus a 30-kata pack (katas-*) governing tool selection, MCP errors, agent loops, context dilution, isolation, prefix caching, session resume, and review/validation patterns — cross-linked from the related capability skills above.

Integrations / MCP (Tranche 6 — arneses)

Capability	Governing skill	MCP/source	Runtime-status
Research & context expansion (NotebookLM)	`notebooklm-research`	`notebooklm` (`.mcp.json`)	no-verificado `[coverage_gap]`
UI prototyping (Google Stitch)	`stitch-prototyping`	`stitch` (HTTP, si conectado)	no-verificado `[coverage_gap]`
Live web (scrape/crawl/search/deep)	`web-research`	firecrawl (CLI)	reference `[DOC]`

Local plugins (homologados)

Catálogo unificado de los plugins locales activos (gobernados/ruteados, no re-implementados). Ver docs/HOMOLOGATION.md.

Plugin	Homolog skill	Wrapper command	Runtime
plugin-dev	`homolog-plugin-dev`	`/claude-native-toolkit:scaffold-plugin`	no-verificado `[coverage_gap]`
skill-creator	`homolog-skill-creator`	`/claude-native-toolkit:create-skill`	no-verificado `[coverage_gap]`
commit-commands	`homolog-commit`	`/claude-native-toolkit:commit`	no-verificado `[coverage_gap]`
claude-md-management	`homolog-claude-md`	`/claude-native-toolkit:revise-claude-md`	no-verificado `[coverage_gap]`
pr-review-toolkit	`homolog-pr-review`	`/claude-native-toolkit:review-pr`	no-verificado `[coverage_gap]`
agent-sdk-dev	`homolog-agent-sdk`	`/claude-native-toolkit:new-sdk-app`	no-verificado `[coverage_gap]`
firecrawl (skills)	`web-research`	—	reference

Solution-Design Foundation (Tranche 11 — ports de foundation source)

El toolkit es especialista en Claude Code, pero ahora con el fundamento para entender el problema, hacer buenos briefs y formular specs socráticamente antes de construir. Ports curados (con propósito) de internal-foundation-import, multi-plataforma (Firebase + Vercel + Hostinger vía Platform Lens), full packet. Lineage foundation source port of …. [CÓDIGO]

Sub-familia	Skills
Discovery	`discovery-orchestrator` (hub), `input-analyst`, `socratic-debate`, `stakeholder-mapping`, `feasibility-validation`
Spec & Análisis	`requirements-engineering`, `functional-spec`, `trade-off-analysis`, `bdd-full-spectrum`
Arquitectura & Datos	`system-architecture`, `solutions-architecture`, `enterprise-architecture`, `domain-driven-design`, `component-architecture`, `api-design`, `database-design`, `data-strategy`, `ai-software-architecture`, `rag-and-embeddings`, `data-privacy-governance`
AI governance	`agent-constitution-creator` (constitución de agentes; complementa `agent-creator`)

Flujo: discovery-orchestrator compone input-analyst → socratic-debate/requirements-engineering → feasibility-validation/trade-off-analysis → system-architecture/database-design/api-design → functional-spec/bdd-full-spectrum. Ver docs/SKILL-GRAPH.md.

Platform & Operations (Tranche 12 — ports de Beta, aplanados)

Mandato de plataforma Firebase + Vercel + Hostinger + capacidades de producción ausentes. Ports curados (con propósito, no todo) de internal-platform-import: routers aplanados a skills first-class (deep content en references/). Lineage platform import (flattened). [CÓDIGO]

Sub-familia	Skills
Platform (mandato)	`firebase-platform`, `vercel-platform` (net-new), `hostinger-platform`, `web-infra-foundation`
Operations	`application-security`, `observability-operations`, `cicd-release-engineering`

Platform Lens: cada requisito se mapea a la plataforma que encaja (firebase-platform ↔ vercel-platform ↔ hostinger-platform), con web-infra-foundation (DNS/SSL/CDN/DR transversal) debajo. Ver docs/SKILL-GRAPH.md.

Architecture/Data/AI/Quality extendido (Tranche 17 — ports de capability source, relicenciados MIT)

Profundidad de arquitectura/datos/IA/calidad. Ports curados (28 net-new) de tu propio capability source (internal-capability-import), relicenciados MIT + de-brandados (capability genérico, sin marca). Lineage capability source port (relicensed MIT). [CÓDIGO]

Sub-familia	Skills
Arquitectura	`security-architecture` (llena ref de `application-security`), `event-architecture`, `infrastructure-architecture`, `integration-architecture`, `cloud-migration`, `cloud-native-architecture`, `devsecops-architecture`, `mobile-architecture`
Datos	`data-engineering`, `data-quality`, `analytics-engineering`, `bi-architecture`
IA	`genai-architecture`, `ai-design-patterns`, `ai-pipeline-architecture`, `ai-conops`, `ai-testing-strategy`, `ai-architecture-audit`, `ai-architecture-implementation`
AWS (Platform)	`aws-architecture-audit`, `aws-architecture-design`, `aws-architecture-implementation`
Quality (Operations)	`performance-engineering`, `testing-strategy`, `accessibility-audit`
Output	`mermaid-diagramming`, `ux-writing`, `documentation-architecture` (PM port)
Validation	`software-viability`
FinOps (Operations)	`finops` (PM port — cloud cost/FinOps)

PM ports (relicenciados MIT) net-new: finops, documentation-architecture (resto de PM = framework-coupled, out-of-scope; técnicas risk-quantification/scenario/cost-estimation robustecieron trade-off-analysis).

Técnicas robustecidas (no importadas, evitan dup): software-architecture→system-architecture, api-architecture→api-design, observability→observability-operations, data-governance→ data-privacy-governance, technical-feasibility→feasibility-validation. Ver docs/SKILL-GRAPH.md.

Fact-Check Status (vs official docs, claude 2.1.x)

Histórico, no re-verificable desde el filesystem en este probe: cada governing skill se comparó contra code.claude.com/docs en su fecha de chequeo. Trátalo como aserción documental fechada, no como evidencia viva [CÓDIGO]; para re-validar, vuelve a correr el fact-check contra docs y registra la fecha en docs/DECISIONS.md. Status registrado:

Skill	Fact-check	Notes
hooks, permissions, worktrees, scheduled-tasks, goal, plugins	confirmado	0 errores.
mcp-engineering, agent-team-orchestration, headless-sdk-automation, agent-creator, deep-link-runbook, artifact-publishing	confirmado (MINOR)	exactos; detalles menores tagueados `[INFERENCIA]`/`coverage_gap`.
custom-tooling-extension	corregido	`allowed-tools` NO restringe (usa `disallowed-tools`); commands fusionados en skills. v1.1.0.
output-style-authoring	corregido	`/output-style` removido en v2.1.91 → `/config`.
channels-integration	parcial	`channels[]` manifest-array sin confirmar en docs → `[SUPUESTO]`.
dynamic-workflow-forge	parcial	gate Agent SDK floor observado `v0.3.149` (2026-06, probe al usar) / helpers `agent()/parallel()/pipeline()` no en workflows.md → `coverage_gap`.

MINOR coverage_gaps registrados: deep-link repo = slug owner/name (no path/URL); agent-teams split-pane excluye VS Code/Windows Terminal/Ghostty; artifacts = beta; agent-creator "supported fields" de plugin son inferencia; headless flags menos comunes pendientes vs cli-reference.

Claude Cowork Notes

Claude Cowork is a declared runtime target alongside Claude Desktop and Claude Code, but it could not be probed from this Claude Code session. Until a Cowork session runs the Runtime Probe, treat every row above as no-verificado for Cowork specifically [coverage_gap]. Capability parity with Claude Code is an [INFERENCIA], not a confirmed fact.

Cómo cerrar el gap (desde una sesión Cowork):

Corre el Runtime Probe arriba; añade la inspección de surfaces que Code no ve: tool list de la sesión, Agent subagent types, mcp__* tools, hooks que disparan, y presencia de Channels/Cron/worktree primitives.
Para cada fila de la Capability Matrix, marca disponible solo con evidencia viva [CÓDIGO]; lo no observable queda no-verificado [coverage_gap].
Registra el resultado como una columna Cowork espejo de Runtime-status (mismo vocabulario), con su propio stamp fecha, versión junto al de Code. Mientras esa columna no exista con evidencia, la cobertura Cowork es cero real, no parcial.

Procedure

Classify the request: inventory, capability lookup, runtime check, or skill routing.
Run the Runtime Probe; record live evidence with [CÓDIGO]/[CONFIG] tags.
Load references/asset-taxonomy.md for asset families and decision vocabulary.
Load references/source-map.md when source traceability is required.
Map the request to the governing skill in the Capability Matrix.
Return availability with status + evidence; flag no-verificado honestly.

Anti-Patterns

Do not claim a capability is disponible without a live probe.
Do not create one skill per documentation URL or per built-in tool.
Do not assume Claude Cowork parity from a Claude Code probe.
Do not treat experimental/feature-gated surfaces as stable.

References

references/asset-taxonomy.md — Canonical asset families and decision rules.
references/source-map.md — Official source map and owner decisions.
../../docs/SKILL-GRAPH.md — Grafo inter-skills (cómo se componen; generado de metadata.relations).
../../docs/ACCEPTANCE.md — Criterios de aceptación por tier · ../../docs/DECISIONS.md — decisiones + trade-offs.

Contract

Aceptación: cada surface devuelta trae governing-skill + runtime-status + evidence-tag; nada se afirma disponible sin probe vivo. [EXPLICIT]
Límites: cataloga y rutea; no ejecuta la capability ni reemplaza a la skill dueña. [EXPLICIT]
Casos borde: runtime no probable (Cowork desde Code) → no-verificado [coverage_gap], nunca disponible. [EXPLICIT]
Supuestos: docs oficiales code.claude.com como índice de comportamiento [DOC]; paridad Cowork es [INFERENCIA]. [SUPUESTO]
Trade-off: dos fuentes de verdad (runtime vivo + docs) — más fricción, pero detecta drift doc↔runtime. [EXPLICIT]

Packet

Capas del packet, cargables bajo demanda (disciplina ICM: una capa por vez, nunca todas juntas): references/ guías de profundidad (cargar UNA por etapa) · assets/ recursos estáticos.

Repository: JaviMontano/claude-plugins
Path: plugins/claude-native-toolkit/skills/claude-native-inventory/SKILL.md
Commit: e8f986b

Last updated: 3 days ago
First committed: 1 month ago

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.