Documentación de DIXI Spec Kit English · Código en GitHub

Esta página está disponible solo en inglés; el resumen en español, si existe, está al final.

v1.14.0 support matrix

Purpose

Declare what DIXI Spec Kit v1.14.0 supports, with what evidence, and which controls are automatic or only instructions. Nothing outside this table is announced as supported.

Systems and Python

PlatformStatusEvidence
macOS + Python 3.15SupportedLocal make verify, clean install from wheel and sdist (B3).
macOS / Linux + Python 3.11 and 3.13Supported when CI passesverify job on ubuntu-latest and macos-latest (ADR-103).
Windows (WSL 2, Linux filesystem)Supported when CI passesUbuntu 24.04 in WSL 2 on windows-latest, project under ~ on ext4: make verify, init --agent claude --controls, check, upgrade --dry-run and blocking hooks, run by the wsl.yml CI job (first run, contract). Projects under /mnt/c (or any Windows drive mounted in WSL) are not supported: Windows processes can create junctions there and its translation layer does not have ext4 link semantics.
Windows (native)Not supportedRequires POSIX operations (O_NOFOLLOW, directory descriptors); dixi exits with code 4 and writes nothing, verified on windows-latest by the windows-refusal CI job. Use WSL 2 (ADR-107).

Integrations

IntegrationStatusWhat it installs
--agent noneSupportedDocumentary base, AGENTS.md and guide. Use it with any editor or agent, without installed commands.
--agent claudeSupported (Claude Code 2.1.319)All of the above plus CLAUDE.md and thirty-three /dixi.* commands in .claude/commands/ (fewer in projects created with earlier versions; see dixi upgrade --dry-run).
--agent codexSupported (Codex CLI 0.167.0; manual run)Baseline plus thirty-three skills in .agents/skills/dixi-<operation>/SKILL.md, each with agents/openai.yaml disabling implicit invocation. Invoke them as $dixi-<operation>. Codex reads AGENTS.md itself.
--agent skillsSupported in OpenCode 1.19.5 (full run with dixi 1.12.0: discovery, explicit invocation as @dixi-<operation>, no implicit start observed); format compatible, not verified, in the other agents (contract)Baseline plus thirty-three Agent Skills in .agents/skills/dixi-<operation>/SKILL.md with disable-model-invocation: true, read by Cursor, Gemini CLI and OpenCode according to their documentation (read on 2026-11-14), and by Goose and Amp according to search results; VS Code/Copilot reads .github/skills/ instead. No hooks: --controls is not available. An agent moves to Supported only with a recorded run.
OthersNot includedHaving AGENTS.md does not amount to support.

Claude Code adapter controls

ControlMechanismDefaultWith --controls
Do not overwrite on installdixi initAutomaticAutomatic
Structure, links and traceabilitydixi check / make verifyAutomaticAutomatic
Frozen tests (committed tests are not edited)Instruction + PR review; hook protect_testsInstructionEnforced by hook
Do not sign, attest, approve or mergeInstruction + branch protection; hook guard_attestationInstructionEnforced by hook
Do not read secretsInstruction; dixi rejects sensitive paths; hook guard_secretsPartialEnforced by hook
No destructive commands in MODE=prodInstruction; hook mode_gateInstructionEnforced by hook
No "done" with make verify in redInstruction; hook verify_on_stopInstructionEnforced by hook (once, then the person decides)
Validate structural decisionsInstruction (the command stops)InstructionInstruction
Tool permissionsYour Claude Code configuration; the kit does not change it——

Hooks are a second layer, not the last one: an agent can reach a file through a path a hook does not inspect. Human review and branch protection remain the effective protection.

Codex adapter controls

ControlMechanismAutomatic?
No overwrite on installdixi initYes
Structure, links and traceabilitydixi check / make verifyYes
No implicit start of a DIXI skillallow_implicit_invocation: false per skillYes (Codex setting)
Frozen tests; no sign, attest, approve or merge; no secrets; MODE=prod safety; no "done" in redInstruction in every skill; with --controls, the same hooks through .codex/hooks.json (Bash, apply_patch, Stop, SubagentStop)Instruction; enforced by hook with --controls once you trust the project and approve the hooks with /hooks
Validate structural decisionsInstruction in every skill + human reviewNo
Sandbox and approvalsYour Codex configuration; the kit does not change it—

Capabilities

Included: init with the minimal, site, app and platform profiles and --regulated, doctor, check, workflow list/show/prepare, upgrade --dry-run, placeholders, pr-size, opt-in --controls, Claude Code subagents, the thirty-three operations constitution, checklist, converge, onboard, bugfix, issues, discover, specify, clarify, plan, tasks, build, verify, review, analyze, status, domain, architecture, contract, threat-model, handoff, change, data, quality, ai, release, promote, operate, finops, compliance, duediligence, interview, risk-review, and es/en documents.

Not included yet: extension and preset catalogs, and a second example language.

Resumen en español

Soportado: macOS y Linux con Python ≥ 3.11 (evidencia local en macOS/3.15; CI en Ubuntu y macOS con 3.11 y 3.13). Windows: soportado a través de WSL 2 con el proyecto en el filesystem de Linux (bajo ~; /mnt/c no está soportado), verificado en CI; Windows nativo sale con código 4. Integraciones: --agent none, --agent claude (Claude Code 2.1.319), --agent codex (Codex CLI 0.167.0, treinta y tres skills con invocación implícita desactivada; ejecución manual registrada) y --agent skills (treinta y tres Agent Skills en .agents/skills/; soportado en OpenCode 1.19.5 con ejecución completa registrada (descubrimiento, invocación explícita como @dixi-<operación>, sin arranque implícito observado); formato compatible, no verificado, en los demás agentes; sin hooks). El adaptador de Claude Code instala treinta y tres comandos, cuatro subagentes y CLAUDE.md, sin permisos ni servidores MCP. Por defecto, las reglas de pruebas congeladas, no firmar/aprobar/integrar, no leer secretos y validar decisiones estructurales son instrucciones; con --controls se instalan hooks (Claude Code y Codex) que hacen cumplir todas salvo la validación de decisiones. Los hooks son una segunda capa, no la última. Aún no incluidos: catálogos de extensiones y presets, y Windows nativo.