convez.ai
SkillDesenvolvimento WebIntermediário
anrylu

InferSpec — Inferência Reversa de Specs OpenSpec

anrylu/infer-spec

Extrai specs estruturadas de código legado + git + docs

Ver no GitHub
InicialPluginScore · de 0 a 100

Ferramenta que reverte a lógica tradicional de desenvolvimento: em vez de escrever specs primeiro e depois código (SDD), ela analisa seu codebase legado, histórico git e documentação para inferir automaticamente specs OpenSpec estruturadas. Ideal para codebases sem especificações formais, marcando lacunas com [GAP]/[TBD] para refinamento iterativo.

Instalação

$npx convez-ai anrylu/infer-spec

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
3
Forks
1
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Documentar APIs e serviços legados sem especificações
  • Gerar spec.md estruturado a partir de codebase existente
  • Transicionar de código para especificação antes de refatoração
  • Auditar e mapear capacidades e requisitos implícitos em produção

Veredito

InferSpec resolve um problema real: legados sem spec. A abordagem DDS (Development-Driven Spec) é pragmática, e o design em camadas (CLI + skills) é elegante. Requer disciplina do usuário para validar GAPs, mas economiza horas de documentação manual.

README

A tradução para o português deste README ainda está na fila.

English | 繁體中文 | 简体中文 | 日本語

InferSpec

Reverse-infer OpenSpec specs from your codebase + git history + docs — designed for legacy code that has no spec.

CI License: MIT

Development-Driven Spec (DDS) — the inverse of Spec-Driven Development. From Code & Context to Clear Specs.

SDD vs DDS

Spec-Driven Development (SDD) is the discipline of writing the spec first, then implementing against it. It works beautifully on greenfield projects.

The real world is mostly brownfield. You inherit a 50K-line Flask service. There is no spec — just a Jira board from three years ago, a Confluence wiki nobody updates, and the git log. SDD has no entry point here.

InferSpec inverts the loop: Development-Driven Spec (DDS). Code already exists; treat it (plus git history, tickets, docs, MCP-attached wikis) as the source of truth and reverse-infer a structured OpenSpec spec from it. Once the spec exists, you can switch back to SDD for new work.

ModeStarting pointOutput
SDD (Spec-Driven Development)A specCode
DDS (Development-Driven Spec)Code + history + docsA spec

Why InferSpec?

InferSpec reads all of it and produces a structured OpenSpec spec — one spec.md per capability — with each Requirement cited back to file:line or a ticket ID. Ambiguities are marked [GAP]/[TBD] so you can fill them interactively in a follow-up pass.

How it works

┌─────────────────────────────────────────────────────────────┐
│  Layer 1 — uvx Python package: installer + CLI               │
│  (never calls an LLM API)                                   │
└─────────────────────────────────────────────────────────────┘
                       │ installs skills into
                       ▼
┌─────────────────────────────────────────────────────────────┐
│  Layer 2 — Skills run inside Claude Code / Codex / Gemini   │
│  / Copilot / OpenCode and use the host's subscription AI    │
└─────────────────────────────────────────────────────────────┘

InferSpec leans on your existing AI subscription. No API keys, no cloud endpoints to configure.

Install

uvx inferspec init --platform claude-code

That drops /inferspec-scan + /inferspec-cap into .claude/skills/ for the current directory. See inferspec platforms for the full list.

Updating

When a new version of the inferspec package ships, the bundled skill files inside each repo's .claude/skills/ (or equivalent) stay frozen at the version that was installed. Refresh them with:

pip install -U inferspec        # or: uvx --refresh inferspec ...
inferspec update                # in each repo where you ran `inferspec init`

inferspec update reads .inferspec.yaml to find the platforms you previously installed into, then re-copies the bundled skills (no prompts). Use inferspec update --check to report drift without writing anything, or inferspec doctor to see the installed-vs-package version per platform.

Usage

Open your AI agent in the target repo. Two skills are available:

/inferspec-scan — bulk-infer specs for every capability:

/inferspec-scan

It runs graphify to cluster files into capabilities, then for each cap reads code + git log + docs/ + (if available) Jira/Confluence via MCP + URLs via the host's WebFetch, and drafts openspec/specs/<cap>/spec.md in OpenSpec format. Drafts may contain [GAP] / [TBD] markers where the AI was unsure.

/inferspec-cap <slug> — single-capability deep dive with interactive Q&A:

/inferspec-cap user-auth
/inferspec-cap "rate limiting"       # fuzzy match
/inferspec-cap                       # interactive picker
/inferspec-cap new-feature --new     # bootstrap a brand-new cap

For one capability, the skill solicits Jira/Confluence/URLs you have, then asks one focused question per [GAP] marker until the spec converges. On exit it offers to commit the result for you.

Multi-source artefacts are picked up automatically — InferSpec detects MCP servers in your host environment rather than shipping its own clients.

Output format

Same convention as OpenSpec:

## Purpose

User authentication for the order portal — replaced the legacy SSO bridge
after incident-1234. See AUTH-456.

## Requirements

### Requirement: Rate Limiting
The system SHALL reject login attempts after 5 failures within 60 seconds.

**Source:** auth.py:18-21, [JIRA AUTH-456]

#### Scenario: Lockout after repeated failures
- **GIVEN** 5 failed attempts in the last minute
- **WHEN** another POST /auth/login arrives
- **THEN** server returns 429

Status

v0.3 alpha. Ships:

  • /inferspec-scan — bulk mode, plus design-doc auto-discovery, OpenAPI/Swagger detection, --since <rev> incremental scan, glossary enforcement (.inferspec-glossary.txt), and removal proposals under openspec/changes/
  • /inferspec-cap <slug> — interactive single-cap mode, also covers iterative gap-fill on existing specs
  • inferspec update — refresh installed skill bundles per repo

License

MIT

Ver o original no GitHub

Bom para

Relacionados