agentic-workflow
apatrem/agentic-workflow
Agentic-coding conventions, decisions (ADRs), and process — a Claude Code plugin that rides on an external interactive engine (currently Superset).
Instalação
npx convez-ai apatrem/agentic-workflow1 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
README
A tradução para o português deste README ainda está na fila.
agentic-workflow
A reusable Claude Code plugin: the conventions, planning, and decisions for agentic coding — not an orchestration engine.
Backbone:
AGENTS.md+ a deterministic gate (CI) + git/PR isolation. LLMs propose. Tools verify. Git isolates. CI decides. Humans merge. Rules remember.
The orchestration engine is an external worktree manager — currently Superset (macOS app plus bundled CLI / SDK / MCP at ~/.superset/bin). It's a pluggable slot: swap it for another manager (e.g. Claude Squad) in one line. You drive the loop; a human merges. This pack is the operating manual + scaffolder that sits on top. (See ADR-0002 for buy-the-engine / build-the-policy and the current pick.)
What it gives you
/agentic-workflow:init— scaffold the baseline conventions into any repo./agentic-workflow:architect— Phase 1: grill-with-docs → ADRs + CONTEXT (grill-me when no domain model yet)./agentic-workflow:plan— Phase 2: tasks with acceptance tests,risk,mode./agentic-workflow:run— Phase 3: drive Superset; human merges./agentic-workflow:review— themedium-tier dual review on a PR (GPT-5.5 xhigh + Fable 5 effort-high → synthesis).docs/adr/*— the decision record (the durable asset).docs/WORKFLOW.md— the one-page model.templates/— AGENTS.md, pre-commit, CLAUDE/.cursor, task template, agent biases.
The exportability contract
Engine = an external interactive manager (currently Superset). Policy/conventions = this plugin. Per-repo specifics = committed in the repo (AGENTS.md, the gate, tasks/, ADRs). Nothing engine-level is copied per repo, so there is no drift — and the engine slot swaps without touching the conventions.
Default posture
A per-task effort/review dial — mode: low | medium | hard, default low (prefer low, justify higher; ADR-0004). low = one implementer + deterministic gate + one adversarial reviewer. medium adds an independent dual review on every PR (GPT-5.5 xhigh + Fable 5 effort-high, synthesized). hard adds competitive best-of-N + a smart-merge (Fable 5 @ xhigh synthesizes N attempts → one diff), then the medium dual review (hard ⊇ medium). Claude-lineage roles are Fable-first; the latest Opus (≥4.8) at high–xhigh is the documented fallback (ADR-0004 Update). Humans merge at every tier — smart-merge ≠ auto-merge; autonomous auto-merge is the separate, orthogonal opt-in tier you graduate into (ADR-0003/0008), not implied by hard.
Requirements
Claude Code · the official CLIs logged in on your subs (claude, codex, cursor-agent) · gh · Superset (macOS; CLI bundled in the app) for the engine.