devflow
yuriyvsavchuk/devflow
Devflow is a framework of **skills** and **worker agents** for Claude Code that enforces disciplined, sequenced software development workflows. It acts as a **router and gatekeeper**: before any code is written, any plan is made, or any review is given, the right worker for the job is selected and announced.
Instalação
npx convez-ai yuriyvsavchuk/devflow1 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.
Devflow
A collection of Claude Code custom skills and worker agent definitions for disciplined, playbook-guided AI-assisted development.
v2 model: playbooks guide, rails verify. A lightweight dispatch skill classifies each new task into one of six playbooks at one of three effort tiers. Playbooks shape the order of work — requirements before plans, reproduction before fixes, measurement before optimization — and produce durable artifacts: specs, plans, ADRs, interface contracts, session records. Deterministic verification (a CLI with run state, Claude Code hooks, CI checks) ships in Phase 2 of the v2 programme; until then the framework is guidance plus artifacts, and says so honestly.
How It Works
- A new task arrives →
devflow-dispatchclassifies it in one line: playbook + effort tier (~800 tokens of routing context, loaded once per task — not per message). - Only the selected playbook loads. It guides the work step by step and produces its artifacts along the way.
- Reviewer and acceptance roles run as fresh-context subagents at standard/full tiers — independence by construction, with findings ranked Blocking / Non-blocking / Question.
- Mid-course corrections are expected: every playbook transition re-checks "still the right playbook and tier?"
Playbooks
| Playbook | Use for | Flow |
|---|---|---|
devflow-shape | Requirements analysis, specification, planning | clarify → research gate → spec with AC-n criteria → size → plan |
devflow-build | Features, behavior changes, refactors, performance | plan → contract → tests → implement → review → accept |
devflow-fix | Bugs, stack traces, wrong behavior | reproduce → failing test → minimal fix → review → accept |
devflow-hotfix | Production incidents | mitigate → verify → ship → record tracked debt → follow-up fix |
devflow-spike | Feasibility questions, approach comparison | hypothesis → throwaway experiment → mandatory retrospective |
devflow-audit | Review requests, security audits | scoped review → severity-ranked findings (security: threat model → dependency scan → review) |
Tests-only and docs-only work goes directly to the test-engineer / docs-updater agents — no playbook ceremony.
Effort Tiers
| Tier | When | What runs |
|---|---|---|
quick | Small, low-risk, no new interfaces | Playbook checklist inline; no subagents |
standard (default) | Everything else | Playbook in main session + one fresh-context reviewer subagent |
full | Public APIs, shared modules, risk keywords (auth, payment, migration…) | Independent reviewer + acceptance subagents; contracts; context maps |
The user's one-word override always wins: quick: fix the date format / full: change the billing API.
What's Inside
skills/— the dispatch skill, six playbooks, and the supporting process skills they call (writing-plans,test-driven-development,systematic-debugging,poc-retrospective,session-continuity,finishing-a-development-branch, and more).agents/— worker subagents with focused roles and per-role model tiers: planner, implementer, test engineer, reviewer, acceptance checker, researcher, triager, profiler, and others. Planning runs on Sonnet by default and escalates to Opus only forfull-tier or XL/low-confidence work.
Every artifact lands in the repo: docs/specs/, docs/decisions/ (ADRs), docs/interfaces/ (contracts), docs/sessions/ (continuity snapshots and hotfix debt), docs/context-maps/.
Installation
# Skills
cp -r skills/* ~/.claude/skills/
# Agents
cp -r agents/* ~/.claude/agents/
Then add the activation line to your ~/.claude/CLAUDE.md (or a project's CLAUDE.md):
When a new development task begins, invoke the devflow-dispatch skill to classify it. Do not re-invoke for mid-task messages or follow-up confirmations.
Migrating from v1: the v1 using-devflow router is now a redirect with a full pipeline→playbook mapping — see skills/using-devflow/SKILL.md. All v1 artifact directories and formats are unchanged.
Team Extension
Devflow makes one developer (or a small team sharing a repo) more effective. The Team Extension (devflow-team) adds coordination for larger teams — structural ledger, decision broadcast, spec-driven QA, cadence/ceremony support. It is private and access-controlled; see team/README.md and team/GETTING-ACCESS.md.
License
MIT