Forging Fire — Swarm Multi-Agent para Claude Code
b0stwickvibes/forging-fire
Motor swarm unificado: rigor DRY + personalidade configurável por repo
Ver no GitHubEngine padronizado e rigoroso de swarm multi-agent para Claude Code, com dois níveis: deep-swarm (por repositório) e deep-swarm-prime (nível de estate). Funciona com um motor invariante e personalidade via configuração JSON, permitindo que melhorias no engine beneficiem todos os projetos simultaneamente.
Instalação
npx convez-ai b0stwickvibes/forging-fire1 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
Casos de uso
- Orquestração de múltiplos agentes especializados em fluxos de trabalho complexos
- Automação de code review com gauntlet de especialistas (arquiteto, revisor, designer UI)
- Forge de novas skills com pesquisa, prototipagem e verificação em triplo-loop
- Auditoria e mapeamento de toda a infraestrutura de desenvolvimento (estate-level census)
- Padronização de processos de desenvolvimento sem divergência de código entre projetos
Veredito

Solução sofisticada para times que precisam orquestrar múltiplos agentes Claude com rigor e DRY. Exige setup cuidadoso de agentes nomeados e compreensão arquitetural clara, mas elimina divergência de orquestração entre projetos. Recomendado para pipelines de desenvolvimento enterprise ou multi-repo.
README
A tradução para o português deste README ainda está na fila.
Forging Fire 🔥
A standardized, rigorous multi-agent swarm for Claude Code. One invariant engine, per-project personality as config — so the rigor improves once and every project benefits.
Three pieces, one dependency chain:
deep-swarm-prime (estate-level conductor)
└── dispatches /deep-swarm (per-repo rigor engine)
└── runs workflows/deep-swarm.js (the agent fan-out)
deep-swarm— run a rigorous swarm on ONE repo/task:research → expert gauntlet → premortem → spec → [user gate] → build → triple-loop verify → capture. Personality lives in a per-repo.claude/deep-swarm.profile.json, never a fork of the engine.skill-forge(forging new skills) is just one preset of it.deep-swarm-prime— the layer above: census your whole setup top-down, find cross-cutting patterns no single-repo view can see, harden the~/.claudemachine itself, then dispatch a scoped/deep-swarmper target. Ends with a ranked roadmap of what to build next.workflows/deep-swarm.js— the engine both skills lean on. Profile-driven, no hardcoded paths.
Install
Copy into your ~/.claude:
# from the repo root
cp workflows/deep-swarm.js ~/.claude/workflows/
cp -r skills/deep-swarm ~/.claude/skills/
cp -r skills/deep-swarm-prime ~/.claude/skills/
(Or run ./install.sh.) Then restart Claude Code so it picks up the new skills.
Requirements (read this — it's the part that bites)
- The engine path.
deep-swarm.jsmust end up at~/.claude/workflows/deep-swarm.js. The skills reference it there. - An agent roster. The swarm's "gauntlet" dispatches subagents by exact display name (e.g.
Software Architect,Code Reviewer,UI Designer,Motion Engineer). Those names must exist in your Agent registry — if they don't, that gauntlet branch silently drops (no error, just less rigor). Either add agents with those names or edit your profile'sexpert_agentsto names you have. - A memory store is optional. The final "capture" step writes a note to a knowledge store (the author uses an MCP-based one). If you don't have one, that step is a no-op — skip it.
Use
/deep-swarm init # bake a .claude/deep-swarm.profile.json for the current repo
/deep-swarm <task> # run the rigorous swarm on a task
/deep-swarm forge skills <list> # forge new skills (the skill-forge preset)
/deep-swarm-prime census # READ-ONLY map of your whole setup (safe first run)
/deep-swarm-prime # full estate run (stops at a user gate before any change)
deep-swarm-prime defaults to AUDIT mode — it maps, analyzes, and recommends, and never mutates anything outside its run directory without explicit per-item approval. Run census first.
Personalize
- Per-repo:
skills/deep-swarm/reference/profile.template.jsonis the schema./deep-swarm initfills one in for you. - Estate-level:
deep-swarm-prime's "Your landmarks" section is the one spot you hand-fill with your own systems. Leave it blank and the census just discovers everything cold.
Design principle
DRY rigor (one engine, improved once-for-all) + bespoke-without-divergence (a config profile per repo) + presets that are data, not separate systems. Forking the orchestration per project is the exact drift this is built to kill.
License
MIT — see LICENSE.
Contributing
Issues and PRs welcome. Keep the engine generic — project-specific behavior belongs in a profile, not a fork of deep-swarm.js.