Spec-driven development with intra-phase parallel planning and Agent Teams parallel execution
npx claudepluginhub divkix/ralph-swarmSpec-driven development with parallel execution via Agent Teams. Plan with research, requirements, design, and tasks phases, then execute sequentially or in parallel with Agent Teams swarm coordination.
Divkix
One command. Full spec. Then build — sequentially or in parallel.
By default, each planning phase pauses for review. Use
--fullto run all phases at once.
Goal → Research → (pause) → Requirements → (pause) → Design → (pause) → Tasks → Execution
ralph-swarm is a Claude Code plugin that takes a development goal, breaks it into a structured spec (research, requirements, design, tasks), then executes each task — either one-at-a-time or in parallel using Agent Teams.
You: "Add user authentication with JWT"
ralph-swarm: *researches codebase, writes requirements, designs architecture, breaks into tasks*
ralph-swarm: *executes each task with specialized agents*
ralph-swarm: SWARM COMPLETE
Two phases:
review first) with an opt-in autopilot path (--yolo)# Add the marketplace
/plugin marketplace add Divkix/ralph-swarm
# Install the plugin
/plugin install ralph-swarm@ralph-swarm
# Restart Claude Code
Note: ralph-swarm hook scripts require jq or python3 to parse .ralph-swarm-state.json safely.
git clone https://github.com/Divkix/ralph-swarm.git
claude --plugin-dir ./ralph-swarm/plugins/ralph-swarm
/ralph-swarm:start "Build an authentication system with JWT tokens"
Runs the research phase and pauses. Continue with /ralph-swarm:requirements, /ralph-swarm:design, and /ralph-swarm:tasks (or use --full to run all phases at once), then review and execute.
/ralph-swarm:start "Build an authentication system with JWT tokens" --swarm
Same incremental planning flow (research first, then review/pause commands), but execution uses multiple agents working in parallel after /ralph-swarm:go.
/ralph-swarm:start "Build an authentication system with JWT tokens" --swarm --yolo
Plans, skips review, immediately fires up a team of agents. No human in the loop until it finishes.
| Command | Description |
|---|---|
/ralph-swarm:start <"goal"> [flags] | Begin planning + execution |
/ralph-swarm:requirements | Run the requirements planning phase |
/ralph-swarm:design | Run the design planning phase |
/ralph-swarm:tasks | Run the task breakdown planning phase |
/ralph-swarm:go | Resume execution after reviewing the spec |
/ralph-swarm:status | Show current progress |
/ralph-swarm:cancel | Abort the session, clean up state |
/ralph-swarm:rollback | Reset to pre-execution state (destructive) |
/ralph-swarm:help | Print usage and flags |
| Flag | Default | Description |
|---|---|---|
--full | false | Run all 4 planning phases in one shot without pausing |
--swarm | false | Enable parallel planning (intra-phase) and parallel execution (Agent Teams) |
--yolo | false | Skip spec review, go straight to execution |
--teammates N | auto | Number of parallel agents (max 10). Auto = min(task_count, 5) |
--agent-type TYPE | auto | Agent type for executors (e.g., typescript-pro, golang-pro) |
--max-iterations N | 30 | Safety cap on execution loop iterations |
--commit | true | Commit after each task (sequential) or after all tasks (swarm) |
--no-commit | — | Disable auto-commits |
Planning consists of four sub-phases that run in strict order. Each delegates to a specialized agent:
Official prompts.chat marketplace - AI prompts, skills, and tools for Claude Code
Open Design — local-first design app exposed to coding agents over MCP. Install once with your agent's plugin command and projects/files/skills are reachable through stdio.
Behavioral guidelines to reduce common LLM coding mistakes, derived from Andrej Karpathy's observations