Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
npx skills add https://github.com/Zate/cc-plugins --skill help
La commande reste sur une seule ligne. Faites défiler horizontalement pour la vérifier avant de la copier.
Vous préférez une copie locale ? Téléchargez les fichiers actuellement disponibles dans SkillsMP.
If $ARGUMENTS specifies a topic, skip to that section. Otherwise:
AskUserQuestion:questions:-question:"What would you like to learn about?"header:"Topic"multiSelect:falseoptions:-label:"Getting Started"description:"New to devloop? Start here"-label:"Commands"description:"What each command does"
-
label:
"The Loop"
description:
"The plan -> run -> fresh cycle"
-
label:
"Skills & Agents"
description:
"On-demand knowledge and parallel work"
-
label:
"Superpowers"
description:
"How devloop works with superpowers plugin"
Topic: Getting Started
What is devloop?
Development workflow where Claude does the work directly. No routine agents. Just you, Claude, and the code.
Quick Start
/devloop:plan "How should we implement feature X?"# Plan (or --deep for exploration)
/devloop:run # Execute tasks
/devloop:fresh && /clear # Save state, clear context (every fresh_threshold tasks (default 10, configurable in .devloop/local.md))
/devloop:run # Continue execution
First Session
Have a task? /devloop:plan [task] (or --deep for detailed exploration)
Claude explores, creates plan
You approve
/devloop:run to implement
Every fresh_threshold tasks (default 10, configurable in .devloop/local.md): /devloop:fresh then /clear then /devloop:run
Done? /devloop:ship
Topic: Commands
Decision Tree
Starting new work?
|-- Unclear requirements -> /devloop:plan --deep
|-- Small, clear task -> /devloop:plan --quick
|-- Normal feature -> /devloop:plan
|-- Have a plan -> /devloop:run
Mid-workflow?
|-- Context heavy -> /devloop:fresh -> /clear -> /devloop:run
|-- Ready to commit -> /devloop:ship
|-- Want review -> /devloop:review
Returning?
|-- /devloop:run
Command Reference
Command
Purpose
/devloop
Start new workflow (smart entry point)
/devloop:plan
Autonomous planning (default mode)
/devloop:plan --deep
Comprehensive exploration with spike report
/devloop:plan --quick
Fast path for small, clear tasks
/devloop:plan --from-issue N
Start from GitHub issue #N
/devloop:run
Autonomous execution
/devloop:run --interactive
With checkpoints
/devloop:run --next-issue
Full issue-to-ship pipeline
/devloop:run-swarm
Swarm for 10+ tasks
/devloop:fresh
Save state for clear
/devloop:review
Code review
/devloop:ship
Commit & PR
/devloop:archive
Archive completed plan
/devloop:new
Create GitHub issue
/devloop:issues
List GitHub issues
/devloop:statusline
Configure statusline
/devloop:help
This guide
Migration from Old Commands
Old Command
New Command
/devloop:spike
/devloop:plan --deep
/devloop:quick
/devloop:plan --quick
/devloop:from-issue N
/devloop:plan --from-issue N
/devloop:continue
/devloop:run --interactive
/devloop:ralph
/devloop:run
Topic: The Loop
Pattern
/devloop:plan [topic] -> Creates plan
|
/devloop:run -> Implements tasks
|
After fresh_threshold tasks (default 10, configurable in .devloop/local.md)?
|-- Yes -> /devloop:fresh -> /clear -> /devloop:run
|-- No -> Continue
Why It Works
Fresh context = better reasoning
Plan preserves progress
Checkpoints keep control
Sustainable pace
When to Fresh
After fresh_threshold tasks (default 10, configurable in .devloop/local.md)
When responses slow
After long exploration
Before taking a break
State Files
File
Purpose
.devloop/plan.md
Current plan (persists)
.devloop/next-action.json
Fresh state (consumed)
.devloop/worklog.md
Work history
Topic: Skills & Agents
Skills: On-Demand Knowledge
Load when needed: Skill: skill-name
Skill
Purpose
plan-management
Working with plan.md
local-config
Project settings via .devloop/local.md
pr-feedback
Integrate PR review comments
devloop-audit
Audit devloop against Claude Code updates
Don't preload. Claude loads automatically when needed.
Agents: Parallel Work Only
Use for: parallel implementations, security scans, large exploration (50+ files)
Don't use for: writing code, tests, git ops, single files, docs
Agent
Purpose
devloop:engineer
Exploration, architecture, code review
devloop:qa-engineer
Test generation
devloop:security-scanner
OWASP, secrets
devloop:doc-generator
READMEs, API docs
devloop:swarm-worker
Parallel task execution
Topic: Troubleshooting
Common Issues
Plan corrupted:rm .devloop/plan.md && /devloopSession ended:/devloop:runAbandon plan:/devloop:archive && /devloopContext heavy:/devloop:fresh && /clear && /devloop:runSkill not loading: Check exact name in skills/INDEX.md
Progress not saving: Use [x] not [X]Statusline missing:/devloop:statusline, restart Claude Code
Claude Code Native Integrations
LSP tools in agents: engineer, qa-engineer, and security-scanner agents use LSP for symbol navigation (goToDefinition, findReferences, documentSymbol). Gracefully falls back to Grep/Glob if no LSP server is configured.
Monitor for long commands: run, run-epic, and run-swarm use Monitor (not Bash) for test suites, builds, and full-codebase linting to stream output in real-time. Falls back to Bash if Monitor is unavailable.
Worktree isolation: run-swarm --worktrees runs each worker in an isolated git worktree. Enable via flag or git.worktree_isolation: true in .devloop/local.md. Off by default.
Token budget: tokens.token_budget in .devloop/local.md caps how many tokens gather-task-context.sh collects per task (default: 4000). Lower for leaner context; raise for large codebases.
Cache-friendly context: tokens.cache_friendly_context: true (default) reorders agent spawn prompts to maximize prompt cache hits, reducing API costs on repeated runs.
Reset Everything
rm -rf .devloop/ && /devloop
Topic: Automation
Autonomous Execution
/devloop:run # Up to 50 iterations
/devloop:run --max-iterations 100 # Custom limit
/devloop:run --interactive # With checkpoints
How It Works
/devloop:run creates ralph-loop state
Claude works on tasks, marks [x]
When all done, outputs <promise>ALL PLAN TASKS COMPLETE</promise>
Loop terminates
Context Guard
Auto-exits at 70% context usage. Configure in .devloop/local.md:
context_threshold:80
Stopping Early
/cancel-ralph
Max iterations reached
Context guard triggered
Topic: Superpowers
Complement, Not Dependency
Devloop and superpowers are complementary plugins with distinct lanes:
Both can coexist. Use devloop for the workflow, superpowers for quality gates within tasks.
Step 3: Offer More
AskUserQuestion:questions:-question:"Learn about another topic?"header:"More"multiSelect:falseoptions:-label:"Yes, show topics"description:"Return to topic menu"-label:"No, I'm good"description:"Exit help"
If yes, return to Step 1. If no:
Ready to start? Try /devloop:plan [task] or /devloop:run if you have a plan.
