DevolaFlow

A guardian's song for your code — orchestrating workflows so nothing is left behind.

Define multi-stage delivery pipelines as declarative YAML. A 4-layer agent hierarchy watches over every task with quality gates, so your code is never unprotected.

Design Architecture Workflow Visualizer Stage Explorer GitHub User Guide

Install in 1 command

# Cursor (project-local)
curl -fsSL https://raw.githubusercontent.com/YoRHa-Agents/DevolaFlow/main/scripts/install.sh | bash -s cursor

# Cursor (user-global, shared across all projects)
curl -fsSL https://raw.githubusercontent.com/YoRHa-Agents/DevolaFlow/main/scripts/install.sh | bash -s cursor --global

# Claude Code (user-global)
curl -fsSL https://raw.githubusercontent.com/YoRHa-Agents/DevolaFlow/main/scripts/install.sh | bash -s claude --global

# Update existing install to latest version
curl -fsSL https://raw.githubusercontent.com/YoRHa-Agents/DevolaFlow/main/scripts/install.sh | bash -s update

Also: ... | bash -s claude · ... | bash -s claude --global · ... | bash -s copilot · ... | bash -s all · ... | bash (auto-detect)

Or just download MVP-SKILL.md and drop it into your tool.

How It Works

Every workflow executes through a strict 4-layer hierarchy. Upper layers dispatch; only the bottom layer works.

L0
Project
picks workflow type, dispatches stages
~3K ctx
▼ dispatches
L1
Stage
decomposes into waves, runs quality gates
~5K ctx
▼ dispatches
L2
Wave
dispatches tasks in parallel, collects results
~4K ctx
▼ dispatches
L3
Task
writes code, runs tests, reviews — the ONLY layer that works
~8K ctx

17 Built-in Workflow Types

Pick the right workflow for your task. Each is a different composition of the same 13 stage primitives.

Workflow When to use Stages
full-pipelineNew feature, greenfield projectdesign - plan - impl - review - test - refine - gate - release
hotfixProduction bug, urgent fixtriage - fix - test - release
refactoringTech debt, restructurescope - plan - impl - test - review
research-onlyCompare alternatives, surveyresearch - compare - report
design-onlyArchitecture, API designresearch - design - review
migrationUpgrade, port systemsassess - plan - impl - validate - cutover
spike-pocPrototype, experimentresearch - prototype - evaluate
documentationDocs, guides, API refssurvey - author - review
security-auditVulnerability scanthreat - scan - analyze - remediate - verify
feature-enhancementExtend existing featuresscope - design - plan - impl - review - test - release
RDRRDesign with researchresearch - design - review - refine (loop)
demo-showcaseBuild demo, presentation, pitchresearch - storyboard - build - review - polish - package
performance-optimizationProfile, benchmark, optimize speedprofile - design - optimize - benchmark - validate
dependency-setupEnvironment, install, configure toolsresearch - plan - configure - verify
onboardingNew contributor, codebase surveyanalyze - document - setup - verify
skill-optimizationOptimize agent skill, benchmark contextsurvey - profile - optimize - benchmark - iterate - document
self-updateSkill refresh, upstream synccheck-refs - research-updates - decompose - integrate - test - evaluate

Interactive Tools

Explore the framework, workflows, and stages visually — no installation required.

ARCHITECTURE

Design Architecture

The complete framework map: every skill file, its design document source, tier, token budget, and the workflows it enables.

Open Architecture Map
INTERACTIVE

Workflow Visualizer

Select any of the 17 workflow types and see its stage pipeline, teams, gate types, and loop structure rendered as a diagram.

Open Visualizer
INTERACTIVE

Stage Explorer

Deep-dive into any of the 13 stage primitives. See input/output types, delegation chains, and context budget allocation.

Open Explorer
DOCS EN

User Guide

Step-by-step: install, configure, run your first workflow, integrate with AI tools.

User Guide
DOCS ZH

Chinese Guide

Quick start: install, configure, run your first workflow, integrate with AI tools.

Chinese Guide
DOCS

Architecture Overview

System architecture: 4-layer hierarchy, 13 stage primitives, gate mechanism (preflight, revision, escalation, abort), context isolation.

Read Overview

What's New in v5.2.0

NineS CLI integration, self-improvement loops, and Karpathy-inspired behavioral norms bring DevolaFlow to a new level of autonomous quality.

INTEGRATION

NineS v2.0.0 CLI Integration

Full CLI migration with 6 breaking changes resolved. Self-improvement loop infrastructure: self-eval, iterate, and benchmark in a unified pipeline.

GitHub
QUALITY

Code Quality: CC 23→7

NineS-guided complexity reduction brought select_context cyclomatic complexity from 23 to 7. NineS findings reduced from 7 to 0 across the codebase.

View Results
NORMS

Karpathy Behavioral Norms

Explicit assumptions documentation, simplicity review gates, and anti-complexity checks inspired by Andrej Karpathy's software engineering principles.

Read More
METRICS

681 Tests, 90.90% Coverage

25 EvoBench scenarios (avg 99.50), 18 context profiles, 18 reference dependencies tracked. Quality gates hold the line — just as Devola would.

View Results

v3.3.0 Highlights

Full workflow coverage: 18 context profiles, 25 benchmark scenarios, optimized budget utilization (99.51 avg composite) — multi-round iterative optimization across all workflows.

BENCHMARKS

EvoBench

Context density benchmark suite with 25 scenarios covering all 18 context profiles. Measures section relevance, information density, noise ratio, and budget utilization. Avg composite: 99.51/100.

View Benchmarks
RULES

34 Repository Rules

Enforceable guardrails codified from iteration lessons: SKILL format constraints, change process rules, context optimization rules, documentation sync rules, and self-improve iteration rules.

View Rules
CONTEXT

Task-Adaptive Selection

18 context profiles for all workflow types. Each profile uses longest-match routing with optimized ~85% budget utilization.

View Profiles
VISUALIZATION

Benchmark Results

Interactive visualization of 6 EvoBench optimization rounds across 25 scenarios. Tracks composite scores, budget utilization, and round-over-round improvement (baseline 80.5 → final 99.51).

View Results

Use with Your AI Tool

DevolaFlow adapts to four AI coding tools from a single source. Run make build-skill and copy the output.

CURSOR

Cursor Skills

Full SKILL.md + references + rules. Copy dist/cursor/ to .cursor/skills/devola-flow/

CODEX

OpenAI Codex

SKILL.md with inline rules + openai.yaml. Copy dist/codex/ to ~/.codex/skills/

CLAUDE

Claude Code

Compressed CLAUDE.md (<200 lines). Copy dist/claude/CLAUDE.md to project root.

COPILOT

GitHub Copilot

Copilot instructions (<4K chars). Copy dist/copilot/.github/ to your repo.