Skip to content

Repository files navigation

DevMuse

中文文档

DevMuse is a risk-proportional software development workflow for Claude Code, Codex, OpenClaw, Hermes Agent, Gemini CLI, and other Agent Skills hosts. It uses a four-layer architecture of rules, skills, agents, and knowledge without forcing every host through the same ceremony.

Based on Superpowers by Jesse Vincent.

How it works

It starts by classifying the work, not by forcing every request through the same ceremony. Read-only code understanding is answered through proportional source inspection. Exact, reversible execution goes straight to the change and verification. Behavior-changing work enters a quick impact probe.

The probe keeps a clear change to an existing flow bounded: 1–3 inline use cases, TDD where behavior changes, and one combined review. The original request is not re-approved unless the probe exposes a real decision.

Cross-system, public-contract, security, migration, or genuinely unresolved work takes the architectural path: approved scope, technical design, implementation plan, TDD, and one final independent review.

If a bounded task grows, DevMuse upgrades it before the risky surface changes. The process can get heavier when evidence justifies it; task size or a trigger verb alone never does.

Installation

Claude Code

# Register marketplace
/plugin marketplace add knotmark-ai/devmuse

# Install plugin
/plugin install devmuse@devmuse

Other hosts

Codex, OpenClaw, Hermes Agent, Gemini CLI, Cursor, GitHub Copilot CLI, OpenCode, and other Agent Skills hosts use native or generated adapters. See the platform support and installation guide for exact commands and capability differences.

Verify installation

Start a new session and ask for something that should trigger a skill (for example, "scope this behavior change" or "let's debug this issue"). Automatic activation is intentionally host-specific: Codex and Gemini auto-activate scope, architecture, and debug when they match; code execution can take over only after the user asks to execute a recognized DevMuse contract or approved plan. Planning and review normally stay with native host capabilities.

Pipeline

DevMuse is a software engineering workflow tool. It routes development work by risk and uncertainty, and provides on-demand product/business analysis tools.

Proportional paths (auto-routed)

direct → bounded → architectural

direct        → verify → end
bounded       → scope → code → combined review → end
architectural → scope → arch → plan → code → review → end
  1. Direct — Read-only inspection with no durable artifact, or exact mechanical/reversible execution. No workflow skill; inspect or execute, verify proportionally, and report.

  2. mu-scope — Probes behavior-changing work and selects a bounded or architectural path. Bounded produces an inline contract; architectural produces an approved Use Case Set.

  3. mu-arch — Turns architectural scope into technical architecture (components, interfaces, data flow, error handling). Proposes 2-3 approaches and presents real decisions for validation.

  4. mu-plan — Breaks approved architecture into implementation tasks with file paths, verification steps, and UC-ID traceability.

  5. mu-code — Executes bounded contracts inline with one combined review, or architectural plans with inline/subagent implementation and one final review. Enforces RED-GREEN-REFACTOR for behavior changes.

  6. mu-review — Full code-quality and requirements-coverage review for the architectural path. Standalone review requests are report-only; fixes and repository integration require the corresponding authority.

Orthogonal skill (auto-routed)

  • mu-debug — Systematic root cause analysis (red loop first, 4-phase process with architecture escalation).

On-demand skills (direct /slash invocation only)

  • mu-mrd — Market requirements: should we build it (premise, quick mode) or full market analysis (competitors, target market, revenue opportunity, MVP scope). Invoke with /mu-mrd.
  • mu-model 🧪 — Domain model: concepts, archetypes, the spine, who produces and maintains what. Runs before PRD and design, writes repo-root CONTEXT.md. Invoke with /mu-model. Its create path has not been validated on a from-zero project — see Validation status.
  • mu-prd — Product requirements: user flows, object lifecycle models, wireframes, per-feature specs, tiering rules. Invoke with /mu-prd.
  • mu-wiki — The single durable home for current architecture documentation, generated from source with citations. Invoke with /mu-wiki generate or /mu-wiki update.
  • mu-retro — Periodic retrospective: git metrics, review patterns, and durable learnings. Invoke with /mu-retro.
  • mu-grill — Relentless plan/design interview until every rework-forcing fork is resolved. Invoke with /mu-grill.

These are NOT auto-routed. The user explicitly invokes them when needed.

Routing

Routing lives in the always-on bootstrap rule. It first excludes out-of-domain messages, then admits read-only inspection and exact low-risk work to Direct; remaining development work routes by intent and repo state. mu-scope uses codebase evidence to choose bounded or architectural ceremony. Persistent architecture documentation is explicit /mu-wiki, not a side effect of understanding code.

Typical Paths

  • Exact mechanical change: Direct → verify → end
  • Bounded feature on an existing flow: mu-scope (inline contract) → mu-code (one combined review)
  • Architectural feature: mu-scope → mu-arch → mu-plan → mu-code → mu-review
  • Greenfield product: /mu-mrd/mu-prd → then feature loop above
  • Bug fix: mu-scope (1 UC) → mu-debug (mu-debug investigates, implements, and verifies the fix)

Sign-off gate: when CODEOWNERS or multi-author git history indicates team-touching work, creative skills (mu-mrd / mu-prd / mu-arch) prompt for stakeholder sign-off at artifact exit. Non-blocking — user can always override.

Architecture

devmuse/
├── plugin/           Claude/OpenClaw/Gemini runtime
│   ├── rules/            Claude always-on routing (SessionStart)
│   ├── skills/           Canonical workflow source
│   ├── agents/           Claude independent roles
│   └── knowledge/        Shared domain knowledge
├── adapters/codex/   Generated Codex plugin and self-contained Agent Skills
├── plugin.yaml       Hermes plugin manifest
├── __init__.py       Hermes namespaced skill registration
├── docs/             Repository documentation; excluded by subtree installs
└── tests/            Development and compatibility tests

Skills

Category Skill Role
Pipeline mu-scope Impact probe, bounded/architectural classification, use cases and conflict detection
Pipeline mu-arch Approved scope → technical architecture spec through collaborative dialogue
Pipeline mu-plan Architecture → detailed implementation plan with UC-ID traceability
Pipeline mu-code Bounded contract or plan → proportional implementation, TDD, self-check, and one path-level review
Pipeline mu-review Report-only standalone review, or authorized review-and-fix with verification
Orthogonal mu-debug Systematic root cause analysis
On-demand mu-mrd Market requirements — worth building? (quick) or full market analysis (competitors, target market, revenue opportunity, MVP scope)
On-demand mu-model 🧪 Domain model — concepts, archetypes, the spine, ownership; written to CONTEXT.md before PRD or design. create path unproven — see Validation
On-demand mu-prd Product requirements — user flows, object lifecycle models, wireframes, feature specs, tiering rules
On-demand mu-wiki Architecture wiki — generates and maintains project-level architecture documentation
On-demand mu-retro Periodic retrospective with git metrics and durable learning capture
On-demand mu-grill Relentless plan/design interview — resolves every rework-forcing fork before work begins
Meta mu-write-skill Create/edit skills using TDD methodology

Agents

Agent Role
mu-reviewer Five-mode reviewer: design doc (review-design), implementation plans (review-plan), code quality (review-code), requirements coverage (review-coverage), security (review-security)
mu-coder Implementation specialist: builds features from task specs

Rules

Rule Role
bootstrap Skill discovery and invocation rules, priority ordering, decision flow

Hooks

Hook Trigger Role
destructive-guard Bash (Claude Code only) Warns before destructive commands (rm -rf, git push -f, DROP TABLE, git reset --hard). Other hosts keep their native safety policy authoritative.

Knowledge

Category Purpose
languages/ Language-specific review criteria (Java, Go, Python, TypeScript)
templates/ Artifact templates (scope Use Case Set)
principles/ Thinking rubrics loaded at decision points — inversion, premise check, stance detection, sign-off gate, grilling, domain glossary, skill quality, and more (see the directory for the current set)
reviews/ Review checklists: security audit (5-phase OWASP), design audit rubric (architecture scoring)

Philosophy

  • Test-Driven Development — Write tests first, always
  • Systematic over ad-hoc — Process over guessing
  • Complexity reduction — Simplicity as primary goal
  • Evidence over claims — Verify before declaring success

Validation status

Skills marked 🧪 carry a path that has not been exercised on real projects yet. They ship anyway — holding them back until perfect is how the validation never happens.

Skill Proven Not proven Report to
mu-model update and sync, derived from two real runs (aflaj restructure, devmuse rebuild) create — never run on a project starting from zero. Its step sequence is reasoned from Event Modeling and Four-Color archetypes, not extracted from practice #47 — carries a report template

Exit criterion: a path drops its 🧪 after running on two independent projects without a process fix. One clean run is luck.

Local Development

Rebuild and validate the generated host adapter first:

npm run build:adapters
npm run test:platforms

Load the Claude plugin directly from a local directory without installation:

claude --plugin-dir /path/to/devmuse/plugin

After making changes, reload without restarting:

/reload-plugins

Optionally add a shell alias for convenience:

alias claude-dev='claude --plugin-dir /path/to/devmuse/plugin'

For Codex, OpenClaw, Hermes, Gemini, and generic Agent Skills local loading, see Platform support.

Updating

Skills update automatically when you update the plugin:

/plugin update devmuse

License

MIT License - see LICENSE file for details

Credits

About

No description, website, or topics provided.

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages