🌐 中文版 | English

devpace

Give your Claude Code projects a steady development pace — requirements change, rhythm stays.

Why devpace

When using Claude Code for product development:

Problem	Without devpace	With devpace
Requirements change = chaos	"Add a feature" triggers cascading confusion, nobody knows the blast radius	Impact analysis + orderly adjustment, Claude won't change the plan on its own
Inconsistent quality	Claude sometimes skips tests, forgets checks	Auto-checks + human approval, quality gates cannot be bypassed
Work drifts from goals	Technical work disconnects from business goals, lots done but unclear value	Goal-to-code traceability at all times
Re-explain everything each session	Manual approach needs 8 user corrections (3 interruption test)	Auto-restores context, 0 corrections

See full walkthrough: from init to done

30-Second Experience

/pace-init                 ← Initialize (once)
"Help me implement user auth" ← Claude auto-tracks tasks, writes code, runs tests, checks quality
"approve" or "reject"      ← You decide whether to merge
"add an export feature"    ← Claude analyzes impact, adjusts plan, waits for your confirmation
"pause auth for now"       ← Impact analysis → pause (preserving work) → adjust plan

Next session, Claude reports: "Last time we stopped at auth module, continue?" — zero manual re-explanation.

Full Development Lifecycle

 Goal         Features       Code Changes      Quality         Ship
You define ──→ Plan together ──→ Claude codes ──→ Auto + You ──→ Optional auto
                │                   │               │              │
             pace-plan          pace-dev        pace-review    pace-release
             pace-change                                       pace-feedback

Requirements can change anytime — /pace-change auto-analyzes impact, adjusts the plan, and waits for your confirmation.

After each cycle, /pace-retro shows quality metrics and improvement trends.

How It Works

devpace builds a goal-to-code traceability chain in your project:

1. Goal alignment — Every code change links back to a business goal. No work without purpose.
2. Auto quality gates — Claude auto-checks code quality and requirement consistency, self-repairs on failure. Human approval cannot be skipped.
3. Change is normal — Requirements changed? Auto impact analysis, orderly adjustment, existing work preserved.
4. Interrupt-proof — Session broke? Auto-resume next time. All state in .devpace/ as plain Markdown.

Under the hood: a Claude Code Plugin using Rules (behavioral guidelines) + Skills (/pace-* commands) + Hooks (auto-triggers at critical moments).

Installation

Prerequisite: Claude Code CLI must be installed.

Marketplace Install (Recommended)

# Step 1: Register marketplace (one-time)
/plugin marketplace add arch-team/devpace

# Step 2: Install
/plugin install devpace@devpace

From Source

git clone https://github.com/arch-team/devpace.git
claude --plugin-dir /path/to/devpace

Verify Installation

After installing, type /pace- in Claude Code. If devpace is loaded, you'll see auto-complete suggestions for /pace-init, /pace-dev, /pace-status, etc.

Commands

Daily Use

Command	Purpose
/pace-init	First-time setup (once, supports --from PRD for auto feature tree)
/pace-dev	Start coding
/pace-status	Check progress
/pace-review	Approve changes
/pace-next	Not sure what's next? Get AI recommendations

When Requirements Change or a Cycle Completes

Command	When to use
/pace-change	Add requirements, pause, change scope, reprioritize
/pace-plan	Cycle done, plan the next one
/pace-retro	Review how things went

Specialized Features (Optional)

Command	Scenario
/pace-test	Requirements-traceable test management
/pace-guard	Risk fabric: Pre-flight scan + Runtime monitoring + Trend analysis + Graduated response
/pace-release	Release orchestration: Changelog + Version bump + Git Tag + GitHub Release
/pace-sync	External tool bridge: sync task state ↔ GitHub Issues (labels + comments)
/pace-role	Switch perspective (PM / Tester / Ops / etc.)
/pace-theory	Learn the methodology behind devpace
/pace-feedback	Collect post-launch feedback
/pace-trace	View full reasoning trace of AI decisions

Most of the time you don't need commands — saying "help me implement X" equals /pace-dev, "where are we" equals /pace-status.

Core Capabilities

Change Management (Core Differentiator)

Capability	Description
Requirement changes	Add features, pause, change scope — auto impact analysis, orderly adjustment, Claude won't change the plan unilaterally
Complexity awareness	Auto-assesses task complexity, small changes fast-track, large changes full process, complexity drift auto-detected

Quality & Traceability

Capability	Description
Quality gates	Code quality + requirement consistency auto-check + adversarial review, human approval cannot be skipped
Goal traceability	From business goals to code changes, always traceable
Test verification	Requirements-driven — strategy generation, coverage analysis, AI acceptance verification, change impact regression

Development Rhythm

Capability	Description
Cross-session restore	Session interrupted? Auto-resume, zero re-explanation, experience persists across sessions
Iteration management	Plan → Execute → Review full cycle, auto-recommends next step
Progressive autonomy	Assisted / Standard / Autonomous — more guidance for new users, less for experienced ones
DORA proxy metrics	Deploy frequency / Lead time / Failure rate / MTTR proxy values, Elite~Low benchmarks + trend comparison
CI/CD awareness	Auto-detects CI tool type, Gate 4 auto-queries CI status, zero config
Risk fabric	Pre-flight 5-dimension risk scan + Runtime monitoring + Graduated autonomous response (High requires human confirmation)
Cross-project insights	High-confidence insights exportable/importable to other projects, reducing redundant learning

Workflow

Two Modes

• Explore mode (default): Freely read code, analyze problems, discuss approaches. No process triggered.
• Advance mode (when changing code): Auto-creates tasks, tracks progress, checks quality. Small changes fast-track, large changes full process.

When unsure, Claude asks: "Ready to start coding, or just exploring?"

Workflow

Normal flow:
Start ──→ In Progress ──→ Pending Review ──→ Done
              │                │
        Auto quality check  You approve      Auto merge + status update
        (Claude handles)   (you decide)

Anytime:
  Requirements changed ──→ Impact analysis ──→ Adjust plan ──→ Continue
  Session interrupted  ──→ Next session auto-resumes from where you left off

Full cycle (optional):
  Plan (pace-plan) → Build (pace-dev) → Review (pace-retro) → Next cycle

Design Principles

Principle	Meaning
Zero friction	Natural language works, no jargon to learn
Progressive disclosure	Default 1-line output, details on demand
Byproducts not prerequisites	Structured data is auto-produced from work, not a required input
Interruption tolerance	Interrupt at any point, seamless resume next time

vs Alternatives

Dimension	GitHub Issues / Manual	devpace
Core model	Task list	Goal → Feature → Code Change traceability
Requirement changes	Manual impact assessment	Auto impact analysis + orderly adjustment
Claude's role	Executor (you direct each step)	Autonomous collaborator (auto-advances, self-checks, waits for your decisions)
Traceability	Task → Code	Business Goal → Feature → Change → Code
Metrics	Completion count	Quality pass rate + value alignment + DORA proxies

What devpace is NOT

• Not a CI/CD pipeline — it works alongside your existing tools (GitHub Actions, Jenkins, etc.)
• Not a project management platform — no web dashboard, no team features, pure CLI
• Not a replacement for git — it creates Markdown state files in .devpace/, your code stays in git

Feedback

Tried devpace? Tell us what you think — one word or a full report, everything helps.

Learn More

• User Guide — Full command reference, modes, state machine details (中文)
• Walkthrough — Complete example from init to finish (中文)
• Contributing — Dev environment, testing, PR process
• Changelog — Version history (English summaries on GitHub Releases)
• Troubleshooting — Search known issues or open a new one

---

MIT