harness-engineering

harness-engineering

Lifecycle orchestrator for AI-native software development.

A top-level Claude Code skill that coordinates the entire software development lifecycle — from requirements capture through release — by applying harness theory (compact machine state, structured artifacts, deterministic transitions) at the phase level instead of the task level.

English · 中文

---

What this is

harness-engineering is not a replacement for harness-plan. It is one layer up:

harness-engineering  ← lifecycle orchestrator (this project)
        │
        └── invokes harness-plan as the implementation-phase executor
                 │
                 └── harness-plan drives feature-level coding

The 7 phases (locked)

discovery → design → architecture → implementation → test → release → ops
                  │                ▲
                  │                │
                  └────────────────┘
                (parallel where upstream allows)

Each phase owns a schema-validated JSON artifact. Transitions are enforced by scripts, not ceremony. Failed validation blocks advance. Stale upstream propagates downstream.

---

Dependency: REQUIRES harness-plan skill

This skill delegates implementation phase to the harness-plan skill. Without harness-plan, phases 1-3 and 5-7 work but phase 4 (implementation) will block.

Why not merge the two skills?

• harness-plan is useful on its own for pure coding tasks
• merging would force all harness-plan users to load engineering's 3500-token lifecycle body on every invocation
• "Compose, do not embed" is principle #8

Where harness-plan is expected:

• ~/.claude/skills/harness-plan/SKILL.md, or
• ~/.codex/skills/harness-plan/SKILL.md, or
• any installed plugin providing harness-plan/skills/harness-plan/SKILL.md

---

Install

Option A: Claude Code

Install harness-plan first, then harness-engineering:

/plugin marketplace add suntao2yl/claude-skill-harness
/plugin install harness-plan@suntao-skills

/plugin marketplace add suntao2yl/claude-skill-engineering
/plugin install harness-engineering@harness-engineering-marketplace

Option B: Codex

Install harness-plan first, then harness-engineering:

python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
  --repo suntao2yl/claude-skill-harness \
  --path plugins/harness-plan/skills/harness-plan

python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py \
  --repo suntao2yl/claude-skill-engineering \
  --path skills/harness-engineering

Restart Codex after installation so both skills appear in the skill list.

Option C: manual (repo clone)

git clone https://github.com/suntao2yl/claude-skill-engineering.git && cd claude-skill-engineering

# install engineering only (checks for harness-plan, warns if missing):
./install.sh

# install engineering + harness-plan from a local source:
./install.sh --with-harness-plan /path/to/harness-plan-skill

# custom skills directory:
./install.sh --prefix /custom/skills/path

Quick start

# after installing both skills:

# in Claude Code:
/harness-engineering init "Build a Python CLI that counts lines in a file, emit JSON, ship v0.1"

# Claude will:
# 1. create .engineering/ with 7 phase subdirs + seed REQ-001
# 2. auto-drive: fill discovery → advance → pause at risk gate → await --confirm
# 3. fill design + architecture in parallel → advance each → pause at arch gate
# 4. enter implementation → delegate to harness-plan → drive features to done
# 5. advance implementation (live-executes verification commands)
# 6. fill test-report → advance (live-executes test commands)
# 7. create git tag → advance release → pause at release gate
# 8. fill ops metrics → terminal

The user only confirms at risk gates (discovery / architecture / release approval). Everything else auto-advances.

---

The 7 phases