Novera-AI-pipeline

four-agent-pipeline — an agent skill that turns one vague requirement into merged, gate-verified code through four specialized agents that keep each other honest: Specifier → Coder → Refactorer → Architect.

Core philosophy: don't review the process — gate the outcome. The human reviews exactly one artifact (the acceptance spec) and trusts a welded-shut set of automated quality gates for everything else.

The Pipeline

Agent	Role	Does	Never does
1. Specifier	Requirements officer	Compiles vague requests into Gherkin acceptance criteria + boundary conditions + DoD	Write a single line of implementation
2. Coder	Implementation officer	Turns every acceptance and unit test green with a minimal implementation	Over-engineer, or sneak in out-of-scope features
3. Refactorer	Structure officer	Improves structure without changing behavior; proves test strength with mutation testing	Change business logic under the name of refactoring
4. Architect	Gatekeeper (veto power)	Emits per-gate PASS/FAIL with measured values; routes failures back upstream	Rewrite code — it only adjudicates

Failures never stall on a human: every FAIL carries the measured value vs. threshold and routes back to the responsible agent automatically (references/failure-routing.md).

The Steel Cage (quality gates)

Gate	Default threshold
Acceptance + unit tests	100% pass
Coverage	≥ 85% line / 75% branch
Mutation score	≥ 80%
Cyclomatic complexity	≤ 10 per function
Duplication	≤ 3%
Dependency direction / lint / security	0 violations / 0 errors / 0 high-critical

Every gate is binary. "Close enough" is a FAIL. Full tool matrix per ecosystem: references/steel-cage.md.

Install

The repository is the skill — clone it into your agent's skills directory:

# Claude Code
git clone https://github.com/whaojie797-design/Novera-AI-pipeline.git ~/.claude/skills/four-agent-pipeline

Then invoke it with phrases like "run the four-agent pipeline on this requirement" or "compile this feature request into acceptance criteria".

Usage

# 1. Scaffold a workspace (optional but handy)
python scripts/init_pipeline.py my-feature --feature user-login

# 2. Feed the requirement to the Specifier (references/specifier.md)
# 3. Human signs off the spec  -> spec is frozen
# 4. Coder (references/coder.md)            -> tests 100% green
# 5. Refactorer (references/refactorer.md)  -> steel-cage metrics met
# 6. Architect (references/architect.md)    -> PASS merges, FAIL routes back

Works with Claude Code subagents, CrewAI, MetaGPT-style orchestrators, or four separate chat sessions with the same model.

Repository Layout

Novera-AI-pipeline/
├── SKILL.md                      # skill entry point (progressive disclosure)
├── references/
│   ├── specifier.md              # Agent 1 full system prompt
│   ├── coder.md                  # Agent 2 full system prompt
│   ├── refactorer.md             # Agent 3 full system prompt
│   ├── architect.md              # Agent 4 full system prompt
│   ├── steel-cage.md             # gate thresholds + tools per ecosystem
│   └── failure-routing.md        # who-fails-goes-back-to-whom
├── assets/templates/             # spec / feature / reports / gate-report / objection
├── scripts/init_pipeline.py      # workspace scaffolder (Python stdlib only)
└── LICENSE                       # MIT

Related

• Novera-AI-skills — eight production-ready agent skills with zero-dependency scripts
• Novera-AI-agent — the original Chinese-language prompt document this skill operationalizes

License

MIT — copy, modify, and use commercially with attribution.