# Elegance Pipeline

> Run a persistent multi-agent workflow scoring source-file code quality: 4 scouts, 2 judges, a planner, a verifier, and 1 gated implementer. Use when you want to rank how clean a repo's files are and convert the weakest into gated refactor work. Triggers on "find ugly files", "review code quality".

- Skill: `ancplua/elegance-pipeline` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ancplua/elegance-pipeline`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ancplua/elegance-pipeline/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: ANcpLua (https://skillmd.com/u/ancplua)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/ancplua/elegance-pipeline

---

# Elegance Pipeline

Multi-agent code-elegance workflow with persistent state and stage gates.

The orchestrator runs in Python (`pipeline.py`); the scoring is language-agnostic
over the target repo, anchored by whatever root file fits it — `CLAUDE.md`,
`package.json` for a Node.js project, or `*.sln` for a .NET solution.

## Pipeline stages

```text
4 Scouts (parallel, opus) -> 2 Judges (parallel, opus) -> 1 Planner (opus)
  -> 1 Verifier (opus) -> 1 Implementer (gated, opus)
```

## State manager

All orchestration goes through the pipeline state manager:

```bash
python ${CLAUDE_PLUGIN_ROOT}/elegance_pipeline/pipeline.py [--state-dir <dir>] <command>
```

Commands: `init`, `status`, `prompt`, `submit`, `signal`

Default state lives at `.claude/elegance_pipeline/state/` and is shared by default.
If you want one dedicated team per spec, give each run its own `--state-dir`.

## How to orchestrate

1. Run `status` to see which slots are ready
2. For each ready slot, run `prompt --role <role> --slot <slot>` to get the task prompt
3. Spawn the appropriate fully qualified subagent (`elegance-pipeline:elegance-scout`,
   `elegance-pipeline:elegance-judge`, `elegance-pipeline:elegance-planner`,
   `elegance-pipeline:elegance-verifier`, or `elegance-pipeline:elegance-implementer`).
   If your runtime exposes a different fully qualified name, use that exact identifier instead of the bare short name.
4. After each agent completes, submit its output via the state manager
5. Run `status` again to see what unlocked
6. Repeat until the pipeline is complete or blocked

## Orchestration rules

- Scouts run in parallel (all 4 at once via background agents)
- Judges run in parallel after ALL scouts are submitted
- Planner runs after BOTH judges are submitted
- Verifier runs after the planner is submitted
- Implementer only runs when the implementation signal is READY
- Never bypass stage gates
- If the verifier says no implementation is warranted, stop

## Agent model mapping

| Role | Agent | Model | Access |
|------|-------|-------|--------|
| Scout | elegance-pipeline:elegance-scout | opus | read-only |
| Judge | elegance-pipeline:elegance-judge | opus | read-only |
| Planner | elegance-pipeline:elegance-planner | opus | read-only |
| Verifier | elegance-pipeline:elegance-verifier | opus | read-only |
| Implementer | elegance-pipeline:elegance-implementer | opus | full edit |

## Manual signal override

```bash
python ${CLAUDE_PLUGIN_ROOT}/elegance_pipeline/pipeline.py [--state-dir <dir>] signal on
python ${CLAUDE_PLUGIN_ROOT}/elegance_pipeline/pipeline.py [--state-dir <dir>] signal off
```