360 Secure SkillHub
360 安全技能中心

dorodango

Dorodango 是代码打磨工作流:测试通过后依次做正确性、清晰性、一致性、打磨四个维度的 pass,每轮由独立子代理执行,0 问题即该维度收敛,状态写入 .attune/dorodango-state.json 支持跨会话续跑,最多 10 轮。

athola
athola
热度 63.59
最近更新 2026-08-26 21:07:32
分类:开发依赖:代码执行 / 读写文件 / 内部调用大模型应用场景:代码审查 / 调试排错 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add nm-attune-dorodango
Metadata
name
dorodango
description
Polishes working code through successive quality passes in fresh subagents. Use after tests pass when code needs multi-dimension refinement before release
version
1.9.8
triggers
["polishing","iterative-refinement","code-quality","convergence"]
metadata
{"openclaw":{"homepage":"https://github.com/athola/claude-night-market/tree/master/plugins/attune","emoji":"🦞"}}
source
claude-night-market
source_plugin
attune

Night Market Skill — ported from claude-night-market/attune. For the full experience with agents, hooks, and commands, install the Claude Code plugin.

Dorodango Polishing Workflow

Named after the Japanese art of polishing a ball of dirt into a high-gloss sphere. Applied to code: take the initial implementation (the "mud ball") and refine it through successive quality passes until it shines.

When To Use

  • After initial implementation is complete and tests pass
  • Code works but needs refinement across multiple quality dimensions
  • Preparing code for review or release
  • Resuming a previous polishing session

When NOT To Use

  • Code does not compile or pass basic tests (fix first)
  • Single-dimension improvement needed (use the specific skill directly: pensive:code-refinement, etc.)
  • Greenfield design phase (use brainstorming instead)

Pass Sequence

Four quality dimensions, each a self-contained pass:

  1. Correctness - run tests, fix failures
  2. Clarity - code readability and structure
  3. Consistency - naming, patterns, style alignment
  4. Polish - documentation, error messages, edges

See modules/pass-definitions.md for detailed scope of each pass type.

Convergence Model

  • Each pass targets one dimension
  • A pass that finds issues_found: 0 marks that dimension as converged
  • Convergence is irreversible per run; a converged dimension is not re-run
  • When all 4 dimensions converge, polishing is complete
  • Maximum 10 total passes (hard limit)
  • If not converged after 10 passes, surface state to human with recommendation to split into smaller units

State Persistence

State tracked in .attune/dorodango-state.json:

{
  "target": "plugins/foo",
  "started_at": "2026-03-18T12:00:00Z",
  "pass_count": 3,
  "passes": [
    {
      "type": "correctness",
      "issues_found": 2,
      "issues_fixed": 2
    },
    {
      "type": "clarity",
      "issues_found": 5,
      "issues_fixed": 5
    },
    {
      "type": "consistency",
      "issues_found": 0
    }
  ],
  "converged_dimensions": ["consistency"],
  "converged": false
}

This file enables resume across sessions. On resume, skip converged dimensions and continue from the next unconverged dimension.

Subagent Isolation

Each pass dispatches a self-contained subagent to prevent context accumulation. The subagent receives:

  • Target directory/files
  • Pass type and scope (from pass-definitions module)
  • Previous pass results (summary only, not full context)

Subagent dispatch is optional for targets under 100 lines of code; in-session review is sufficient for small files.

Workflow

  1. Initialize state file (or load existing)
  2. Determine next unconverged dimension
  3. Dispatch subagent for that dimension
  4. Record results in state file
  5. If dimension converged (0 issues), mark it
  6. If all dimensions converged or 10 passes reached, stop
  7. Otherwise, proceed to next dimension

Cross-References

  • pensive:code-refinement - used in clarity pass
  • conserve:code-quality-principles - KISS/YAGNI/SOLID
  • imbue:latent-space-engineering - frame pass prompts with emotional framing for better results