DeepSeek HarnessHub
플러그인 제출

zevorn/dsh-humanize ↗★ 2

@humanize/dsh-humanize

Humanize RLCR bundle for the DeepSeek Harness: DSH skills, Codex review, and the Humanize trajectory view.

AI 분석

核心用途是运行 RLCR 闭环进行代码优化，并由 Codex 进行独立审查。适合追求高质量代码生成与轨迹回放的开发任务。

skills
패키지
@humanize/dsh-humanize
버전
1.18.0
최근 업데이트
2026. 8. 13.

설치

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:zevorn/dsh-humanize

문서

전체 README 읽기 ↗

Install


# Then install humanize plugin
/plugin install humanize@PolyArch

Requires codex CLI for review. See the full Installation Guide for prerequisites and alternative setup options.

DeepSeek Harness

Humanize is also available as a standard DeepSeek Harness profile bundle. The DeepSeek V4 Flash Max builder agent runs the RLCR loop inside a DSH session while the Codex review agent independently gates progress. The bundle registers these skills: humanize, humanize-rlcr, ask-codex, humanize-gen-plan, and humanize-refine-plan; it also mounts the Humanize trajectory view in the latest DSH web client.


# Install the standard bundle into the web profile.
dsh plugin --profile web add github:dsh-external/dsh-humanize#

Git installs build the web client through the bundle's prepare script. If pnpm blocks that build, add the exact package key it prints to $DSH_HOME/profiles/web/pnpm-workspace.yaml under allowBuilds, then rerun the command. Configure the builder model (deepseek-v4-flash-max) in the DSH model settings — the full walkthrough is in the Installation Guide for DeepSeek Harness.

Quick Start

  1. Generate an idea draft from a loose thought (optional — skip if you already have a draft):

    /humanize:gen-idea "add undo/redo to the editor"

    Output goes to .humanize/ideas/-.md and a companion directions.json artifact. Pass a .md path to expand existing rough notes. --n controls how many parallel directions explore the idea (default 6).

  2. Explore directions as parallel prototypes (optional — skip if you want to go straight to planning):

    /humanize:explore-idea .humanize/ideas/-.directions.json

    Dispatches bounded parallel prototype workers (one per direction), each running in an isolated git worktree. After all workers complete, writes .humanize/explore//explore-report.md for audit/ranking details and .humanize/explore//final-idea.md as the plan-ready synthesis. Worker worktrees are optional prototype fast paths; the default follow-up is to generate a clean plan from final-idea.md.

  3. Generate a plan from your draft or explored final idea:

    /humanize:gen-plan --input .humanize/explore//final-idea.md --output docs/plan.md

    Add --coach to run mandatory short-answer stage quizzes after each planning stage. Normal plan decision questions stay separate; quiz mismatches are treated as design drift, AI design correction, or background gaps before the agent expands the next planning layer. Generated plans include a Feature Map / Capability Map before the task breakdown so each task carries its global capability context.

  4. Refine an annotated plan before implementation when reviewers add comments (CMT: ... ENDCMT, ..., or ...):

    /humanize:refine-plan --input docs/plan.md

  5. Run the loop:

    /humanize:start-rlcr-loop docs/plan.md

    When the plan has a capability map, RLCR records a Capability Anchor in each round contract and Goal Tracker active task so Claude coding and Codex review stay aligned with the map.

  6. Consult Gemini for deep web research (requires Gemini CLI):

    /humanize:ask-gemini What are the latest best practices for X?

  7. Monitor progress (in another terminal, not inside Claude Code):

    source

/scripts/humanize.sh # Or just add it into your .bashec or .zshrc humanize monitor rlcr # RLCR loop humanize monitor skill # All skill invocations (codex + gemini) humanize monitor codex # Codex invocations only humanize monitor gemini # Gemini invocations only


## Documentation

- [Usage Guide](https://github.com/zevorn/dsh-humanize/blob/f39c19d5a596735c9dfaa1cbb0b95c02152b039c/docs/usage.md) -- Commands, options, environment variables
- [Install for Claude Code](https://github.com/zevorn/dsh-humanize/blob/f39c19d5a596735c9dfaa1cbb0b95c02152b039c/docs/install-for-claude.md) -- Full installation instructions
- [Install for Codex](https://github.com/zevorn/dsh-humanize/blob/f39c19d5a596735c9dfaa1cbb0b95c02152b039c/docs/install-for-codex.md) -- Codex skill runtime setup
- [Install for Kimi](https://github.com/zevorn/dsh-humanize/blob/f39c19d5a596735c9dfaa1cbb0b95c02152b039c/docs/install-for-kimi.md) -- Kimi CLI skill setup
- [Configuration](https://github.com/zevorn/dsh-humanize/blob/f39c19d5a596735c9dfaa1cbb0b95c02152b039c/docs/usage.md) -- Shared config hierarchy and override rules
- [Bitter Lesson Workflow](https://github.com/zevorn/dsh-humanize/blob/f39c19d5a596735c9dfaa1cbb0b95c02152b039c/docs/bitlesson.md) -- Project memory, selector routing, and delta validation

## License

MIT

## Quick Start

1. **Generate an idea draft** from a loose thought (optional — skip if you already have a draft):
```bash
/humanize:gen-idea "add undo/redo to the editor"

Output goes to .humanize/ideas/-.md and a companion directions.json artifact. Pass a .md path to expand existing rough notes. --n controls how many parallel directions explore the idea (default 6).

  1. Explore directions as parallel prototypes (optional — skip if you want to go straight to planning):

    /humanize:explore-idea .humanize/ideas/-.directions.json

    Dispatches bounded parallel prototype workers (one per direction), each running in an isolated git worktree. After all workers complete, writes .humanize/explore//explore-report.md for audit/ranking details and .humanize/explore//final-idea.md as the plan-ready synthesis. Worker worktrees are optional prototype fast paths; the default follow-up is to generate a clean plan from final-idea.md.

  2. Generate a plan from your draft or explored final idea:

    /humanize:gen-plan --input .humanize/explore//final-idea.md --output docs/plan.md

    Add --coach to run mandatory short-answer stage quizzes after each planning stage. Normal plan decision questions stay separate; quiz mismatches are treated as design drift, AI design correction, or background gaps before the agent expands the next planning layer. Generated plans include a Feature Map / Capability Map before the task breakdown so each task carries its global capability context.

  3. Refine an annotated plan before implementation when reviewers add comments (CMT: ... ENDCMT, ..., or ...):

    /humanize:refine-plan --input docs/plan.md

  4. Run the loop:

    /humanize:start-rlcr-loop docs/plan.md

    When the plan has a capability map, RLCR records a Capability Anchor in each round contract and Goal Tracker active task so Claude coding and Codex review stay aligned with the map.

  5. Consult Gemini for deep web research (requires Gemini CLI):

    /humanize:ask-gemini What are the latest best practices for X?

  6. Monitor progress (in another terminal, not inside Claude Code):

    source

/scripts/humanize.sh # Or just add it into your .bashec or .zshrc humanize monitor rlcr # RLCR loop humanize monitor skill # All skill invocations (codex + gemini) humanize monitor codex # Codex invocations only humanize monitor gemini # Gemini invocations only