harness

Harness — Router (BE / FE)

This skill is a thin router. It scans the project to decide whether it is a backend or a frontend, then delegates the actual harness build to the harness-be (backend) or harness-fe (frontend) sub-skill.

Previously this skill produced the core 6 agents directly, but that baked backend assumptions like NestJS/DDD/Hex into the core agent prompts. Detection now happens first, and the detected project type plus concrete evidence (framework / architecture / ORM / test stack) is injected into the agent prompts.

Language Policy

All user-facing output must be in Korean. This file is written in English for token efficiency (Claude reads it as instructions), but anything the end user sees — AskUserQuestion text, progress messages, error messages, final summaries, CLAUDE.md blocks written into the project — must be rendered in natural Korean.

• Claude-directed instructions in this file: English.
• Embedded Korean strings marked <!-- user-facing -->: literal templates, never translate.
• Frontmatter description Korean triggers: literal, never translate.
• When you narrate progress to the user or write a summary, translate to natural Korean first — do not surface English reasoning text directly.

Core Principles

1. Detect project type first — read backend / frontend signals from package.json or other manifest files.
2. Ask the user when ambiguous — full-stack projects require the user to declare intent.
3. Delegate to a sub-skill once detected — actual harness build is owned by harness-be or harness-fe.
4. Both paths can be offered — for full-stack projects, run BE first, then FE in sequence.

Invocation Patterns

Input	Behavior
/oh-my-harness:harness	Auto-detect → delegate to sub-skill
/oh-my-harness:harness {feature}	Detect + delegate + work on the feature
/oh-my-harness:harness-be / /oh-my-harness:harness-fe (direct invocation)	Bypass router, run the sub-skill directly

Workflow

Phase 0: Check for existing harness

Read the project's .claude/agents/ and CLAUDE.md. If a <!-- harness-fingerprint v1 --> block exists, you can already tell which side it is:

• skill: harness-be present → re-invoke harness-be (maintenance mode)
• skill: harness-fe present → re-invoke harness-fe (maintenance mode)
• Both present → ask the user which one to maintain and call the matching sub-skill

If no fingerprint is found, proceed to Phase 1.

Phase 1: Project type detection

Collect backend and frontend signals from the project root.

Backend signals:

Signal	Weight
package.json contains @nestjs/*, express, fastify, koa, hono, @trpc/server	+3
pyproject.toml / requirements.txt contains fastapi, flask, django, starlette	+3
go.mod exists + gin / echo / fiber import	+3
build.gradle(.kts) / pom.xml contains spring-boot, micronaut, quarkus	+3
Gemfile contains rails	+3
composer.json contains laravel/framework	+3
nest-cli.json exists	+2
migrations/, prisma/schema.prisma, ormconfig.* exist	+1
controller/ / service/ / repository/ directories repeat	+1

Frontend signals:

Signal	Weight
package.json contains react + react-dom (without next, remix, gatsby)	+3
package.json contains vue (and no Nuxt)	+3
package.json contains @angular/core + angular.json	+3
package.json contains svelte (and no SvelteKit)	+3
package.json contains solid-js	+3
next + next.config.*	+4 (full-stack signal)
nuxt + nuxt.config.*	+4 (full-stack signal)
@sveltejs/kit + svelte.config.*	+4 (full-stack signal)
@remix-run/*	+4 (full-stack signal)
vite.config.* configures a React/Vue/Svelte plugin	+2
src/components/, components/, app/pages/ etc. component directories	+1
.storybook/ exists	+1

Phase 2: Branch decision

Sum the scores and decide.

Situation	Action
BE score ≥ 3 AND FE score < 3	Delegate to harness-be
FE score ≥ 3 AND BE score < 3	Delegate to harness-fe
BE score ≥ 3 AND FE score ≥ 3 (full-stack)	Confirm via AskUserQuestion (see below)
BE score < 3 AND FE score < 3 (insufficient signal)	Empty project — ask the user about intent

Full-stack confirmation prompt:

이 프로젝트는 백엔드와 프론트엔드 시그널이 모두 감지되었습니다.
어떤 관점으로 하네스를 구축할까요?

[1] Backend 먼저 (harness-be) — 이후 harness-fe로 확장 가능
[2] Frontend 먼저 (harness-fe) — 이후 harness-be로 확장 가능
[3] 둘 다 순차 실행 — harness-be → harness-fe 연달아 (권장: 진짜 풀스택 모노레포)
[4] 이 번에는 {하나만} — 나머지는 나중에

Phase 3: Sub-skill delegation

Run the sub-skill matching the chosen path. Use the Skill tool to invoke one of:

• skill: "oh-my-harness:harness-be" (forward the user's feature argument if any)
• skill: "oh-my-harness:harness-fe" (same)

When the user picks "both" in full-stack: run harness-be first. After harness-be completes, run harness-fe. harness-fe's Phase 0 will detect the existing harness-be fingerprint and operate in "second-harness extension mode" — it will not overwrite the existing core 6 agents and leaders, and will only add frontend-specific agents (ui-reviewer, a11y-auditor, etc.) and skills. The fingerprint block records both skills.

<!-- harness-fingerprint v1 -->
skill: harness-be, harness-fe  # when both have been run
...
<!-- /harness-fingerprint -->

Phase 4: Empty project handling

If BE score < 3 AND FE score < 3 AND there are almost no source files, treat it as a brand-new project. AskUserQuestion:

프로젝트가 비어있거나 시그널이 부족합니다. 어떤 종류의 프로젝트를 만들고 싶으신가요?

[1] Backend (NestJS, FastAPI, Spring, Go 등) → harness-be
[2] Frontend (React SPA, Next.js, Vue/Nuxt 등) → harness-fe
[3] 풀스택 (둘 다) → harness-be + harness-fe 순차
[4] 나중에 — 지금은 하네스 빌드 안 함

Run the chosen sub-skill based on the answer. In an empty project the sub-skill may follow up with a tech-stack question to the user (e.g., "NestJS / Express / Fastify?").

---

Execution constraints

• This router skill must not generate the core 6 agents or the leaders directly. Agent generation belongs entirely to the sub-skills.
• <Project_Context> injection logic, detection trees, and prompt templates all live under the sub-skills' references/.
• The router's decision logic only chooses which sub-skill receives control — the criteria are defined here, but detailed detection is repeated by the sub-skill itself.

Routing matrix summary

Detection result	Delegate to
Backend signals dominant	harness-be
Frontend signals dominant	harness-fe
Full-stack — user picks "BE first"	harness-be → optional follow-up harness-fe
Full-stack — user picks "FE first"	harness-fe → optional follow-up harness-be
Full-stack — user picks "both"	harness-be → harness-fe in sequence
Empty project — user picks "Backend"	harness-be
Empty project — user picks "Frontend"	harness-fe
Empty project — user picks "later"	Exit (no harness build)

References

• Backend harness builder: ../harness-be/SKILL.md
• Frontend harness builder: ../harness-fe/SKILL.md
• Shared references (repeated module detection, domain term extraction, agent design patterns): references/ in this directory

State of the references/ directory

The existing reference documents in this directory are still imported and used by the sub-skills:

• references/model-selection-protocol.md — sequential model-selection flow for the core 6 agents — must be Read by harness-be / harness-fe in Phase 2
• references/project-analysis-protocol.md — repeated module detection, domain term extraction (§3) — reused by harness-be / harness-fe detection protocols
• references/agent-design-patterns.md — agent design principles
• references/skill-generation-guide.md, references/skill-writing-guide.md — skill authoring guides
• references/orchestrator-template.md — orchestrator template
• references/harness-examples.md — harness examples
• references/qa-agent-guide.md — QA agent guide

These files are not used directly by this router skill, but harness-be and harness-fe import them, so they must not be deleted.