| name | yui-dev-workflow |
| description | Use when writing, changing, reviewing, testing, or delegating any code in the YUI repo — any feature, bugfix, refactor, UI, schema, or chore beyond a trivial single-file edit. |
YUI Development Workflow
How development happens in YUI: the mandatory work rules, delegation decisions, the review and verification gates, and the client-side anti-patterns. Load this before touching code. Project identity and the on-demand doc index stay in AGENTS.md.
Before you start
REQUIRED SUB-SKILL: load karpathy-guidelines — surface assumptions, keep changes surgical, define verifiable success criteria before writing code.
Work Rules (user directive, mandatory)
- Worktree → PR. All work happens in a git worktree and lands via PR;
main requires a PR and green CI (PreToolUse(Bash) guard denies git commit/git push on main — the agent cannot commit/push to main and must request the user to run it directly). New-worktree setup: run bash scripts/worktree-setup.sh <worktree> after a manual git worktree add (Claude-created worktrees run this automatically via WorktreeCreate hook).
- GitHub tracker in English. Issues, issue comments, and PR titles/bodies are written in English (chat with the user is any language); enforced by the
pr-title CI job.
- UI: review existing → propose text structure → mock HTML → implement. Read
src/ui/, DESIGN.md, PRODUCT.md before any UI work; propose structure, get confirmation, create a standalone mock HTML, then implement (detail: docs/agent-guide/design-context.md).
- Tests accompany behavior. New or changed behavior ships its test in the same PR; the
test-guard CI job enforces this (skip-tests label bypasses). Write the failing test first (test:), then implementation (feat:), then refactor if needed (refactor:).
- Verify what you can verify before asking the user. Anything observable (UI rendering / DOM state / logs) — verify yourself and attach proof to the PR's Runtime-evidence section; ask the user only for things that genuinely require them (audio playback, physical input feel).
- Comments: minimal, present-tense only. Comment only what the code cannot say itself, in one line; no decision-history, spec-citation, or issue-number breadcrumbs.
- Docs: current-state only. Write what the system is, declaratively, matching the code — no change-narrative, no PR/issue numbers as prose, no dated changelogs, no future/unbuilt work (
PostToolUse docs guard enforces this).
Main Agent Role
- Requirements clarification — understand user intent and define work scope
- Task delegation — distribute work to sub-agents (ensure failing tests precede implementation — TDD ordering)
- Integration verification — confirm
pnpm test + pnpm typecheck:test + cargo test + pnpm build + pnpm lint pass (vitest runs test files without typechecking them; only typecheck:test catches stale fixtures)
- Orchestration — manage task ordering and dependencies
- Chore work - You can directly commit and push to the main branch if the work is chore work. chore work means: 3~4 files edit and less than 100 line edit. In this case you don't need to worktree either. just edit, stage, then report back to user.
When to delegate
Implementation work goes to general-purpose sub-agents (model: opus) driven by a written spec — the prompt carries the target paths, the rules that apply, and the TDD ordering. Two project agents are gates rather than implementers: Code Reviewer reviews every non-trivial diff, and Reality Checker certifies runtime/UI evidence.
Default by rule; ask only on the boundary.
- Delegate to a sub-agent if ANY: adds/changes behavior (needs TDD), or spans multiple files/steps.
- Delegate to an external CLI agent (codex/opencode via
agent-send) — load the agent-team:agent-team skill first; roles, session naming, and the spec→review→implement→verify cycle live there. External agents see none of this conversation, so the prompt must carry the YUI rules that apply (worktree→PR, TDD ordering, English tracker) and absolute paths.
- Do it directly if ALL: no new behavior (typo·doc·comment·rename·config), ~20 lines or fewer, single file, mechanical.
- Genuinely borderline? Ask the user one line: "이거 ○○ 작업인데 직접 할까요, 서브에이전트로 위임할까요?"
Anti-patterns (do not do)
- No brain in the client. Judgment / persona state / mode branching belongs to the backend.
- No inline control tags. Emotion/motion goes through
generate_express tool-call arguments only — not inline tokens in speech text.
- No unverified assumptions. Consult docs first. If not in docs, record the decision in docs before implementing.
- No hardcoding. Endpoints/models/VRM paths/motion sets go in
configs/.
