Front Agent

Skill: front-agent

Trigger: /front-agent [request] Purpose: Handle setup, intent detection, planning, execution, and verification with minimal context overhead.

Canonical Rules

• This file is the canonical runtime workflow
• README.md is documentation only and should not be loaded during execution
• front-agent never implements directly
• Use only active merged agents: component-auditor, developer, ui-builder, api-integrator, test-runner, reviewer, refactor-architect
• Do not use legacy agents: implementer, test-writer, figma-builder, style-matcher, orchestrator

Request Gate

0. RTK Mode Gate

Resolve the rtk token-filter mode before any agent spawn. This is plugin-scoped — no global Claude Code hook is installed, so every other project is unaffected.

Resolution order:

1. Inline arg — if the user message contains --rtk=<mode> or --no-rtk, persist it with bash hooks/rtk-wrap.sh --set <mode> (use off for --no-rtk). Skip UI.
2. Env var — if FE_COPILOT_RTK is set, do nothing. It is a hard override.
3. Session flag — if .fe-copilot-cache/rtk-session.flag already exists, reuse it.
4. Otherwise — launch the AskUserQuestion tool with a single question:
   • Question: 이 작업에서 rtk로 명령 출력을 압축할까요?
   • Options:
     • Off — raw 명령 그대로 (기본 동작)
     • On — standard (git/tsc/lint/test 압축) (Recommended)
     • On — aggressive (+ ultra-compact)
     • On — git-only (git 계열만 압축)
   • After answer, call bash hooks/rtk-wrap.sh --set <mode> where mode is off | standard | aggressive | git-only.

Valid mode strings: off, standard, aggressive, git-only. Anything else must be rejected with a one-line error.

The user can re-run /rtk at any time to change the mode mid-session.

1. Lazy Setup

If knowledge/index.md is missing, initialize project knowledge automatically. Do not ask the user to run /setup.

2. Ambiguity Check

Ask a clarification question before planning if any of the following apply:

• The target is unclear
• Scope is overly broad
• The request mixes multiple major tasks
• "Optimize" or "improve" does not specify what dimension matters

Rules:

• Ask at most 2 questions
• Ask only for missing information that blocks planning

3. Intent Classification

4. Figma Check

If intent is ui and no Figma URL is present, ask for one in the user's language.

• URL provided -> figma
• No URL -> ui

5. API Spec Check

Trigger: intent is feature AND request contains any API keyword: fetch, axios, API, GraphQL, query, mutation, Apollo, WebSocket, endpoint, 서버, 백엔드, 연결, 데이터, 불러와, 저장, 실시간

If triggered, ask once in the user's language:

"Do you have an API spec? (Swagger/OpenAPI URL, GraphQL schema, .md/.yaml file, or paste the endpoint directly)"

Token budget rules:

• Pass API spec only to api-integrator — never to developer, ui-builder, or other agents
• If a URL was provided, api-integrator fetches only the relevant endpoint(s) on demand — do not load the full document upfront

6. Review Fast-Path

If intent is review, skip the Plan Gate entirely and spawn reviewer directly on the current changes. No plan.md, no user approval wait — pure review tasks do not need a plan. The reviewer verdict (PASS/FAIL) is the final output; no commit/push follows.

Applies only when intent is exactly review. Any other intent still goes through the Plan Gate below.

7. Plan Gate

Create plan.md, then wait for explicit user approval before any execution.

Use this format:

markdownCopy code
# Plan: [task name]

## Goal
[what and why]

## Intent
[figma / ui / feature / refactor / review]

## Inputs
- Figma: [URL or none]
- API: [spec, endpoint, or none]

## Affected Files
- path/to/file.tsx - [change]

## Design Check
- Reuse: [component/hook/util to reuse, or "none — new primitive justified because X"]
- Responsibility: [how logic splits across files — e.g. "fetch in hook, render in component, format in util"]
- Risk: [conflict with an existing pattern or convention, or "none"]

## Execution Steps
- Step 1: [agent] - [task]
- Step 2: [agent] - [task]

## Branch
[feat|fix|ui|refactor]/[task-name]

## Commit Units
- [type]: [description]

Design Check rules:

• Required for ui, figma, feature, refactor. Omit for review.
• Max 1 line per bullet. No prose paragraphs.
• If Reuse says "new primitive", the justification must be one concrete reason — not a vague "for flexibility".
• Authoring order:
  1. At plan creation, fill Reuse / Responsibility / Risk from a quick scan + project knowledge.
  2. After component-auditor runs, update the Reuse line in place with the concrete candidates it surfaced. This is a small in-place edit, not a new section.
  3. developer / ui-builder / api-integrator read the updated plan.md as their canonical Design Check.
• Do not duplicate the audit list. The Reuse line is the single source of truth.
• This section exists to catch design mistakes BEFORE implementation. It is read by the user during plan approval and re-read by reviewer during the design category check.

Ask: Please review plan.md. Approve it and I will execute.

Context Budget

For each agent call, include only:

1. The relevant agents/*.md file
2. The relevant constraints.md sections — obtained via bash hooks/extract-constraints.sh <agent-name>. Do not pass the full file.
3. The minimal plan.md excerpt needed for that step
4. A compact handoff block
5. Only the exact files needed for execution

Compact Handoff Schema

markdownCopy code
## Handoff
- changed_files: path1, path2
- reusable_components: Button, Card
- decisions: keep existing filter sidebar pattern
- blockers: failing `cart.spec.ts`
- test_status: not_run | passed:npm run test | failed:npm run test

Rules:

• Omit empty fields
• Max 5 bullets
• One line per bullet
• No prose summaries
• Prefer paths, identifiers, and commands

Workflow Selection

|| = run in parallel — launch both agents simultaneously and wait for both before proceeding.

Parallel review gate: [reviewer || codex-review] launches both reviewers concurrently on the same changed files. git-commit proceeds only if BOTH return PASS (or the user explicitly overrides codex-review). Any FAIL blocks commit and its fix list feeds the retry loop.

Shared Skip Rules

• Skip search-knowledge if stored knowledge is empty or irrelevant. Run bash hooks/knowledge-has-content.sh; exit != 0 means placeholder-only — skip the agent spawn entirely and proceed.
• component-auditor is MANDATORY for any intent that creates or modifies UI: ui, figma, and any feature whose Affected Files include .tsx/.jsx components, hooks, or styles. Skipping in these cases is a workflow violation.
• Skip component-auditor only when ALL of the following hold:
  • intent is review (review fast-path), OR
  • intent is feature AND the change touches zero component/hook/style files (pure service/util/API wiring)
• Skip api-integrator unless UI data fetching or mutation behavior changes
• Skip save-knowledge if the task produced no durable learning

Feature

Copy code
[search-knowledge? || component-auditor?] -> developer
-> test-runner -> api-integrator? -> [reviewer || codex-review]
-> git-branch -> git-commit -> git-pr -> save-knowledge?

Figma

Copy code
[search-knowledge? || component-auditor] -> ui-builder
-> [pixel-check || a11y-check] -> [reviewer || codex-review]
-> git-branch -> git-commit -> git-pr -> save-knowledge?

UI Without Design

Copy code
[search-knowledge? || component-auditor] -> ui-builder
-> a11y-check -> [reviewer || codex-review]
-> git-branch -> git-commit -> git-pr -> save-knowledge?

Refactor

Copy code
search-knowledge? -> refactor-architect -> user re-approval
-> component-auditor? -> developer -> test-runner -> [reviewer || codex-review]
-> git-branch -> git-commit -> git-pr -> save-knowledge?

Review

Fast-path: bypasses the Plan Gate (see Request Gate step 6).

Copy code
reviewer

Harness Loop

Use the retry loop for feature, ui, figma, and refactor implementation steps.

textCopy code
MAX_ATTEMPTS = 3

run implementation agent
run test-runner when applicable

if tests fail:
  retry with only the failing error context

if attempts exceed MAX_ATTEMPTS:
  create a git issue
  record a repeatable failure pattern if appropriate
  stop and report

Retry context must contain only:

• failing command
• failing file/test identifier
• shortest useful error excerpt

Never resend the full previous transcript.

Completion

• Never declare completion without test evidence when tests are applicable
• Never move to git-commit without reviewer PASS
• Save knowledge only if a reusable pattern, decision, or repeatable issue emerged