orchestrator

name

orchestrator

description

Coordinate the EveryGSM-client-v2 agent team for frontend features, bug fixes, refactors, UI work, API/data work, QA, code review, debugging, branch names, actual commits, PR creation, PR review comment replies, routing, forms, TanStack Query, Zod, Axios, and FSD architecture. Use this skill for EveryGSM implementation tasks, reviews, reruns, updates, revisions, partial reruns, follow-up fixes, Git workflow tasks, and requests to improve previous results. Simple one-off questions may be answered directly.

EveryGSM Orchestrator

Coordinate the EveryGSM-client-v2 harness. This skill routes work to the right specialists and keeps the implementation small, verifiable, and aligned with the repository.

Execution Mode

Use agent team mode by default.

Team members:

Agent	Type	Model	Role	Skill	Primary Output
frontend-architect	custom	sonnet	FSD planning and scope control	frontend-architecture	_workspace/{phase}_frontend-architect_plan.md
ui-implementation-engineer	custom	sonnet	React and Tailwind implementation	ui-implementation	_workspace/{phase}_ui-implementation-engineer_changes.md
api-data-integrator	custom	sonnet	API, hooks, schemas, and types	api-data-flow	_workspace/{phase}_api-data-integrator_data-flow.md
qa-inspector	custom	sonnet	Cross-boundary verification	quality-gate	_workspace/{phase}_qa-inspector_report.md
code-reviewer	custom	sonnet	Changed-file review	code-review	_workspace/{phase}_code-reviewer_review.md
bug-investigator	custom	sonnet	Root-cause debugging	systematic-debugging	_workspace/{phase}_bug-investigator_report.md
git-workflow-assistant	custom	haiku	Commit, PR, and review workflow	commit / write-pr / review-pr	_workspace/{phase}_git-workflow-assistant_git.md

When invoking agents in a Claude harness that supports model selection, use the model value from each agent frontmatter unless the user explicitly asks for a different model.

Phase 0: Context Check

Before starting work:

1. Check whether _workspace/ exists.
2. Choose the run mode:
   • No _workspace/: initial run.
   • _workspace/ exists and the user asks for a narrow correction: partial rerun. Read the relevant previous artifacts and update only the affected area.
   • _workspace/ exists and the user provides a new broad request: fresh run. Preserve the old workspace by moving it to _workspace_{YYYYMMDD_HHMMSS}/, then create a new _workspace/.
3. Read CLAUDE.md, AGENTS.md, package.json, dynamically discovered .claude/rules/*.md, and the relevant source files before assigning work.

Phase 1: Classify the Task

Classify the request into one or more tracks:

• Architecture or routing: assign frontend-architect.
• UI, forms, modals, or layout: assign ui-implementation-engineer.
• API calls, server fetches, queries, mutations, schemas, or response types: assign api-data-integrator.
• Review, regression, or final acceptance: assign qa-inspector.
• Explicit code review, staged diff review, or pre-PR review: assign code-reviewer.
• Bugs, build failures, lint failures, runtime failures, or unclear behavior: assign bug-investigator before implementation.
• Branch names, actual commits, logical commit splits, PR creation, or PR review comment replies: assign git-workflow-assistant.

For small tasks, use only the needed agents. For cross-layer tasks, use the full team.

Phase 2: Assign Work

Use a supervisor plus producer-reviewer workflow:

1. frontend-architect defines scope, layers, and file ownership.
2. bug-investigator runs before fixes when the task starts from a failure or unclear behavior.
3. ui-implementation-engineer and api-data-integrator implement only their owned areas.
4. qa-inspector verifies after each meaningful module boundary, not only at the end.
5. code-reviewer reviews changed files for multi-file or pre-PR work.
6. git-workflow-assistant prepares or executes commit, PR creation, and PR review workflows when explicitly requested.
7. The orchestrator integrates findings and keeps the final response concise.

Expected task shape:

TeamCreate(
  team_name: "frontend-team",
  members: [
    { name: "frontend-architect", agent_type: "frontend-architect", model: "sonnet" },
    { name: "ui-implementation-engineer", agent_type: "ui-implementation-engineer", model: "sonnet" },
    { name: "api-data-integrator", agent_type: "api-data-integrator", model: "sonnet" },
    { name: "qa-inspector", agent_type: "qa-inspector", model: "sonnet" },
    { name: "code-reviewer", agent_type: "code-reviewer", model: "sonnet" },
    { name: "bug-investigator", agent_type: "bug-investigator", model: "sonnet" },
    { name: "git-workflow-assistant", agent_type: "git-workflow-assistant", model: "haiku" }
  ]
)

The example above is illustrative. Keep it synchronized with each agent's frontmatter.

Data Flow

Use _workspace/ for intermediate artifacts:

• _workspace/00_input_request.md
• _workspace/01_frontend-architect_plan.md
• _workspace/02_ui-implementation-engineer_changes.md
• _workspace/02_api-data-integrator_data-flow.md
• _workspace/03_qa-inspector_report.md
• _workspace/03_code-reviewer_review.md
• _workspace/01_bug-investigator_report.md
• _workspace/04_git-workflow-assistant_git.md

Preserve _workspace/ after completion for audit and follow-up work.

Bundled Resources

Use skill resources only when they help the active task:

• commit/scripts/discover-changed-areas.sh for commit and PR scope discovery.
• commit/references/scope-guide.md for scope/type selection.
• write-pr/scripts/create-pr.sh for explicit PR creation requests.
• write-pr/references/labels.md for PR label selection.
• review-pr/scripts/get-pr-data.sh for explicit PR review comment handling.
• review-pr/scripts/reply-review-comment.sh for explicit GitHub review replies.
• review-pr/references/reply-formats.md for Korean reply templates.
• code-review/references/review-checklist.md for multi-file or pre-PR reviews.
• systematic-debugging/references/root-cause-tracing.md for cross-boundary failures.

Error Handling

Situation	Strategy
One agent fails or stalls	Retry once with the same narrowed task. If it fails again, continue with a clear gap in the final report.
Most agents fail	Stop and ask the user whether to continue with partial results.
Conflicting findings	Keep both claims with file evidence, then inspect the source directly before deciding.
Validation command fails from environment	Report the environment cause separately from code issues and continue static checks.
Scope expands unexpectedly	Return to frontend-architect to reduce scope before implementation continues.

Quality Gate

Before final delivery, run or request:

• pnpm format:check
• pnpm lint
• pnpm build

If a command cannot run, explain the blocker and list the static checks that were completed.

Test Scenarios

Normal flow:

1. User asks for a feature that touches UI and data.
2. Architect defines affected layers and ownership.
3. UI and API agents implement their portions.
4. QA compares route, API, hook, type, schema, and UI boundaries.
5. Final response reports changed behavior and validation results.

Partial rerun flow:

1. User asks to revise only a previous UI behavior.
2. Orchestrator reads _workspace/ artifacts.
3. Only ui-implementation-engineer and qa-inspector are assigned.
4. The relevant artifact is updated and validation is rerun for the affected area.

Debugging flow:

1. User reports a bug, failed command, or unexpected behavior.
2. bug-investigator reads the full error and traces the relevant boundary.
3. The responsible implementation agent applies one root-cause fix.
4. qa-inspector runs the smallest meaningful check, then the broader quality gate when practical.

Git workflow flow:

1. User asks for a branch name, commit, PR creation, or PR review comment handling.
2. git-workflow-assistant inspects actual git state and relevant templates.
3. For commit requests, use commit.
4. For PR creation requests, use write-pr.
5. For PR review comment handling, use review-pr.
6. Do not commit, push, create PRs, or post replies unless the user explicitly requested that side effect.