한국어

This repository contains a compact operating workflow for coding agents. The source of truth is AGENTS.md: it defines how an agent should inspect context, record requirements, plan work, request execution approval, make changes, verify results, and archive workflow state.

AGENTS.md is designed for projects where agent work should stay controlled and traceable. It favors a short approval-based workflow over immediate edits to project files.

The workflow is built around these rules:

• Inspect the active workflow and worktree before starting non-trivial work.
• Record confirmed requirements before planning complex or approval-gated changes.
• Plan before implementation, and connect each plan item back to a requirement.
• Ask for execution approval before modifying durable project files.
• Keep task state current while work is being executed.
• Verify the result before marking work complete.
• Archive workflow files when there is no remaining actionable work.
• Never overwrite, discard, commit, push, rebase, or reset user changes unless explicitly asked.

Active workflow files live under .agent-workflow/. They are working state, not project deliverables, and should stay short, practical, and current.

• .agent-workflow/request.md: confirmed requirements, assumptions, open questions, exclusions, and references.
• .agent-workflow/plan.md: objective, constraints, selected approach, affected files, execution order, risks, validation, and approval items.
• .agent-workflow/task.md: current executable tasks with TASK-* IDs and explicit states.
• .agent-workflow/task.recent.md: recently completed tasks and verification summary.
• .agent-workflow/speedwagon.md: decision-relevant external findings.
• .agent-workflow/archive/<timestamp>-<task-slug>/: archived workflow history with a concise summary.

Use only the files needed for the current task. When active workflow files are stale or complete, archive them instead of letting them become logs.

The workflow uses stable IDs to keep decisions traceable:

• REQ-* for confirmed requirements.
• SPEC-* for plan items that trace back to requirements.
• TASK-* for executable work that traces back to plan items.

These IDs are not ceremony. They make it clear what was requested, what approach was approved, and what was actually executed.

• AGENTS.md: the workflow and operating rules.
• README.md: English overview.
• README.ko.md: Korean overview.

For a project, place AGENTS.md at the project root so agents can read it before working. For a global default, copy or merge the relevant rules into your agent's global instructions.

The workflow does not require every file for every request. Small questions may need no workflow files. Non-trivial or approval-gated work should use the minimum set needed to keep requirements, planning, tasks, and verification clear.

This repository treats an agent as a careful collaborator, not an autonomous executor. The expected pattern is: read the current state, clarify scope, plan the change, ask before editing durable files, verify the result, and report what changed.