forge

Generates a project's domain agents — conversational grilling derives the recurring roles, then writes standard Claude Code subagent cards to .claude/agents/<role>.md (each `description` carrying a "when to use" so fg-run can dispatch a matching role as agentType). Cards load only at session start, so RESTART the session after generating them before fg-run can use them. On-demand utility outside the loop. Use in contexts like 'forge agents', '도메인 에이전트 만들어', '에이전트 팀 구성', 'create project agents', 'domain agents', 'agent roles'.

Grilling session that stress-tests your plan against the project's domain model, terminology, and documented decisions, updating CONTEXT.md/ADRs inline. The forge loop's entry point (the plan-grilling stage), always conversational. Use to start or refine a task — "start a new task", "start with forge", "let's work on this", "refine the plan", '새 작업 시작', '계획 다듬자', '이 계획 그릴링해줘'.

Seals a finished task — confirms the retro, marks STATUS.md done, archives it, empties the active .forge state, and closes the loop (which blocks the same plan from re-running). `fg-done all` batch-seals every already-executed task at once (retros skipped, backlog untouched, verification gate intact). Use when a task's retro is done — '작업 완료', '봉인', '이거 마무리', '작업 정리', 'forge complete', 'fg-done all', '봉인 all', '모두 봉인'. (Note: 'forge cleanup' routes to fg-cleanup, not here.)

After execution, classify learnings and promote them to CONTEXT.md, ADRs, and the retro log (.forge/retro), then surface the next inquiry. Always conversational; respects the promotion discipline. Use to record what you learned after a task — 'forge learn', '회고하자', '이번 작업 정리해줘'.

Integrates a merged branch's forge content (.forge/branch/<branch>/) into .forge/. Two modes: `fg-merge` (no arg) integrates only; `fg-merge <branch>` also runs `git merge <branch>` for you first, then integrates (default branch only; on conflict it stops, leaving the conflict in place). A deterministic script (forge-merge.sh/.js) does the git-free integration — usable AI-free in CI. Use in contexts like 'forge merge', 'fg-merge <branch>', '브랜치 통합', '브랜치 forge 합쳐줘', '머지하고 통합'.

Runs a refined plan (.forge/plan.md, or a waiting plan in .forge/backlog/) as a Claude Code Dynamic Workflow — one unexecuted plan runs right away; several show a priority-sorted selection (with 'Run all'), then the chosen task is promoted into the active slot. Warns before re-running a plan that already ran. Use in contexts like 'forge run', 'forge execute', '계획 실행', '이거 워크플로우로 돌려줘'.

Opens a browser-based visual companion in any conversation — a local zero-dependency server shows mockups, diagrams, and visual A/B options you push as HTML, and records the user's clicks as events you read back; `fg-visual stop` shuts it down. During fg-ask grilling the companion is offered automatically just-in-time (this skill is the standalone entry point). On-demand utility outside the loop. Use in contexts like 'forge visual', 'visual companion', '시각적으로 보여줘', '목업 보여줘', '브라우저로 보여줘', '화면으로 비교해줘'.

Optional adversarial review between fg-run and fg-learn — a reviewer assumes the result is WRONG and hunts for evidence across six lenses (failure points, hidden assumptions, misread requirements, security/perf/data-loss, unexpected misuse, weak decisions), fanning out as parallel subagents; fix-needed findings become a fix-forward plan on your approval. Outside the loop, never a seal gate, always skipped in fg-next all / fg-loop. Use in contexts like 'adversarial review', '적대적 리뷰', '이 결과 공격적으로 검토', '허점 찾아줘', 'red team this'.

Retires stale or superseded ADRs out of the active decision set — proposes candidates with rationale, and on your approval moves each to .forge/adr/retired/ (IDs never reused, nothing deleted). fg-ask stops reading retired/ as source of truth, so retired decisions drop out of grilling fuel. On-demand utility outside the loop (sealing a task is fg-done, not this). Use in contexts like 'forge cleanup', 'ADR 정리', '오래된 ADR 치워'.

A read-only integrity health check for forge — a deterministic script (forge-doctor.sh/.js) surveys the .forge/ state contract and docs/manifests and reports inconsistencies with severity + fix hints, exiting 0/1/2 so it doubles as an AI-free CI gate. Catches orphaned/broken state, half-sealed done/, an un-integrated branch root, version drift, README bilingual drift, ADR-ID issues, missing script twins. Writes nothing, never auto-fixes (fg-status: where am I; fg-doctor: is the state healthy). Use in contexts like 'forge doctor', '무결성 검사', '상태 점검', 'health check', 'check forge state'.

Discards incomplete (not-yet-sealed) forge work you no longer want — backlog plans, the active slot, awaiting-retro tasks, or a halted goal loop. Presents the items with a per-item risk level, then hard-delete (default) or archive to .forge/dropped/, behind a confirmation gate (already-run code is NOT reverted). Removes forge state only — never git or your code. Outside the loop (sealing a task is fg-done). Use in contexts like 'forge drop', 'fg-drop', '작업 버리기', '계획 지워', '백로그 비워', 'discard plan'.

Toggle forge's eco mode on/off — when on: caps fg-run subagents at sonnet, and activates the embedded Eco laziness-first discipline (ECO.md — code simplicity + terse output) in fg-run subagents, fg-ask grilling, and the current session. Does NOT switch the main session's model. `fg-eco on`/`off`; no arg shows state. On-demand utility outside the loop. Use in contexts like 'forge eco', 'eco on', 'eco off', '에코 모드', '경제 모드', 'lazy mode'.

Goal-driven momentum loop with bounded replan — an initial conversational inquiry pins a machine-verifiable stop condition, a fix-forward scope, and a replan cap (default 3) into .forge/loop.md, then drives tasks unattended (run → UAT → auto-skip retro → seal) until the checks pass, halting only at the walls (unverifiable UAT, genuine fork, cap exhausted, no progress). On-demand orchestrator outside the loop. Use in contexts like 'forge loop', '루프 시작', '조건 충족까지 반복', 'goal loop'.

Maps the codebase with parallel subagents into structured documents under .forge/codebase/, so later grilling reads a map instead of re-exploring the code (cuts context rot). On-demand utility outside the loop. Use in contexts like "map the codebase", "analyze the codebase", "코드베이스 분석", "코드베이스 지도".

Derives the single next step of the forge loop (reusing fg-status's state machine) and runs it — announces the step, then invokes that skill (fg-status reports; fg-next acts). One-shot by default; with the 'all' argument it drives the backlog to completion, auto-skipping retros and halting at the walls (failed/unverifiable UAT, a genuine fork, empty state). The cold-re-entry entry point. Use in contexts like 'forge next', '다음 단계', '이어서 해줘', '계속 진행', 'fg-next all', '다음 전부 진행'.

A lightweight lane outside the forge loop for trivial tasks (typo fixes, path tweaks, version bumps, small renames) — grills lightly, then runs the change directly with no formal artifacts (no ADR/plan/run/STATUS/retro). Bails to fg-ask if the task turns out non-trivial. Use in contexts like 'forge quick', '빠르게 처리', '이거 빨리 해줘', 'quick task'.

A read-only status reporter for the forge loop — surveys .forge/ and shows where every task stands (active slot, backlog, awaiting-retro, done history, quick lane) plus the single next step you need, with its trigger. Writes nothing, never auto-runs (fg-status reports; fg-next acts). On-demand utility outside the loop. Use in contexts like 'forge status', '상태', '어디까지 했지', '진행 상황', 'where am I'.

Set up (or refresh) a Claude Code statusline showing forge's loop progress, in one of two modes — method 1 (append) wraps your existing statusLine as extra rows (original preserved); method 2 (merge) installs a forge-owned unified script with daleseo-style system info (model/dir/git/context/usage bars) + forge progress, with a compact/full density toggle. On-demand setup utility outside the loop. Use in contexts like 'forge statusline', 'statusline 설정', '상태바', '상태 표시줄', 'add forge to statusline'.

Toggle forge's TDD mode on/off — a persistent setting in .forge/config.json. `fg-tdd on`/`off` set it; no arg shows the current state. When on, fg-ask defaults to asking per task and fg-run runs test-first. On-demand utility outside the loop. Use in contexts like 'forge tdd', 'tdd on', 'tdd off', 'TDD 켜', 'TDD 꺼', 'tdd 상태'.

Review the open GitHub issues of the current repo and triage them — classify, summarize, and score each one's fix-worthiness against the project's own docs — then print a ranked report so you can decide which to act on. Read-only: it fetches via the `gh` CLI and writes nothing (no files, no labels, no comments). It never fixes, comments, closes, or opens PRs — judgment and the fix stay with you. Use when you want to review/triage GitHub issues or decide which are worth fixing, in contexts like 'issue triage', 'triage issues', 'which issues to fix', 'review open issues', '이슈 트리아지', '이슈 검토', 'gh 이슈 정리', '어떤 이슈 고칠지'.