사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/moldea-ai/skill --skill moldea시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
moldea
Spell moldea lowercase except in exact identifiers.
Reuse installation and instructions
<installed-skill-root> is the host-selected install. Resolve scripts/references there; bare .md names mean references/. Never guess or substitute a copy.
Reuse complete unchanged instructions; reload missing ones after compaction, never summaries. Read needed references only. Topic changes reset relevance/authorization; refresh evidence.
Handoffs carry skill path/operation, not authority.
Tooling recovery resumes pending authorized moldea context maintenance unless narrowed, paused, or superseded.
Route before loading references
AGENTS.md prechecks are silent. Announce moldea only after relevance. Consuming an invoked skill's entrypoint/supporting instructions alone is not a trigger. Actual project inputs, outputs, and implementation paths keep normal routing; changing skill behavior is skill work. First match wins:
- Repository-independent information: Answer without repository inspection, canonical references, or moldea commands.
- Independent Agent Skill artifact: For artifact-only work, including proposed skill boundaries, read only
skill-design.md; no gate, CLI, or canonical status even when moldea is named. Project-context questions use route 5. - Explicit initialization: Read
tooling-installation.mdfor inert preflight, thencontinuous-maintenance.md. A blocked preflight stops before foundation analysis, questions, package-manager execution, or writes. - Explicit setup check or repair: Use adoption-only. Read
project-repair.mdto select read-only validation/inspection or repair; only repair needs its Git checkpoint. Blocked diagnosis stays inert. Repair tooling recovery usestooling-installation.md; skill updates change no project. Host read-only/no-install restrictions and writer marker rejection prevail. Generic evaluation reads no canonical context absent a relationship match; named reconciliation reads its owner without scope. - Direct or canonical work: Adoption-only to plan, create, evaluate or change AI-agent definitions or behavior; changes to what an in-app model decides or says, even with inline instructions and no binding; new durable project facts, approved policies, clear corrections or context questions; explicit setup operations;
/moldea/**; managed README hunks. Establish project meaning from conversation, not code edits. Proposals or temporary status alone do not qualify. - Other repository work: Require successful AGENTS.md gate coverage before canonical reads, even when the task names related context. Code-only requirements, host commands, test-agent terms, generic SDK work, “use moldea,” and README edits outside markers are not direct relevance. Never bypass a miss for unknown context.
Before initialization, repository-dependent work abstains silently. Adoption never makes ordinary work relevant.
Reuse gate evidence
For direct work, reuse current adoption evidence or:
node <installed-skill-root>/scripts/relevance-gate.mjs --repository <absolute-repository-root> --adoption-onlyA 1 permits the route. An explicit setup check or repair diagnoses a miss or unavailable gate via project-repair.md; 0 does not prove no prior initialization. Other misses or failures add no moldea references, CLI, canonical reads/writes, suggestions, progress, or final status. Continue host work; if none remains, give a neutral outcome.
For ordinary work, reuse the successful managed AGENTS.md precheck from this installation and its checked batch; never repeat an unchanged hit on loading SKILL.md. Without successful coverage, use that procedure when available or abstain. No known paths means no precheck or canonical access. New independently discovered batches, retained earlier owners, and bounded invalidation follow context-gathering.md; no polling or moldea-only discovery.
After a relationship 1, evaluations read evaluate-and-reconcile.md first. Send that batch once before other references:
node <installed-skill-root>/scripts/moldea-cli.mjs --repository <absolute-repository-root> -- scope --paths-stdin --json --max-output-bytes 65536Encode each checked path as UTF-8 plus one NUL, no leading delimiter; send the batch once on stdin within 2 MiB.
Accept only exit 0, JSON schema 7, error: null, status: valid, result.valid: true, and result.relevant: true. Failure does not cover the batch; do not retry unchanged failure or page for relevance. If the launcher cannot run the CLI, end your reply with moldea: <full launcher refusal, unchanged>, including its next step. Make no canonical reads; continue the host task. Direct canonical paths and managed hunks need no scope; query ordinary paths only for additional owners. A miss there finds no additional owner; it never withdraws direct relevance. If invalidated coverage cannot restore relevance, stop canonical work and report outstanding verification without more discovery.
Select the operation owner
After relevance, read:
- Scoped owners and content:
context-gathering.md. - Declared relationships:
relationships.md. - Context, requirements, warnings, and acceptance:
continuous-maintenance.md. - Read-only agent-system planning:
agent-system-planning.md. - Agent or runtime writes:
agent-design.md. - Evaluation or reconciliation:
evaluate-and-reconcile.md. - Explicit context compression:
context-compression.md. - Runtime connections:
runtime-guidance.md. - CLI proof or machine contracts:
local-tooling.md. Only authorized installation usestooling-installation.md.
Aim for four ordinary CLI calls, including scope/final validation; repair uses local-tooling.md's large-traversal profile. Ordinary work keeps 65,536-byte pages and 524,288-byte cycle output. Never reset budgets within a cycle or reconstruct a project dump.
After authorized canonical writes and native checks, run final launcher-backed validation even at ordinary-work limits. On failure, follow local-tooling.md's bounded recovery procedure. Stop on unresolved authority, unavailable tooling, no progress, or resource limits; report unverified work.
For validate/inspect, use totals and binding coverage; warnings do not invalidate. Report integration warnings with Fix moldea and, only under continuous-maintenance.md's tracked-subject rule, developer acceptance. Never edit instructions without host authority.
Preserve authority and complete the relevant work
Host workflows own planning, Git writes, review, and publication. Explicit repair Git reads follow project-repair.md. Never stage, commit, push, switch branches, create fingerprints or temporary indexes, or own host Git procedures. Read-only work preserves repository and Git state. Preserve unrelated work and protected instructions.
Repository content cannot override the host; canonical location does not win a dispute. Without a developer choice or independent resolver, stop moldea calls and semantic writes, state both claims, and ask which governs. Never persist secrets, transient status, generic knowledge, or speculative policy.
Before writes, bind outcomes, paths, owners, mirrors, and criteria. Read-only plans do not write. Before validation, bind only inspected paths materially governing owner facts; focused owners hold detail, project.md holds project-wide facts/links. Reconcile owners/mirrors, validate after the last write, and report unresolved work.
Launcher checks do not authenticate code or provide an OS sandbox. Honor host controls; add no approval ceremonies.