agent-creator

Agent Creator

Создавайте агентские пакеты как переносимые системы, а не как промпты, которые работают только в текущем чате.

Core rule

Модель предлагает. Runtime, tools, tests, permissions, artifacts и reviewers доказывают.

Не доверяйте safety, tool execution, validation, memory hygiene и final acceptance только инструкции в промпте.

Перед началом

1. Прочитайте ../grill-me/SKILL.md.
2. Зафиксируйте objective, target users, maturity, scope, autonomy, risk, allowed/forbidden actions, data sources, target platforms, export shape, validation и done condition.
3. Прочитайте ../workflow-loop-me/SKILL.md и соберите workflow notes, candidate loops и выбранный workflow spec.
4. Если пользователь дал ссылки на репозитории или инструменты, прочитайте ../repo-tool-librarian/SKILL.md и сохраните их как catalog candidates без интеграции.
5. Для production/library/governed пакетов прочитайте references/production-skill-os.md.

Обязательная форма серьезного agent package

• router / instruction entrypoint;
• Agent/Skill IR;
• versioned system identity and harness boundary for session, harness, sandbox, artifact store and credential broker;
• default-deny permission policy with approval binding, provenance-aware data flow and runtime enforcement;
• append-only run-event contract for effects, permissions, transfers, state revisions and budgets;
• workflow discovery notes, workflow spec и discovery ledger;
• repo/tool library для идей и reusable building blocks, не для автоматической интеграции;
• platform adapters;
• birth contract, runtime profile, birth plan, environment readiness и project context templates;
• IDE/runtime adaptation policy;
• focused skills или workflows;
• typed tool registry;
• permission matrix;
• hook registry, lifecycle event mappings и hook validation для правил, требующих автоматического enforcement;
• subagent topology decision, role registry, delegation plan, task/result contracts, run ledger, isolation, lifecycle и Subagent Eval Lab;
• memory and state model;
• workspace hygiene rules;
• checker/reviewer layer;
• context packaging policy;
• Trigger Lab;
• eval-validity report with task inventory, feasibility witnesses, grader mapping, contamination checks and frozen holdout;
• Output Eval Lab with outcome, trajectory, boundary and stability graders, repeated trials and infrastructure calibration;
• Harness Assumption Registry and matched-budget Ablation Lab;
• workflow/negative fixtures;
• skill training lab and skill-candidate lifecycle;
• target conformance matrix;
• final evidence contract;
• release review;
• machine-readable release decision that reconciles required gates and derives allowed claims;
• evidence ledger and claim guard;
• package verification;
• install simulation from final zip;
• post-birth cleanup;
• runtime/devkit boundary.