state-verification — скилл для Claude Code: проверка состояний
Скилл state-verification для Claude Code проверяет состояния объектов и связи по различимым наблюдениям. Он нужен, если следующий шаг зависит от удержания объекта, размещения на опоре или завершения действия.
MIT ★ 488 обновлён 2026-11-15
npx skills add Open-RUA/uni-robot-agent --skill state-verification -a claude-codeЧто делает state-verification
Скилл помогает агенту проверить, удерживается ли нужный объект, правильно ли он размещён и выполнено ли требуемое действие. Он предлагает выбрать один вопрос, от ответа на который зависит следующий шаг, и отличить подтверждённое состояние от опровергнутого или неразрешённого. Положение руки само по себе не доказывает, что объект удерживается.
Для проверки используются уже полученные изображения и записи наблюдений; новый ракурс нужен, только если он различит возможные варианты. view_env_state(step=...) выбирает индекс записи, а не шаг физической симуляции. После столкновения, проскальзывания или смены опоры затронутые связи нужно перепроверить. Для единиц измерения, систем координат и ограничений робота нужны описание робота и опубликованное руководство текущей среды.
Какие задачи решает
- проверить, удерживает ли захват нужный объект
- подтвердить, что объект остался на целевой опоре
- отличить вход объекта в отверстие от перекрытия силуэтов
- сверить требуемую ориентацию и учесть все объекты
- найти связи, которые могли нарушиться после столкновения
Примеры запросов агенту
- Проверь, удерживает ли захват нужный объект, прежде чем переносить его.
- Установи, остался ли объект на опоре после отпускания захвата.
- Проверь, завершена ли расстановка, и перечисли связи, которые пока нельзя подтвердить.
Как установить state-verification
1Через Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill state-verification -a claude-code2Вручную: положить файлы
~/.claude/skills/state-verification/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/state-verification/
1Через Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill state-verification -a codex2Вручную: положить файлы
.agents/skills/state-verification/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill state-verification -a cursor2Вручную: положить файлы
.cursor/skills/state-verification/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install Open-RUA/uni-robot-agent uni_robot_agent/resources/skills/general/gripper/state-verification2Вручную: положить файлы
.github/skills/state-verification/SKILL.md1Через Skills CLI (Vercel)
npx skills add Open-RUA/uni-robot-agent --skill state-verification -a gemini-cli2Командой платформы
gemini skills install https://github.com/Open-RUA/uni-robot-agent --path uni_robot_agent/resources/skills/general/gripper/state-verificationВручную: положить файлы
.devin/skills/state-verification/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о state-verification
Как установить state-verification в Claude Code?
Выполни npx skills add Open-RUA/uni-robot-agent --skill state-verification -a claude-code. Файл скилла находится по пути ~/.claude/skills/state-verification/SKILL.md.
Что делает state-verification?
Скилл помогает проверять отношения между объектами по наблюдаемым признакам: например, удерживается ли объект захватом и остался ли он на опоре после отпускания. Он разделяет подтверждённые, опровергнутые и неразрешённые состояния.
Что означает параметр step в view_env_state?
Он выбирает индекс записи наблюдения, а не шаг физической симуляции. Повторное чтение последней записи не продвигает симуляцию и не ждёт завершения текущего действия.
Исходный текст
SKILL.md · MIT
--- name: state-verification description: Verify object states and relationships from distinguishing evidence. Use when the next action depends on an unresolved state or when assessing completion across occlusion and multiple objects. --- # Verify the relation that changes the next decision ## Select one unresolved question Identify the unresolved relation that matters for the next dependent action: attachment, support, containment, orientation, identity or a required effect. Derive requirements from the instruction; add no preferred pose, count or order. Choose an observation whose possible outcomes would change the next decision. Use existing images and feedback first. Select a new view only when it can distinguish the remaining outcomes; no written proof is required per call. Keep three outcomes: confirmed, contradicted and unresolved. A tool return or absence of visible failure does not by itself confirm an object relationship. | Question | Distinguishing evidence | Insufficient alone | |---|---|---| | Is the intended object controlled? | A safe test motion preserves its relation to the hand, with consistent identity and source-view change. | Closed fingers, opening residual, or a centered wrist image. | | Is it supported after placement? | After release and clear withdrawal, it remains on the intended support. | The hand reached the placement pose. | | Is it contained or engaged? | Visible opening, edge and depth relationships support entry rather than bridging or leaning. | Overlapping silhouettes. | | Is its required orientation correct? | Identifiable ends, faces or axes distinguish the requested orientation. | Correct center position or an ambiguous long axis. | | Are all required objects accounted for? | Each identity has an observed destination or a still-valid chain of confirmed transfers and placements. | Attempt counts or one final merged silhouette. | A verification lift is appropriate when lifting is intended and observed clearance permits it. For constrained or fragile objects, choose a compatible test; do not impose a lift on every operation. Inspect the resulting object motion rather than treating the test command as evidence. For an uncertain contact effect, use `general/gripper/contact-effect/SKILL.md`. ## Compare the right records and views Choose the camera that exposes the deciding feature rather than requesting all views by default. For attachment, compare the original support and the hand; for engagement, expose the remaining gap or entering feature. A second view is useful when it resolves identity, occlusion or competing explanations. view_env_state(step=...) selects a recorded observation index, not a native physics step. Use returned record_index and step to identify the before/after states. Re-reading the latest record advances no physics and does not wait for an in-flight action. Resolve that action before checking its final effect. When a physical process needs time, use a short hold for a specific observable purpose, obtain its completed result, and inspect the attached image. Track the same object across records and account for camera movement. An object centered in a moving wrist image may still rest on its original support; image displacement alone cannot separate object motion from camera motion. After a verification lift, compare the same object against its original support and the hand. If the hand rises but the object stays at its source and becomes smaller in the wrist view, the grasp did not retain it. A centered image is not counter-evidence: the camera moved with the hand. When the views cannot resolve height or attachment, use a distinguishing view or a calibrated measurement of the same visible feature before dependent transport. Do not measure again when existing observations already establish the relation. If lifting is inappropriate, choose a compatible probe. For a press, inspect target displacement and release rather than counting reached hand poses; use the contact-effect skill for event counts. ## Preserve a small evidence chain Retain only the identity, required relation, its status and the observation or transition supporting it. Update affected entries; do not narrate a full ledger after every call. A brief record such as "A released onto B; support observed" can preserve evidence that will later be occluded. Include the observation's step or time when needed to distinguish it from an older view. For nested or stacked objects, retain each established support relation; a final merged outline cannot retroactively confirm an unverified entry or repair a disturbed base. A confirmed relation can remain valid after occlusion. Later placements need not expose every earlier object again if identities and transfers were tracked and subsequent actions give no reason to suspect disturbance. An unverified relation does not become confirmed merely because the object disappeared. A collision, slip, changed support or ambiguous identity can invalidate affected relations and those depending on them. A relation depending on a displaced base needs review even if its own object was not directly touched. Recheck those relations rather than restarting the whole arrangement. During occluded actions that could change a relation, preserve uncertainty until useful evidence returns. ## End the check with a decision - Confirmed: continue the dependent action without duplicate views. - Contradicted: correct the observed mismatch and update the affected relation. - Unresolved: choose one useful view or safe probe that separates the alternatives before committing to a dependent action. Prefer the least disruptive check. If the same view adds no information, change the view, probe or hypothesis. If available tools cannot distinguish the alternatives, keep the relation unresolved and pursue a different feasible observation, approach or independent subgoal when order permits. Longer explanations do not strengthen the evidence. For completion, combine current observations with earlier evidence that remains valid. Check required relationships and their dependencies, not whether all planned actions were attempted. Report remaining uncertainty faithfully; follow the system prompt's stopping conditions. Uncertainty does not authorize success or an early finish, and does not justify repeating an uninformative check. When exposed, read `state.arms[arm].metadata.gripper_opening_m` for actual signed pad gap and `gripper_target_opening_m` for the effective drive target. Unavailable measurements remain unknown; a residual gap alone does not establish attachment. Read the current environment's robot description and published manual for units, frames and capability limits.
© Open-RUA · MIT · Open-RUA/uni-robot-agent