antithesis-review-inputs — скилл для Claude Code: обзор входов
Скилл antithesis-review-inputs для Claude Code оценивает, как генерация входов помогает исследовать разные состояния системы. Запускается по просьбе проверить, не ограничивает ли генерация входов поиск ошибок.
Apache-2.0 152 установок ★ 313 обновлён 2026-10-22
npx skills add antithesishq/antithesis-skills --skill antithesis-review-inputs -a claude-codeЧто делает antithesis-review-inputs
Скилл анализирует код нагрузки и документацию, чтобы оценить, какие области поведения системы могут оставаться недостижимыми. Он ищет фиксированные веса действий, узкие диапазоны данных, постоянный уровень параллелизма и другие ограничения случайного выбора. Перед разбором читает references/anti-patterns.md.
В отчёте есть краткое описание нагрузки и выводы по уровням значимости: с указанием файла и строки, недостигнутых состояний и уже работающих приёмов. Скилл доступен только для чтения: он не меняет код, не создаёт PR и не предлагает конкретные правки. Он оценивает генерацию входов, а не корректность системы, нагрузочного кода, проверок или инфраструктуры Antithesis.
Какие задачи решает
- найти фиксированные веса действий в нагрузке
- выявить диапазоны данных, которые оставляют состояния недостижимыми
- разделить ограничения генерации входов по значимости
- понять, какие приёмы разнообразия уже использует нагрузка
Примеры запросов агенту
- Проверь, ограничивает ли генерация входов исследование состояний в этой нагрузке.
- Разбери, как нагрузка использует случайность, и покажи недостижимые области поведения.
- Оцени разнообразие действий и данных в нагрузке Antithesis.
Как установить antithesis-review-inputs
1Через Skills CLI (Vercel)
npx skills add antithesishq/antithesis-skills --skill antithesis-review-inputs -a claude-code2Вручную: положить файлы
~/.claude/skills/antithesis-review-inputs/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/antithesis-review-inputs/
1Через Skills CLI (Vercel)
npx skills add antithesishq/antithesis-skills --skill antithesis-review-inputs -a codex2Вручную: положить файлы
.agents/skills/antithesis-review-inputs/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add antithesishq/antithesis-skills --skill antithesis-review-inputs -a cursor2Вручную: положить файлы
.cursor/skills/antithesis-review-inputs/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install antithesishq/antithesis-skills antithesis-review-inputs2Вручную: положить файлы
.github/skills/antithesis-review-inputs/SKILL.md1Через Skills CLI (Vercel)
npx skills add antithesishq/antithesis-skills --skill antithesis-review-inputs -a gemini-cli2Командой платформы
gemini skills install https://github.com/antithesishq/antithesis-skills --path antithesis-review-inputsВручную: положить файлы
.devin/skills/antithesis-review-inputs/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add antithesishq/antithesis-skills --skill antithesis-review-inputs -a opencodeClaude.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 об antithesis-review-inputs
Как установить antithesis-review-inputs в Claude Code?
Выполни npx skills add antithesishq/antithesis-skills --skill antithesis-review-inputs -a claude-code. Файл скилла находится по пути ~/.claude/skills/antithesis-review-inputs/SKILL.md.
Что делает antithesis-review-inputs?
Он оценивает, насколько генерация входов помогает нагрузке исследовать разные состояния системы. Результат — отчёт с находками, их значимостью и ссылками на код.
Что именно проверяет antithesis-review-inputs?
Скилл проверяет структуру генерации входов и её влияние на исследование состояний. Он не проводит глубокую проверку корректности системы, нагрузочного кода, утверждений или инфраструктуры.
Исходный текст
SKILL.md · Apache-2.0
--- name: antithesis-review-inputs description: > Review a workload's input generation for state space exploration: how well it uses randomness to drive the system under test into diverse regions of behavior. metadata: version: "2026-11-12 6ba10e2" --- # Antithesis Review Inputs **Skill version:** `2026-11-12 6ba10e2` ## Purpose and Goal Examine a workload and report how well its input generation explores state space. The system under test is a black box — the review works entirely from the workload code and any documentation the user provides. This skill is read-only. It produces a report of findings. It does not modify the workload, open PRs, or suggest specific code changes. The framing is "untapped exploration" — findings are ceilings on what the workload can reach, not defects. A workload exhibiting any anti-pattern may still be productive. The question is whether it could reach more interesting behavior with structural changes to how it uses randomness. ## When to Use When a user isn't finding bugs and wants to understand whether the workload's input generation is limiting what Antithesis can explore. Also useful as a periodic health check on workloads — both human-written and agent-generated. ## Core Principle Randomize the *policy* per run, not just the *choices under a fixed policy*. Each run should explore a different region of behavior by biasing its configuration differently. The skill reviews how well the workload does this. ## Workflow ### 1. Locate the workload Find the test commands and workload code. Common locations: - `antithesis/test/` — test command scripts - Workload source files referenced by test commands - Client libraries or helpers used by the workload If the workload location isn't obvious, ask the user. ### 2. Read the anti-patterns reference Read `references/anti-patterns.md` before examining the workload. It contains the full catalog of input generation anti-patterns organized by category, with explanations of why each limits exploration and what better looks like. ### 3. Examine the workload Read the workload code and look for structural patterns that match the anti-patterns. Focus on what is visible in the code: **Action selection** — How are operations chosen? - Is there a fixed set that's always fully active, or per-run subsetting? - Are action weights fixed, uniform, or varied per-run? - Is ordering imposed by the harness or emergent from randomness? - Are there symmetric pairs always active together? **Data generation** — How are inputs produced? - Are values hardcoded, drawn from fixed pools, or generated? - Do ranges vary per-run, or are bounds hardcoded? - Does input quality vary (valid, invalid, mixed), or is it always valid? - Do payload sizes span orders of magnitude, or stay in a narrow band? **Structure** — How is the workload organized? - Is the concurrency level fixed or varied per-run? - Does operation pacing vary, or is it always steady/fast? - Does the workload always start from empty state? ### 4. Build the findings For each anti-pattern you identify in the workload: - Name the pattern - Point to where in the code you see it (file and line) - Explain what state space it leaves unexplored - Describe what the workload could reach if this changed The explanation should be concrete and specific to this workload, not a generic restatement of the anti-pattern. "Your three actions (put, get, delete) are always all active with equal weight — runs where only deletes fire would test behavior with zero keys" is better than "all actions are always active." ### 5. Tier the findings Not all findings deserve equal weight. Tier them: - **High signal** — The workload has a clear structural limitation that prevents it from reaching entire regions of state space. Example: all actions always active with no subsetting, all values hardcoded. - **Worth considering** — The workload could explore more space with changes, but the limitation is more contextual or less severe. Example: no null/absent value variation, fixed concurrency level. - **Observation** — Something the reviewer noticed that might matter depending on the system. Example: encoding variation, seed data identical across runs. When you can't assess a finding's severity without domain knowledge (e.g., "this range is hardcoded, but I don't know what the system accepts"), say so. Present the structural fact and let the user supply the domain judgment. ### 6. Present the report Structure the report as: 1. **What the workload does** — A brief summary of what the reviewer sees: what actions the workload takes, how it generates data, how it's structured. This is the reviewer's model of the workload, presented so the user can correct misunderstandings before reading findings. 2. **Findings** — Organized by tier (high signal first), grouped by category when multiple findings cluster. Each finding names the pattern, points to the code, and explains the unexplored state space in terms specific to this workload. 3. **What's working well** — Patterns the workload already does right. If the workload has per-run action subsetting, or varies payload sizes, say so. Findings-only reports feel like a critique; acknowledging what works frames the report as an assessment. ## Tone The report must read as "here's what you could explore that you're not reaching" — never as "your workload is bad." A workload exhibiting multiple anti-patterns may still be finding bugs; the goal is to help it find more. Don't frame findings as "you need to do this." Frame them as "this is state space you're not reaching, and here's what reaching it would look like." The user decides what's worth changing. ## Scope Boundaries This skill reviews **input generation** — how the workload drives the system into diverse states. It does not review: - **Assertion quality** — Whether the workload would notice a bug if it reached one. That's a separate concern. - **System correctness** — Whether the system under test is behaving correctly. The SUT is a black box. - **Workload correctness** — Whether the workload code has bugs. The reviewer reads the workload's structure, not its correctness. - **Infrastructure** — Whether the Antithesis setup, Docker configuration, or test templates are correct. If the reviewer notices something outside its scope that seems important (e.g., the workload has no assertions at all), it can mention it briefly as an observation, but should not review it in depth. ## Reference Files | Reference | When to read | | ----------------------- | --------------------------------------------------- | | `references/anti-patterns.md` | Before examining any workload — the full catalog |
© antithesishq · Apache-2.0 · antithesishq/antithesis-skills