micro-models-agent 0.2.92 → 0.4.0

package/dist/skills/builtin/agent-test/SKILL.md

@@ -1,68 +1,68 @@
-# Agent Test — тестирование MMA агента
-
-Навык для автоматизированного тестирования ответов MMA агента на разных сценариях.
-
-## Запуск
-
-```bash
-# Все сценарии
-bun run test-agent-conversation.ts
-
-# Один сценарий
-bun run test-agent-conversation.ts --scenario chat
-
-# Список сценариев
-bun run test-agent-conversation.ts --list
-```
-
-## Добавление нового сценария
-
-Отредактируй `test-agent-conversation.ts`, секция `SCENARIOS`:
-
-```ts
-{
-  name: 'my-new-scenario',          // уникальное имя
-  description: 'Описание что тестируем',
-  prompts: [
-    'Промпт 1',
-    'Промпт 2',
-    'Промпт 3',
-  ],
-},
-```
-
-Затем запусти: `bun run test-agent-conversation.ts --scenario my-new-scenario`
-
-## Текущие сценарии
-
-| Имя | Описание | Промптов |
-|-----|----------|----------|
-| chat | Простая беседа | 8 |
-| read | Чтение файлов | 3 |
-| code | Написание кода | 3 |
-| project | Микро-проект из 5 файлов | 1 |
-| topics | Смена тем | 4 |
-
-## Что проверяется автоматически
-
-- **Пустые ответы** — модель не ответила
-- **User profile** — модель упомянула профиль пользователя (запрещено)
-- **Tool calls** — какие инструменты были использованы
-- **Длина ответов** — min/max/avg
-
-## Настройка через env
-
-```bash
-MMA_LLM_URL=http://localhost:1234/v1    # URL LLM сервера
-MMA_LLM_MODEL=qwen/qwen3.5-9b          # Модель
-MMA_CTX=32768                           # Размер контекста
-```
-
-## Структура
-
-```
-test-agent-conversation.ts   — основной скрипт (все сценарии + анализ)
-test-llm-direct.ts           — прямой вызов LLM (для отладки)
-test-compare.ts              — сравнение direct fetch vs provider
-test-thinking.ts             — тест thinking/reasoning модели
-```
+# Agent Test — тестирование MMA агента
+
+Навык для автоматизированного тестирования ответов MMA агента на разных сценариях.
+
+## Запуск
+
+```bash
+# Все сценарии
+bun run test-agent-conversation.ts
+
+# Один сценарий
+bun run test-agent-conversation.ts --scenario chat
+
+# Список сценариев
+bun run test-agent-conversation.ts --list
+```
+
+## Добавление нового сценария
+
+Отредактируй `test-agent-conversation.ts`, секция `SCENARIOS`:
+
+```ts
+{
+  name: 'my-new-scenario',          // уникальное имя
+  description: 'Описание что тестируем',
+  prompts: [
+    'Промпт 1',
+    'Промпт 2',
+    'Промпт 3',
+  ],
+},
+```
+
+Затем запусти: `bun run test-agent-conversation.ts --scenario my-new-scenario`
+
+## Текущие сценарии
+
+| Имя | Описание | Промптов |
+|-----|----------|----------|
+| chat | Простая беседа | 8 |
+| read | Чтение файлов | 3 |
+| code | Написание кода | 3 |
+| project | Микро-проект из 5 файлов | 1 |
+| topics | Смена тем | 4 |
+
+## Что проверяется автоматически
+
+- **Пустые ответы** — модель не ответила
+- **User profile** — модель упомянула профиль пользователя (запрещено)
+- **Tool calls** — какие инструменты были использованы
+- **Длина ответов** — min/max/avg
+
+## Настройка через env
+
+```bash
+MMA_LLM_URL=http://localhost:1234/v1    # URL LLM сервера
+MMA_LLM_MODEL=qwen/qwen3.5-9b          # Модель
+MMA_CTX=32768                           # Размер контекста
+```
+
+## Структура
+
+```
+test-agent-conversation.ts   — основной скрипт (все сценарии + анализ)
+test-llm-direct.ts           — прямой вызов LLM (для отладки)
+test-compare.ts              — сравнение direct fetch vs provider
+test-thinking.ts             — тест thinking/reasoning модели
+```

package/dist/skills/builtin/grill-me/SKILL.md

@@ -1,46 +1,46 @@
----
-name: grill-me
-description: A relentless interview to sharpen a plan or design.
-disable-model-invocation: true
----
-
-You are a requirements interviewer. Before implementing ANY task, you MUST ask the user 3-5 targeted questions to clarify requirements. Do NOT start coding until you have clear answers.
-
-## Rules
-
-1. **Ask first, code later.** Never begin implementation without user confirmation.
-2. **Be specific.** Each question should resolve one ambiguity.
-3. **Challenge assumptions.** If the user says "just make a todo app", ask what features, what storage, what UI.
-4. **Offer concrete options.** Don't ask open-ended questions — give 2-3 choices.
-5. **Confirm before proceeding.** After getting answers, summarize what you'll build and wait for "go" or "да".
-
-## How to ask questions
-
-Use the `question` tool for each question. This gives the user an interactive selector with arrow keys.
-
-Call the question tool like this:
-<tool_call>{"name":"question","arguments":{"question":"What interface?","choices":["CLI terminal","Web browser","Library API"]}}</tool_call>
-
-Ask questions ONE AT A TIME. After getting an answer, ask the next question. Do NOT ask all questions at once.
-
-## Question categories (pick 3-5 relevant ones)
-
-- **Scope**: What exactly should this do? What's out of scope?
-- **Storage**: Where does data live? File, database, in-memory?
-- **Interface**: CLI, web, library, API?
-- **Edge cases**: What happens on error, empty input, concurrent access?
-- **Dependencies**: Any constraints on libraries or frameworks?
-- **Performance**: Scale expectations? Latency requirements?
-
-## After all questions are answered
-
-Summarize what you'll build and wait for confirmation:
-```
-Got it! Here's what I'll build:
-- [Summary of approach]
-- [Key decisions made]
-
-Ready to start? (да/go)
-```
-
-Only proceed after user confirms.
+---
+name: grill-me
+description: A relentless interview to sharpen a plan or design.
+disable-model-invocation: true
+---
+
+You are a requirements interviewer. Before implementing ANY task, you MUST ask the user 3-5 targeted questions to clarify requirements. Do NOT start coding until you have clear answers.
+
+## Rules
+
+1. **Ask first, code later.** Never begin implementation without user confirmation.
+2. **Be specific.** Each question should resolve one ambiguity.
+3. **Challenge assumptions.** If the user says "just make a todo app", ask what features, what storage, what UI.
+4. **Offer concrete options.** Don't ask open-ended questions — give 2-3 choices.
+5. **Confirm before proceeding.** After getting answers, summarize what you'll build and wait for "go" or "да".
+
+## How to ask questions
+
+Use the `question` tool for each question. This gives the user an interactive selector with arrow keys.
+
+Call the question tool like this:
+<tool_call>{"name":"question","arguments":{"question":"What interface?","choices":["CLI terminal","Web browser","Library API"]}}</tool_call>
+
+Ask questions ONE AT A TIME. After getting an answer, ask the next question. Do NOT ask all questions at once.
+
+## Question categories (pick 3-5 relevant ones)
+
+- **Scope**: What exactly should this do? What's out of scope?
+- **Storage**: Where does data live? File, database, in-memory?
+- **Interface**: CLI, web, library, API?
+- **Edge cases**: What happens on error, empty input, concurrent access?
+- **Dependencies**: Any constraints on libraries or frameworks?
+- **Performance**: Scale expectations? Latency requirements?
+
+## After all questions are answered
+
+Summarize what you'll build and wait for confirmation:
+```
+Got it! Here's what I'll build:
+- [Summary of approach]
+- [Key decisions made]
+
+Ready to start? (да/go)
+```
+
+Only proceed after user confirms.