Operium Orchestra
Operium Orchestra is an AI coding agent for VS Code focused on local-first development, project context, and visible agent diagnostics.
It works with local models through Ollama, LM Studio, and Orchestra Runtime, plus cloud providers through Operium integrations. The default workflow is agentic: the model can inspect files, search the workspace, use RAG context, run allowed tools, and explain what it verified.
Highlights
- Agent by default: Chat, Plan, and IDE context are no longer separate products. Orchestra runs as an agent workflow by default; Plan is a read-only planning toggle.
- Permission presets: Default, Auto-check, and Full access control what the agent may do.
- Local provider support: Ollama, LM Studio, and Orchestra Runtime for local models; Operium cloud integrations for remote access.
- RAG for code and docs:
@codebase, @folder, @docs, and the read-only codebase_search tool.
- Hybrid retrieval: text retrieval works without embeddings; embeddings can be added through OpenAI-compatible
/v1/embeddings or Ollama native embeddings.
- AgentRunState: visible working memory for checked files, searches, verified facts, open questions, active skills, provider/tool protocol, and stop reasons.
- Run diagnostics: debug panel shows provider, tool protocol, repeated tool calls, readiness, guardrails, and why the agent stopped.
- Agent Skills: lazy workflow instructions from built-ins,
.orchestra/skills/**/SKILL.md, and ~/.orchestra/skills/**/SKILL.md.
- Project config:
.orchestra/config.yaml, project rules, snippets, RAG settings, and tool policies.
- Operium Hub install links: one-click model install through
vscode://.../install-model?manifest_url=... with origin allowlist and checksum verification.
Supported Providers
| Provider |
Use case |
| Ollama |
Recommended local baseline for fast, stable local agent runs. |
| LM Studio |
OpenAI-compatible local models, including native tool-capable models when supported. |
| Orchestra Runtime |
Local GGUF models through the Operium runtime layer. |
| Polza / Operium cloud |
Cloud models and hosted integrations where configured. |
Agent Modes And Access
Orchestra uses one agent workflow with toggles:
- Default: agent workflow with workspace tools available according to permissions.
- Plan: read-only investigation and implementation planning. Write, terminal, and network actions are blocked.
- IDE context: optional editor/file context attachment.
- Permissions:
- Default: balanced local agent access.
- Auto-check: asks before higher-risk actions.
- Full access: allows all built-in tool groups permitted by the extension.
The debug UI shows the actual provider and tool protocol used for the run: native, fallback, or none.
RAG And Workspace Context
Orchestra keeps a local retrieval index in:
.orchestra/index
The index is local cache data and is safe to delete or gitignore.
Use:
@codebase provider timeout
@folder src/providers timeout
@docs runtime manifest
The agent can also call:
codebase_search
RAG supports:
- codebase search;
- folder-scoped retrieval;
- documentation retrieval;
- text retrieval without embeddings;
- optional semantic retrieval with embeddings;
- generated/cache/vendor path filtering;
- eval scenarios through
npm run rag:eval.
Agent Skills
Skills are Markdown workflow instructions that are selected lazily for the current task.
Built-in skills include:
- context engineering;
- incremental implementation;
- test-driven debugging;
- code review and quality;
- planning and task breakdown;
- security and hardening.
Project skills live here:
.orchestra/skills/<skill-name>/SKILL.md
Global skills live here:
~/.orchestra/skills/<skill-name>/SKILL.md
Skills are shown in Settings -> Skills and in AgentRunState diagnostics when activated.
Project Config
Orchestra reads project-local config from:
.orchestra/config.yaml
Typical project config areas:
- provider URLs and default model;
- sampling options;
- RAG paths and retrieval mode;
- tool policies;
- snippets;
- rules;
- skills in
.orchestra/skills.
Rules are always-on instructions. Skills are task-specific workflow instructions. Keep persistent project policy in .orchestra/rules; keep reusable workflows in .orchestra/skills.
Important Settings
| Setting |
Default |
Description |
orchestra.ollamaUrl |
http://localhost:11434 |
Ollama server URL. |
orchestra.lmStudioUrl |
http://localhost:1234 |
LM Studio / OpenAI-compatible base URL. |
orchestra.runtimeUrl |
http://localhost:8100 |
Orchestra Runtime URL. |
orchestra.defaultModel |
empty |
Default selected model; empty means auto-pick. |
orchestra.requestTimeoutMs |
7200000 |
HTTP/model request timeout in milliseconds. |
orchestra.codebaseAutoIndex |
true |
Automatically update the local RAG index. |
orchestra.codebasePaths |
[] |
Workspace-relative roots for @codebase; empty means whole workspace. |
orchestra.embeddingBaseUrl |
LM Studio URL |
OpenAI-compatible embeddings base URL. |
orchestra.embeddingProvider |
auto |
auto, openai, or native ollama. |
orchestra.ragRetrievalMode |
auto |
auto, hybrid, text, or embeddings. |
orchestra.docsPaths |
["docs", "README.md", ".orchestra/docs"] |
Paths indexed by @docs. |
orchestra.agentSkillsEnabled |
true |
Enables lazy agent skill injection. |
orchestra.autoCompactOnHighContext |
true |
Auto-compact conversation when context is near the threshold. |
Runtime And Model Install
Operium Orchestra does not install Orchestra Runtime automatically during extension installation.
Runtime download happens on demand from the extension UI. If a local runtime binary is already available, Orchestra reuses it.
Operium Hub can install models through a deep link:
vscode://OperiumControlHub.operium-orchestra/install-model?manifest_url=<url-encoded manifest URL>
The extension validates manifest_url against orchestra.hubAllowedOrigins, downloads the artifact, verifies checksum, and stores it under orchestra.modelsPath or ~/.orchestra/models.
QA And Development
Useful commands:
npm run compile
npm run test
npm run build
npm run rag:eval
Manual QA checklist:
docs/manual-qa.md
Before packaging VSIX:
npm run compile
npm run test
npm run build
npm run rag:eval
npm run package
Support
Operium Orchestra на русском
Operium Orchestra — AI coding agent для VS Code с упором на локальные модели, контекст проекта и прозрачную диагностику работы агента.
Главное
- Агентный поток по умолчанию.
- Переключатель План для read-only исследования.
- Уровни доступа: По умолчанию, Автопроверка, Полный доступ.
- Локальные провайдеры: Ollama, LM Studio, Orchestra Runtime.
- RAG по коду и документации:
@codebase, @folder, @docs, codebase_search.
- AgentRunState: проверенные файлы, поиски, факты, открытые вопросы, активные skills и причины остановки.
- Отдельная run/debug диагностика: provider, tool protocol, повторы tools, readiness и host guardrails.
- Agent Skills из встроенного набора,
.orchestra/skills и ~/.orchestra/skills.
- Проектный config:
.orchestra/config.yaml.
Быстрый старт
- Установите расширение.
- Запустите Ollama, LM Studio или Orchestra Runtime.
- Откройте боковую панель Operium Orchestra.
- Выберите модель.
- Для контекста проекта используйте:
@codebase provider timeout
@folder src/providers timeout
@docs runtime manifest
RAG
Индекс хранится локально:
.orchestra/index
Это cache, его не нужно коммитить. Text retrieval работает без embeddings. Для semantic retrieval настройте embeddings endpoint и модель.
Skills
Проектные skills:
.orchestra/skills/<skill-name>/SKILL.md
Глобальные skills:
~/.orchestra/skills/<skill-name>/SKILL.md
Rules — это постоянные инструкции. Skills — это workflow-навыки, которые Orchestra подключает только под конкретную задачу.
Проверки
npm run compile
npm run test
npm run build
npm run rag:eval
Актуальный ручной чеклист находится в docs/manual-qa.md.
