Полная документация ниже — автоматически из репозитория trashchenkov/gigachat-skills (ветка main). Оригинальные материалы проекта распространяются на условиях его лицензии.

GigaChat Skills

Portable skill folders for AI coding agents that need to write code against the GigaChat API and the GigaChain set of solutions.

This repository is intentionally tool-agnostic. Different agent products expose reusable instructions in different ways:

• some support repository or team-level skills
• some rely on project instruction files
• some use custom modes, prompts, or MCP-connected workflows

This repository packages the GigaChat knowledge as self-contained skill folders so it can be adapted to any of those environments.

What is included

• gigachat-navigation
• gigachat-setup
• gigachat-sdk-chat
• gigachat-sdk-functions
• gigachat-sdk-files-embeddings
• langchain-gigachat
• gpt2giga
• deepagents-gigachat

Each folder contains:

• SKILL.md for trigger logic, workflow, and default behavior
• references/ for detailed guidance loaded on demand

The repository includes lightweight verification smoke scripts under verification/ for claims summarized in the feature matrix. It does not include workspace-specific state files or a full benchmark suite; the skill folders remain the reusable instruction layer.

Skills

gigachat-navigation

Choose the right integration layer:

• native gigachat SDK
• langchain-gigachat
• gpt2giga
• deepagents-gigachat

Use this first when the agent should decide which path fits the task.

gigachat-setup

Use for:

• credentials
• OAuth scope selection
• TLS certificates
• environment variables
• connectivity troubleshooting

gigachat-sdk-chat

Use for direct Python integration with the official gigachat SDK:

• chat
• streaming
• message history
• model selection
• token counting

gigachat-sdk-functions

Use for native tool and function calling with the official SDK:

• function schemas
• tool selection
• second-turn tool result flow
• built-in function caveats

gigachat-sdk-files-embeddings

Use for:

• file upload and lifecycle
• file-backed prompts
• multimodal constraints
• direct embeddings

langchain-gigachat

Use for LangChain-based work with GigaChat:

• chat models
• tools
• structured output
• embeddings
• RAG

gpt2giga

Use when an existing client must keep an OpenAI-style or Anthropic-style API and talk to GigaChat through the compatibility proxy.

deepagents-gigachat

Use when building or configuring Deep Agents / deepagents-code with GigaChat harness profiles:

• profile package installation
• provider keys and auto-registration
• deepagents-code model config
• profile behavior and benchmark caveats

Verification smoke scripts

The verification/ folder contains lightweight scripts referenced by gigachat-navigation/references/feature-matrix.md. They are not a full benchmark suite. By default they avoid live API calls; pass --live when credentials or a local gpt2giga proxy are configured.

Examples:

python verification/gigachat_chat.py --live --json
python verification/langchain_chat.py --live --json
python verification/gpt2giga_openai_http.py --live --json

See verification/LIVE_TESTS.md for the latest public-safe live smoke evidence.

How to use this repository

There is no universal installation path across agent products.

Instead, use the repository in the way your agent environment expects reusable instructions to be provided:

1. Copy one or more skill folders into the place where your agent loads project or team skills.
2. If your tool does not support native skill folders, adapt the contents of SKILL.md and references/ into that tool’s project instruction mechanism.
3. Keep each skill folder intact so the workflow guidance and references stay together.

Adapting to different agent tools

Typical patterns across modern coding agents include:

• repository-level instruction files
• reusable skill or module folders
• custom modes with tool selection and long-form instructions
• team-level shared prompts or agent bundles

This repository maps best to tools that support reusable instruction folders directly. For tools that do not, the content is still usable as a structured source of project guidance.

Confidence labels

The skills use three confidence labels:

• smoke-covered: covered by a lightweight script in verification/; run with --live to reproduce against configured services
• source-backed: supported by upstream documentation, release notes, or source code
• caution: supported by evidence but unsafe as a default recommendation
• inference: useful recommendation, but not a hard guarantee

Important cross-skill rule

For GigaChat file workflows, keep one request limited to one file modality.

Allowed:

• image + image
• audio + audio
• document + document

Avoid:

• image + audio
• image + pdf
• audio + pdf

If a user task spans multiple modalities, split the work into separate requests and combine results in application code.

Notes

• gpt2giga is documented as a compatibility layer, not as a byte-for-byte replacement for OpenAI or Anthropic APIs.
• langchain-gigachat is for LangChain-native workflows; use the native SDK skills when LangChain is not needed.
• deepagents-gigachat is for Deep Agents harness/profile behavior; use it only when Deep Agents or deepagents-code is in scope.
• gigachat-setup should usually come before implementation-specific skills.