llm_provider

System module that manages LLM provider instances. Automatically configured from agent brain: definitions in YAML. Supports any OpenAI-compatible API and the Anthropic native SDK.

Property	Value
Module ID	llm_provider
Version	1.0.0
Type	system (auto-loaded, hidden from agents)
Dependencies	openai (for OpenAI-compatible providers), anthropic (for Anthropic)

---

Role in the Architecture

The LLM provider module is the bridge between Digitorn agents and LLM APIs. It:

1. Auto-configures from the brain: section of each agent definition.
2. Resolves provider URLs automatically -- provider: deepseek maps to https://api.deepseek.com/v1.
3. Manages connections -- creates and reuses async HTTP clients with connection pooling.
4. Handles tool calling -- detects whether the model supports native tool calling or requires text-based recovery.
5. Normalizes responses -- provides a consistent response format regardless of provider.

---

Supported Providers

The module auto-resolves base URLs for common providers:

Provider	Base URL	Backend
openai	https://api.openai.com/v1	openai_compat
deepseek	https://api.deepseek.com/v1	openai_compat
anthropic	https://api.anthropic.com	anthropic
groq	https://api.groq.com/openai/v1	openai_compat
mistral	https://api.mistral.ai/v1	openai_compat
together	https://api.together.xyz/v1	openai_compat
ollama	http://localhost:11434/v1	openai_compat
lm_studio	http://localhost:1234/v1	openai_compat
vllm	http://localhost:8000/v1	openai_compat
openrouter	https://openrouter.ai/api/v1	openai_compat

Custom providers work by specifying base_url directly in the brain config.

---

Configuration

The module is configured implicitly through agent brain definitions:

agents:
  - id: assistant
    brain:
      provider: deepseek
      model: deepseek-chat
      temperature: 0.2
      max_tokens: 4096
      config:
        api_key: "{{env.DEEPSEEK_API_KEY}}"

No explicit modules: llm_provider: declaration is needed.

---

Text-Based Tool Calling Recovery

When a model does not support native tool calling (or produces malformed tool calls), the module includes a multi-format parser that recovers tool calls from text:

1. Native JSON tool call format
2. XML-wrapped tool calls
3. Markdown code block JSON
4. Inline JSON with function names
5. Unicode quote normalization

This makes Digitorn compatible with local models (Ollama, vLLM) that have imperfect tool calling support.

---

Actions (6)

These actions are for internal use and are hidden from agents:

Action	Risk	Description
configure	medium	Register a named provider instance
chat	low	Send a chat completion request
remove	low	Remove a provider instance
list_providers	low	List configured providers
update_defaults	low	Update default generation parameters