Models and providers • codeagent

codeagent wraps any ellmer Chat, so it is provider-agnostic.

Any ellmer backend

chat <- ellmer::chat_openai_compatible(   # Databricks / Azure / vLLM / custom
  base_url    = Sys.getenv("CODEAGENT_BASE_URL"),
  model       = Sys.getenv("CODEAGENT_MODEL"),
  credentials = function() Sys.getenv("CODEAGENT_API_KEY")
)
client <- codeagent_client(chat, permission_mode = "bypass")

Supported provider values include openai_compatible, anthropic, openai, google_gemini, deepseek, groq, github, ollama, posit, databricks, aws_bedrock, and azure_openai.

Switching models mid-session

switch_model(client, "openai/deepseek-r1")   # CLI / one-shot
# In the Shiny app: the Settings panel model picker, or /model

The Shiny path swaps the provider in place (Route A) so the Chat object identity — and every callback/closure bound to it — stays valid.

Reasoning / thinking

Models with extended thinking are controlled natively via ellmer:

# settings.json: "effortLevel": "high"  ->  params(reasoning_effort = "high")

For OpenAI-compatible endpoints (Databricks Claude, DeepSeek), codeagent sets preserve_thinking = TRUE, so reasoning_content is parsed into ellmer ContentThinking and rendered distinctly (dimmed in the REPL, a collapsible block in the Shiny UI).

Small/fast model

Compaction and permission classification use a cheaper model. Set it via CODEAGENT_FAST_MODEL (the "fast" tier).