Eai.LLM.Direct (eai v0.1.15)

Direct LLM API orchestration using Converse-based internal message IR.

All internal history is [Eai.Message.t()]. Before sending to an LLM provider, the appropriate adapter converts messages to provider-specific wire format. Responses are parsed back into Eai.Message.t(). LLM requests flow through Pipeline.llm_pre_hooks/4 and llm_post_hooks/5 for pre/post request interception.

Summary

Functions

run(messages, pty_session_id \\ "default", opts \\ %{})

Run LLM tool-calling loop with messages.

Functions

run(messages, pty_session_id \\ "default", opts \\ %{})

Run LLM tool-calling loop with messages.

Internal API called by Chat.talk/1. Handles model selection, adapter routing, tool execution, and result polling.

Options (map)

:model (atom) — Model name: :deepseek, :claude_opus, etc.

              Default: from config `:default_model`

:prompt (atom) — System prompt: :momoka, :coder, :analyst, etc.
```
               Default: `:momoka`
```

:chat_session_id (string) — Session for multi-session isolation.

                           Default: `"default"`

:card_system_prompt (string) — Extra system prompt (role layer, appended).
```
                              Default: nil
```
:card_tools (list) — Allowlist of tool names. Default: all tools
:card_pre_context (list) — Pre-loaded messages (for prefix caching).
```
                          Default: nil
```

Internal fields (filled automatically by Chat.talk, do NOT set externally):

:provider — Adapter provider (:anthropic, :openai_compat, etc.)
:api_key — LLM API key (from env)
:url — LLM endpoint URL
:receive_timeout — HTTP timeout in ms
:reasoning_effort — Model-specific (e.g., "high" for DeepSeek)
:model_str — Actual model string for API

Returns

`{:ok, final_reply}` or `{:error, reason}`

Example (internal, called by Chat.talk)

iex> Eai.LLM.Direct.run(
  messages,
  "default",
  %{model: :deepseek, prompt: :coder, chat_session_id: "work"}
)
{:ok, "Here's the refactored code..."}