API Reference

BilgeAgent

AI-powered coding copilot agent. Supports both OpenAI-compatible and Ollama backends.

BilgeAgent(config, working_dir)

Create a BilgeAgent with the given configuration and working directory.

BilgeConfig

Configuration for the Bilge coding copilot.

BilgeState

Mutable state maintained across the coding session.

LLMConfig

Configuration for LLM API calls.

Message

A message in the conversation.

OllamaConfig

Configuration for Ollama API calls.

Supports both the native Ollama API (/api/chat) and the OpenAI-compatible endpoint (/v1/chat/completions).

• model::String - Model name (e.g., "llama3.1", "qwen2.5", "mistral")
• host::String - Ollama server address (default: "http://localhost:11434")
• max_tokens::Int - Maximum tokens in response (default: 32768)
• temperature::Float64 - Sampling temperature (default: 0.1)
• use_openai_compat::Bool - Use OpenAI-compatible endpoint (default: false)

Tool

Represents a callable tool that the LLM can invoke.

ToolCall

Represents a tool call request from the LLM.

ToolExecution

Record of a single tool execution.

TurnResult

Result of a single conversation turn.

_create_tools(state, max_output_chars)

Create all coding tools for the agent.

_extract_tool_calls_from_text(content)

Fallback parser: extract tool calls from text content when models embed them in their output instead of using the native tool calling API. Handles common patterns like <toolcall>...</toolcall> and json {"name":...} blocks.

Returns (cleanedcontent, toolcalls) where tool_calls may be nothing.

_glob_to_regex(pattern)

Convert a glob pattern to a Regex. Supports: * (any non-/ chars), ** (any chars including /), ? (single char).

_handle_slash_command(agent, input) -> Bool

Handle a slash command. Returns true to continue the REPL, false to exit.

_print_tool_summary(exec)

Print a brief summary of a tool execution.

_read_input(state)

Read user input with history navigation and multi-line support. Returns nothing on EOF or Ctrl-C.

_read_line_raw(prompt, color, history) -> Union{String, Nothing}

Read a single line from a TTY with arrow key history navigation. Supports: Up/Down (history), Left/Right (cursor), Home/End, Backspace, Delete, Ctrl-A/E/K/U/W/L, and UTF-8 input. Returns nothing on Ctrl-C or Ctrl-D.

_resolve_path(state, path)

Resolve a path relative to the working directory. Absolute paths pass through.

_show_history(agent)

Display a summary of the conversation history.

_strip_think_tags(text)

Remove <think>...</think> blocks from model output (common in reasoning models).

_truncate_output(text, max_chars)

Truncate output to max_chars, adding a notice if truncated.

bilge(; api_key, model, base_url, ollama, host, use_openai_compat, working_dir)

Start the Bilge interactive coding copilot.

• api_key::String - OpenAI API key (default: ENV["OPENAIAPIKEY"])
• model::String - Model name (default: "gpt-4o" or "llama3.1" for Ollama)
• base_url::String - API base URL (default: "https://api.openai.com/v1")
• ollama::Bool - Use Ollama backend (default: false)
• host::String - Ollama host (default: "http://localhost:11434")
• use_openai_compat::Bool - Use Ollama's OpenAI-compatible endpoint (default: false)
• working_dir::String - Working directory (default: pwd())

using Bilge

bilge(ollama=true, model="qwen3")

bilge(api_key="sk-...")

bilge(api_key="key", base_url="https://api.example.com/v1", model="my-model")

build_system_prompt(working_dir, model_name)

Build the system prompt for the Bilge coding copilot.

call_llm(config, messages, tools)

Make an API call to the LLM.

call_ollama(config, messages, tools)

Make an API call to Ollama using the native /api/chat endpoint.

call_ollama_openai_compat(config, messages, tools)

Make an API call using Ollama's OpenAI-compatible endpoint (/v1/chat/completions).

check_ollama_connection(; host="http://localhost:11434")

Check if Ollama server is running and accessible.

• Bool - true if server is reachable

execute_tool(agent, name, args)

Execute a tool by name with the given arguments.

format_tool_result(tool_call_id, result)

Create a tool result message.

list_ollama_models(; host="http://localhost:11434")

List available models on the Ollama server.

• Vector{String} of model names

messages_to_ollama_format(messages)

Convert Message objects to Ollama's native message format.

messages_to_openai_format(messages)

Convert Message objects to OpenAI API format.

parse_llm_response(response)

Parse LLM response into Message and tool calls.

parse_ollama_response(config, response)

Parse Ollama response into Message and tool calls. Handles both native and OpenAI-compatible formats. Falls back to text parsing if the model embeds tool calls in content.

process_turn(agent, user_input; on_event=nothing) -> TurnResult

Process a single conversation turn. Sends the user message to the LLM, executes any tool calls, and returns the final response.

The optional on_event callback receives status updates:

• (:thinking,) — LLM is generating a response
• (:tool_start, name, args) — a tool is about to execute
• (:tool_done, exec) — a tool finished (ToolExecution)

tools_to_ollama_format(tools)

Convert Tool objects to Ollama's native tool calling format.

tools_to_openai_format(tools)

Convert Tool objects to OpenAI function calling format.