Struct Config 

#[non_exhaustive]
pub struct Config {

    pub model: String,
    pub base_url: String,
    pub api_key: Option<String>,
    pub api_key_env: String,
    pub system_prompt: String,
    pub temperature: Option<f32>,
    pub max_tokens: Option<u32>,
    pub max_iterations: usize,
    pub effort: Option<String>,
    pub response_format: Option<Value>,
    pub extra_body: Map<String, Value>,
    pub max_total_output_tokens: Option<u64>,
    pub max_tool_output_bytes: Option<usize>,
    pub cwd: PathBuf,
    pub additional_dirs: Vec<PathBuf>,
    pub load_project_context: bool,
    pub sandbox: SandboxPolicy,
    pub approval: ApprovalPolicy,
    pub auto_approved_tools: HashSet<String>,
    pub approval_handler: Option<Box<dyn Fn(&ToolCall) -> bool + Send + Sync>>,
    pub pre_tool_hook: Option<Box<dyn Fn(&str, &Value) -> Option<String> + Send + Sync>>,
    pub post_tool_hook: Option<Box<dyn Fn(&str, &str, bool) + Send + Sync>>,
    pub prompts: HashMap<String, String>,
    pub compact_after_messages: Option<usize>,
    pub tool_overrides: HashMap<String, ToolOverride>,
    pub extra_headers: HashMap<String, String>,
    pub event_sink: Option<EventSink>,
}

Everything that shapes an crate::Agent: the model and endpoint, the credentials, sampling parameters, the system prompt, and per-tool overrides.

Build one with Config::builder.

Fields (Non-exhaustive)

Non-exhaustive structs could have additional fields added in future. Therefore, non-exhaustive structs cannot be constructed in external crates using the traditional Struct { .. } syntax; cannot be matched against without a wildcard ..; and struct update syntax will not work.

model: String

Model identifier as understood by the endpoint, e.g. anthropic/claude-opus-4-8 or openai/gpt-5 on OpenRouter.

base_url: String

Base URL of the OpenAI-compatible endpoint (no trailing /chat/...).

api_key: Option<String>

Explicit API key. If None, Self::api_key_env is consulted.

api_key_env: String

Environment variable to read the API key from when Self::api_key is unset.

system_prompt: String

System prompt prepended to every conversation.

temperature: Option<f32>

Optional sampling temperature.

max_tokens: Option<u32>

Optional output token cap.

max_iterations: usize

Maximum number of model/tool iterations per crate::Agent::send call.

effort: Option<String>

Reasoning/effort level sent to the model (reasoning_effort).

response_format: Option<Value>

Structured-output constraint (response_format), e.g. a json_schema.

extra_body: Map<String, Value>

Extra request-body fields merged in (provider-native passthrough: prompt-cache controls, provider-specific knobs).

max_total_output_tokens: Option<u64>

Optional cap on cumulative output tokens across one crate::Agent::send loop; the loop stops once exceeded.

max_tool_output_bytes: Option<usize>

Max bytes of a single tool result fed back into the conversation. Output beyond this is truncated with a notice, so one runaway command (a huge log, a binary dump) can’t explode the context window. None disables the cap. Defaults to 100 KB.

cwd: PathBuf

Working directory tools operate within.

additional_dirs: Vec<PathBuf>

Additional roots beyond cwd (the analog of --add-dir / multi-root): searched for project-context files and available to tools.

load_project_context: bool

Whether to auto-load CLAUDE.md / AGENTS.md into the system prompt.

sandbox: SandboxPolicy

Filesystem confinement applied to write-capable tools.

approval: ApprovalPolicy

When the agent must seek approval before running a tool.

auto_approved_tools: HashSet<String>

Tools that never require approval under ApprovalPolicy::OnRequest.

approval_handler: Option<Box<dyn Fn(&ToolCall) -> bool + Send + Sync>>

Consulted when a tool call needs approval; None denies by default.

pre_tool_hook: Option<Box<dyn Fn(&str, &Value) -> Option<String> + Send + Sync>>

Runs before each tool executes; may block the call.

post_tool_hook: Option<Box<dyn Fn(&str, &str, bool) + Send + Sync>>

Runs after each tool executes (observational).

prompts: HashMap<String, String>

Named prompt templates (skills / slash commands). A user message of the form /<name> <args> is expanded to the template with {args} filled.

compact_after_messages: Option<usize>

If set, the conversation is compacted once it grows beyond this many messages (older middle turns are summarized into one marker), keeping the system prompt and the most recent turns.

tool_overrides: HashMap<String, ToolOverride>

Per-tool enable/disable + description overrides, keyed by tool name.

extra_headers: HashMap<String, String>

Extra HTTP headers sent with every request (e.g. OpenRouter’s HTTP-Referer / X-Title attribution headers).

event_sink: Option<EventSink>

Optional sink for streaming crate::AgentEvents.

Implementations

impl Config

pub fn builder() -> ConfigBuilder

Start building a Config from defaults.

pub fn tool_enabled(&self, name: &str) -> bool

Whether a tool is enabled given the overrides (defaults to enabled).

pub fn needs_approval(&self, tool: &str) -> bool

Whether a tool call requires approval before it runs, given the policy and the auto-approve allowlist.

pub fn tool_description<'a>(&'a self, name: &str, builtin: &'a str) -> &'a str

The effective description for a tool, applying any override.

impl Config

pub fn from_profile_file( path: impl AsRef<Path>, profile: &str, ) -> Result<ConfigBuilder>

Load a named profile from a JSON config file into a builder. Layered: start from defaults, then apply the named profile’s set fields.

Trait Implementations

impl Default for Config

fn default() -> Self

Returns the “default value” for a type.