Struct Agent 

pub struct Agent {

    pub model: Model,
    pub thinking_level: ThinkingLevel,
    pub tools: ToolRegistry,
    pub messages: Vec<Message>,
    pub system_prompt: String,
    pub cwd: PathBuf,
    pub max_turns: u32,
    pub max_tokens: Option<u32>,
    pub role: Option<Role>,
    pub hooks: HookRunner,
    pub api_key: String,
    pub auth_store: Option<Arc<Mutex<AuthStore>>>,
    pub ui: Arc<dyn UserInterface>,
    pub context_config: ContextConfig,
    pub retry_policy: RetryPolicy,
    pub mode: AgentMode,
    pub has_mana_skill: bool,
    pub has_mana_basics_skill: bool,
    pub has_mana_delegation_skill: bool,
    pub guardrail_config: GuardrailConfig,
    pub guardrail_profile: Option<GuardrailProfile>,
    pub lua_tool_loader: Option<LuaToolLoader>,
    pub file_cache: Arc<FileCache>,
    pub checkpoint_state: Arc<CheckpointState>,
    pub file_tracker: Arc<Mutex<FileTracker>>,
    pub anchor_store: Arc<AnchorStore>,
    pub read_max_lines: usize,
    pub cache_options: CacheOptions,
    pub continue_policy: ContinuePolicy,
    pub config: Arc<Config>,
    /* private fields */
}

The core agent — runs the ReAct loop (reason, act, observe).

Fields

model: Model

thinking_level: ThinkingLevel

tools: ToolRegistry

messages: Vec<Message>

system_prompt: String

cwd: PathBuf

max_turns: u32

max_tokens: Option<u32>

role: Option<Role>

hooks: HookRunner

api_key: String

auth_store: Option<Arc<Mutex<AuthStore>>>

Optional auth store for automatic OAuth token refresh before LLM calls. Optional auth store for automatic OAuth token refresh before LLM calls.

ui: Arc<dyn UserInterface>

context_config: ContextConfig

Context management thresholds (wired from Config via AgentBuilder).

retry_policy: RetryPolicy

Retry policy for transient LLM stream failures.

mode: AgentMode

Active agent mode — controls which tools are permitted.

has_mana_skill: bool

Whether a mana skill is available in discovered resources.

has_mana_basics_skill: bool

Whether a mana-basics skill is available in discovered resources.

has_mana_delegation_skill: bool

Whether a mana-delegation skill is available in discovered resources.

guardrail_config: GuardrailConfig

Engineering guardrails config.

guardrail_profile: Option<GuardrailProfile>

Resolved guardrail profile (None = disabled).

lua_tool_loader: Option<LuaToolLoader>

Cloneable Lua extension tool loader inherited from the session/builder.

file_cache: Arc<FileCache>

In-session file content cache, shared across tool calls.

checkpoint_state: Arc<CheckpointState>

Shared checkpoint/file-history state, used to capture destructive edit restore points.

file_tracker: Arc<Mutex<FileTracker>>

Tracks which files have been read; used for staleness and unread-edit warnings.

anchor_store: Arc<AnchorStore>

Session-local anchors emitted by read and consumed by anchored edit mode.

read_max_lines: usize

Max lines the read tool may return before truncating. 0 means unlimited.

cache_options: CacheOptions

Cache options for LLM requests.

continue_policy: ContinuePolicy

Policy for imp-local visible auto-continuation after high-confidence turns.

config: Arc<Config>

Resolved runtime config for tool-specific policy checks.

Implementations

impl Agent

pub fn new(model: Model, cwd: PathBuf) -> (Self, AgentHandle)

pub async fn run(&mut self, prompt: String) -> Result<()>

Run the agent loop with an initial prompt.