Struct ServerConfig 

pub struct ServerConfig {
    pub bind: String,
    pub port: u16,
    pub gateway_url: String,
    pub gateway_key: Option<String>,
    pub model: String,
    pub seed_kb: bool,
    pub max_iterations: u32,
    pub max_tokens: u32,
    pub storage: StorageBackend,
    pub widget_auth_strict: bool,
    pub confirm_tools: Vec<String>,
    pub judge_model: String,
}

Fully-resolved server configuration.

Fields

bind: String

IP address to bind (127.0.0.1 for local dev, 0.0.0.0 in containers).

port: u16

Port to bind.

gateway_url: String

LLM gateway base URL.

gateway_key: Option<String>

Optional gateway API key. None means LLM turns are unavailable and send_message returns a clean error.

model: String

Model id.

seed_kb: bool

Whether to seed the knowledge base with demo docs on startup.

max_iterations: u32

Agent-loop iteration cap per turn.

max_tokens: u32

max_tokens per LLM call.

storage: StorageBackend

Storage backend (drives both the storage adapter and the matching durable admin stores). Defaults to StorageBackend::Memory.

widget_auth_strict: bool

Fail-closed embeddable-widget auth: when true, a session for an agent the WidgetAuthProvider has no policy for is rejected (unknown/unregistered agents can’t be embedded). When false (default), an absent policy is allowed — so the permissive default provider leaves /ws open. Set WIDGET_AUTH_STRICT=1 in front of a real provider. Origin + authContext are always enforced for agents that do have a policy, regardless of this flag.

confirm_tools: Vec<String>

Write-confirmation HITL: tool-name substrings that require human approval before the agent may run them. When non-empty, a turn that calls a matching tool parks and emits a confirm_tool_action_required event; the client resumes it with confirm_tool_action. Read from SMOOTH_AGENT_CONFIRM_TOOLS (comma-separated). Empty (the default) means no tool ever requires confirmation — no turn parks, byte-for-byte unchanged from before HITL. Matched by core’s ConfirmationHook (contains).

judge_model: String

Cheap fast-tier model for the post-turn conversation-workflow judge (SMOODEV-590). Independent of model so the judge stays cheap even when a turn runs on a bigger model. Read from SMOOTH_AGENT_JUDGE_MODEL; defaults to DEFAULT_MODEL (haiku-tier).

Implementations

impl ServerConfig

pub fn from_env() -> Self

Read configuration from the environment, applying documented defaults.

pub fn confirmation_tool_patterns(&self) -> Option<Vec<String>>

The configured write-confirmation tool patterns, or None when none are configured (so the runner installs no ConfirmationHook and the turn behaves exactly as before HITL). Some only when at least one non-empty pattern is set.

pub fn has_llm(&self) -> bool

true when a gateway key is present, so LLM turns can actually run.

pub fn llm_config(&self) -> Option<LlmConfig>

Build the smooth-operator LlmConfig for live turns using the server’s configured (env) gateway key.

Returns None when no gateway key is configured (callers should emit a clean protocol error rather than attempting a turn).

In a multi-tenant flavor the per-turn key comes from a GatewayKeyResolver instead; use llm_config_with_key once the per-org key is resolved.

pub fn llm_config_with_key(&self, key: String) -> LlmConfig

Build the smooth-operator LlmConfig for live turns with an explicit gateway key (gateway URL, model, and limits still come from this config).

This is the per-org seam’s entry point: a GatewayKeyResolver resolves the key for the turn’s org (falling back to the env key), and the resolved key is threaded through here. With the default env resolver this produces exactly the same config as llm_config.

pub fn placeholder_llm_config(&self) -> LlmConfig

Build an LlmConfig without requiring a gateway key, for the test-only path where a MockLlmClient is injected (the scenario-parity corpus). The mock replaces the client built from this config, so its url/key/model are never used to make a network call — this just satisfies the engine’s LlmConfig argument so a keyless deterministic turn can run. Not reachable on the production path (only consulted when chat_provider is Some).

Trait Implementations

impl Clone for ServerConfig

fn clone(&self) -> ServerConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ServerConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.