Steering

Steering is TraitClaw’s safety and guidance system. It consists of three complementary traits that work together to keep agents safe, guided, and observable.

The Three Pillars

Component	Purpose	When it runs
Guard	Prevent dangerous actions	Before tool execution
Hint	Inject guidance into context	Before LLM generation
Tracker	Collect metrics and adapt	After each step

One-Line Setup

use traitclaw_steering::Steering;

let agent = Agent::builder()
    .provider(provider)
    .steering(Steering::auto())  // All guards + hints + tracking
    .build()?;

Built-in Guards

Guard	Protection
ShellDenyGuard	Blocks shell command execution
LoopDetectionGuard	Detects infinite reasoning loops
RateLimitGuard	Limits requests per time window
ToolBudgetGuard	Limits total tool calls per session
ContentFilterGuard	Filters harmful content patterns
PromptInjectionGuard	Detects prompt injection attempts
WorkspaceBoundaryGuard	Restricts file access to workspace

Built-in Hints

Hint	Guidance
BudgetHint	Injects remaining token budget awareness
SystemPromptReminder	Re-injects system prompt periodically
TruncationHint	Warns when context is being truncated
TeamProgressHint	Reports multi-agent team progress

Custom Guards

use traitclaw_core::Guard;

struct MaxCostGuard {
    max_dollars: f64,
    spent: AtomicF64,
}

#[async_trait]
impl Guard for MaxCostGuard {
    async fn check(&self, action: &AgentAction) -> GuardResult {
        if self.spent.load() > self.max_dollars {
            GuardResult::Deny("Cost budget exceeded".into())
        } else {
            GuardResult::Allow
        }
    }
}