Summary

The before_tool_call hook is defined in plugins/hooks.js and exported via createHookRunner(), but it's never called in the actual tool execution flow. This prevents plugins from intercepting tool calls before execution.

Use Case

Security and guardrail plugins need to intercept tool calls to:

• Block operations that exceed trust/budget thresholds
• Enforce risk-based autonomy limits
• Log all tool invocations for audit trails
• Implement zero-trust patterns for AI agents

Current Behavior

• Plugin registers hook via api.registerHook(["before_tool_call"], handler)
• Hook appears in hookNames array when listing plugins
• Hook is never invoked — tools execute without calling runBeforeToolCall

Expected Behavior

Before tool.execute() is called, the hook should fire and allow plugins to:

• Block execution ({ block: true, blockReason: "..." })
• Modify parameters ({ params: {...} })
• Log/observe (return {})

Proposed Fix

In steerable-agent-loop.js (or .ts), inside executeToolCalls(), before the tool.execute() call (~line 254):

// Before tool.execute(), check plugin hooks
if (hookRunner?.hasHooks?.("before_tool_call")) {
    const hookResult = await hookRunner.runBeforeToolCall({
        toolName: toolCall.name,
        toolCallId: toolCall.id,
        params: validatedArgs,
    }, ctx);
    
    if (hookResult?.block) {
        throw new Error(hookResult.blockReason ?? `Tool ${toolCall.name} blocked by plugin`);
    }
    
    if (hookResult?.params) {
        Object.assign(validatedArgs, hookResult.params);
    }
}

result = await tool.execute(toolCall.id, validatedArgs, signal, ...);
Context
Working on [Equilibrium Guard](https://github.com/rizqcon/equilibrium-guard), a zero-trust security layer for AI agents. The plugin infrastructure is excellent — just needs this hook wired up to enable enforcement.