Summary

The before_tool_call plugin hook currently receives toolName and params, but not the model's output that preceded the tool call. This limits the ability to implement compliance patterns.

Use Case

I want to enforce a behavioral pattern where the model must output a classification statement (e.g., CLASSIFY: [task] → Model: [cheap/expensive]) before every tool call. A plugin could:

1. Listen to before_tool_call
2. Check if the preceding model output contains the required pattern
3. Block the tool call with a helpful message if missing

Current Limitation

// Current event shape
PluginHookBeforeToolCallEvent = {
  toolName: string;
  params: Record<string, unknown>;
};

No access to what the model said before the tool call.

Proposed Enhancement

Add optional context about the preceding model output:

PluginHookBeforeToolCallEvent = {
  toolName: string;
  params: Record<string, unknown>;
  // New fields
  precedingText?: string;      // Model output before this tool call
  messageId?: string;          // Reference to the full message if needed
};

Alternatives Considered

• Behavioral enforcement only: Works but relies on the model remembering the rule
• System prompt injection: Already possible via before_agent_start, but can't enforce

Impact

This would enable plugins to implement various compliance/governance patterns:

• Cost classification enforcement
• Tool usage audit trails with context
• Custom approval workflows based on what the model is trying to do

Happy to discuss implementation approach if helpful.