Model prompt and agent hook handler types

[bguidolim]

Phase 2: Prompt & Agent Hook Handler Types

Context

Claude Code now supports two LLM-based hook handler types:

• type: "prompt" — single-turn LLM evaluation returning a yes/no decision
• type: "agent" — spawns a subagent with tool access (Read, Grep, Glob) to verify conditions

New Fields (shared by both types)

Field	Type	Description
type	"prompt" or "agent"	Handler type
prompt	String	Prompt text. $ARGUMENTS placeholder for hook input JSON
model	String?	Model to use (defaults to a fast model)

Timeout Defaults

• prompt: 30 seconds
• agent: 60 seconds

Example (prompt)

hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: prompt
          prompt: "Review this bash command for safety: $ARGUMENTS"
          model: haiku

Files to Modify

• Sources/mcs/Core/Settings.swift — extend HookEntry to support prompt and model fields
• Sources/mcs/ExternalPack/ExternalPackManifest.swift — allow packs to define prompt/agent hooks
• Validation: if type == "prompt" or type == "agent", prompt is required
• Tests

Acceptance Criteria

• HookEntry can represent prompt and agent hooks with all fields
• Pack authors can define prompt/agent hooks in techpack.yaml
• Validation ensures prompt is present for these types
• Backward compatible

Reference

• Hooks reference — Prompt and agent hook fields