Agent-oriented task metadata

[8bitAlex]

Add optional metadata to task definitions so AI agents can reason about task safety and decide what they can run autonomously vs. what requires user approval.

Proposed schema

tasks:
  test:
    cmd: go test ./...
    agent:
      safe: true           # idempotent, no side effects
      reads: ["./..."]     # paths/globs read
      writes: []           # paths/globs written
      description: "Run unit tests"
  deploy:
    cmd: ./scripts/deploy.sh
    agent:
      safe: false          # has side effects — agent should prompt
      description: "Deploy to production"

Behavior

• Agents (via MCP) check agent.safe before auto-executing
• Unsafe tasks require explicit user confirmation in agent clients that respect the hint
• Metadata mirrors Claude Code's permission model for consistency

Requirements

• Add to JSON Schema (Issue Add Claude Code GitHub Workflow #4)
• MCP server (Issue Initial Structure & Prototype #1) surfaces the metadata in tool descriptions
• Backwards compatible — tasks without agent block default to requiring confirmation

Acceptance criteria

• Schema updated
• MCP tool descriptions include safety metadata
• Documentation explains the safety model
• Example raid.yaml files demonstrate usage
• Tasks lacking an agent: block are surfaced to MCP clients as safe: false by default