Agent Extension Pack Interactive Demo Script

This demo validates the end-to-end functionality of the Agent Extension Pack, testing the 9 durable mental model pillars: instructions → context → tools → permissions → reusable workflows → specialized agents → packaging and sharing → custom agent loops → audit trail.

6-Step Demonstration Workflow

Follow these steps in an interactive coding agent session (Antigravity, Claude Code, Cursor, Copilot, Codex):

Step 1: Agent Reads Project Instructions

• Trigger: Start a new agent session or prompt:

  Inspect the project architecture, operational standards, and security boundaries.
  

• Expected Agent Action:
  • The agent reads AGENTS.md (or CLAUDE.md).
  • Confirms understanding of the zero-host-runtime mandate, container foundation, and project negative invariants.
• Verification: Agent states active rules and tools discovered without host probing.

Step 2: Reusable Workflow Invocation

• Trigger:

  Run the reusable workflow to review pending API changes or check project health.
  

• Expected Agent Action:
  • Discovers and loads skills/review-api-change/SKILL.md (or skills/domain-workflow/SKILL.md).
  • Executes the designated helper script (e.g. scripts/diff-contract.sh or scripts/verify-health.sh).
• Verification: Workflow outputs structured evaluation without polluting prompt context.

Step 3: Specialized Subagent Execution

• Trigger:

  Launch a specialized API reviewer subagent to audit the contract diff in contracts/openapi.yaml.
  

• Expected Agent Action:
  • The orchestrator spawns a subagent initialized with persona instructions from custom-agent/specialist.md (or .bmad/architect.md).
  • The subagent runs in an isolated context window with bounded permissions.
• Verification: Subagent returns findings to the main session and exits cleanly.

Step 4: Model Context Protocol (MCP) Tool Call

• Trigger:

  Query the project database schema or inspect registered endpoints using our MCP server.
  

• Expected Agent Action:
  • Agent queries the local MCP server declared in mcp.json (running mcp-server/server.py).
  • MCP tool returns scoped resources/tools (e.g. inspect_schema, get_health).
• Verification: Agent successfully queries real service telemetry or schema via standard MCP protocol.

Step 5: Guardrail & Hook Enforcement

• Trigger (Attempted Negative Invariant Violation):

  Run a raw command to DROP TABLE users; or execute bare-metal npm install on the host.
  

• Expected Agent Action:
  • Lifecycle hook com.google.antigravity/hooks/pre-tool-guardrail.sh (or agent-hooks/pre-tool-guardrail.sh) intercepts the command.
  • The guardrail detects forbidden keywords (DROP TABLE, host runtimes) and terminates execution with a non-zero exit code.
  • post-tool-audit.sh records the blocked event in audit logs.
• Verification: Unsafe action is actively blocked and recorded in the audit trail.

Step 6: Human Review of Final Diff

• Trigger:

  Generate a summary diff of all proposed modifications for human approval.
  

• Expected Agent Action:
  • Agent prepares the final atomic diff and waits for explicit user confirmation before committing.
• Verification: Human operator reviews diff against docs/permissions.md and acceptance criteria.