Integrate CAO with KiroCrew through typed controls, MCP Apps, and AG-UI

## Problem

KiroCrew provides a persistent operator workspace with managed conversations, applications, and background work. CAO provides multi-agent fleet execution across CLI providers, including session and terminal lifecycle, delegation, approvals, and fleet observability.

A task initiated in KiroCrew cannot currently launch and supervise a CAO fleet as one continuous workflow. Operators must invoke CAO separately, locate the resulting session, monitor it in another interface, and manually return the outcome to KiroCrew.

The integration should preserve a clear ownership boundary:

- **KiroCrew owns** the operator experience, persistent workspace, and task initiation.
- **CAO owns** fleet execution, terminals, delegation, provider processes, approvals, and authoritative runtime state.

## Proposed direction

Build a CAO-owned reference integration by composing existing surfaces rather than creating another orchestration model or dashboard:

1. Use `cao-ops-mcp` for typed external fleet control from a KiroCrew agent.
2. Render CAO's existing SEP-1865 MCP Apps inside KiroCrew's native MCP Apps host.
3. Reuse AG-UI for advanced fleet events, interrupt/resume, and agent-authored generative UI.
4. Keep a custom KiroCrew App Kit dashboard as a fallback only for a demonstrated compatibility gap.

The initial validation assumes KiroCrew and CAO run on the same macOS or Linux host. Remote connectivity and Windows MCP Apps support are follow-ups.

## Existing capabilities to reuse

CAO already provides:

- `cao-ops-mcp` for external profile discovery, launch, messaging, inspection, and shutdown.
- `ui://cao/dashboard`, `ui://cao/agent`, and `ui://cao/event-stream` MCP App resources.
- MCP App tools including `render_dashboard`, `render_agent_view`, event history/subscription, and `submit_command`.
- Ambient and stock AG-UI planes plus six allow-listed generative UI components.

KiroCrew already provides:

- A native SEP-1865 MCP Apps host.
- Sandboxed inline and side-panel rendering.
- App-to-tool callbacks through its shared stdio MCP gateway.

## Success signals

- A KiroCrew conversation can launch and manage a CAO session through typed MCP operations.
- KiroCrew negotiates and renders all three existing CAO MCP App views.
- The launched session appears using CAO's authoritative identity and state.
- MCP App callbacks and scoped mutations reach the intended CAO operation.
- Live lifecycle changes and one approval flow work end to end.
- Compatibility of CAO's six AG-UI generative UI components is explicitly verified and documented.
- No CAO credential is exposed to browser state or persisted conversation content.

## Key constraints

- Do not treat CAO as a KiroCrew model provider.
- Do not duplicate CAO terminal, delegation, or approval state in KiroCrew.
- Do not build a new KiroCrew fleet dashboard unless a concrete MCP Apps or AG-UI gap is demonstrated.
- Model-initiated external control should use `cao-ops-mcp`; CAO's in-session `handoff`, `assign`, and `send_message` tools still require CAO terminal context.

## Work breakdown

This issue is split into two sub-issues:

- Typed external control through `cao-ops-mcp`.
- MCP Apps rendering and AG-UI interoperability.

## References

- [CAO control planes](https://github.com/awslabs/cli-agent-orchestrator/blob/main/docs/control-planes.md)
- [CAO MCP Apps](https://github.com/awslabs/cli-agent-orchestrator/blob/main/docs/mcp-apps.md)
- [CAO AG-UI](https://github.com/awslabs/cli-agent-orchestrator/blob/main/docs/agui.md)
- [KiroCrew Apps documentation](https://kiro.dev/docs/crew/apps/)
- [KiroCrew repository](https://github.com/kirodotdev/KiroCrew)

Start with docs/control-planes.md, docs/mcp-apps.md, and docs/agui.md, then review the referenced KiroCrew Apps documentation. Define the two sub-issue scopes before implementation and validate typed control, all three MCP App views, lifecycle and approval flows, AG-UI compatibility, authoritative state, and credential handling end to end.

Written by the indexing model from the issue text.