Feature Request: Support AG-UI or ChatCompletion streaming format for InvokeHarness API

[waltmayf]

Description

Currently the managed harness InvokeHarness API returns a bespoke Bedrock Converse-shaped SSE format
(HarnessContentBlockStartEvent, etc.). Front-end integrations (chat UIs, etc.) expect AG-UI or
ChatCompletion streaming formats. This requires a custom transport adapter on every client.

Request: Add a --protocol or similar option to InvokeHarness to emit ag-ui or ChatCompletion
events natively, removing the need for client-side format translation.

Acceptance Criteria

• InvokeHarness accepts a protocol parameter with valid values: bedrock-converse (default, existing behavior), ag-ui, openai-chat-completions
• When protocol=ag-ui is specified, the streaming response emits valid AG-UI SSE events (TextMessageStart, TextMessageContent, TextMessageEnd, ToolCallStart, ToolCallEnd, RunFinished, etc.) conforming to the AG-UI spec
• When protocol=openai-chat-completions is specified, the streaming response emits OpenAI-compatible chat.completion.chunk SSE events (data: {"choices":[{"delta":{...}}]})
• When protocol is omitted or set to bedrock-converse, behavior is identical to current (no regression)
• Tool call events (MCP or inline) are correctly represented in each protocol format
• A front-end chat UI (e.g., Vercel AI SDK, AG-UI client) can connect directly to InvokeHarness with protocol=ag-ui or protocol=openai-chat-completions without a custom transport adapter layer
• Documentation updated with protocol parameter options, examples, and a migration guide from the current bespoke format
• SDK clients (Python, TypeScript) expose the protocol parameter in their invoke_harness / invokeHarness wrappers

Additional Context

No response