Phase 1 — Bug fixes:
- Fix generic type extraction (List[str], Optional[int], Dict) in @app.tool
- Fix Pydantic Callable crash with ConfigDict(arbitrary_types_allowed=True)
- Replace deprecated on_event("startup") with lifespan context manager
- Replace print() with structured logging
- Fix missing setup_mcp_server export from mcp module
- Delete empty orchestration module (contradicts Direction A)
- Expand public API: UCAgentRegistry, UCAgentSpec, MCPServerConfig, etc.

Phase 2 — Test coverage (71 tests, all mocked):
- test_agent_app: tools, generics, MCP toggle, UC toggle, endpoints
- test_a2a_client: fetch, send, stream, errors, timeouts
- test_agent_discovery: workspace scanning, capabilities parsing
- test_uc_registry: register, get, list, delete, tag filtering
- test_mcp_server: JSON-RPC tools/list, tools/call, server/info
- test_uc_functions: type mapping, function conversion, discovery

Phase 3 — Documentation (17 pages):
- 4 essential pages: installation, first-agent, agent-app guide, tools guide
- 11 stub pages for remaining mkdocs nav entries
- README: UC/MCP marked done, orchestration removed from roadmap

Phase 4 — Polish:
- Add hello_agent.py minimal example
- Update full_featured_agent.py to lifespan pattern
- Add py.typed PEP 561 marker
- Wire __version__ to importlib.metadata
- Add respx dev dep and pytest asyncio_mode config

…and runtime lineage

Adds the complete dashboard SPA (React + FastAPI) with three new features:
- Batch UC registration: one-click "Register All in UC" from GovernanceTab
- Runtime lineage: observed edges from chat traces shown as cyan dashed lines
- Auto-refresh: lineage and governance tabs refresh automatically after scan

Backend: register_all_agents(), ingest_trace(), POST /api/uc/register-all,
POST /api/lineage/observe, enhanced scan response.

Frontend: new types, API functions, fire-and-forget observe in useChat,
lastScanTimestamp in useAgents for auto-refresh, updated LineageGraph
with observed edge rendering, legend entries, and result banners.

Three bugs discovered during live E2E testing with real UC catalog:

1. registered_models.create() requires short name, not fully-qualified
   3-part name — changed name=full_name to name=spec.name with separate
   catalog_name/schema_name params.

2. RegisteredModelsAPI has no set_tag/list_tags methods — replaced
   tag-based metadata with comment-based storage using
   ---AGENT_META--- JSON marker in the model comment field.

3. list_agents() failed with "Cannot have empty schema if catalog is
   set" — fixed to iterate schemas first, then list models per schema.

Supervisor agents now report routing decisions via _routing metadata
in their MCP tool responses. The dashboard extracts this to create
observed_calls_agent edges in the lineage graph (dashed cyan lines).

Changes:
- supervisor agent.py: track _last_routing with sub_agent name,
  fix mlflow compatibility (OutputItem, async tool invocation)
- supervisor app.py: include _routing in route_query response
- scanner.py: extract _routing from MCP tool responses into trace
- governance.py: parse routing.sub_agent in ingest_trace() to
  create observed agent-to-agent edges

Supervisor agent now reports tables_accessed per sub-agent route,
creating observed_reads_table edges in the lineage graph. Chat
endpoint auto-ingests traces so edges appear without frontend
round-trip. Merge function creates missing table/agent nodes.

Complete rewrite to agent platform architecture:

- SDK: add_agent_card() + add_mcp_endpoints() helpers (replaces AgentApp-only pattern)
- CLI: databricks-agents deploy/status/destroy/dashboard/platform commands
- Deploy: multi-agent deploy from agents.yaml (topo sort, wiring, permissions)
- Dashboard: discovery, governance, lineage, system builder wizard
- System Builder: 4-step wizard (select, wire, configure, deploy) with dagre
  auto-layout, custom ReactFlow nodes/edges, async deploy with polling
- Registry: Unity Catalog registration with comment-based metadata
- MCP: decoupled JSON-RPC server with tool exposure
- Examples: hello-world, research-agent, data-tools, supervisor patterns
- Tests: 109 passing (SDK, dashboard, system builder, deploy)