Add experimental Session Store SDK adapters

[annzhang-db]

Summary

This draft prototypes how third-party agent SDKs can use the proposed Databricks Session Store REST API.

• adds a shared async REST client with Databricks authentication, pagination, LiteSwap routing, and Session/SessionEvent operations
• adds databricks_openai.agents.DatabricksSession, implementing the OpenAI Agents SDK Session protocol
• adds DatabricksClaudeSessionStore, implementing Claude Agent SDK transcript mirroring, listing, deletion, and subagent transcript discovery
• supports a separate Conversations API base URL for the Databricks OpenAI client
• documents branch installation and runnable OpenAI/Claude integration shapes

Why this is a draft

The REST resource names, endpoint prefix, and idempotency contract are still being finalized. The public classes are marked experimental so we can use this branch in Databricks Apps while continuing to make breaking prototype changes.

API design findings

• OpenAI maps directly to list, append, pop, and clear operations.
• Claude storage identity is (project_key, session_id, optional subpath); the adapter includes the project hash in the physical Databricks session ID to avoid cross-project collisions.
• Claude subagent subpath values map naturally to child Sessions.
• Claude retries UUID-bearing transcript entries. The prototype performs best-effort read-before-append deduplication, but atomic retry safety still requires a caller idempotency key per event or equivalent conditional append semantics.

How do you know it works?

The new tests cover authenticated request construction, URL encoding, pagination, errors, all event mutations, OpenAI item mapping and protocol conformance, Claude UUID deduplication, project isolation, subagent child sessions, listing, and cascade deletion.

• Core suite: 104 passed, 54 skipped, 1 environment-bound credential test deselected.
• OpenAI unit suite: 179 passed, 1 skipped, 2 environment-bound credential tests deselected.
• Ruff formatting and lint pass for the core and OpenAI packages.
• Targeted type checks pass; full package checks report only pre-existing unused-ignore warnings.
• A live LiteSwap smoke test successfully created, appended, listed, popped, cleared, and deleted a session.
• Live adapter smoke tests round-tripped an OpenAI item and verified Claude UUID deduplication, project listing, and subagent discovery.

[aravind-segu]

we should use DatabricksOpenAI here right so workspace client authorization flows through?