[appkit plugin] [2/x] Add @databricks/appkit-agent package

[hubertzub-db]

🥞 Stacked PR

• stack/agent-plugin-p2
• stack/agent-plugin-p3 (diff)
• stack/agent-plugin-p4 (diff)
• stack/agent-plugin-p5 (diff)
• stack/agent-plugin-p6 (diff)
• stack/agent-plugin-p7 (diff)
• stack/agent-plugin-p8 (diff)

---

Summary

Extracts the agent plugin from the AppKit monorepo into a standalone, publishable package at integrations/appkit-agent.

This package provides:

1. AgentInterface — a contract for writing custom agent implementations that speak the OpenAI Responses API format (streaming + non-streaming).
2. StandardAgent — a ready-to-use LangGraph-based ReAct agent with function tools and Databricks-hosted tool integration (Genie, Vector Search, MCP servers).

Key decisions:

• @databricks/appkit is a peer dependency — the plugin integrates via AppKit's public Plugin/toPlugin/PluginManifest API
• LangChain packages (@databricks/langchainjs, @langchain/core, @langchain/langgraph, @langchain/mcp-adapters) are optional peer dependencies — only needed when using StandardAgent, not needed for custom agentInstance implementations
• Internal createLogger replaced with a local minimal console logger to avoid coupling to AppKit internals
• Added CI workflow (appkit-agent.yml) and release workflow (release-appkit-agent.yml)

[smurching]

Looks pretty good (mostly just a copy-over of the code from databricks/appkit#166), but should we also include tests here/those seem missing in this PR?

[hubertzub-db]

but should we also include tests here/those seem missing in this PR?

@smurching agree, let's bring over tests in this PR. Initially I wanted to bring them over in latter parts but there's no guarantee we'll get there soon - so it sounds good to colocate them here.
I've added appkit test harness, moved unit tests and added one larger integration-ish test.

[smurching]

@hubertzub-db some findings from Claude Code review of the latest:

Gaps / Issues

1. invoke-handler.ts:126 — chatHistory always includes the last message before the final user turn, not "all but last". If the input has
   only one message (a single user turn), chatHistory is [] — that's fine. But if the last message is an assistant turn (e.g. [user,
   assistant]), the assistant message gets passed as history and userMessages.at(-1) is the only user message. There's no test for an input
   that ends with an assistant message — the handler returns 400 in that case actually (no user message check passes), but the chatHistory
   still includes all items. Worth a test.
2. agent.ts — abortActiveOperations() has no test. The MCP client close path (both success and error) is untested.
3. agent.ts — exports().invoke uses [...messages].reverse().find(m => m.role === "user") which picks the last user message, but
   chatHistory is messages.slice(0, -1) which slices off the last message (not necessarily the last user message). For example, with [user,
   user, assistant], invoke would pick the second user message correctly, but chatHistory would be [user, user] — the last assistant turn is
   excluded. This looks like a bug, and there's no test for this case.
4. standard-agent.ts:178 — on_chat_model_stream with non-string content is silently skipped (e.g. array content from tool-use chunks).
   There's no test verifying this doesn't leak unexpected events.
5. invoke-handler.test.ts — non-streaming agent error path is untested. The outer catch at line 155 (res.status(500)) is never exercised.
6. hosted-tools.test.ts — resolveHostedTools([]) empty array case is missing, though trivially covered by the "resolves multiple" test
   implicitly.

The most significant issues are #3 (likely bug in exports().invoke history slicing) and #2 (no abortActiveOperations coverage). I'd recommend requesting fixes for those before merging.

I think the conclusions generally make sense, WDYT? Case #3 seems unlikely but could happen e.g. if the agent loop is interrupted, etc, and doesn't hurt to have a test for #2. But otherwise this looks good!

[hubertzub-db]

@smurching
Thanks for the feedback, addressed following issues in a36b8f8:

1. this repeats pattern from the original code, however let me reinforce the logic there + add tests
2. added test
3. fixed and added test
4. added test