Durable execution: minimal advanced-template + chatbot integration

[dhruv0811]

Summary

Wires the advanced templates (agent-langgraph-advanced, agent-openai-advanced) and the e2e chatbot up to the bridge's LongRunningAgentServer so an in-flight agent run survives a pod restart without losing the user's turn. Companion to bridge PR databricks-ai-bridge#425 (merged).

This is a re-imagined, minimal version of #204 — addressing the feedback that the prior PR's 506 lines (and especially the 272-line Express proxy) were too clunky. The keystone simplification: the express proxy is gone. Vercel AI SDK's databricksFetch (in providers-server.ts) is already the single boundary every agent request flows through, so all the durable-execution glue lives there now.

What the chatbot does

All durable-execution work consolidated into e2e-chatbot-app-next/packages/ai-sdk-providers/src/providers-server.ts:

• Background-mode injection. When API_PROXY points at a long-running server, set body.background = true on streaming requests so the server persists every SSE frame to its durable store and the retrieve endpoint can resume mid-stream.
• Rotated-conversation alias capture. Sniff response.resumed SSE sentinels to capture the bridge's rotated conversation_id (e.g. chat-123::attempt-2), keyed by the original chat id in a module-scope Map. On the next user turn, databricksFetch substitutes the captured alias into the outgoing body.context.conversation_id so the request lands on the post-resume session, not the orphan-poisoned original.
• Auto-resume on stream close. If the SSE stream closes without [DONE], transparently re-stream from GET /responses/{id}?stream=true&starting_after=<seq> (capped at 5 retries). Bytes pass through untouched; we only sniff data frames to track response_id / sequence_number / DONE state.

All three sit inside databricksFetch — no new server route, no separate proxy layer, no new module.

What the templates do

• Cross-turn UI-echo dedup. Both advanced templates' chat handlers dedupe the chatbot's full-history echo against the SDK's own state. LangGraph asks the checkpointer (agent.aget_state(config)); OpenAI treats the session as authoritative whenever non-empty (session.get_items()). When the SDK already holds the prior turns, only the latest user message is forwarded — no synthetic tool events, no SDK-specific wrappers, no bridge involvement. Lives in agent_server/utils.py::deduplicate_input per template.
• Bridge log surfacing. Uvicorn's default logging config drops INFO from non-uvicorn loggers, silently swallowing the durable-execution lifecycle breadcrumbs. Both advanced templates' start_server.py now attach a StreamHandler to the databricks_ai_bridge logger so [durable] resume…, [durable] stale-scan loop start…, and friends reach app stdout.

What is NOT in the templates

• No express proxy. The previous iteration of this PR added a 272-line /invocations proxy in server/src/index.ts to do background-mode injection + alias capture + auto-resume. All three responsibilities collapsed into databricksFetch; the express layer is back to its main-branch shape.
• No per-SDK adapter wrappers. No AsyncDatabricksSession.aget_tuple overrides, no DatabricksSaver.aget_tuple overrides. The bridge contract is "agent owns its session/checkpointer state"; templates abide by it.
• No bridge dependency pin. LongRunningAgentServer already exists in databricks-ai-bridge 0.19.0 (older flavor), so without the new prose-recovery release the templates degrade gracefully — the chatbot's alias capture stays dormant because no response.resumed sentinels arrive. We'll bump the dependency floor when 0.20.0 (containing #425) ships.
• No bundled databricks.yml customizations. Bundle names and app names stay at the template defaults; per-deploy customizations live outside the PR.

Files

Component	File
Background injection + alias capture + auto-resume	e2e-chatbot-app-next/packages/ai-sdk-providers/src/providers-server.ts
LangGraph UI-echo dedup helper	agent-langgraph-advanced/agent_server/utils.py::deduplicate_input
LangGraph dedup wired into stream_handler	agent-langgraph-advanced/agent_server/agent.py
OpenAI UI-echo dedup heuristic (session is authoritative)	agent-openai-advanced/agent_server/utils.py::deduplicate_input
Bridge log surfacing	agent-{langgraph,openai}-advanced/agent_server/start_server.py

Settings

Setting	Default	Rationale
MAX_RESUME_ATTEMPTS (chatbot)	5	Cap auto-resume attempts on SSE close-without-DONE so a permanently-down server doesn't loop forever.

Testing the durable path before 0.20.0 ships

Two pieces of pre-release setup are needed if you want to exercise the new prose-recovery flow end-to-end (after 0.20.0 they go away):

1. Pin to bridge main. Add a temporary [tool.uv.sources] block to either advanced template's pyproject.toml:

   [tool.uv.sources]
   databricks-ai-bridge = { git = "https://github.com/databricks/databricks-ai-bridge.git", branch = "main" }

   Without this, databricks-ai-bridge resolves to the 0.19.0 PyPI release (which still has LongRunningAgentServer but lacks #425's prose-recovery, heartbeat, and attempt_number rotation).

2. Enable the debug kill endpoint. Set LONG_RUNNING_ENABLE_DEBUG_KILL=1 on the deployed app (env block in databricks.yml). This gates POST /_debug/kill_task/{response_id}, the test-only endpoint that simulates a pod crash mid-stream. Leave it unset in production.

Test plan

• agent-langgraph-advanced UI testing on dhruv-lg-claude-durable (Claude Sonnet 4.5) — multi-tool kill mid-deep_research, multi-turn ✅
• agent-openai-advanced UI testing on dhruv-oai-gpt-durable (GPT-5) — same matrix ✅
• agent-openai-advanced UI testing on dhruv-oai-claude-durable (Claude Sonnet 4.5) — same matrix ✅
• Base-template regression: agent-openai-agents-sdk deployed as dhruv-oai-base-newui against the same chatbot branch — confirms the new providers-server.ts doesn't break templates that don't use durable execution ✅
• e2e integration tests still green for all advanced templates

Companion PR

databricks-ai-bridge#425 — LongRunningAgentServer durable prose-recovery + always-rotate (merged).

Known follow-ups (non-blocking)

• Move bridge log surfacing into LongRunningAgentServer.__init__ itself so individual templates don't each need to wire it up. Currently lives in each start_server.py.
• Bump the databricks-ai-bridge dependency floor in both advanced templates once a release containing #425 ships on PyPI.
• Lift the chatbot's auto-resume + alias capture into the upstream @databricks/ai-sdk-provider package once the durable contract stabilizes.

Text.Only.-.Prose.mov

Tool.Calling.Multiturn.-.Prose.mov

[dhruv0811]

@bbqiu do we want to maybe gate this behind some debug env var? (my pr desc mentions the env var used to enable the debug kill endpoint, maybe we can bar the logging behind that too?)