feat: replace LangChainProvider with Runner protocol implementation (AIC-2388)

[jsonbailey]

Summary

Replaces the legacy LangChainProvider extension model with the Runner protocol for the LangChain provider package. This is a breaking change — LangChainProvider is removed.

Removed:

• LangChainProvider class and its tests — callers must migrate to the new factory/runner pattern

Added:

• LangChainModelRunner — Runner for chat completions; run(input: string | LDMessage[], outputType?) returns RunnerResult { content, metrics, raw, parsed? }. Structured output is parsed via withStructuredOutput.
• LangChainAgentRunner — Runner for a single agent. Runs a tool-calling loop using LangChain's native bindTools and ToolMessage. Tool definitions come from config.model.parameters.tools; tool callables come from a caller-supplied ToolRegistry. Populates LDAIMetrics.toolCalls with the names of tools invoked.
• LangChainRunnerFactory — extends AIProvider; exposes createModel(config) and createAgent(config, tools?). Applies OpenTelemetry instrumentation (via @traceloop/instrumentation-langchain) fire-and-forget in the constructor if available.

Renamed / moved:

• langchainHelper.ts → LangChainHelper.ts (casing fix)
• createLangChainModel moved from LangChainRunnerFactory into LangChainHelper and exported from the package

Renamed:

• mapProvider → mapProviderName in helper exports

Test plan

• yarn workspace @launchdarkly/server-sdk-ai-langchain test — all tests pass
• yarn workspace @launchdarkly/server-sdk-ai-langchain lint — clean
• yarn workspace @launchdarkly/server-sdk-ai-langchain run build — clean

🤖 Generated with Claude Code

---

Note

High Risk
High risk because it removes the exported LangChainProvider API and replaces it with new Runner/factory entrypoints, requiring downstream migration and potentially affecting tool-calling and metrics behavior.

Overview
Breaking change: removes LangChainProvider (and its tests) and switches the package API to the LaunchDarkly Runner protocol.

Adds LangChainModelRunner for chat completions (string + optional structured output) and LangChainAgentRunner for tool-calling agents (binds tools from config, executes a caller-supplied ToolRegistry, aggregates token usage, and records invoked tool names). Introduces LangChainRunnerFactory to build these runners and keep the existing optional OpenTelemetry instrumentation, and exports new helpers (createLangChainModel, message conversion, provider-name mapping, and usage/metrics extraction) from index.ts.

^{Reviewed by Cursor Bugbot for commit 2479fb2. Bugbot is set up for automated code reviews on this repo. Configure here.}

---

accidental

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 2479fb2. Configure here.}

[cursor]

Agent executes tools on final iteration without consuming results

Medium Severity

When MAX_ITERATIONS is exhausted, the loop still executes all tool calls on the final iteration (i=24) even though their results are never sent back to the model. Tools with side effects (database writes, external API calls, mutations) will be triggered unnecessarily, and the latency of those calls is wasted. The loop structure doesn't check whether it's on the last iteration before calling _executeTool. This was flagged in PR discussion by @joker23.

 

^{Reviewed by Cursor Bugbot for commit 2479fb2. Configure here.}