feat: support conversation history directly in AI Provider model runners

[jsonbailey]

Summary

Port the multi-turn conversation history pattern from the Python AI SDK provider runners to the js-core AI SDK provider runners. Mirrors the behaviour shipped in python-server-sdk-ai #166.

Each provider model runner now keeps an internal conversation history that is seeded from the AI config's messages on construction and grows with each successful call. On every run() invocation:

1. The user prompt is appended to the existing history before being sent to the model.
2. If the call succeeds and produces non-empty content, the user prompt and the assistant's reply are persisted to the history.
3. If the call fails or returns empty content, the history is left unchanged so the next call can retry cleanly.

The public run(input) interface is unchanged and still returns the same RunnerResult. The history is purely internal state on the runner instance.

Per-provider approach

• OpenAI (OpenAIModelRunner.ts) — private LDMessage[] history, mirroring the Python OpenAI runner. The OpenAI Chat Completions API has no native conversation state.
• LangChain (LangChainModelRunner.ts) — uses InMemoryChatMessageHistory from @langchain/core/chat_history, the JS equivalent of langchain_core.chat_history. This mirrors the Python LangChain runner.
• Vercel (VercelModelRunner.ts) — private ModelMessage[] history (Vercel AI SDK native v5 type). No analogue exists in the Python SDK; chose ModelMessage[] to avoid re-converting on every call.

Behaviour change for multi-turn callers (medium risk)

Callers that invoke run() multiple times on the same runner instance will now accumulate conversation history across calls. Token usage will grow with each turn since the full history is sent on every call. Callers that want stateless behaviour should construct a new runner per call.

Test plan

• Multi-turn accumulation across two successful run() calls (per provider)
• No accumulation when run() throws (per provider)
• No accumulation when run() returns empty content (per provider)
• Config / system messages still prepended ahead of accumulated history on every call (per provider)
• Existing tests still pass (43 OpenAI / 42 LangChain / 24 Vercel)
• yarn workspaces foreach -pR --topological-dev --from '@launchdarkly/server-sdk-ai' run build succeeds
• Lint passes for each touched package

---

Note

Medium Risk
Changes run() semantics for repeated calls by accumulating prior prompts/responses, which can increase token usage and alter model outputs; failures/empty content are explicitly excluded from history to reduce runaway state.

Overview
Adds stateful multi-turn support to the LangChainModelRunner, OpenAIModelRunner, and VercelModelRunner by keeping an internal conversation history seeded from config.messages and appending it to every run() call.

History is only persisted after successful calls with non-empty content; exceptions or empty/multimodal responses leave history unchanged to allow clean retries. New unit tests in each provider verify accumulation behavior and that config/system messages stay prepended ahead of the growing history.

^{Reviewed by Cursor Bugbot for commit b74c851. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-actions]

@launchdarkly/js-sdk-common size report
This is the brotli compressed size of the ESM build.
Compressed size: 26281 bytes
Compressed size limit: 29000
Uncompressed size: 128971 bytes

[github-actions]

@launchdarkly/js-client-sdk-common size report
This is the brotli compressed size of the ESM build.
Compressed size: 38487 bytes
Compressed size limit: 39000
Uncompressed size: 211236 bytes

[github-actions]

@launchdarkly/js-client-sdk size report
This is the brotli compressed size of the ESM build.
Compressed size: 31906 bytes
Compressed size limit: 34000
Uncompressed size: 113658 bytes

[github-actions]

@launchdarkly/browser size report
This is the brotli compressed size of the ESM build.
Compressed size: 179498 bytes
Compressed size limit: 200000
Uncompressed size: 830837 bytes