fix(vercel-ai): prevent tool call span map memory leak

[lithdew]

Tool calls were only cleaned up on tool errors, causing unbounded retention in tool-heavy apps (and potential OOMs when inputs/outputs were recorded). Store only span context in the global map and clean up on successful tool results; add tests for caching/eviction.

Before submitting a pull request, please take a look at our
Contributing guidelines and verify:

• If you've added code that should be tested, please add tests.
• Ensure your code lints and the test suite passes (yarn lint) & (yarn test).
• Link an issue if there is one related to your pull request. If no issue is linked, one will be auto-generated and linked.

Closes #issue_link_here

[lithdew]

Any chance for a review? cc @nicohrubec @sergical @RulaKhaled

[nicohrubec]

Thanks for contributing! Generally looks good to me. The cleanup on successful tool results is definitely missing and there also seems to be no reason to store the full spans in the map. Do you think we really need the LRUCache? I get that the idea is to cap the amount of context we store, but I would rather not impose any arbitrary limits.

Would you like to work on this or should we take over?

[lithdew]

The LRU cache isn't strictly necessary — it was just a safeguard to cap memory usage in case spans aren't cleaned up for some reason. Happy to drop it in favor of a plain map with proper cleanup on both success and error paths. I can make the change, or feel free to take it over — either works for me.

[nicohrubec]

@lithdew If you could get rid of the LRUCache that would be great. I'll give it another look then

[lithdew]

Made the changes — would appreciate the re-review 🙏. cc @nicohrubec

[lithdew]

Hi apologies, wanted to ping on an update for this - it has been causing OOM's for a number of developers I work with.

[nicohrubec]

@lithdew I am on it. I will push a few updates and then we should be good to go.

[sentry]

Bug: The onSuccess callback, which cleans up toolCallSpanContextMap, is not called when a wrapped Vercel AI SDK method throws an error, causing a memory leak.
_{Severity: CRITICAL}

Suggested Fix

The cleanup logic should be executed regardless of whether the operation succeeds or fails. Move the cleanup logic to a finally-equivalent path, such as the onFinally callback in handleCallbackErrors, to ensure it runs in both success and error scenarios. Alternatively, call the cleanup logic from the onError handler as well.

Prompt for AI Agent

Review the code at the location below. A potential bug has been identified by an AI
agent.
Verify if this is a real issue. If it is, propose a fix; if not, explain why it's not
valid.

Location: packages/node/src/integrations/tracing/vercelai/instrumentation.ts#L265-L271

Potential issue: The cleanup logic for tool call span contexts,
`processToolCallResults`, is only executed via the `onSuccess` callback of
`handleCallbackErrors`. When a wrapped Vercel AI SDK method (e.g., `generateText`,
`streamText`) throws an error, the `onSuccess` callback is never invoked. As a result,
the span contexts for any tool calls associated with the failed operation are never
removed from the `toolCallSpanContextMap`. In applications with frequent tool call
errors, this leads to unbounded memory growth as the map retains stale context objects
indefinitely.