Provider failures during chat can't be attributed to a provider or model from the logs

[drioatbu]

When a provider call fails during a normal chat, the log entry is:

error: [api/server/controllers/agents/client.js #sendCompletion] Unhandled error type

No provider, model, or error name. We hit this during a provider capacity incident: users were getting errors, the logs showed only that line, and we had to go to the vendor console to find out which provider and model were rejecting requests.

We patched our instances to log the failure with context. Same event, taken from our production logs:

error: [ProviderError:chat] provider_error source=chat status=429 provider=azureOpenAI model=gpt-5.5 latencyMs=100860 errorName=Error errorCode=rate_limit_exceeded

Every field is known at the catch site. The latency alone told us the SDK had been retrying for 100 seconds before the user saw an error.

Summarization failures already log this via describeProviderError in @librechat/agents (#12737), for example:

error: [agents:summarize] Summarization LLM call failed [bedrock/claude-sonnet-4]: ThrottlingException

with provider, model, errorName and status attached as metadata for JSON output.

Could chat completion, title generation, and the OpenAI-compatible API failures use the same shape? Kept within the getSafeErrorMetadata rules (no response bodies or headers). Format is your call; we used key=value in the message because the default console drops metadata, but the same fields as metadata would work with CONSOLE_JSON.

Happy to open PRs if that helps.