Asuran published a DEV Summer Bug Smash submission on August 12, 2026, describing a fix for Sentry's JavaScript SDK after Gemini chat traces stopped recording the configuration used for model calls. The bug affected Sentry's Google GenAI auto-instrumentation, which is meant to attach the model, generation config and system instruction to each Gemini span (a timed trace record for one operation).

The regression came from getsentry/sentry-javascript#20086 and followed refactor #19990. Google GenAI chat usage starts with `chats.create({ model, config })`, where `config` can include temperature, top_p, top_k, max output tokens, penalties, tool list and system instruction; actual model calls then happen through `chat.sendMessage(...)` or `chat.sendMessageStream(...)`. Sentry removed the `chats.create()` span because it was not a real model call, but that also removed the place where the create-time config had been captured.

After the refactor, `chat.sendMessage()` and `chat.sendMessageStream()` spans only saw the per-call message. Attributes including `gen_ai.request.temperature`, `top_p`, `top_k`, `max_tokens`, penalties, `available_tools` and `gen_ai.system_instructions` disappeared from traces. The root cause was a deep proxy that re-proxied the chat object returned by `chats.create()` but discarded that call's arguments, leaving the message-span code without the chat config.

The proposed fix is PR getsentry/sentry-javascript#23316 on branch `fix/google-genai-chat-config-attrs`. It adds `mergeChatCreateParams`, passes the `chats.create()` arguments through the proxy, and builds message span attributes from merged parameters. The create-time config is used as the default, while a per-message config replaces it wholesale for that request, matching `@google/genai` behavior as `params.config ?? chat.config`.

The real call to Google still uses its original arguments, so the change does not alter Gemini requests. Non-chat calls such as `models.generateContent` do not receive chat context and are left unchanged. The author also deliberately left create-time `history` off message spans because it is conversation seed rather than per-message input.

Testing added a new file with four cases: config attached to `sendMessage` spans, config attached to `sendMessageStream` spans, per-message config replacing create config, and no leakage into `models.generateContent` spans. On the unfixed code, three of the four cases failed because attributes were `undefined`; with the fix, all four passed. The `@sentry/server-utils` suite increased from 38 files / 335 tests to 39 files / 339 tests, all passing. `oxlint --type-aware`, `oxfmt --check` and source-type `tsc` also passed, while unrelated pre-existing test-config `tsc` errors remained.

The author verified the fix against the live Gemini API using `@google/genai` v1.20.0, real Sentry tracing and model `gemini-2.5-flash`. A test call used temperature 0.8, topP 0.9, topK 40, maxOutputTokens 512, and system instruction "You are a friendly robot who likes to be funny." Before the fix, `gen_ai.request.model` was present but temperature, top_p, top_k, max_tokens and system instructions were missing. After the fix, the same span recorded model `gemini-2.5-flash`, temperature 0.8, top_p 0.9, top_k 40, max_tokens 512 and the system instruction. The author disclosed that Claude from Anthropic assisted development, while design, review and verification were done by the author.

Asuran published a DEV Summer Bug Smash submission on August 12, 2026, describing a fix for Sentry's JavaScript SDK after Gemini chat traces stopped recording the configuration used for model calls. The bug affected Sentry's Google GenAI auto-instrumentation, which is meant to attach the model, generation config and system instruction to each Gemini span (a timed trace record for one operation).

The regression came from getsentry/sentry-javascript#20086 and followed refactor #19990. Google GenAI chat usage starts with `chats.create({ model, config })`, where `config` can include temperature, top_p, top_k, max output tokens, penalties, tool list and system instruction; actual model calls then happen through `chat.sendMessage(...)` or `chat.sendMessageStream(...)`. Sentry removed the `chats.create()` span because it was not a real model call, but that also removed the place where the create-time config had been captured.

After the refactor, `chat.sendMessage()` and `chat.sendMessageStream()` spans only saw the per-call message. Attributes including `gen_ai.request.temperature`, `top_p`, `top_k`, `max_tokens`, penalties, `available_tools` and `gen_ai.system_instructions` disappeared from traces. The root cause was a deep proxy that re-proxied the chat object returned by `chats.create()` but discarded that call's arguments, leaving the message-span code without the chat config.

The proposed fix is PR getsentry/sentry-javascript#23316 on branch `fix/google-genai-chat-config-attrs`. It adds `mergeChatCreateParams`, passes the `chats.create()` arguments through the proxy, and builds message span attributes from merged parameters. The create-time config is used as the default, while a per-message config replaces it wholesale for that request, matching `@google/genai` behavior as `params.config ?? chat.config`.

The real call to Google still uses its original arguments, so the change does not alter Gemini requests. Non-chat calls such as `models.generateContent` do not receive chat context and are left unchanged. The author also deliberately left create-time `history` off message spans because it is conversation seed rather than per-message input.

Testing added a new file with four cases: config attached to `sendMessage` spans, config attached to `sendMessageStream` spans, per-message config replacing create config, and no leakage into `models.generateContent` spans. On the unfixed code, three of the four cases failed because attributes were `undefined`; with the fix, all four passed. The `@sentry/server-utils` suite increased from 38 files / 335 tests to 39 files / 339 tests, all passing. `oxlint --type-aware`, `oxfmt --check` and source-type `tsc` also passed, while unrelated pre-existing test-config `tsc` errors remained.

The author verified the fix against the live Gemini API using `@google/genai` v1.20.0, real Sentry tracing and model `gemini-2.5-flash`. A test call used temperature 0.8, topP 0.9, topK 40, maxOutputTokens 512, and system instruction "You are a friendly robot who likes to be funny." Before the fix, `gen_ai.request.model` was present but temperature, top_p, top_k, max_tokens and system instructions were missing. After the fix, the same span recorded model `gemini-2.5-flash`, temperature 0.8, top_p 0.9, top_k 40, max_tokens 512 and the system instruction. The author disclosed that Claude from Anthropic assisted development, while design, review and verification were done by the author.