fix(bedrock): Log text and usage for streamed InvokeModel calls to OpenAI models

[alexcastgonz (pcbeingused333)]

OpenAI models on Bedrock (openai.gpt-oss-20b-1:0, openai.gpt-oss-120b-1:0, …) stream Chat Completions chunks from InvokeModelWithResponseStream. The span for such a call ends up with no token metrics, and its output is a dump of the first 20 raw chunks instead of the text.

Two things in aggregateInvokeModelResponseStreamChunks cause this:

• Usage. gpt-oss chunks have no usage field (unless the caller asks for stream_options.include_usage). The counts only arrive in amazon-bedrock-invocationMetrics on the last chunk, and the aggregator reads usage, message.usage and metadata.usage, but never that.
• Text. extractTextFromJsonLike has no case for choices[].delta.content, so text is empty and the output falls back to {chunk_count, chunks}.

Here's a real openai.gpt-oss-20b-1:0 stream (27 chunks, us-east-1) replayed through wrapBedrockRuntime from braintrust 3.37.2:

logged output:  {"chunk_count":27,"chunks":[{"choices":[{"delta":{"content":"","role":"assistant"}, ...
logged metrics: {"time_to_first_token":0.014}
last chunk:     "amazon-bedrock-invocationMetrics": {"inputTokenCount":126,"outputTokenCount":125, ...}

Changes

• extractTextFromJsonLike reads choices[].delta.content.
• When no usage / metadata.usage is present, the last chunk's amazon-bedrock-invocationMetrics is used. parseBedrockRuntimeMetrics already maps inputTokenCount / outputTokenCount.
• It's only a fallback: Nova also sends amazon-bedrock-invocationMetrics, but with inputTokenCount: 0, outputTokenCount: 0 (seen on nova-micro, lite, pro and 2-lite), so its metadata.usage has to keep winning. A test pins that.

Tests

The two new tests in bedrock-runtime-plugin.test.ts use chunk shapes taken from real Bedrock streams:

• gpt-oss: text + invocationMetrics → prompt_tokens: 126, completion_tokens: 125, tokens: 251.
• Nova: metadata.usage 6/5 alongside zeroed invocationMetrics → 6/5.

$ cd js && npx vitest run src/instrumentation/plugins/bedrock-runtime-plugin.test.ts
Tests  19 passed (19)
# main 5a8b3c5 with only the tests added: 1 failed (the gpt-oss test), 18 passed
$ npx vitest run src/instrumentation
Tests  1006 passed (1006)
$ npx tsc --noEmit -p .        # no errors
$ npx eslint <changed files>   # 0 errors (8 existing warnings, none on changed lines)
$ npx prettier --check ...     # clean

Includes a patch changeset.

Not in this PR: tool calls in streamed InvokeModel output (gpt-oss delta.tool_calls, Nova toolUse blocks) still aren't reconstructed; only the text is logged. Happy to follow up if that's wanted.

🤖 Generated with Claude Code