🐛 Bug Report: gen_ai.bedrock.guardrail.activation counts every Bedrock call

Component: Bedrock Instrumentation

📜 Description

is_guardrail_activated() in guardrail.py ends with:

return response.get("amazon-bedrock-guardrailAction") != "NONE"

When no guardrail is configured, Bedrock omits that key, so .get() returns None and None != "NONE" is True. Every ordinary response is treated as a guardrail activation, and gen_ai.bedrock.guardrail.activation is incremented on every Bedrock call.

Defaulting the key fixes it: response.get("amazon-bedrock-guardrailAction", "NONE") != "NONE".

👟 Reproduction steps

from opentelemetry.instrumentation.bedrock.guardrail import is_guardrail_activated

print(is_guardrail_activated({"stopReason": "end_turn"}))  # True, expected False

👍 Expected behavior

A response with no amazon-bedrock-guardrailAction key is not an activation.

👎 Actual Behavior

It counts as one, so the metric reports activations for users with no guardrails configured.

🤖 Python Version

3.12

📃 Additional context

The same line crashes when metrics are disabled, since _instrument() sets the counters to None but guardrail.py dereferences them unconditionally:

guardrail.py", line 179, in guardrail_converse
    metric_params.guardrail_activation.add(1, attrs)
AttributeError: 'NoneType' object has no attribute 'add'

Present from 0.39.0 through 0.62.3.

👀 Duplicate check

Checked, no existing issue.

Start in guardrail.py with is_guardrail_activated() and the guardrail_converse path described in the issue. Run the provided reproduction and inspect how metrics-disabled instrumentation reaches guardrail_activation.add. Done means responses without the guardrail action key are not activations and disabled metrics do not dereference a missing counter.