Decode base64 meta_struct payloads

Smoke tests inspecting AppSec or other structured span metadata will see the wrong type and can fail or silently miss the metadata they are intended to validate.

Was this helpful? React 👍 or 👎
_{🤖 Datadog Autotest · What is Autotest? · @DataDog review to ask questions · Any feedback? Reach out in #autotest}

#12089 (comment)

Decode binary meta_struct entries before exposing them

When a trace contains meta_struct, each v0.4 entry is a MessagePack-encoded binary payload (see TraceMapperV0_4.MetaStructWriter), which /test/traces represents as a base64 JSON string. Deserializing directly into Map<String, Object> therefore exposes strings instead of the structured objects returned by the existing v0.4 decoder, so AppSec/LLM smoke tests that inspect or cast these values will fail or evaluate the wrong data. Decode each base64 value and unpack its MessagePack payload; the current synthetic test should also use the wire representation rather than an already-decoded object.

Useful? React with 👍 / 👎.

It does not apply here. From Claude:

Thanks — this is the right instinct in general (in the raw msgpack v0.4/v1 wire formats, meta_struct values are binary MessagePack blobs, which is exactly why SpanV04/SpanV1 decode them). But it doesn't apply to the JSON this decoder consumes: the dd-apm-test-agent decodes meta_struct server-side before serving /test/traces and /test/session/traces, so the JSON already carries nested objects, not base64.

From ddapm_test_agent/trace.py at the pinned v1.44.0:
meta_struct: NotRequired[Dict[str, Dict[str, Any]]] # decoded dicts, not strings

def _parse_meta_struct(value):
return {key: msgpack.unpackb(val_bytes) for key, val_bytes in value.items()}
_parse_meta_struct is applied via _flexible_decode_meta_struct(payload) inside decode_v04(...), and handle_session_traces / handle_test_traces simply web.json_response(traces) those already-decoded traces. Source: https://github.com/DataDog/dd-apm-test-agent/blob/v1.44.0/ddapm_test_agent/trace.py#L363-L367

So typing meta_struct as base64 Map<String,String> and re-decoding would actually break against the real agent (the value is a nested JSON object, not a string). Keeping Map<String,Object> and letting the value deserialize directly is correct here. Not making this change; verified against the agent source (and the existing metaStructEmptyWhenAbsentAndAMapWhenPresent fixture uses the real nested shape).