feat: add OTel test semantic convention attributes to Experiment spans

[anirudha]

Description

Adds test.* span attributes from the OTel test semantic conventions proposal to existing Experiment evaluation spans. All changes are strictly additive: no wrapper spans, no OTel events, all existing gen_ai.evaluation.* attributes preserved unchanged.

Related: open-telemetry/semantic-conventions-genai#79

What changed

src/strands_evals/experiment.py (+34 lines)

Change	Details
name parameter	Optional str on __init__, defaults to "unnamed_experiment". Stored as self._name, exposed via @property.
to_dict / from_dict	Serializes and restores name. from_dict falls back to "unnamed_experiment" for legacy dicts without the key.
import uuid	Added at module top.
run_evaluations (sync)	Generates run_id = str(uuid.uuid4()) at method start. Adds test.suite.name, test.suite.run.id, test.case.name, test.case.id to eval_case span initial attributes. Adds test.case.result.status to each evaluator span's set_attributes.
run_evaluations_async	Same run_id generation. Passes run_id to _worker.
_worker (async)	New run_id parameter. Adds same test.* attributes to execute_case and evaluator spans.

No new classes, modules, or architectural changes. The diff is ~34 lines of production code.

pyproject.toml (+3 lines)

Added hypothesis>=6.0.0,<7.0.0 to three dependency sections:

• [project.optional-dependencies] test
• [tool.hatch.envs.hatch-test] dependencies
• [tool.hatch.envs.default] dependencies

tests/strands_evals/test_experiment.py (+11 lines)

Updated 11 existing to_dict test assertions to include the new "name": "unnamed_experiment" key in expected dictionaries. No test logic changed.

tests/strands_evals/test_experiment_otel_conventions.py (new, 741 lines)

6 property-based tests (hypothesis, 100 iterations each) + 10 unit tests:

Test	What it validates
Property 1	Experiment(name=s).name == s for any string
Property 2	to_dict → from_dict round-trip preserves name
Property 3 (sync)	eval_case spans have all 4 test.* attributes with correct values
Property 3 (async)	execute_case spans have all 4 test.* attributes with correct values
Property 4	Evaluator span test.case.result.status matches aggregate_pass boolean
Property 5	Existing gen_ai.evaluation.* attributes preserved on async spans
Property 6	Experiments with/without name produce identical evaluation reports
Unit: default name	"unnamed_experiment" when name not provided
Unit: legacy from_dict	Dict without name key defaults correctly
Unit: no wrapper span	No test_suite_run span created (sync + async)
Unit: no add_event	No OTel events used for test.* data (sync + async)
Unit: run_id format	test.suite.run.id is valid UUID4 (sync + async)

src/strands_evals/evaluators/coherence_evaluator.py (whitespace only)

Trailing whitespace cleanup on 2 docstring lines. Likely from formatter.

Design decisions worth reviewing

1. No wrapper span: run_id is a flat attribute on each case span rather than derived from a parent test_suite_run span. This preserves the flat trace structure that backends like Langfuse/Jaeger expect for session_id-based grouping (important for ActorSimulator multi-turn conversations).

2. Span attributes, not events: All test.* metadata uses span.set_attributes(). Maximizes backend compatibility since not all backends support event attributes.

3. run_id per invocation, not per instance: Each call to run_evaluations/run_evaluations_async gets a fresh UUID4. Concurrent calls on the same Experiment instance get distinct IDs.

4. Backward compatibility: name defaults to "unnamed_experiment", from_dict handles missing key gracefully. Constructor accepts all previously valid argument combinations.

Span attribute schema

eval_case <name>  /  execute_case <name>
├── test.suite.name      = experiment.name
├── test.suite.run.id    = UUID4 (unique per run_evaluations call)
├── test.case.name       = case.name
├── test.case.id         = case.session_id
├── gen_ai.evaluation.*  = (unchanged)
│
└── evaluator <Name>
    ├── test.case.result.status = "pass" | "fail"
    └── gen_ai.evaluation.*     = (unchanged)

How to verify

hatch test tests/ -vv

All 400 tests pass (69 existing + 17 new OTel convention tests + 314 other project tests).

[poshinchen]

Now I remember why reverted this grouping logic....

In strands-evals we have the ActorSimulator which allows user to execute multi-turn conversation for evaluations.

With this current changes, the multi-turn conversation will be wrapped together into a single case span. But in general those are different requests, and it contradicts what other OTEL-supported backend (Langfuse / Jaeger...) has. Users should group the traces based on the session_id, instead of wrapping all the executions into one span.

I had a long discussion with @jjbuck and this was the final decision that we made.

We can debate it again to whether group all of the multi-turn conversation into a single span. But to start simple, I'll be fine with having those test* and so on as the attributes. I think they are from OTEL Test Attributes?

[poshinchen]

And I also intentionally made them into span_attributes instead of event_attributes because some sources do not support event.

And from the documentation, it seems like they can both be span_attributes and the event_attributes.

[poshinchen]

Isn't this the same as session.id?

[poshinchen]

same here

[poshinchen]

Do you really need test_experiment_otel_conventions.py? =

It's basically mocks and verify the mock spans. This can reduce the need of hypothesis dependency

[anirudha]

out on vacation, i'll address the comments next week

[poshinchen]

Do you need us to adjust the PR?