clean up: drop redundant comments

The SKILL.md paragraph carries the "why" for dspy.lm; variable names carry
the rest.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: starfolkbot

py/src/braintrust/integrations/dspy/test_dspy.py

@@ -42,36 +42,29 @@ def test_dspy_callback(memory_logger):
 
     spans_by_name = {span["span_attributes"]["name"]: span for span in spans}
 
-    # DSPy modules are pipeline steps → task per spec.
-    module_span = spans_by_name.get("dspy.module.ChainOfThought")
-    assert module_span is not None
+    module_span = spans_by_name["dspy.module.ChainOfThought"]
     assert module_span["span_attributes"]["type"] == "task"
     assert module_span["metadata"]["module_class"].endswith("ChainOfThought")
 
     lm_span = spans_by_name["dspy.lm"]
     assert lm_span["context"]["span_origin"]["instrumentation"]["name"] == "dspy-auto"
-    # dspy.lm intentionally has no `llm` type — the underlying provider integration
-    # (litellm/openai/anthropic) owns the `llm` leaf and its token accounting.
     assert lm_span["span_attributes"].get("type") != "llm"
     assert "metadata" in lm_span
     assert "model" in lm_span["metadata"]
     assert MODEL in lm_span["metadata"]["model"]
-    # Still record provider so downstream tools can attribute the call correctly;
-    # derived from the litellm-style "openai/..." prefix.
     assert lm_span["metadata"]["provider"] == "openai"
     assert "input" in lm_span
     assert "output" in lm_span
 
     format_span = spans_by_name["dspy.adapter.format"]
-    parse_span = spans_by_name["dspy.adapter.parse"]
-
     assert format_span["span_attributes"]["type"] == "task"
     assert format_span["metadata"]["adapter_class"].endswith("ChatAdapter")
     assert "signature" in format_span["input"]
     assert "demos" in format_span["input"]
     assert "inputs" in format_span["input"]
     assert isinstance(format_span["output"], list)
 
+    parse_span = spans_by_name["dspy.adapter.parse"]
     assert parse_span["span_attributes"]["type"] == "task"
     assert parse_span["metadata"]["adapter_class"].endswith("ChatAdapter")
     assert "signature" in parse_span["input"]

py/src/braintrust/integrations/dspy/tracing.py

@@ -8,12 +8,7 @@
 
 _INSTRUMENTATION = "dspy-auto"
 
-# LiteLLM-style model strings are "<provider>/<model>" (e.g. "openai/gpt-4o-mini").
-# The prefix identifies the provider whose pricing applies.
 _LM_METADATA_PARAM_ALLOWLIST = ("temperature", "max_tokens", "top_p", "top_k", "stop")
-
-# Aggregate eval stats that DSPy Evaluate surfaces on its return dict. These are
-# domain-level scores, not spec-listed span `metrics` keys, so they go in metadata.
 _EVALUATE_METADATA_ALLOWLIST = ("accuracy", "score", "total", "correct")
 
 
@@ -139,13 +134,6 @@ def on_lm_start(
             if key in inputs:
                 metadata[key] = inputs[key]
 
-        # dspy.lm is intentionally NOT typed as `llm`. DSPy delegates the actual
-        # provider call to LiteLLM (or a patched provider client), which owns the
-        # `llm` leaf span and its token accounting. Typing this parent as `llm`
-        # would produce two `llm` spans per model call when the underlying
-        # provider is also instrumented, and this span cannot supply the tokens
-        # required for `llm` because DSPy's callback contract does not expose
-        # usage in `outputs`.
         span = start_span(
             name="dspy.lm",
             input=inputs,