feat(langgraph): Usage attributes on invocation spans

[alexander-alderman-webb]

Description

Add prompt, response, and total token counts to LangGraph invocation spans.

Issues

Contributes to #5170

Reminders

• Please add tests to validate your changes, and lint your code using tox -e linters.
• Add GH Issue ID & Linear ID (if applicable)
• PR title should use conventional commit style (feat:, fix:, ref:, meta:)
• For external contributors: CONTRIBUTING.md, Sentry SDK development docs, Discord community

[codecov]

Codecov Report

❌ Patch coverage is 92.30769% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.22%. Comparing base (1b7085d) to head (4f3fab3).
⚠️ Report is 6 commits behind head on master.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
sentry_sdk/integrations/langgraph.py	92.30%	1 Missing and 1 partial ⚠️

Additional details and impacted files

@@ Coverage Diff @@## master #5211 +/- ## ========================================== + Coverage 84.17% 84.22% +0.04%  ========================================== Files 181 181 Lines 18443 18486 +43 Branches 3283 3295 +12 ========================================== + Hits 15524 15569 +45 + Misses 1904 1899 -5 - Partials 1015 1018 +3 

Files with missing lines	Coverage Δ
sentry_sdk/integrations/langgraph.py	76.27% <92.30%> (-0.81%)	⬇️

... and 5 files with indirect coverage changes

[cursor]

Bug: Usage data miscounted when PII collection disabled

When should_send_default_pii() is false or include_prompts is false, input_messages remains None. This causes _get_new_messages in _set_response_attributes to return all output messages instead of just the new ones. Since LangGraph state accumulates messages, usage data will include tokens from all messages in the response rather than only the new messages added during this invocation. The input messages need to be parsed unconditionally (at least for _get_new_messages) to correctly calculate usage data regardless of PII settings.

sentry_sdk/integrations/langgraph.py#L184-L204

sentry-python/sentry_sdk/integrations/langgraph.py

Lines 184 to 204 in 4f3fab3

	# Store input messages to later compare with output
	input_messages=None
	if (
	len(args) >0
	andshould_send_default_pii()
	andintegration.include_prompts
	):
	input_messages=_parse_langgraph_messages(args[0])
	ifinput_messages:
	normalized_input_messages=normalize_message_roles(input_messages)
	scope=sentry_sdk.get_current_scope()
	messages_data=truncate_and_annotate_messages(
	normalized_input_messages, span, scope
	)
	ifmessages_dataisnotNone:
	set_data_normalized(
	span,
	SPANDATA.GEN_AI_REQUEST_MESSAGES,
	messages_data,
	unpack=False,
	)

sentry_sdk/integrations/langgraph.py#L240-L260

sentry-python/sentry_sdk/integrations/langgraph.py

Lines 240 to 260 in 4f3fab3

	input_messages=None
	if (
	len(args) >0
	andshould_send_default_pii()
	andintegration.include_prompts
	):
	input_messages=_parse_langgraph_messages(args[0])
	ifinput_messages:
	normalized_input_messages=normalize_message_roles(input_messages)
	scope=sentry_sdk.get_current_scope()
	messages_data=truncate_and_annotate_messages(
	normalized_input_messages, span, scope
	)
	ifmessages_dataisnotNone:
	set_data_normalized(
	span,
	SPANDATA.GEN_AI_REQUEST_MESSAGES,
	messages_data,
	unpack=False,
	)

---

Bug: Usage data miscounted when PII collection disabled

When should_send_default_pii() is false or include_prompts is false, input_messages remains None. This causes _get_new_messages in _set_response_attributes to return all output messages instead of just the new ones. Since LangGraph state accumulates messages, usage data will include tokens from all messages in the response rather than only the new messages added during this invocation. The input messages need to be parsed unconditionally (at least for _get_new_messages) to correctly calculate usage data regardless of PII settings.

sentry_sdk/integrations/langgraph.py#L184-L204

sentry-python/sentry_sdk/integrations/langgraph.py

Lines 184 to 204 in 4f3fab3

	# Store input messages to later compare with output
	input_messages=None
	if (
	len(args) >0
	andshould_send_default_pii()
	andintegration.include_prompts
	):
	input_messages=_parse_langgraph_messages(args[0])
	ifinput_messages:
	normalized_input_messages=normalize_message_roles(input_messages)
	scope=sentry_sdk.get_current_scope()
	messages_data=truncate_and_annotate_messages(
	normalized_input_messages, span, scope
	)
	ifmessages_dataisnotNone:
	set_data_normalized(
	span,
	SPANDATA.GEN_AI_REQUEST_MESSAGES,
	messages_data,
	unpack=False,
	)

sentry_sdk/integrations/langgraph.py#L240-L260

sentry-python/sentry_sdk/integrations/langgraph.py

Lines 240 to 260 in 4f3fab3

	input_messages=None
	if (
	len(args) >0
	andshould_send_default_pii()
	andintegration.include_prompts
	):
	input_messages=_parse_langgraph_messages(args[0])
	ifinput_messages:
	normalized_input_messages=normalize_message_roles(input_messages)
	scope=sentry_sdk.get_current_scope()
	messages_data=truncate_and_annotate_messages(
	normalized_input_messages, span, scope
	)
	ifmessages_dataisnotNone:
	set_data_normalized(
	span,
	SPANDATA.GEN_AI_REQUEST_MESSAGES,
	messages_data,
	unpack=False,
	)

---

[alexander-alderman-webb]

This is okay, input messages do not have token info attached to them.

usage data will include tokens from all messages in the response rather than only the new messages added during this invocation.