Next Python SDK major

[sentrivana]

We're preparing our next major on this branch.

The project is tracked in Linear. If you don't have access, we'll try to tag issues belonging to the project with the SDK 3.0 label on GitHub so that you can follow along.

Notable changes

• Transaction-based tracing will be removed. Span streaming will be the default tracing model.
• Python 3.6 support will be removed.

Context

You might have read this announcement about us discontinuing work on a 3.0. This is referring to the work done on the potel-base branch, which included two types of changes: a huge refactor of our tracing code on the one hand, and various unrelated changes, improvements and fixes on the other. We're dropping the huge refactor part, and only porting the rest, to a new branch and eventually a new 3.0 release.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.76%. Comparing base (14aff96) to head (d864ba0).
⚠️ Report is 4 commits behind head on master.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #5005       +/-   ##
===========================================
+ Coverage   70.55%   83.76%   +13.21%     
===========================================
  Files         180      180               
  Lines       18077    18080        +3     
  Branches     3008     3009        +1     
===========================================
+ Hits        12754    15145     +2391     
+ Misses       4432     1943     -2489     
- Partials      891      992      +101     

Files with missing lines	Coverage Δ
sentry_sdk/integrations/__init__.py	88.42% <100.00%> (+0.37%)	⬆️

... and 61 files with indirect coverage changes

[github-actions]

Codecov Results 📊

✅ 66913 passed | ❌ 1 failed | ⏭️ 3019 skipped | Total: 69933 | Pass Rate: 95.68% | Execution Time: 189m 32s

📊 Comparison with Base Branch

Metric	Change
Total Tests	📉 -66968
Passed Tests	📉 -62840
Failed Tests	📈 +1
Skipped Tests	📉 -4129

➕ New Tests (1)

View new tests

• test_input_attributes_nonstreaming_chat[True-messages0-expected_system_instructions0-expected_input_messages0]
  • File: tests.integrations.mistral.test_mistral
  • Status: ❌ Failing

❌ Failed Tests

test_input_attributes_nonstreaming_chat[True-messages0-expected_system_instructions0-expected_input_messages0]

File: tests.integrations.mistral.test_mistral
Suite: py3.10-mistral-v2.0.5
Error: KeyError: 140715163947008

Stack Trace

tests/integrations/mistral/test_mistral.py:380: in test_input_attributes_nonstreaming_chat
    sentry_init(
tests/conftest.py:349: in inner
    client = sentry_sdk.Client(*a, **kw)
sentry_sdk/client.py:278: in __init__
    self._init_impl()
sentry_sdk/client.py:428: in _init_impl
    self.integrations = setup_integrations(
sentry_sdk/integrations/__init__.py:216: in setup_integrations
    for integration_cls in iter_default_integrations(
sentry_sdk/integrations/__init__.py:44: in iter_default_integrations
    yield getattr(import_module(module), cls)
/usr/lib/python3.10/importlib/__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
<frozen importlib._bootstrap>:1050: in _gcd_import
    ???
<frozen importlib._bootstrap>:1024: in _find_and_load
    ???
<frozen importlib._bootstrap>:171: in __enter__
    ???
<frozen importlib._bootstrap>:123: in acquire
    ???
E   KeyError: 140715163947008

✅ Patch coverage is 90.13%. Project has 2088 uncovered lines.
❌ Project coverage is 90.13%. Comparing base (c3a7b52) to head (35f70d2).

Coverage diff

@@            Coverage Diff             @@
##        master       #PR       +/-##
==========================================
- Coverage    90.20%    90.13%    -0.07%
==========================================
  Files          194       180       -14
  Lines        25997     21155     -4842
  Branches      9662      7656     -2006
==========================================
+ Hits         23448     19067     -4381
- Misses        2549      2088      -461
- Partials      1450      1199      -251

---

Generated by Codecov Action

[github-actions]

Semver Impact of This PR

⚪ None (no version bump detected)

📋 Changelog Preview

This is how your changes will appear in the changelog.
Entries from this PR are highlighted with a left border (blockquote style).

---

New Features ✨

• (logging) Separate ignore lists for events/breadcrumbs and sentry logs by sl0thentr0py in #5698

Bug Fixes 🐛

Anthropic

• Set exception info on streaming span when applicable by alexander-alderman-webb in #5683
• Patch AsyncStream.close() and AsyncMessageStream.close() to finish spans by alexander-alderman-webb in #5675
• Patch Stream.close() and MessageStream.close() to finish spans by alexander-alderman-webb in #5674

Documentation 📚

• Add note on AI PRs to CONTRIBUTING.md by sentrivana in #5696

Internal Changes 🔧

• Add -latest alias for each integration test suite by sentrivana in #5706
• Use date-based branch names for toxgen PRs by sentrivana in #5704
• 🤖 Update test matrix with new releases (03/19) by github-actions in #5703
• Add client report tests for span streaming by sentrivana in #5677

Other

• Next Python SDK major by sentrivana in #5005

• Update CHANGELOG.md by sentrivana in #5685

---

_{🤖 This preview updates automatically when you update the PR.}

[sentry-warden]

Tool input/output always sent without PII/data_collection guards

wrapped_tool always serializes full tool args and return values onto spans, bypassing the gen_ai inputs/outputs and include_prompts checks used elsewhere in this integration. Gate GEN_AI_TOOL_INPUT/OUTPUT the same way prompts and response text are gated.

Evidence

• In async_wrapped and sync_wrapped, _capture_tool_input(args, kwargs, tool) and the tool result are passed to span.set_attribute(SPANDATA.GEN_AI_TOOL_INPUT/OUTPUT, safe_serialize(...)) with no privacy check.
• The same file gates prompts, response text, and related gen_ai content behind has_data_collection_enabled(...gen_ai.inputs/outputs) or should_send_default_pii() and integration.include_prompts (e.g. set_span_data_for_request / set_span_data_for_response).
• Sibling AI integrations (LangChain, Pydantic AI, OpenAI Agents) only set tool input/output when those same guards allow it.
• Tool args/results are application-controlled and commonly contain user content, secrets, or PII, so they are sent to Sentry even when send_default_pii=False or data_collection.gen_ai.*=False.

_{Identified by Warden · security-review, find-bugs, code-review · 59T-VTZ}

[sentry-warden]

Migration guide omits transaction-to-span-streaming API break

Document removal of start_transaction/Transaction and the start_span signature change (op/description → name/attributes), with concrete before/after examples—this is the headline 3.0 break and currently has no migration path here.

Evidence

• PR states transaction-based tracing is removed and span streaming is the default tracing model.
• sentry_sdk/api.py exports start_span, continue_trace, and new_trace only—no start_transaction.
• sentry_sdk/traces.py start_span(name=..., attributes=...) replaces the old op/description API; tracing.py/Transaction are gone.
• MIGRATION_GUIDE.md Changed/Removed sections (lines 1–119+) never mention start_transaction, Transaction, or how to rewrite existing tracing calls.

Also found at 1 additional location

• sentry_sdk/__init__.py:48-49

_{Identified by Warden · code-review · YNX-E65}

[sentry-warden]

Async stream early-return drops await and returns a coroutine

When _agent_framework_chat_generation_entered is set, new_async_generate_content_stream returns f(...) instead of await f(...), so callers get a coroutine instead of the stream.

Evidence

• The surrounding path at line 140 correctly does return await f(self, *args, **kwargs) when the integration is missing.
• The new early-return at lines 142-143 does return f(self, *args, **kwargs) with no await.
• Sibling async integrations (openai.py, anthropic.py) await on the same _agent_framework_chat_generation_entered guard.

Also found at 1 additional location

• sentry_sdk/integrations/google_genai/__init__.py:245-246

_{Identified by Warden · code-review, find-bugs · QMM-3HM}

[sentry-warden]

Chat generation exit skips context cleanup and leaves suppression flag set

_exit_span only ends the inner Span, so successful on_chat_model_end paths never call _AgentFrameworkChatGenerationContext.__exit__() and leave _agent_framework_chat_generation_entered stuck true, suppressing later OpenAI/Anthropic/Google GenAI spans on that scope. Exit the stored context (as _handle_error already does) instead of unwrapping to the raw span.

Evidence

• _create_generation_span() stores an _AgentFrameworkChatGenerationContext in span_map; that class sets span._scope._agent_framework_chat_generation_entered = True in __init__ and only clears it in __exit__.
• _exit_span() was changed to accept only a Span and call span.__exit__(), which ends the span but does not clear the flag.
• on_chat_model_end unwraps context.span and calls _exit_span(span, run_id), so the success path never invokes context.__exit__().
• _handle_error() still calls context.__exit__() correctly, so only successful chat completions leak the flag.
• OpenAI/Anthropic/Google GenAI integrations skip instrumentation when get_current_scope()._agent_framework_chat_generation_entered is true.

Also found at 2 additional locations

• sentry_sdk/integrations/langchain.py:537-561
• sentry_sdk/integrations/langchain.py:18-18

_{Identified by Warden · code-review, find-bugs · 7R5-TEG}

[sentry-warden]

Removed before_send_transaction option is undocumented and has no equivalent migration guidance

Existing applications that pass before_send_transaction to sentry_sdk.init() now fail during initialization with TypeError: Unknown option 'before_send_transaction'. The 3.0 migration guide does not document this removal or explain that transaction-level filtering is no longer available; before_send_span is not a direct equivalent because streamed spans cannot be dropped.

Evidence

• ClientConstructor.__init__ no longer declares before_send_transaction, so it is absent from DEFAULT_OPTIONS generated by _get_default_options().
• _get_options() rejects the legacy keyword with TypeError("Unknown option %r" % (key,)) before the client can initialize.
• The migration guide lists other removed options but does not mention before_send_transaction or the loss of transaction-level filtering.
• Client._capture_telemetry() routes the replacement callback through before_send_span, whose implementation explicitly keeps spans even when the callback returns None.

_{Identified by Warden · code-review · YR8-HD6}

[sentry-warden]

Langgraph agent spans lose usage and response model attributes

Removing _set_usage_data and _set_response_model_name drops gen_ai.usage.* and gen_ai.response.model from invoke_agent spans; if this was only StreamedSpan cleanup, keep setting them via set_attribute/set_data_normalized.

Evidence

• The hunk deletes _set_usage_data and _set_response_model_name, and removes their calls from _set_response_attributes.
• Those helpers previously aggregated response_metadata.token_usage and model_name onto the langgraph invoke_agent span.
• _set_response_attributes still only records response text/tool calls, so usage and model attributes are no longer set on this path.
• Sibling AI integrations (e.g. langchain callbacks) still record usage/model on their own spans, so this is a langgraph-agent-span behavior change, not a global AI attribute removal.

_{Identified by Warden · code-review, find-bugs · GAY-67G}

[sentry-warden]

Streaming iterator swallows user exceptions via capture_internal_exceptions

Wrap the stream loop so capture_internal_exceptions only guards SDK collection, not yield; otherwise stream errors are swallowed and the caller never sees them.

Evidence

• new_iterator() wraps for x in old_iterator: ... yield x in with capture_internal_exceptions().
• capture_internal_exceptions().__exit__ returns True, so any exception from the Cohere stream or from the generator is suppressed.
• span.end() still runs after the suppressed exception, so the caller gets a silent truncated stream and an OK span.
• Sibling integrations (google_genai, langchain) re-raise stream exceptions and only use capture_internal_exceptions around telemetry.

Also found at 1 additional location

• sentry_sdk/integrations/cohere.py:228-234

_{Identified by Warden · find-bugs, code-review · HCZ-A5T}

[sentry-warden]

Celery control-flow exceptions not registered, so task segment still ends as error

setup_once() never calls _register_control_flow_exception() for Celery Retry/Ignore/Reject, so the outer task segment still becomes error on span exit even when the exception is treated as non-error control flow.

Evidence

• Celery setup_once() only patches tracers/workers and ignores loggers; it does not register CELERY_CONTROL_FLOW_EXCEPTIONS.
• Span.__exit__ sets SpanStatus.ERROR whenever should_be_treated_as_error(ty, value) is true.
• should_be_treated_as_error() only skips types previously passed to _register_control_flow_exception().
• Huey and ARQ both register their control-flow exceptions in setup_once(); Celery does not, so uncaught Retry/Ignore/Reject still fail the outer queue.task.celery segment.

Also found at 1 additional location

• sentry_sdk/integrations/celery/__init__.py:95-100

_{Identified by Warden · find-bugs, code-review · GQL-YD4}

[sentry-warden]

AgentExecutor.stream leaves invoke_agent span active if setup raises

Wrap result = f(self, *args, **kwargs) in try/except and call span.__exit__(*exc_info) on failure so a setup error does not leave the active span on the scope (see google_genai stream wrappers).

Evidence

• new_stream() calls sentry_sdk.traces.start_span(...), and Span._start() immediately sets the span as the current scope span when active=True.
• result = f(self, *args, **kwargs) is unguarded; only new_iterator / new_iterator_async call span.__exit__.
• If AgentExecutor.stream raises before returning an iterator, the span is never finished and remains the active span for later work.
• google_genai stream wrappers catch exceptions around the underlying call and finish the span; this path does not.

_{Identified by Warden · code-review · GUS-JBC}

[sentry-warden]

Transaction events now sent as error events

After removing transaction handling, events with type "transaction" still pass through and are sampled as errors then envelope.add_event()'d; drop or reject them explicitly instead of mis-categorizing.

Evidence

• Previously is_transaction skipped _should_sample_error and used envelope.add_transaction(event_opt).
• That branch is gone; only is_checkin is special-cased, so type == "transaction" falls into the else path.
• envelope.add_event() always sets item type to "event", so a transaction payload is emitted as an error event.
• Call sites still exist that pass {"type": "transaction"} into capture_event (e.g. tests/test_transport.py), and Event typing still allows "transaction".

_{Identified by Warden · code-review · 3PJ-F9H}

[sentry-warden]

trace_lifecycle removed without migration note

Removing trace_lifecycle will hard-fail existing init(trace_lifecycle=...) callers, and MIGRATION_GUIDE.md does not document the removal or that stream mode is now the only behavior.

Evidence

• The hunk deletes trace_lifecycle: Optional[Literal['static', 'stream']] from the public constructor signature.
• DEFAULT_OPTIONS is generated from ClientConstructor.__init__, so the option disappears from accepted config keys.
• _get_options() rejects unknown keys with TypeError, so old configs break at init time.
• MIGRATION_GUIDE.md has no entry for trace_lifecycle, while related removals like stream_gen_ai_spans are explicitly listed.

_{Identified by Warden · code-review · KAM-DSE}

[sentry-warden]

Boto3 streaming span remains active and hijacks subsequent span parenting

Create the streaming span with active=False (or explicitly scope it only around body reads). The current implementation leaves it as the active span for the entire lifetime of the returned StreamingBody, causing unrelated spans and trace propagation to use the HTTP streaming span as their parent and allowing delayed cleanup to overwrite the current span stack.

Evidence

• _sentry_after_call() ends the request span at line 173, then starts streaming_span with the default active=True at lines 179-186.
• Span._start() replaces scope.span with the streaming span, so spans created while the body remains open are parented to http.client.stream rather than the caller's span.
• streaming_span.end() runs later from read() or close() and restores the span saved at creation; if cleanup occurs while another span is active, _end() unconditionally overwrites that current scope entry, corrupting the span stack.
• Existing boto3 streaming tests assert emitted span metadata and parent IDs but do not verify the current span or nested spans while the body is open.

_{Identified by Warden · find-bugs · L6S-EH8}