feat(closes OPEN-10364): make tracers init-based

[gustavocidornelas]

Pull Request

Summary

Turns openlayer.lib.init() into an auto-instrumenting entry point.

Calling init() with no extra arguments now detects every installed Openlayer-supported LLM SDK (openai, anthropic, mistral, groq, gemini, oci, azure_content_understanding, litellm, portkey, google_adk) and patches it so that every newly-constructed client is auto-traced.

Existing trace_<x>(client) calls keep working. They're now also idempotent, so mixing auto-instrumentation with explicit wraps no longer double-wraps methods.

Changes

• Add auto_instrument: bool | list[str] = True parameter to init() in src/openlayer/lib/tracing/tracer.py.
  Procedural (not stored in _tracer_config), evaluated fresh on every call. False short-circuits, a list patches
  only that subset.
• New module src/openlayer/lib/integrations/_auto.py with: IntegrationSpec dataclass, 10-entry REGISTRY,
  auto_instrument(targets) and unpatch_all() orchestrators, plus shared _patch_class_init / _unpatch_class_init
  helpers that wrap a client class __init__ via functools.wraps (so __wrapped__ makes unpatch trivially reversible).
• Per-integration _patch_<x>() / _unpatch_<x>() helpers added to 7 tracers (openai_tracer.py, anthropic_tracer.py, mistral_tracer.py, groq_tracer.py, gemini_tracer.py, oci_tracer.py, azure_content_understanding_tracer.py). Each wraps the SDK's client class __init__ so new instances auto-call the
  existing trace_<x>(client) function (the portkey_tracer.py pattern). Covers sync + async + Azure variants where
  applicable.
• Bundled idempotency fix: every per-instance trace_<x>(client) in 9 tracer files gained a getattr(client, "_openlayer_patched", False) is True entry guard plus a marker assignment before return. Pre-existing correctness
  improvement — without it, calling trace_openai(client) twice nested the wrappers and double-counted every call. The
  strict is True comparison avoids false positives from MagicMock's auto-attribute behavior (caught by the bedrock
  test suite).
• configure() deprecated alias updated to default auto_instrument=False via kwargs.setdefault. Existing
  configure() callers see no surprise patching on upgrade.
• Re-export auto_instrument and unpatch_all from openlayer.lib.
• _is_installed probes the full dotted path via importlib.util.find_spec (not just the top-level segment).
  Fixes a namespace-package collision where google.adk falsely reported installed because google.generativeai was
  installed under the same namespace.

Context

OPEN-10364: Make tracers init-based

Testing

• Manual testing

Monitoring

• No expected impact

Notes

• Behavior change: anyone calling init() today (no args) will now auto-instrument every installed supported SDK
  on their next upgrade. This is the intended behavior — the "init-based" tracer model — but worth a callout in the
  changelog. Users can opt out with init(auto_instrument=False) or by switching to the deprecated configure()
  (which preserves the no-patching default).
• Pre-existing client instances created before init() are NOT retroactively auto-traced. The patch wraps the
  class __init__, so only future construction goes through it. Class-method-level patching for catching pre-existing
  instances is deferred.
• Out of scope (deferred to a follow-up): Bedrock auto-instrumentation (boto3 service-client shape needs a
  different mechanism), LangChain auto-instrumentation (callback-based, no clean global hook), and the
  auto_instrument env-var convenience.
• Patch failure isolation: if one integration's patcher raises, the orchestrator logs a warning and continues
  with the rest. The returned dict[str, bool] reflects per-integration success.

[whoseoyster]

@gustavocidornelas did you want me to review?

[viniciusdsmello]

trace_portkey() doesn't have a class-level idempotency guard — it re-assigns Portkey.__init__ = traced_init unconditionally on every call. If init() is called twice with auto_instrument=True (a supported pattern), Portkey.__init__ gets stacked: each call wraps the previous traced_init. The per-instance _openlayer_portkey_patched marker prevents double-tracing of new instances, but the __init__ wrapping itself accumulates a layer per call.

Either add a _CLASS_PATCHED_ATTR-style guard at the top of trace_portkey(), or refactor it to use the shared _patch_class_init(Portkey, _wrap_portkey_instance) helper so it benefits from the same idempotency contract as the other integrations.

Same defect confirmed in _patch_google_adk(): it calls wrapt.wrap_function_wrapper(...) unconditionally for each target (BaseAgent.run_async, BaseLlmFlow._call_llm_async, etc.), and the _wrapped_methods.append(...) tracker is only used for unpatch — it doesn't dedupe on re-patch. A second init() call stacks wrappers on every ADK method too. Worth fixing both in the same pass.

[viniciusdsmello]

google.generativeai (package google-generativeai) is the legacy SDK. Google's new "Google Gen AI SDK" is google.genai (package google-genai), and they're actively migrating users to it — the legacy SDK is in maintenance mode.

Worth a follow-up to either:

• Register a second entry for google.genai pointing at a new tracer, or
• At minimum, leave an inline TODO so this doesn't get lost.

Not a blocker for this PR, but users on the new SDK will silently get no auto-instrumentation today.

[gustavocidornelas]

@viniciusdsmello, I addressed your comments. Can you review again? Thanks!

[viniciusdsmello]

LGTM. Clean design — the IntegrationSpec registry + lazy _patch_via lambdas + two-level idempotency markers (_CLASS_PATCHED_ATTR on the class, _openlayer_patched on the instance) is the right shape, and configure() preserving the no-patch default is a thoughtful backward-compat call.

Approving — the inline comments (portkey/google_adk class-level idempotency, Gemini SDK migration) are non-blocking; happy to see them addressed here or as follow-ups.