trodo-python 2.23.1__tar.gz → 2.23.3__tar.gz

{trodo_python-2.23.1 → trodo_python-2.23.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: trodo-python
-Version: 2.23.1
+Version: 2.23.3
 Summary: Trodo Analytics SDK for Python — server-side event tracking
 License: ISC
 Keywords: analytics,tracking,trodo,server-side

{trodo_python-2.23.1 → trodo_python-2.23.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "trodo-python"
-version = "2.23.1"
+version = "2.23.3"
 description = "Trodo Analytics SDK for Python — server-side event tracking"
 readme = "README.md"
 license = { text = "ISC" }

trodo_python-2.23.3/tests/test_framework_deference.py

@@ -0,0 +1,41 @@
+"""One model call must produce one llm span.
+
+langchain-openai depends on the `openai` package, so a chain invocation IS a
+provider client call. With both instrumentors registered that single call emits
+two sibling llm spans carrying identical tokens, and the run's totals double.
+Measured on the Node side against real traffic: a 45/2 call reported 90/4. Cost
+escaped doubling only by luck -- the framework's duplicate reported
+model=unknown and so could not be priced at all.
+
+The framework instrumentor therefore defers to the provider one, which knows the
+real model id and can be priced.
+"""
+from trodo.otel.auto_instrument import _superseding_providers
+
+
+def test_langchain_defers_to_openai():
+    assert "openai" in _superseding_providers("langchain", ["anthropic", "openai"])
+
+
+def test_langchain_defers_to_anthropic():
+    assert "anthropic" in _superseding_providers("langchain", ["anthropic"])
+
+
+def test_langchain_registers_when_no_provider_it_routes_through_is_active():
+    # LangChain over a provider we do not instrument: its span is then the ONLY
+    # capture, and suppressing it would lose the call entirely.
+    assert _superseding_providers("langchain", ["google_generativeai"]) == []
+    assert _superseding_providers("langchain", []) == []
+
+
+def test_providers_never_defer_to_each_other():
+    # Two providers are two different clients; there is nothing to double.
+    assert _superseding_providers("openai", ["anthropic"]) == []
+    assert _superseding_providers("anthropic", ["openai"]) == []
+
+
+def test_only_lists_providers_registered_before_the_framework():
+    # The check reads `active`, built in registration order (anthropic, openai,
+    # openai_v2, langchain, ...). Listing bedrock would read right and never fire.
+    got = _superseding_providers("langchain", ["openai", "anthropic", "bedrock", "cohere"])
+    assert "bedrock" not in got and "cohere" not in got

{trodo_python-2.23.1 → trodo_python-2.23.3}/tests/test_prompt_trace.py

@@ -259,3 +259,80 @@ def test_run_set_input_with_ordinary_input_records_nothing(processor, http):
 
     attrs = http.run_ingest[0]["run"].get("attributes") or {}
     assert "trodo.prompts" not in attrs
+
+
+# ── Prompts are SPAN-level, not run-level ───────────────────────────────────
+#
+# One run's spans routinely use different prompts, different versions, or none.
+# The run's "most recently compiled" prompt is a fallback, and applying it to
+# every span meant compiling anywhere in a run attributed that prompt to every
+# LATER span. Caught in production: a span named 'no-prompt' was labelled with
+# whichever prompt had been compiled last.
+
+
+def test_does_not_leak_last_prompt_onto_a_later_unrelated_span(processor, http):
+    mgr = PromptManager(PromptHttp(prompt=PROMPT_WIRE))
+    with wrap_agent(processor=processor, team_site_id="site-x", agent_name="chat"):
+        p = mgr.get("refund-agent", cache_ttl_seconds=0)
+        with span("used-it", kind="llm"):
+            p.compile(q="x")
+        with span("unrelated"):
+            pass
+
+    spans = http.run_ingest[0]["spans"]
+    attrs = next(s for s in spans if s["name"] == "unrelated").get("attributes") or {}
+    assert "trodo.prompt.name" not in attrs
+
+
+def test_still_labels_a_model_span_following_a_compile(processor, http):
+    mgr = PromptManager(PromptHttp(prompt=PROMPT_WIRE))
+    with wrap_agent(processor=processor, team_site_id="site-x", agent_name="chat"):
+        p = mgr.get("refund-agent", cache_ttl_seconds=0)
+        p.compile(q="x")
+        with span("model-call", kind="llm"):
+            pass
+
+    spans = http.run_ingest[0]["spans"]
+    attrs = next(s for s in spans if s["name"] == "model-call")["attributes"]
+    assert attrs["trodo.prompt.name"] == "refund-agent"
+
+
+def test_labels_the_span_the_prompt_was_compiled_inside(processor, http):
+    """A plain span wrapping a hand-rolled model call plainly used the prompt
+    it compiled; gating purely on kind would have lost it."""
+    mgr = PromptManager(PromptHttp(prompt=PROMPT_WIRE))
+    with wrap_agent(processor=processor, team_site_id="site-x", agent_name="chat"):
+        p = mgr.get("refund-agent", cache_ttl_seconds=0)
+        with span("hand-rolled"):
+            p.compile(q="x")
+
+    spans = http.run_ingest[0]["spans"]
+    attrs = next(s for s in spans if s["name"] == "hand-rolled")["attributes"]
+    assert attrs["trodo.prompt.name"] == "refund-agent"
+
+
+def test_each_span_gets_its_own_prompt_and_the_run_rolls_up_both(processor, http):
+    other = dict(PROMPT_WIRE, name="other-agent", version_hash="bbbb2222")
+
+    class Two:
+        def get_prompt(self, name, label=None, version=None):
+            return {"prompt": dict(other if name == "other-agent" else PROMPT_WIRE)}
+
+    mgr = PromptManager(Two())
+    with wrap_agent(processor=processor, team_site_id="site-x", agent_name="chat"):
+        a = mgr.get("refund-agent", cache_ttl_seconds=0)
+        b = mgr.get("other-agent", cache_ttl_seconds=0)
+        with span("span-a", kind="llm"):
+            a.compile(q="x")
+        with span("span-b", kind="llm"):
+            b.compile(q="y")
+        with span("span-c"):
+            pass
+
+    spans = http.run_ingest[0]["spans"]
+    at = lambda n: (next(s for s in spans if s["name"] == n).get("attributes") or {})
+    assert at("span-a")["trodo.prompt.name"] == "refund-agent"
+    assert at("span-b")["trodo.prompt.name"] == "other-agent"
+    assert "trodo.prompt.name" not in at("span-c")
+    used = http.run_ingest[0]["run"]["attributes"]["trodo.prompts"]
+    assert sorted(r["name"] for r in used) == ["other-agent", "refund-agent"]

{trodo_python-2.23.1 → trodo_python-2.23.3}/trodo/otel/auto_instrument.py

@@ -554,6 +554,33 @@ def _register_instrumentors() -> None:
     ]
 
 
+# Framework instrumentors that call a provider client we ALSO instrument.
+#
+# langchain-openai depends on the `openai` package, langchain-anthropic on
+# `anthropic` -- so one chain invocation is also a genuine provider client call.
+# With both instrumentors registered that single call emits TWO sibling llm
+# spans carrying the same tokens, and the run's totals come out doubled.
+# Measured on the Node side against real traffic: a 45/2 call reported 90/4.
+#
+# The provider span is the better of the two: it carries the real model id, so
+# it can be priced, while the framework's reported model=unknown and could not
+# be. So the framework instrumentor defers.
+#
+# Only providers registered BEFORE the framework in _INSTRUMENTORS may appear
+# here -- the check reads `active`, which is built in registration order. That
+# order is anthropic, openai, openai_v2, langchain, ... so those are covered;
+# listing bedrock or cohere would read correctly and never fire.
+_ROUTES_THROUGH: dict[str, tuple[str, ...]] = {
+    "langchain": ("openai", "openai_v2", "anthropic"),
+    "llama_index": ("openai", "openai_v2", "anthropic"),
+}
+
+
+def _superseding_providers(name: str, active: List[str]) -> List[str]:
+    """Provider instrumentors already active that `name` would double-count."""
+    return [p for p in _ROUTES_THROUGH.get(name, ()) if p in active]
+
+
 def enable_auto_instrument(
     processor: TrodoSpanProcessor,
     disable: Optional[Iterable[str]] = None,
@@ -590,6 +617,24 @@ def enable_auto_instrument(
     for name, register in _INSTRUMENTORS:
         if name in disabled:
             continue
+        # BEFORE registering: a framework instrumentor whose model calls are
+        # already captured by an active provider instrumentor would double every
+        # token. Checked here rather than after, because instrumentors patch on
+        # instrument() and undoing it afterwards is too late.
+        superseded = _superseding_providers(name, active)
+        if superseded:
+            _warn_once(
+                f"superseded-{name}",
+                f"auto-instrument: '{name}' was NOT loaded because "
+                f"{', '.join(superseded)} instrumentation is already active and "
+                f"{name} calls those clients directly -- running both makes one "
+                f"model call emit two llm spans and doubles its tokens and cost. "
+                f"Your {name} calls are still captured, by the provider "
+                f"instrumentation, which also records the real model id so the "
+                f"span can be priced. To choose the other way round, pass "
+                f"disable=['{superseded[0]}'].",
+            )
+            continue
         try:
             register()
             active.append(name)

{trodo_python-2.23.1 → trodo_python-2.23.3}/trodo/otel/prompt_trace.py

@@ -25,7 +25,7 @@ PromptRef = Dict[str, Any]
 
 def new_prompt_state() -> Dict[str, Any]:
     """Per-run accumulator, shared by reference across the run's span tree."""
-    return {"current": None, "all": []}
+    return {"current": None, "current_span_id": None, "all": []}
 
 
 # Compiled prompt (or its messages list) -> the version it came from.
@@ -93,6 +93,7 @@ def record_compiled_prompt(ref: PromptRef) -> None:
     state = getattr(active, "prompt_state", None) if active is not None else None
     if isinstance(state, dict):
         state["current"] = ref
+        state["current_span_id"] = getattr(active, "span_id", None)
         key = _ref_key(ref)
         used: List[PromptRef] = state["all"]
         if not any(_ref_key(r) == key for r in used):
@@ -116,7 +117,11 @@ def prompt_attributes(ref: Optional[PromptRef]) -> Dict[str, str]:
 
 
 def merge_prompt_attrs(
-    active: Any, attrs: Optional[Dict[str, Any]], from_input: Optional[PromptRef] = None
+    active: Any,
+    attrs: Optional[Dict[str, Any]],
+    from_input: Optional[PromptRef] = None,
+    kind: Optional[str] = None,
+    span_id: Optional[str] = None,
 ) -> Optional[Dict[str, Any]]:
     """Merge the prompt this span used into its attributes.
 
@@ -134,7 +139,24 @@ def merge_prompt_attrs(
         remember_prompt(state, from_input)
         merged.update(prompt_attributes(from_input))
     elif isinstance(state, dict):
-        merged.update(prompt_attributes(state.get("current")))
+        # The run's "most recently compiled" prompt is a FALLBACK, used only
+        # where it is defensible. Applying it to every span meant compiling a
+        # prompt anywhere in a run attributed it to every LATER span -- a tool,
+        # a retrieval, a bit of plain work all claiming a prompt they never ran.
+        # Prompts belong to SPANS: one run's spans routinely use different
+        # prompts, different versions, or none at all.
+        #
+        # Two cases survive: a MODEL call (kind 'llm'), the only thing that runs
+        # a prompt and the one shape whose input we cannot read when it comes
+        # from a provider instrumentation; and the very span the prompt was
+        # compiled INSIDE, which plainly used it whatever its kind.
+        compiled_here = (
+            state.get("current") is not None
+            and span_id is not None
+            and state.get("current_span_id") == span_id
+        )
+        if kind == "llm" or compiled_here:
+            merged.update(prompt_attributes(state.get("current")))
     return merged or None
 
 

{trodo_python-2.23.1 → trodo_python-2.23.3}/trodo/otel/wrap_agent.py

@@ -876,7 +876,8 @@ class join_run:
             temperature=self.handle.temperature,
             tool_name=self.handle.tool_name,
             attributes=merge_prompt_attrs(
-                self._ctx, self.handle.attributes, self.handle.prompt_ref
+                self._ctx, self.handle.attributes, self.handle.prompt_ref,
+                self._kind, self._span_id,
             ),
         )
         try:
@@ -988,7 +989,8 @@ class span:
             temperature=self.handle.temperature,
             tool_name=self.handle.tool_name,
             attributes=merge_prompt_attrs(
-                self._active, self.handle.attributes, self.handle.prompt_ref
+                self._active, self.handle.attributes, self.handle.prompt_ref,
+                self._kind, self._span_id,
             ),
         )
         processor: TrodoSpanProcessor = self._active.processor  # type: ignore[assignment]

{trodo_python-2.23.1 → trodo_python-2.23.3}/trodo_python.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: trodo-python
-Version: 2.23.1
+Version: 2.23.3
 Summary: Trodo Analytics SDK for Python — server-side event tracking
 License: ISC
 Keywords: analytics,tracking,trodo,server-side

{trodo_python-2.23.1 → trodo_python-2.23.3}/trodo_python.egg-info/SOURCES.txt

@@ -7,6 +7,7 @@ tests/test_datasets.py
 tests/test_end_run.py
 tests/test_error_enrichment.py
 tests/test_experiments.py
+tests/test_framework_deference.py
 tests/test_get_prompt_version_types.py
 tests/test_instrumentor_resolution.py
 tests/test_llm_usage_cost.py