agent-trajectory-diff 0.2.0__tar.gz → 0.2.2__tar.gz

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/.gitignore

@@ -2,6 +2,11 @@
 #              AGENTDIFF GITIGNORE
 # ==========================================
 
+# 0. Secrets — never commit API keys
+.env
+.env.*
+!.env.example
+
 # 1. Project Context (User Request)
 /context/
 

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: agent-trajectory-diff
-Version: 0.2.0
+Version: 0.2.2
 Summary: A developer-first package to evaluate and find regressions in agent trajectories
 Project-URL: Homepage, https://github.com/lostmartian/agentdiff
 Project-URL: Repository, https://github.com/lostmartian/agentdiff
@@ -153,7 +153,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - uses: lostmartian/agentdiff/.github/actions/agentdiff-check@v0.2.0
+      - uses: lostmartian/agentdiff/.github/actions/agentdiff-check@v0.2.2
         with:
           baseline: traces/baseline.json   # committed baseline trace
           candidate: traces/candidate.json # generated by an earlier step

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/README.md

@@ -133,7 +133,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - uses: lostmartian/agentdiff/.github/actions/agentdiff-check@v0.2.0
+      - uses: lostmartian/agentdiff/.github/actions/agentdiff-check@v0.2.2
         with:
           baseline: traces/baseline.json   # committed baseline trace
           candidate: traces/candidate.json # generated by an earlier step

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/src/agentdiff/__init__.py

@@ -29,7 +29,7 @@ Typical flow::
     assert_no_regressions(report)
 """
 
-__version__ = "0.2.0"
+__version__ = "0.2.2"
 
 from agentdiff.adapters import (
     BaseAdapter,

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/src/agentdiff/adapters/langfuse.py

@@ -6,6 +6,14 @@ from agentdiff.models.step import StepStatus, StepType, TokenUsage, TraceStep
 from agentdiff.models.trace import AgentTrace
 
 
+def _pick(mapping: dict[str, Any], *keys: str) -> Any:
+    """Return the first present value among the given key variants."""
+    for key in keys:
+        if mapping.get(key) is not None:
+            return mapping[key]
+    return None
+
+
 class LangfuseAdapter(BaseAdapter):
     @classmethod
     def from_dict(cls, data: dict[str, Any]) -> AgentTrace:
@@ -26,6 +34,21 @@ class LangfuseAdapter(BaseAdapter):
         observations = data.get("observations")
         if not isinstance(observations, list):
             observations = []
+
+        # The API may return observations in non-chronological order; sort by
+        # start time so step order reflects the actual execution timeline.
+        def _start_key(obs):
+            raw = _pick(obs, "startTime", "start_time")
+            ts = parse_iso_timestamp(raw)
+            if ts is not None:
+                return ts
+            return None
+
+        keyed = [(obs, _start_key(obs)) for obs in observations]
+        keyed = [(obs, k) for obs, k in keyed if isinstance(obs, dict)]
+        keyed.sort(key=lambda pair: (pair[1] is None, pair[1] or 0))
+        observations = [obs for obs, _ in keyed]
+
         steps: list[TraceStep] = []
 
         for idx, obs in enumerate(observations):
@@ -34,12 +57,16 @@ class LangfuseAdapter(BaseAdapter):
                     f"Langfuse observation at index {idx} is not an object"
                 )
             obs_id = obs.get("id") or f"obs_{idx}"
-            parent_id = obs.get("parentObservationId")
+            parent_id = _pick(obs, "parentObservationId", "parent_observation_id")
 
             # Map Langfuse type to StepType
             obs_type = str(obs.get("type", "")).upper()
-            if obs_type == "GENERATION":
+            if obs_type in ("GENERATION", "EMBEDDING"):
                 step_type = StepType.LLM_CALL
+            elif obs_type in ("TOOL", "RETRIEVER"):
+                step_type = StepType.TOOL_CALL
+            elif obs_type in ("AGENT", "CHAIN", "GUARDRAIL", "EVALUATOR"):
+                step_type = StepType.ROUTING
             elif obs_type == "SPAN":
                 name_lower = str(obs.get("name", "")).lower()
                 if "tool" in name_lower or "call" in name_lower:
@@ -83,8 +110,8 @@ class LangfuseAdapter(BaseAdapter):
 
             # Latency
             latency_ms = 0.0
-            start_str = obs.get("startTime")
-            end_str = obs.get("endTime")
+            start_str = _pick(obs, "startTime", "start_time")
+            end_str = _pick(obs, "endTime", "end_time")
             if start_str and end_str:
                 t1 = parse_iso_timestamp(start_str)
                 t2 = parse_iso_timestamp(end_str)
@@ -92,7 +119,7 @@ class LangfuseAdapter(BaseAdapter):
                     latency_ms = max(0.0, (t2 - t1).total_seconds() * 1000.0)
             if latency_ms == 0.0:
                 latency_ms = (
-                    cls._as_float(obs.get("latency_ms"))
+                    cls._as_float(_pick(obs, "latency_ms", "latency"))
                     or cls._as_float(obs.get("duration") or 0.0) * 1000.0
                 )
 
@@ -100,9 +127,9 @@ class LangfuseAdapter(BaseAdapter):
             level = str(obs.get("level", "")).upper()
             status = StepStatus.SUCCESS
             error_message = None
-            if "ERROR" in level or obs.get("statusMessage"):
+            if "ERROR" in level or _pick(obs, "statusMessage", "status_message"):
                 status = StepStatus.ERROR
-                error_message = obs.get("statusMessage")
+                error_message = _pick(obs, "statusMessage", "status_message")
 
             step = TraceStep(
                 step_id=obs_id,

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/src/agentdiff/adapters/openai_agents.py

@@ -127,7 +127,11 @@ class OpenAIAgentsAdapter(BaseAdapter):
         if not isinstance(span_data, dict):
             span_data = {}
         stype = str(span_data.get("type") or span.get("type") or "").lower()
-        if stype in _SKIPPED_TYPES:
+        sname = str(span_data.get("name") or "").lower()
+        # Real OpenAI Agents traces mark the task/turn bookkeeping spans as
+        # type="custom" with name="task"/"turn"; older/fixture shapes used a
+        # dedicated type. Skip both so they don't surface as steps.
+        if stype in _SKIPPED_TYPES or (stype == "custom" and sname in _SKIPPED_TYPES):
             return None
 
         step_type = _TYPE_MAP.get(stype, StepType.THOUGHT)

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/src/agentdiff/adapters/openinference.py

@@ -31,8 +31,12 @@ class OpenInferenceAdapter(BaseAdapter):
             context = span.get("context")
             if not isinstance(context, dict):
                 context = {}
-            span_id = context.get("span_id") or span.get("span_id") or f"span_{idx}"
+            span_id = str(
+                context.get("span_id") or span.get("span_id") or f"span_{idx}"
+            )
             parent_id = span.get("parent_span_id") or span.get("parent_id")
+            if parent_id is not None:
+                parent_id = str(parent_id)
 
             attrs = span.get("attributes")
             if not isinstance(attrs, dict):
@@ -100,7 +104,12 @@ class OpenInferenceAdapter(BaseAdapter):
                     if isinstance(start_time, (int, float)) and isinstance(
                         end_time, (int, float)
                     ):
-                        latency_ms = (end_time - start_time) * 1000.0
+                        # OTel spans carry nanoseconds (~1e18); the legacy shape
+                        # uses seconds. Detect ns by magnitude and normalize.
+                        scale = 1.0
+                        if abs(end_time) > 1e11:
+                            scale = 1e-9
+                        latency_ms = max(0.0, (end_time - start_time) * scale * 1000.0)
                     else:
                         t1 = parse_iso_timestamp(start_time)
                         t2 = parse_iso_timestamp(end_time)
@@ -155,6 +164,7 @@ class OpenInferenceAdapter(BaseAdapter):
                 or first_span.get("trace_id")
                 or trace_id
             )
+            trace_id = str(trace_id)
 
         # Task input / Final output from root span if available
         task_input = {}

{agent_trajectory_diff-0.2.0 → agent_trajectory_diff-0.2.2}/src/agentdiff/testing/assertions.py

@@ -51,6 +51,7 @@ def assert_no_regressions(
         )
 
     if errors:
+        report.passed = False
         errors_str = "\n".join(f"  - {err}" for err in errors)
         summary_str = report.summary()
         raise AssertionError(