From db95a89ba179568f33e0c1cd9d188f5bdd460021 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Mon, 17 Aug 2026 16:07:08 -0500
Subject: [PATCH 01/12] fix(converter): keep only last message from
 full-history agent output (HYBIM-988)

---
 src/splunk_ao/converter/attribute_mapping.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index 64b782d3..90ec0561 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -447,6 +447,10 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
         return
     if full_history and input_messages is not None and output_messages[: len(input_messages)] == input_messages:
         output_messages = output_messages[len(input_messages) :]
+    # Full history includes all messages in the run, not just the final response.
+    # Keep only the last message — it is always the agent's final output.
+    if full_history and len(output_messages) > 1:
+        output_messages = [output_messages[-1]]
     attrs["gen_ai.output.messages"] = _json_string(_with_finish_reasons(output_messages))
 
 

From 2ad296269a97fef9c8e94d74a1b61869b8399ad1 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Mon, 17 Aug 2026 17:11:23 -0500
Subject: [PATCH 02/12] test(converter): add edge case tests for full-history
 agent output reduction (HYBIM-988)

---
 tests/test_attribute_mapping.py | 84 +++++++++++++++++++++++++++++++++
 1 file changed, 84 insertions(+)

diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index 31571bfa..1b629b81 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -418,6 +418,90 @@ def test_orchestration_output_omits_repeated_input_history() -> None:
     ]
 
 
+def test_orchestration_full_history_with_tool_call_keeps_last_message() -> None:
+    # LangGraph accumulated state: user → tool-call AI (empty content) → tool response → final AI
+    # The first post-dedup message has empty content; the UI would show "—" without the fix.
+    user = {"role": "user", "content": "What is the dosage of Lisinopril?"}
+    ai_toolcall = {"role": "assistant", "content": "", "tool_calls": [{"id": "tc1", "function": {"name": "rag_search", "arguments": '{"query":"Lisinopril dosage"}'}}]}
+    tool_resp = {"role": "tool", "content": "Lisinopril: 10mg daily", "tool_call_id": "tc1"}
+    ai_final = {"role": "assistant", "content": "Common dosage is 10mg once daily."}
+
+    span = AgentSpan(
+        name="Agent",
+        agent_type=AgentType.default,
+        input=json.dumps({"messages": [user]}),
+        output=json.dumps({"messages": [user, ai_toolcall, tool_resp, ai_final]}),
+    )
+
+    attrs = build_span_attributes(span)
+
+    output_messages = json.loads(attrs["gen_ai.output.messages"])
+    assert len(output_messages) == 1
+    assert output_messages[0]["role"] == "assistant"
+    assert output_messages[0]["parts"][0]["content"] == "Common dosage is 10mg once daily."
+
+
+def test_orchestration_full_history_multi_turn_keeps_last_message() -> None:
+    # Multi-turn: output contains the full conversation history after multiple exchanges.
+    # Only the last message should be kept regardless of role.
+    user1 = {"role": "user", "content": "Hello"}
+    ai1 = {"role": "assistant", "content": "Hi, how can I help?"}
+    user2 = {"role": "user", "content": "What is Lisinopril?"}
+    ai2 = {"role": "assistant", "content": "Lisinopril is a blood pressure medication."}
+
+    span = AgentSpan(
+        name="Agent",
+        agent_type=AgentType.default,
+        input=json.dumps({"messages": [user1]}),
+        output=json.dumps({"messages": [user1, ai1, user2, ai2]}),
+    )
+
+    attrs = build_span_attributes(span)
+
+    output_messages = json.loads(attrs["gen_ai.output.messages"])
+    assert len(output_messages) == 1
+    assert output_messages[0]["parts"][0]["content"] == "Lisinopril is a blood pressure medication."
+
+
+def test_orchestration_full_history_multiple_tool_rounds_keeps_last_message() -> None:
+    # Two tool call rounds before the final answer — last message is still the only output.
+    user = {"role": "user", "content": "Compare Lisinopril and Amlodipine"}
+    tc1_ai = {"role": "assistant", "content": "", "tool_calls": [{"id": "tc1", "function": {"name": "search", "arguments": '{"query":"Lisinopril"}'}}]}
+    tc1_resp = {"role": "tool", "content": "Lisinopril: ACE inhibitor", "tool_call_id": "tc1"}
+    tc2_ai = {"role": "assistant", "content": "", "tool_calls": [{"id": "tc2", "function": {"name": "search", "arguments": '{"query":"Amlodipine"}'}}]}
+    tc2_resp = {"role": "tool", "content": "Amlodipine: calcium channel blocker", "tool_call_id": "tc2"}
+    ai_final = {"role": "assistant", "content": "Lisinopril is an ACE inhibitor; Amlodipine is a calcium channel blocker."}
+
+    span = AgentSpan(
+        name="Agent",
+        agent_type=AgentType.default,
+        input=json.dumps({"messages": [user]}),
+        output=json.dumps({"messages": [user, tc1_ai, tc1_resp, tc2_ai, tc2_resp, ai_final]}),
+    )
+
+    attrs = build_span_attributes(span)
+
+    output_messages = json.loads(attrs["gen_ai.output.messages"])
+    assert len(output_messages) == 1
+    assert "Amlodipine" in output_messages[0]["parts"][0]["content"]
+
+
+def test_orchestration_non_full_history_output_not_reduced() -> None:
+    # Plain string output (full_history=False) — the last-message reduction must NOT fire.
+    span = AgentSpan(
+        name="Agent",
+        agent_type=AgentType.default,
+        input="What is Lisinopril?",
+        output="Lisinopril is a blood pressure medication.",
+    )
+
+    attrs = build_span_attributes(span)
+
+    output_messages = json.loads(attrs["gen_ai.output.messages"])
+    assert len(output_messages) == 1
+    assert output_messages[0]["parts"][0]["content"] == "Lisinopril is a blood pressure medication."
+
+
 def test_orchestration_preserves_schema_valid_parts_and_tool_calls() -> None:
     span = WorkflowSpan(
         name="tool-workflow",

From b727b9180871f1772ec30477e7969e70ae95ce88 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Tue, 18 Aug 2026 10:22:30 -0500
Subject: [PATCH 03/12] fix(converter): preserve agent and workflow outputs

---
 CHANGELOG.md                                 |   7 +
 src/splunk_ao/converter/attribute_mapping.py |  18 ++-
 tests/test_attribute_mapping.py              | 144 ++++++++++++++++++-
 3 files changed, 158 insertions(+), 11 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1b750ee4..de91d4bc 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Fixed
+
+- Agent and workflow output conversion now removes only confirmed repeated input
+  history, preserves multiple terminal assistant messages, and reports the
+  standard `tool_call` finish reason when an output requests a tool and no
+  source finish reason is available.
+
 ## [0.2.1] - 2026-08-07
 
 ### Fixed
diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index 90ec0561..f3929270 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -268,7 +268,12 @@ def _orchestration_messages(value: Any, default_role: str) -> tuple[list[dict[st
 def _with_finish_reasons(messages: list[dict[str, Any]], finish_reason: str | None = None) -> list[dict[str, Any]]:
     for message in messages:
         source_finish_reason = message.get("finish_reason")
-        message["finish_reason"] = finish_reason or source_finish_reason or "unknown"
+        inferred_finish_reason = (
+            "tool_call"
+            if any(part.get("type") == "tool_call" for part in message.get("parts", []) if isinstance(part, Mapping))
+            else "unknown"
+        )
+        message["finish_reason"] = finish_reason or source_finish_reason or inferred_finish_reason
     return messages
 
 
@@ -445,12 +450,13 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
     output_messages, full_history = _orchestration_messages(span.output, "assistant")
     if output_messages is None:
         return
-    if full_history and input_messages is not None and output_messages[: len(input_messages)] == input_messages:
+    if full_history and input_messages and output_messages[: len(input_messages)] == input_messages:
         output_messages = output_messages[len(input_messages) :]
-    # Full history includes all messages in the run, not just the final response.
-    # Keep only the last message — it is always the agent's final output.
-    if full_history and len(output_messages) > 1:
-        output_messages = [output_messages[-1]]
+
+        terminal_start = len(output_messages)
+        while terminal_start > 0 and output_messages[terminal_start - 1].get("role") == "assistant":
+            terminal_start -= 1
+        output_messages = output_messages[terminal_start:]
     attrs["gen_ai.output.messages"] = _json_string(_with_finish_reasons(output_messages))
 
 
diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index 1b629b81..5e0b01de 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -1,6 +1,6 @@
 import json
 from types import SimpleNamespace
-from typing import cast
+from typing import Any, cast
 from uuid import uuid4
 
 import pytest
@@ -418,11 +418,134 @@ def test_orchestration_output_omits_repeated_input_history() -> None:
     ]
 
 
+@pytest.mark.parametrize("span_type", [WorkflowSpan, AgentSpan])
+def test_orchestration_full_history_preserves_all_terminal_assistant_messages(
+    span_type: type[WorkflowSpan] | type[AgentSpan],
+) -> None:
+    # Given: a full-history result with two terminal assistant outputs after the exact input history.
+    user = {"role": "user", "content": "Give me two alternatives"}
+    first = {"role": "assistant", "content": "First alternative"}
+    second = {"role": "assistant", "content": "Second alternative"}
+    span_kwargs: dict[str, Any] = {
+        "name": "planner",
+        "input": json.dumps({"messages": [user]}),
+        "output": json.dumps({"messages": [user, first, second]}),
+    }
+    if span_type is AgentSpan:
+        span_kwargs["agent_type"] = AgentType.planner
+
+    # When: the orchestration content is converted.
+    attrs = build_span_attributes(span_type(**span_kwargs))
+
+    # Then: the repeated input prefix is removed without reducing the terminal outputs to one message.
+    assert json.loads(attrs["gen_ai.output.messages"]) == [
+        _text_message("assistant", "First alternative", finish_reason="unknown"),
+        _text_message("assistant", "Second alternative", finish_reason="unknown"),
+    ]
+
+
+@pytest.mark.parametrize("span_type", [WorkflowSpan, AgentSpan])
+def test_orchestration_full_history_removes_confirmed_input_prefix(
+    span_type: type[WorkflowSpan] | type[AgentSpan],
+) -> None:
+    # Given: the input is the complete history immediately before the final assistant response.
+    user = {"role": "user", "content": "What is the dosage?"}
+    tool_call = {
+        "role": "assistant",
+        "content": "",
+        "tool_calls": [{"id": "call-1", "function": {"name": "search", "arguments": '{"query":"dosage"}'}}],
+    }
+    tool_response = {"role": "tool", "content": "10 mg daily", "tool_call_id": "call-1"}
+    final = {"role": "assistant", "content": "The common dosage is 10 mg daily."}
+    input_history = [user, tool_call, tool_response]
+    span_kwargs: dict[str, Any] = {
+        "name": "healthcare",
+        "input": json.dumps({"messages": input_history}),
+        "output": json.dumps({"messages": [*input_history, final]}),
+    }
+    if span_type is AgentSpan:
+        span_kwargs["agent_type"] = AgentType.default
+
+    # When: the orchestration content is converted.
+    attrs = build_span_attributes(span_type(**span_kwargs))
+
+    # Then: only the newly produced terminal response is exported as output.
+    assert json.loads(attrs["gen_ai.output.messages"]) == [
+        _text_message("assistant", "The common dosage is 10 mg daily.", finish_reason="unknown")
+    ]
+
+
+def test_orchestration_infers_tool_call_finish_reason_when_absent() -> None:
+    # Given: a workflow emits an assistant tool call without a source finish reason.
+    output = {
+        "update": {
+            "messages": [
+                {
+                    "role": "assistant",
+                    "content": "",
+                    "tool_calls": [{"id": "call-1", "function": {"name": "search", "arguments": '{"query":"dosage"}'}}],
+                }
+            ]
+        }
+    }
+
+    # When: the workflow output is converted.
+    attrs = build_span_attributes(WorkflowSpan(name="tools", output=json.dumps(output)))
+
+    # Then: the standard tool-call finish reason is inferred from the output part.
+    output_message = json.loads(attrs["gen_ai.output.messages"])[0]
+    assert output_message["finish_reason"] == "tool_call"
+    assert output_message["parts"][0]["type"] == "tool_call"
+
+
+def test_orchestration_tool_response_uses_unknown_finish_reason() -> None:
+    # Given: a workflow emits a tool response, which has no model-generation finish reason.
+    output = {
+        "update": {"messages": [{"role": "tool", "content": {"dosage": "10 mg daily"}, "tool_call_id": "call-1"}]}
+    }
+
+    # When: the workflow output is converted.
+    attrs = build_span_attributes(WorkflowSpan(name="tools", output=json.dumps(output)))
+
+    # Then: its valid tool response structure is preserved without inventing a model finish reason.
+    output_message = json.loads(attrs["gen_ai.output.messages"])[0]
+    assert output_message["finish_reason"] == "unknown"
+    assert output_message["parts"] == [
+        {"type": "tool_call_response", "id": "call-1", "response": {"dosage": "10 mg daily"}}
+    ]
+
+
+def test_orchestration_preserves_explicit_finish_reason_for_tool_call() -> None:
+    # Given: the source supplies its own finish reason for a message containing a tool call.
+    output = {
+        "update": {
+            "messages": [
+                {
+                    "role": "assistant",
+                    "content": "",
+                    "finish_reason": "provider_tool_calls",
+                    "tool_calls": [{"id": "call-1", "function": {"name": "search", "arguments": '{"query":"dosage"}'}}],
+                }
+            ]
+        }
+    }
+
+    # When: the workflow output is converted.
+    attrs = build_span_attributes(WorkflowSpan(name="tools", output=json.dumps(output)))
+
+    # Then: inference does not overwrite source telemetry.
+    assert json.loads(attrs["gen_ai.output.messages"])[0]["finish_reason"] == "provider_tool_calls"
+
+
 def test_orchestration_full_history_with_tool_call_keeps_last_message() -> None:
     # LangGraph accumulated state: user → tool-call AI (empty content) → tool response → final AI
     # The first post-dedup message has empty content; the UI would show "—" without the fix.
     user = {"role": "user", "content": "What is the dosage of Lisinopril?"}
-    ai_toolcall = {"role": "assistant", "content": "", "tool_calls": [{"id": "tc1", "function": {"name": "rag_search", "arguments": '{"query":"Lisinopril dosage"}'}}]}
+    ai_toolcall = {
+        "role": "assistant",
+        "content": "",
+        "tool_calls": [{"id": "tc1", "function": {"name": "rag_search", "arguments": '{"query":"Lisinopril dosage"}'}}],
+    }
     tool_resp = {"role": "tool", "content": "Lisinopril: 10mg daily", "tool_call_id": "tc1"}
     ai_final = {"role": "assistant", "content": "Common dosage is 10mg once daily."}
 
@@ -466,11 +589,22 @@ def test_orchestration_full_history_multi_turn_keeps_last_message() -> None:
 def test_orchestration_full_history_multiple_tool_rounds_keeps_last_message() -> None:
     # Two tool call rounds before the final answer — last message is still the only output.
     user = {"role": "user", "content": "Compare Lisinopril and Amlodipine"}
-    tc1_ai = {"role": "assistant", "content": "", "tool_calls": [{"id": "tc1", "function": {"name": "search", "arguments": '{"query":"Lisinopril"}'}}]}
+    tc1_ai = {
+        "role": "assistant",
+        "content": "",
+        "tool_calls": [{"id": "tc1", "function": {"name": "search", "arguments": '{"query":"Lisinopril"}'}}],
+    }
     tc1_resp = {"role": "tool", "content": "Lisinopril: ACE inhibitor", "tool_call_id": "tc1"}
-    tc2_ai = {"role": "assistant", "content": "", "tool_calls": [{"id": "tc2", "function": {"name": "search", "arguments": '{"query":"Amlodipine"}'}}]}
+    tc2_ai = {
+        "role": "assistant",
+        "content": "",
+        "tool_calls": [{"id": "tc2", "function": {"name": "search", "arguments": '{"query":"Amlodipine"}'}}],
+    }
     tc2_resp = {"role": "tool", "content": "Amlodipine: calcium channel blocker", "tool_call_id": "tc2"}
-    ai_final = {"role": "assistant", "content": "Lisinopril is an ACE inhibitor; Amlodipine is a calcium channel blocker."}
+    ai_final = {
+        "role": "assistant",
+        "content": "Lisinopril is an ACE inhibitor; Amlodipine is a calcium channel blocker.",
+    }
 
     span = AgentSpan(
         name="Agent",

From 766d70795004e16d55f05e1b6df65a5a9524a612 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Tue, 18 Aug 2026 10:46:00 -0500
Subject: [PATCH 04/12] test(converter): add parallel tool node output coverage

---
 tests/test_attribute_mapping.py | 22 +++++++++++++---------
 1 file changed, 13 insertions(+), 9 deletions(-)

diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index 5e0b01de..a0864e90 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -620,20 +620,24 @@ def test_orchestration_full_history_multiple_tool_rounds_keeps_last_message() ->
     assert "Amlodipine" in output_messages[0]["parts"][0]["content"]
 
 
-def test_orchestration_non_full_history_output_not_reduced() -> None:
-    # Plain string output (full_history=False) — the last-message reduction must NOT fire.
-    span = AgentSpan(
-        name="Agent",
-        agent_type=AgentType.default,
-        input="What is Lisinopril?",
-        output="Lisinopril is a blood pressure medication.",
+def test_orchestration_message_container_without_input_prefix_match_not_reduced() -> None:
+    # A WorkflowSpan (e.g. ToolNode) returning multiple messages whose output does NOT
+    # prefix-match the input state — dedup gate never fires, so all messages must survive.
+    # This is the parallel-tool-call shape: two ToolMessages from a single ToolNode invocation.
+    tool_msg_1 = {"role": "tool", "content": "Lisinopril: 10 mg daily", "tool_call_id": "tc1"}
+    tool_msg_2 = {"role": "tool", "content": "Amlodipine: 5 mg daily", "tool_call_id": "tc2"}
+    span = WorkflowSpan(
+        name="tools",
+        input=json.dumps({"messages": [{"role": "user", "content": "Compare dosages"}]}),
+        output=json.dumps({"messages": [tool_msg_1, tool_msg_2]}),
     )
 
     attrs = build_span_attributes(span)
 
     output_messages = json.loads(attrs["gen_ai.output.messages"])
-    assert len(output_messages) == 1
-    assert output_messages[0]["parts"][0]["content"] == "Lisinopril is a blood pressure medication."
+    assert len(output_messages) == 2
+    assert output_messages[0]["parts"][0]["response"] == "Lisinopril: 10 mg daily"
+    assert output_messages[1]["parts"][0]["response"] == "Amlodipine: 5 mg daily"
 
 
 def test_orchestration_preserves_schema_valid_parts_and_tool_calls() -> None:

From ce3d54c62dc052e1c534d7a1a43d8a134a8b30a9 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Tue, 18 Aug 2026 15:37:47 -0500
Subject: [PATCH 05/12] fix(converter): simplify full-history trim to last
 message and add edge case tests

---
 CHANGELOG.md                                 | 10 ++--
 src/splunk_ao/converter/attribute_mapping.py |  7 +--
 tests/test_attribute_mapping.py              | 57 +++++++++++++++++---
 3 files changed, 58 insertions(+), 16 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index de91d4bc..7f33e391 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,10 +9,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
-- Agent and workflow output conversion now removes only confirmed repeated input
-  history, preserves multiple terminal assistant messages, and reports the
-  standard `tool_call` finish reason when an output requests a tool and no
-  source finish reason is available.
+- Agent and workflow output conversion now removes confirmed repeated input
+  history and keeps only the last message as the terminal output; intermediate
+  tool-call and tool-response messages are no longer included in
+  `gen_ai.output.messages` for full-history spans. The `tool_call` finish reason
+  is inferred when an output requests a tool and no source finish reason is
+  available.
 
 ## [0.2.1] - 2026-08-07
 
diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index f3929270..03be3c3e 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -451,12 +451,7 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
     if output_messages is None:
         return
     if full_history and input_messages and output_messages[: len(input_messages)] == input_messages:
-        output_messages = output_messages[len(input_messages) :]
-
-        terminal_start = len(output_messages)
-        while terminal_start > 0 and output_messages[terminal_start - 1].get("role") == "assistant":
-            terminal_start -= 1
-        output_messages = output_messages[terminal_start:]
+        output_messages = output_messages[len(input_messages) :][-1:]
     attrs["gen_ai.output.messages"] = _json_string(_with_finish_reasons(output_messages))
 
 
diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index a0864e90..f9151e00 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -419,10 +419,9 @@ def test_orchestration_output_omits_repeated_input_history() -> None:
 
 
 @pytest.mark.parametrize("span_type", [WorkflowSpan, AgentSpan])
-def test_orchestration_full_history_preserves_all_terminal_assistant_messages(
+def test_orchestration_full_history_keeps_last_terminal_message(
     span_type: type[WorkflowSpan] | type[AgentSpan],
 ) -> None:
-    # Given: a full-history result with two terminal assistant outputs after the exact input history.
     user = {"role": "user", "content": "Give me two alternatives"}
     first = {"role": "assistant", "content": "First alternative"}
     second = {"role": "assistant", "content": "Second alternative"}
@@ -434,13 +433,10 @@ def test_orchestration_full_history_preserves_all_terminal_assistant_messages(
     if span_type is AgentSpan:
         span_kwargs["agent_type"] = AgentType.planner
 
-    # When: the orchestration content is converted.
     attrs = build_span_attributes(span_type(**span_kwargs))
 
-    # Then: the repeated input prefix is removed without reducing the terminal outputs to one message.
     assert json.loads(attrs["gen_ai.output.messages"]) == [
-        _text_message("assistant", "First alternative", finish_reason="unknown"),
-        _text_message("assistant", "Second alternative", finish_reason="unknown"),
+        _text_message("assistant", "Second alternative", finish_reason="unknown")
     ]
 
 
@@ -640,6 +636,55 @@ def test_orchestration_message_container_without_input_prefix_match_not_reduced(
     assert output_messages[1]["parts"][0]["response"] == "Amlodipine: 5 mg daily"
 
 
+def test_orchestration_full_history_ends_on_tool_message_keeps_last() -> None:
+    # return_direct=True tool: run ends on a tool response, no final assistant message.
+    # The dedup gate fires (prefix matches) but the last message is a tool, not assistant.
+    # Must return the tool message rather than an empty list.
+    user = {"role": "user", "content": "Get patient P001"}
+    ai_toolcall = {
+        "role": "assistant",
+        "content": "",
+        "tool_calls": [{"id": "tc1", "function": {"name": "get_patient", "arguments": '{"id":"P001"}'}}],
+    }
+    tool_resp = {"role": "tool", "content": "George Rivera, Lisinopril 10mg", "tool_call_id": "tc1"}
+    span = AgentSpan(
+        name="Agent",
+        agent_type=AgentType.default,
+        input=json.dumps({"messages": [user, ai_toolcall]}),
+        output=json.dumps({"messages": [user, ai_toolcall, tool_resp]}),
+    )
+
+    attrs = build_span_attributes(span)
+
+    output_messages = json.loads(attrs["gen_ai.output.messages"])
+    assert len(output_messages) == 1
+    assert output_messages[0]["parts"][0]["response"] == "George Rivera, Lisinopril 10mg"
+
+
+def test_orchestration_full_history_ends_on_tool_call_ai_message_keeps_last() -> None:
+    # interrupt_before=["tools"]: run ends on a tool-call AIMessage with empty content.
+    # The last message is assistant role but content="" — must not return empty list.
+    user = {"role": "user", "content": "Search for Lisinopril"}
+    ai_toolcall = {
+        "role": "assistant",
+        "content": "",
+        "tool_calls": [{"id": "tc1", "function": {"name": "search", "arguments": '{"query":"Lisinopril"}'}}],
+    }
+    span = AgentSpan(
+        name="Agent",
+        agent_type=AgentType.default,
+        input=json.dumps({"messages": [user]}),
+        output=json.dumps({"messages": [user, ai_toolcall]}),
+    )
+
+    attrs = build_span_attributes(span)
+
+    output_messages = json.loads(attrs["gen_ai.output.messages"])
+    assert len(output_messages) == 1
+    assert output_messages[0]["role"] == "assistant"
+    assert output_messages[0]["finish_reason"] == "tool_call"
+
+
 def test_orchestration_preserves_schema_valid_parts_and_tool_calls() -> None:
     span = WorkflowSpan(
         name="tool-workflow",

From 86c3be1a5912561a59118c4340a01e1b9fa48ae5 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Tue, 18 Aug 2026 22:45:36 -0500
Subject: [PATCH 06/12] refactor(converter): drop unrelated tool_call
 finish-reason inference and add WorkflowSpan trim test
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Removes the tool_call finish-reason inference added to _with_finish_reasons —
it is unrelated to HYBIM-988 and widens blast radius to LLM spans. Updates
CHANGELOG and adds WorkflowSpan coverage for the full-history trim path.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 CHANGELOG.md                                 |  4 +---
 src/splunk_ao/converter/attribute_mapping.py |  7 +-----
 tests/test_attribute_mapping.py              | 25 ++++++++++++++++----
 3 files changed, 23 insertions(+), 13 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7f33e391..0f1bf14a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -12,9 +12,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Agent and workflow output conversion now removes confirmed repeated input
   history and keeps only the last message as the terminal output; intermediate
   tool-call and tool-response messages are no longer included in
-  `gen_ai.output.messages` for full-history spans. The `tool_call` finish reason
-  is inferred when an output requests a tool and no source finish reason is
-  available.
+  `gen_ai.output.messages` for full-history spans.
 
 ## [0.2.1] - 2026-08-07
 
diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index 03be3c3e..d14b8f6e 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -268,12 +268,7 @@ def _orchestration_messages(value: Any, default_role: str) -> tuple[list[dict[st
 def _with_finish_reasons(messages: list[dict[str, Any]], finish_reason: str | None = None) -> list[dict[str, Any]]:
     for message in messages:
         source_finish_reason = message.get("finish_reason")
-        inferred_finish_reason = (
-            "tool_call"
-            if any(part.get("type") == "tool_call" for part in message.get("parts", []) if isinstance(part, Mapping))
-            else "unknown"
-        )
-        message["finish_reason"] = finish_reason or source_finish_reason or inferred_finish_reason
+        message["finish_reason"] = finish_reason or source_finish_reason or "unknown"
     return messages
 
 
diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index f9151e00..4c19098a 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -471,7 +471,7 @@ def test_orchestration_full_history_removes_confirmed_input_prefix(
     ]
 
 
-def test_orchestration_infers_tool_call_finish_reason_when_absent() -> None:
+def test_orchestration_tool_call_assistant_message_finish_reason_unknown() -> None:
     # Given: a workflow emits an assistant tool call without a source finish reason.
     output = {
         "update": {
@@ -488,9 +488,9 @@ def test_orchestration_infers_tool_call_finish_reason_when_absent() -> None:
     # When: the workflow output is converted.
     attrs = build_span_attributes(WorkflowSpan(name="tools", output=json.dumps(output)))
 
-    # Then: the standard tool-call finish reason is inferred from the output part.
+    # Then: finish_reason defaults to "unknown" — no inference from parts.
     output_message = json.loads(attrs["gen_ai.output.messages"])[0]
-    assert output_message["finish_reason"] == "tool_call"
+    assert output_message["finish_reason"] == "unknown"
     assert output_message["parts"][0]["type"] == "tool_call"
 
 
@@ -682,7 +682,24 @@ def test_orchestration_full_history_ends_on_tool_call_ai_message_keeps_last() ->
     output_messages = json.loads(attrs["gen_ai.output.messages"])
     assert len(output_messages) == 1
     assert output_messages[0]["role"] == "assistant"
-    assert output_messages[0]["finish_reason"] == "tool_call"
+    assert output_messages[0]["finish_reason"] == "unknown"
+
+
+def test_orchestration_full_history_workflow_span_trim() -> None:
+    # WorkflowSpan (non-root LangGraph node) that carries full state: the trim
+    # must fire the same way it does for AgentSpan when the prefix matches.
+    user = {"role": "user", "content": "What is Lisinopril?"}
+    assistant = {"role": "assistant", "content": "Lisinopril is an ACE inhibitor."}
+    span = WorkflowSpan(
+        name="summarise", input=json.dumps({"messages": [user]}), output=json.dumps({"messages": [user, assistant]})
+    )
+
+    attrs = build_span_attributes(span)
+
+    output_messages = json.loads(attrs["gen_ai.output.messages"])
+    assert len(output_messages) == 1
+    assert output_messages[0]["role"] == "assistant"
+    assert output_messages[0]["parts"][0]["content"] == "Lisinopril is an ACE inhibitor."
 
 
 def test_orchestration_preserves_schema_valid_parts_and_tool_calls() -> None:

From 6439657f967392ef86803b739f60e57ea0386c6d Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Tue, 18 Aug 2026 22:49:38 -0500
Subject: [PATCH 07/12] docs(converter): document prefix-match gate and
 threaded-history limitation

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 src/splunk_ao/converter/attribute_mapping.py | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index d14b8f6e..25035f01 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -445,6 +445,11 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
     output_messages, full_history = _orchestration_messages(span.output, "assistant")
     if output_messages is None:
         return
+    # Reduction is intentionally tied to a confirmed prefix match: we only trim when the
+    # output starts with an exact copy of the input, which is the LangGraph stateless
+    # (no checkpointer) full-history shape. With a checkpointer the input is the new turn
+    # only while the output carries the whole persisted thread, so the prefix never matches
+    # and no reduction fires — known limitation, tracked separately.
     if full_history and input_messages and output_messages[: len(input_messages)] == input_messages:
         output_messages = output_messages[len(input_messages) :][-1:]
     attrs["gen_ai.output.messages"] = _json_string(_with_finish_reasons(output_messages))

From 58a1f79b957c9081c730784cc8f82629b12ab659 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Wed, 19 Aug 2026 19:36:16 -0500
Subject: [PATCH 08/12] fix(converter): restore tool_call finish-reason
 inference and gate trim on history_stripped

- _with_finish_reasons: infer "tool_call" finish reason when a message has
  tool_call parts but no explicit source finish reason; "unknown" remains
  the fallback for all other cases
- _set_orchestration_content: split dedup and [-1:] trim using a
  history_stripped flag so parallel tool-call outputs (ToolNode) are never
  collapsed to one message when the output does not echo the input prefix
- Update two test assertions from "unknown" to "tool_call" for tool-call
  assistant message cases
- CHANGELOG updated with both fixes
---
 CHANGELOG.md                                 |  9 ++++++++-
 src/splunk_ao/converter/attribute_mapping.py | 17 +++++++++++++++--
 tests/test_attribute_mapping.py              |  6 +++---
 3 files changed, 26 insertions(+), 6 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0f1bf14a..10651f77 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -12,7 +12,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Agent and workflow output conversion now removes confirmed repeated input
   history and keeps only the last message as the terminal output; intermediate
   tool-call and tool-response messages are no longer included in
-  `gen_ai.output.messages` for full-history spans.
+  `gen_ai.output.messages` for full-history spans. The trim is now gated on the
+  prefix-match dedup check so parallel tool-call outputs (e.g. a ToolNode with
+  multiple simultaneous calls) are never collapsed to one message when the output
+  does not echo the input.
+- `gen_ai.output.messages` assistant entries whose parts contain a `tool_call`
+  but carry no explicit `finish_reason` now receive `"tool_call"` instead of
+  `"unknown"`. Explicit source finish reasons and tool-response messages are
+  unaffected.
 
 ## [0.2.1] - 2026-08-07
 
diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index 25035f01..357fdba1 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -268,7 +268,13 @@ def _orchestration_messages(value: Any, default_role: str) -> tuple[list[dict[st
 def _with_finish_reasons(messages: list[dict[str, Any]], finish_reason: str | None = None) -> list[dict[str, Any]]:
     for message in messages:
         source_finish_reason = message.get("finish_reason")
-        message["finish_reason"] = finish_reason or source_finish_reason or "unknown"
+        # Infer "tool_call" when the message has tool_call parts but no explicit finish reason.
+        inferred_finish_reason = (
+            "tool_call"
+            if any(part.get("type") == "tool_call" for part in message.get("parts", []) if isinstance(part, Mapping))
+            else "unknown"
+        )
+        message["finish_reason"] = finish_reason or source_finish_reason or inferred_finish_reason
     return messages
 
 
@@ -450,8 +456,15 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
     # (no checkpointer) full-history shape. With a checkpointer the input is the new turn
     # only while the output carries the whole persisted thread, so the prefix never matches
     # and no reduction fires — known limitation, tracked separately.
+    # The [-1:] trim is gated on history_stripped so that parallel tool-call outputs (e.g.
+    # a ToolNode with multiple simultaneous calls) are never collapsed to one message when
+    # the output is not a full-history echo of the input.
+    history_stripped = False
     if full_history and input_messages and output_messages[: len(input_messages)] == input_messages:
-        output_messages = output_messages[len(input_messages) :][-1:]
+        output_messages = output_messages[len(input_messages) :]
+        history_stripped = True
+    if history_stripped:
+        output_messages = output_messages[-1:]
     attrs["gen_ai.output.messages"] = _json_string(_with_finish_reasons(output_messages))
 
 
diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index 4c19098a..ca5946ec 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -488,9 +488,9 @@ def test_orchestration_tool_call_assistant_message_finish_reason_unknown() -> No
     # When: the workflow output is converted.
     attrs = build_span_attributes(WorkflowSpan(name="tools", output=json.dumps(output)))
 
-    # Then: finish_reason defaults to "unknown" — no inference from parts.
+    # Then: finish_reason is inferred as "tool_call" from the parts.
     output_message = json.loads(attrs["gen_ai.output.messages"])[0]
-    assert output_message["finish_reason"] == "unknown"
+    assert output_message["finish_reason"] == "tool_call"
     assert output_message["parts"][0]["type"] == "tool_call"
 
 
@@ -682,7 +682,7 @@ def test_orchestration_full_history_ends_on_tool_call_ai_message_keeps_last() ->
     output_messages = json.loads(attrs["gen_ai.output.messages"])
     assert len(output_messages) == 1
     assert output_messages[0]["role"] == "assistant"
-    assert output_messages[0]["finish_reason"] == "unknown"
+    assert output_messages[0]["finish_reason"] == "tool_call"
 
 
 def test_orchestration_full_history_workflow_span_trim() -> None:

From 73421d7dce4d0ac9d6af3c88bfa8eb178ade9100 Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Thu, 20 Aug 2026 09:36:01 -0500
Subject: [PATCH 09/12] =?UTF-8?q?fix(converter):=20apply=20reviewer=20sugg?=
 =?UTF-8?q?estions=20=E2=80=94=20replace=20block=20comment=20and=20drop=20?=
 =?UTF-8?q?redundant=20test?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/splunk_ao/converter/attribute_mapping.py | 15 +++++++--------
 tests/test_attribute_mapping.py              | 16 ----------------
 2 files changed, 7 insertions(+), 24 deletions(-)

diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index 357fdba1..76e50384 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -451,14 +451,13 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
     output_messages, full_history = _orchestration_messages(span.output, "assistant")
     if output_messages is None:
         return
-    # Reduction is intentionally tied to a confirmed prefix match: we only trim when the
-    # output starts with an exact copy of the input, which is the LangGraph stateless
-    # (no checkpointer) full-history shape. With a checkpointer the input is the new turn
-    # only while the output carries the whole persisted thread, so the prefix never matches
-    # and no reduction fires — known limitation, tracked separately.
-    # The [-1:] trim is gated on history_stripped so that parallel tool-call outputs (e.g.
-    # a ToolNode with multiple simultaneous calls) are never collapsed to one message when
-    # the output is not a full-history echo of the input.
+    # The trim is gated on a confirmed prefix match so that it only fires when the output
+    # is genuinely accumulated input history. Without this gate, any span whose output is a
+    # top-level {"messages": [...]} container would be reduced to a single message — a
+    # LangGraph ToolNode emitting one ToolMessage per parallel tool call would lose all but
+    # the last. Consequence: with a checkpointer the input is only the new turn while the
+    # output carries the whole persisted thread, so the prefix never matches and no
+    # reduction fires. That case remains unhandled.
     history_stripped = False
     if full_history and input_messages and output_messages[: len(input_messages)] == input_messages:
         output_messages = output_messages[len(input_messages) :]
diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index ca5946ec..31a01e7c 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -685,22 +685,6 @@ def test_orchestration_full_history_ends_on_tool_call_ai_message_keeps_last() ->
     assert output_messages[0]["finish_reason"] == "tool_call"
 
 
-def test_orchestration_full_history_workflow_span_trim() -> None:
-    # WorkflowSpan (non-root LangGraph node) that carries full state: the trim
-    # must fire the same way it does for AgentSpan when the prefix matches.
-    user = {"role": "user", "content": "What is Lisinopril?"}
-    assistant = {"role": "assistant", "content": "Lisinopril is an ACE inhibitor."}
-    span = WorkflowSpan(
-        name="summarise", input=json.dumps({"messages": [user]}), output=json.dumps({"messages": [user, assistant]})
-    )
-
-    attrs = build_span_attributes(span)
-
-    output_messages = json.loads(attrs["gen_ai.output.messages"])
-    assert len(output_messages) == 1
-    assert output_messages[0]["role"] == "assistant"
-    assert output_messages[0]["parts"][0]["content"] == "Lisinopril is an ACE inhibitor."
-
 
 def test_orchestration_preserves_schema_valid_parts_and_tool_calls() -> None:
     span = WorkflowSpan(

From 7fdd4739702bb538b05d2aaa981584f15647a711 Mon Sep 17 00:00:00 2001
From: Erdenesaikhan Tserendavga <105012329+etserend@users.noreply.github.com>
Date: Thu, 20 Aug 2026 14:31:04 -0500
Subject: [PATCH 10/12] Update src/splunk_ao/converter/attribute_mapping.py

Co-authored-by: Fernando Correia <fercor@cisco.com>
---
 src/splunk_ao/converter/attribute_mapping.py | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index 76e50384..d8a7b05c 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -463,7 +463,15 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
         output_messages = output_messages[len(input_messages) :]
         history_stripped = True
     if history_stripped:
-        output_messages = output_messages[-1:]
+        # Prefer the last non-user message: the terminal message may legitimately be a tool
+        # response (return_direct) or a tool-call AIMessage, but a trailing user turn is an
+        # input, not this span's output.
+        terminal = next(
+            (index for index in reversed(range(len(output_messages))) if output_messages[index].get("role") != "user"),
+            None,
+        )
+        if terminal is not None:
+            output_messages = output_messages[terminal : terminal + 1]
     attrs["gen_ai.output.messages"] = _json_string(_with_finish_reasons(output_messages))
 
 

From 7c19f7514a08e79c741b97ca7e1ac4c97487919e Mon Sep 17 00:00:00 2001
From: Erdenesaikhan Tserendavga <105012329+etserend@users.noreply.github.com>
Date: Thu, 20 Aug 2026 14:31:15 -0500
Subject: [PATCH 11/12] Update tests/test_attribute_mapping.py

Co-authored-by: Fernando Correia <fercor@cisco.com>
---
 tests/test_attribute_mapping.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index 31a01e7c..689473c1 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -471,7 +471,7 @@ def test_orchestration_full_history_removes_confirmed_input_prefix(
     ]
 
 
-def test_orchestration_tool_call_assistant_message_finish_reason_unknown() -> None:
+def test_orchestration_tool_call_assistant_message_infers_tool_call_finish_reason() -> None:
     # Given: a workflow emits an assistant tool call without a source finish reason.
     output = {
         "update": {

From f1d4ef620a3d8803f2c4a4c196c7a6719aa3d76f Mon Sep 17 00:00:00 2001
From: etserend <ertserendavga@splunk.com>
Date: Thu, 20 Aug 2026 18:07:26 -0500
Subject: [PATCH 12/12] =?UTF-8?q?fix(converter):=20apply=20reviewer=20sugg?=
 =?UTF-8?q?estions=20=E2=80=94=20fold=20history=5Fstripped=20and=20prefer?=
 =?UTF-8?q?=20last=20non-user=20message?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- Remove history_stripped flag; fold non-user-message trim into the single if block (r3824609327)
- Replace [-1:] with last non-user-message search to avoid surfacing trailing user turns (r3824609186)
- Rename test to reflect tool_call inference rather than unknown finish reason (r3824609453)
- Remove extra blank line at line 688 (r3824609548)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 src/splunk_ao/converter/attribute_mapping.py | 3 ---
 tests/test_attribute_mapping.py              | 1 -
 2 files changed, 4 deletions(-)

diff --git a/src/splunk_ao/converter/attribute_mapping.py b/src/splunk_ao/converter/attribute_mapping.py
index d8a7b05c..39b5dde9 100644
--- a/src/splunk_ao/converter/attribute_mapping.py
+++ b/src/splunk_ao/converter/attribute_mapping.py
@@ -458,11 +458,8 @@ def _set_orchestration_content(attrs: MutableMapping[str, AttributeValue], span:
     # the last. Consequence: with a checkpointer the input is only the new turn while the
     # output carries the whole persisted thread, so the prefix never matches and no
     # reduction fires. That case remains unhandled.
-    history_stripped = False
     if full_history and input_messages and output_messages[: len(input_messages)] == input_messages:
         output_messages = output_messages[len(input_messages) :]
-        history_stripped = True
-    if history_stripped:
         # Prefer the last non-user message: the terminal message may legitimately be a tool
         # response (return_direct) or a tool-call AIMessage, but a trailing user turn is an
         # input, not this span's output.
diff --git a/tests/test_attribute_mapping.py b/tests/test_attribute_mapping.py
index 689473c1..7cdd6662 100644
--- a/tests/test_attribute_mapping.py
+++ b/tests/test_attribute_mapping.py
@@ -685,7 +685,6 @@ def test_orchestration_full_history_ends_on_tool_call_ai_message_keeps_last() ->
     assert output_messages[0]["finish_reason"] == "tool_call"
 
 
-
 def test_orchestration_preserves_schema_valid_parts_and_tool_calls() -> None:
     span = WorkflowSpan(
         name="tool-workflow",