fix: set_conversational_output leaking CAS event, phantom CAS events for OpenAI envelope chunk

[maxduu]

Summary

Two independent fixes that stop the runtime from emitting phantom messageStart / toolCall events to CAS for internal machinery that shouldn't be user-visible.

1. OpenAI Responses API envelope chunk

Issue: OpenAI's Responses API (e.g. gpt-5.4) fronts a stream with an empty envelope chunk (id=resp_..., from the SSE response.created event), then streams the actual content under a different id (lc_run--...). The mapper treated the envelope as a new message and emitted a phantom messageStart + contentPartStart pair to CAS.

Fix: src/uipath_langchain/runtime/messages.py — shape-guard that skips streaming message-start/content-part-start events for chunks with no content, no tool calls, and chunk_position != "last".

Before:

After:
(With temporary prints). Note the "Skipping phantom message start..." print.

2. set_conversational_output tool call leaking to CAS

cc @andrewwan-uipath

Why: GENERATE_CONVERSATIONAL_OUTPUT runs an internal LLM call with tool_choice="any" bound only to the set_conversational_output tool. TAG_NOSTREAM on the invoke config correctly suppressed the LLM callback path, but LangGraph's messages-stream mode also streams message-outputs from the return of {"messages": [response]}. So the chat-mapper emitted toolCallStart + executingToolCall events to CAS for the synthetic tool call.

Fix: Extract the tool call args inside the node and route them via inner_state.conversational_output instead of messages. The AIMessage never enters the messages channel, so there's nothing for LangGraph's sweep to pick up. This refactors and simplifies a lot of the terminate_node logic since it simply just obtains the output from the inner_state of the graph rather than parsing the last message.

Before:
Note the temporary log Mapping AIMessage to events: resp_...

After:

Test plan

• Added test_returns_extracted_args_via_inner_state and test_raises_when_llm_omits_set_conversational_output_call in test_conversational_output_node.py
• Updated test_conversational_extracts_custom_output_fields to pass args via inner_state
• Added envelope-chunk coverage in test_chat_message_mapper.py
• All 2388 tests pass, lint clean

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Prevents the runtime from emitting phantom CAS events for (1) OpenAI Responses API “envelope-only” chunks and (2) the internal set_conversational_output tool call used for conversational structured output extraction.

Changes:

• Add a shape-guard to skip envelope-only AIMessageChunks (no content / no tool calls / not chunk_position="last") to avoid phantom messageStart events.
• Route set_conversational_output tool-call args via inner_state.conversational_output (instead of the messages channel) so LangGraph’s messages-stream sweep can’t leak synthetic tool calls to CAS.
• Update/extend tests to cover envelope-chunk skipping and the new conversational output plumbing + error path.

Reviewed changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
uv.lock	Bumps package version to 0.14.4.
pyproject.toml	Bumps project version to 0.14.4.
src/uipath_langchain/runtime/messages.py	Skips envelope-only chunks to prevent phantom CAS message start/content-part start events.
tests/runtime/test_chat_message_mapper.py	Updates existing mapper tests and adds coverage for envelope-only chunk skipping + empty-last-chunk behavior.
src/uipath_langchain/agent/react/types.py	Adds InnerAgentGraphState.conversational_output to carry extracted structured output args.
src/uipath_langchain/agent/react/conversational_output_node.py	Extracts set_conversational_output args from the internal LLM response and stores them in inner_state; raises LLM_INVALID_RESPONSE if the tool call is missing.
src/uipath_langchain/agent/react/terminate_node.py	Reads structured custom output fields from inner_state.conversational_output and removes graph-shape-dependent logic.
src/uipath_langchain/agent/react/agent.py	Simplifies terminate node creation (no longer passes with_conversational_output_node).
tests/agent/react/test_conversational_output_node.py	Updates expectations to assert inner_state updates (no messages) and adds test for missing tool call -> structured error.
tests/agent/react/test_terminate_node.py	Updates termination behavior tests to supply conversational output via inner_state.
tests/agent/react/test_create_agent.py	Updates mocks/assertions for the new create_terminate_node signature usage.

[maxduu]

Removed outdated comment.

[sonarqubecloud]

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud