fix(llm): normalize message part order for Vertex AI Gemini tool turns (fixes #156)

[VG31OP]

Fixes #156.

Problem

In ARTEMIS's mobile workflows using Vertex AI Gemini (e.g. gemini-3.8-flash), the first model turn succeeds and executes a UI action. However, when ARTEMIS sends the follow-up request with the tool result and the next screen observation, Vertex AI returns:

400 Requests ending with a model turn are not supported

This occurs because ARTEMIS appends a ToolMessage (tool execution result) followed by a HumanMessage (new screen observation) to its conversation history. LangChain's Vertex serializer (_parse_chat_history_gemini) converts ToolMessage to functionResponse and merges consecutive user-turn messages into a single Content(role="user") object with parts ordered as [functionResponse, text]. Vertex Gemini 3.x rejects this part ordering.

Solution

1. Outbound Normalization (normalize_vertex_messages):
   • In artemis/services/llm.py, produces a normalized copy of the message list for Vertex AI invocations where HumanMessage observations in a tool-calling turn are positioned before ToolMessage results.
   • Populates ToolMessage.name from the preceding AIMessage.tool_calls map if missing.
   • Preserves the underlying stored conversation history in TranscriptLedger and State without mutation.
2. Gateway Integration:
   • Updated RobustChatModelWrapper (complete(), ainvoke(), invoke()) and acomplete() to automatically detect Vertex AI endpoints (is_vertex_provider or ChatVertexAI class name) and apply the normalization before request execution.
   • Other providers (Google Developer API, OpenAI, Anthropic, etc.) remain completely untouched.
3. Provider Helper:
   • Added is_vertex_provider() in artemis.llm.google to cleanly check for Vertex AI provider configurations.

Validation

• New Regression Tests: Added tests/unit/test_vertex_message_normalizer.py (7 tests) covering:
  • Preserving empty and non-tool conversation histories.
  • Reordering screen observation text before tool responses.
  • Multi-tool bursts and multi-turn workflows.
  • LangChain _parse_chat_history_gemini part ordering verification ([text, function_response]).
  • RobustChatModelWrapper and acomplete gateway integration.
  • Non-Vertex provider isolation.
• Unit Test Suite: 61/61 tests passed across test_vertex_message_normalizer.py, test_google_provider.py, and test_llm_gateway.py.
• Linting & Formatting: ruff check and ruff format --check passed cleanly across all changed files.

[gemini-code-assist]

Code Review

This pull request introduces message normalization for Vertex AI Gemini models to prevent HTTP 400 errors caused by incorrect ordering of tool responses and user observations. It adds a normalize_vertex_messages utility, integrates it into the robust chat model wrapper and completion functions, and includes comprehensive unit tests. The review feedback suggests using getattr to safely access tool_call_id on messages to avoid potential AttributeErrors, and extracting the duplicated input preparation logic into a static helper method to improve maintainability.

[VG31OP]

Thank you for the review! I've addressed the feedback in commit 83b4445:

1. Extracted Static Helper: Extracted RobustChatModelWrapper._prepare_vertex_inputs to eliminate duplicated input preparation logic between RobustChatModelWrapper._prepare_inputs and acomplete().
2. Safe Attribute Access: Used getattr(m, 'tool_call_id', None) when inspecting ToolMessages in normalize_vertex_messages to prevent potential AttributeErrors on mock or custom message objects.
3. Validation: All 61/61 unit tests pass cleanly alongside ruff check and ruff format --check.

[VG31OP]

@gemini-code-assist please review the latest updates in commit 83b4445.