[16.0] llm_anthropic: KeyError 'role' — missing format_messages implementation

[Excauboi]

Bug Description

The llm_anthropic module is missing the format_messages / format_message infrastructure that other providers (e.g., llm_openai) implement. This causes a KeyError: 'role' when anthropic_chat() receives mail.message records instead of pre-formatted dicts.

Steps to Reproduce

1. Install llm, llm_anthropic, llm_thread, llm_assistant, llm_tool
2. Configure an Anthropic provider with a valid API key
3. Create a model, assistant, prompt, and thread
4. Send a message via the chat UI or SSE endpoint GET /llm/thread/generate?thread_id=X&message=Y

Expected Behavior

Assistant responds with LLM-generated content.

Actual Behavior

SSE returns:

data: {"type": "error", "error": "'role'"}
data: {"type": "done"}

Root Cause

The flow is:

1. llm_assistant/models/llm_thread.py → _generate_assistant_response() calls self.get_llm_messages() which returns a mail.message recordset
2. This recordset is passed through model.chat() → provider.chat() → anthropic_chat()
3. anthropic_chat() iterates messages doing msg["role"] — but mail.message records have the field llm_role, not role

The llm_openai module handles this correctly by implementing:

• openai_format_messages() on llm.provider (converts recordset to list of dicts)
• openai_format_message() on mail.message (converts a single record)
• openai_chat() uses _prepare_chat_params() which calls format_messages()

The llm_anthropic module has none of this — anthropic_chat() directly iterates the raw recordset assuming dict-like role/content keys.

Proposed Fix

Add anthropic_format_message() to mail.message and anthropic_format_messages() to llm.provider, following the same pattern as llm_openai. Then update anthropic_chat() to use _prepare_chat_params_base() or call format_messages() before processing.

Environment

• Odoo 16 Enterprise (Clodofy hosting)
• llm v16.0.1.4.2
• llm_anthropic v16.0.1.1.0
• llm_thread v16.0.1.3.2
• llm_assistant v16.0.1.5.3
• llm_tool v16.0.3.0.0

[ayushin]

Thank you for reporting. The main branch is 18.0 and e need to back-port some of the functionality into 16.0 Any takers?

[Excauboi]

We are on it :)