fix: preserve system message during multi-function registration in AgentBuilder

[roli-lpci]

Summary

Fixes #5037.

When AgentBuilder._build_agents() registers multiple functions to a single agent, the system message is overwritten on each iteration rather than accumulated. The loop reads from the static agent_configs dict (which holds the raw LLM-generated config and never mutates) instead of the live agent's .system_message property (which reflects prior update_system_message() calls).

Result: After registering N functions, the agent's system message contains only the last function's description. All prior function descriptions are lost, and the GROUP_CHAT_DESCRIPTION wrapper — including TERMINATE instructions — is destroyed.

Root cause

# Before (line 759): reads from STATIC config — always returns original raw message
agents_current_system_message = [
    agent["system_message"] for agent in agent_configs if agent["name"] == perferred_agent_name
][0]

# After: reads from LIVE agent — reflects all prior update_system_message() calls
preferred_agent.system_message

What the fix preserves

• All function descriptions across loop iterations
• GROUP_CHAT_DESCRIPTION wrapper (TERMINATE instruction, group chat context, member list)
• CODING_AND_TASK_SKILL_INSTRUCTION block

Also extracts preferred_agent as a local variable, eliminating 3 repeated self.agent_procs_assign[perferred_agent_name][0] lookups.

Test plan

• New regression test test_multi_function_system_message_preserved() — exercises actual _build_agents() with 3 functions assigned to one agent. 9 assertions verify all function names, descriptions, TERMINATE instruction, and GROUP_CHAT_DESCRIPTION structure are preserved. Runs without API keys (mocks config_list_from_json and builder_model.create).
• Result: 1 passed, 7 skipped (skipped tests are OpenAI-gated integration tests)
• Existing tests unaffected

🤖 Generated with Claude Code