Version: ruby_llm 1.15.0

Description

RubyLLM::Providers::Anthropic::Chat#build_system_content logs a warning whenever a chat has more than one :system-role message:

# lib/ruby_llm/providers/anthropic/chat.rb:34-39
if system_messages.length > 1
  RubyLLM.logger.warn(
    "Anthropic's Claude implementation only supports a single system message. " \
    'Multiple system messages will be combined into one.'
  )
end

The warning is misleading: Anthropic's system parameter accepts an array of text content blocks (each optionally with cache_control), and the gem already correctly serializes multiple system messages into that array via flat_map:

system_messages.flat_map do |msg|
  content = msg.content
  if content.is_a?(RubyLLM::Content::Raw)
    content.value
  else
    Media.format_content(content)
  end
end

So passing two :system messages via chat.with_instructions(a).with_instructions(b, append: true) produces a valid two-element system: [...] payload, with each block's cache_control preserved. The warning's claim that the gem is "combining them into one" is no longer accurate (and Anthropic's API has accepted multi-block system content for years).

Impact

Applications that intentionally split system content into separate blocks — e.g., a static prompt + a per-conversation context envelope — get spurious log spew on every request. In high-throughput agent settings, this floods logs.

Reproduction

chat = RubyLLM.chat(model: 'claude-sonnet-4-6')
chat.with_instructions("Static prompt.")
chat.with_instructions("Per-session context.", append: true)
chat.ask("Hello")
# Logs: "Anthropic's Claude implementation only supports a single system message. ..."
# Actual API payload:
#   system: [
#     { type: "text", text: "Static prompt." },
#     { type: "text", text: "Per-session context." }
#   ]
# — correctly multi-block.

Suggested fix

Remove the length > 1 warning block. Optional: replace it with a brief comment documenting that multi-block system content is supported and is forwarded faithfully to Anthropic's system array.

If there's appetite, also add a test asserting that given two :system messages the rendered Anthropic payload contains two distinct text blocks (and per-block cache_control is preserved through the round-trip).

References

• Anthropic Messages API: https://docs.anthropic.com/en/api/messages — system accepts a string or an array of text content blocks.
• Prompt caching docs with cache_control on system blocks: https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching
• Multi-block system content has been supported since the prompt-caching launch in August 2024 (announcement: https://www.anthropic.com/news/prompt-caching).
• anthropic-sdk-python ships system: Union[str, Iterable[TextBlockParam]] natively.
• The gem's own Bedrock provider already handles multiple :system messages without warning — see lib/ruby_llm/providers/bedrock/chat.rb.

Happy to PR the fix.