fix: translate OpenAI image_url content into RubyLLM attachments (#51)

[obie]

Fixes #51.

Problem

A multimodal user message added to the transcript in OpenAI content-array form is silently dropped on the RubyLLM backend:

c.transcript << { user: [{ type: "image_url", image_url: { url: "data:image/png;base64,..." } }] }
c.chat_completion   # vision model gets TEXT ONLY → confabulates an answer, no error raised

ruby_llm_request passed the array straight to chat.add_message(role: :user, content:). RubyLLM treats a raw array of OpenAI content hashes as plain text — it never recognizes the image_url part as an image, so nothing reaches the provider.

Fix

New Raix::MultimodalContentAdapter translates OpenAI-style content arrays into a RubyLLM::Content with attachments before add_message:

• data: URIs are base64-decoded into a binary StringIO. RubyLLM's Attachment does not recognize data: URIs — it would classify one as a filesystem path and blow up on File.binread. Decoding to IO is the same trick RubyLLM uses internally (providers/gemini/media.rb).
• http(s) URLs pass through unchanged (Attachment fetches them).
• Anything that isn't an array of hashes with at least one image_url part is returned untouched, so text completions and the Anthropic cache_control multipart shape are unaffected.

chat.add_message(role: :user, content: MultimodalContentAdapter.translate(content))

Tests

New spec/raix/multimodal_content_adapter_spec.rb (6 examples) verifies a data: URI decodes to a real image/png attachment (Marcel content-sniffing), http URLs become URL attachments, text survives alongside the image, string-keyed parts work, and non-image content passes through unchanged.

Full suite: 113 examples, 0 failures, 1 pending. RuboCop clean.

Note on base branch

This PR is stacked on fix-function-tool-adapter-rich-schema (PR #52) so both fixes share the same [2.0.5] CHANGELOG section. Merge #52 first; GitHub will retarget this PR to main automatically.

🤖 Generated with Claude Code