fix: apply cache_control for Claude 4.x models (not just claude-3)

[sumleo]

Problem

Raix::MessageAdapters::Base#content only emits the Anthropic cache_control multipart shape when the model id contains the literal anthropic/claude-3:

if model.to_s.include?("anthropic/claude-3") && cache_at && msg[:content].to_s.length > cache_at.to_i

The current Claude 4.x ids (anthropic/claude-sonnet-4, anthropic/claude-opus-4-7, anthropic/claude-haiku-4-5, …) don't contain the substring claude-3, so the guard never matches them. Setting cache_at on a Claude 4.x model silently does nothing: the content is sent as a plain string, no cache_control block is attached, and the prompt-cache discount is never applied. The gem already targets Claude 4.x elsewhere (e.g. lib/raix/configuration.rb references the Claude 4.7 family), so the caching path is the only place still pinned to claude-3.

Fix

Broaden the guard in lib/raix/message_adapters/base.rb to match any Anthropic Claude model:

if model.to_s.include?("anthropic/claude") && cache_at && msg[:content].to_s.length > cache_at.to_i

anthropic/claude is a prefix of every Anthropic Claude id, so this covers claude-3 and the *-4 family alike. Non-Claude models still fall through untouched — behavior is unchanged for anything that isn't an Anthropic Claude.

Tests

Added an example in spec/raix/message_adapters/base_spec.rb that stubs a Claude 4 id (anthropic/claude-sonnet-4) and asserts large content is wrapped with cache_control. It fails against the old claude-3-only guard (content stays a bare string) and passes with the fix. The existing claude-3 example is retained.

CHANGELOG entry added under ## [Unreleased].

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 44be866308

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Preserve multimodal content before adding cache_control

When a Claude 4 request uses abbreviated multimodal content (for example { user: [{ type: "text", ... }, { type: "image_url", ... }] }) and cache_at is below the array's stringified length, this broadened guard now wraps the entire array as the text: value of a single cache-control block. That prevents MultimodalContentAdapter.translate in chat_completion from seeing the original image_url parts, so the image is dropped/serialized instead of sent; the old claude-3 guard did not affect Claude 4 vision requests.

Useful? React with 👍 / 👎.

[sumleo]

Good catch - fixed in 805d8b9. The wrapping is now limited to plain String content, so multimodal arrays (image_url parts) pass through to MultimodalContentAdapter untouched. Added a spec asserting an array-valued Claude 4 message is left as-is.

[sumleo]

Hi @obie, gentle nudge on this when you have a moment. It's a small, self-contained prompt-caching fix, and I'm happy to rebase or tweak anything if that would make review easier. Thanks for the project and your time!