[FEATURE] Surface Anthropic compaction blocks

[fvaleye]

Scope check

• This is core LLM communication (not application logic)
• This benefits most users (not just my use case)
• This can't be solved in application code with current RubyLLM
• I read the Contributing Guide

Due diligence

• I searched existing issues
• I checked the documentation

What problem does this solve?

Anthropic shipped a server-side context-compaction feature (compact_20260112, docs) that summarizes earlier conversation turns when input tokens cross a configurable threshold. The summary is returned in the response as a dedicated compaction content block, separate from the assistant's normal text reply.

When streaming, the response stream contains a new event type:

event: content_block_start
data: {"type":"content_block_start","index":0,"content_block":{"type":"compaction"}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"compaction_delta","summary":[...]}}

lib/ruby_llm/providers/anthropic/streaming.rb#build_chunk only recognizes text_delta, thinking_delta, and signature_delta:

def extract_content_delta(data, delta_type)
  return data.dig('delta', 'text') if delta_type == 'text_delta'
  nil
end

So compaction_delta events are silently dropped.

Result: using RubyLLM::Chat#ask against claude-sonnet-4-6 or claude-opus-4-7 with context_management.edits = [{ type: 'compact_20260112', ... }] have no way to:

1. Detect that compaction actually fired server-side
2. Read the compaction summary content (needed for follow-up turns per Anthropic's recommended pattern: "Keep the original messages in your list and let the API handle removing the compacted content". But the summary itself is what needs to round-trip back to keep Anthropic from re-running compaction every turn)
3. Persist the summary for observability

Proposed solution

Mirror how thinking is handled. Concretely:
1.

• Recognize compaction_delta (and any compaction_* deltas) in build_chunk / extract_*_delta helpers
• Add compaction: accumulation alongside content/thinking

• Add extract_compaction_content(content_blocks) that selects c['type'] == 'compaction'
• Pass to Message.new via a new keyword

• Add compaction attribute similar to thinking
• Default nil for non-Anthropic providers

4: Update format_message to round-trip compaction blocks back to Anthropic on subsequent turns (so the gem also handles the cost-saving "skip re-compaction" path Anthropic's docs describe).

This is symmetric with the thinking-block plumbing already present, so the change should be local and additive.

Why this belongs in RubyLLM

• It's a first-class feature of the Anthropic Messages API.
• The data is only available at the streaming layer: compaction_delta events live in the SSE stream; without gem support, an application has to either fork the gem or re-implement Anthropic SSE parsing externally.
• It's not actually Anthropic-specific in concept. OpenAI shipped an equivalent server-side compaction feature on its Responses API (docs) with the same shape: a context_management config, a token threshold, and a server-returned compaction item meant to round-trip on subsequent turns. So a generic compaction accessor on RubyLLM::Message will eventually be populated by multiple providers, same evolution thinking went through.