[FEATURE] Add an overridable Model.compact() and route summarization through it #5464

Human Overview

Summarization should go through the model, so each provider can decide how to compact. Model.compact() gets a working default that does today's client-side summary, and providers that support native compaction override it, the same way countTokens() works.

SDK Language

Both / Not language-specific

Problem Statement

SummarizeStrategy builds the summary itself, on the client. A provider that can compact server-side, keeping preserved thinking valid, has no way to take over. The strategy also has nothing to tell it where a safe cut point is (a user-prompt boundary, never between a toolUse and its toolResult).

Proposed Solution

Part 1 of design 0022:

• Add a concrete Model.compact(), not an abstract one. The default is today's client-side summary, and it returns [summary, ...preserved] or undefined.
  • TS: compact(messages, options?: CompactOptions) with preservedMessages, systemPrompt, toolSpecs, instructions.
  • Python: compact(messages, *, preserved_messages=0, tool_specs=None, system_prompt=None, instructions=None).
• Add a shared compactionSplitIndex helper. It moves the cut back to a user-prompt boundary, never splits a toolUse/toolResult pair, and never summarizes pinned messages.
• The strategy decides when to compact and the model decides how. Message-level summaries in SummarizeStrategy always call model.compact() (through _compact()), passing the live system prompt and tool specs, and SummarizeConfig.systemPrompt as instructions.
• BaseOffloadStrategy gains _isRetroactive (_isMessageLevel || _preserveRecent > 0). A retroactive per-block summary calls compact() only when the model reports serverCompaction === true. Otherwise it keeps today's behavior.

Done when:

• Python and TypeScript have unit tests for the default compact(), compactionSplitIndex (tool pairs, pinned messages, user-prompt boundaries), and the strategy routing.
• Summarization output and behavior are unchanged for models that don't override compact().

Use Case

A provider adds native compaction by overriding one method, and every summarize strategy (context_manager="auto", agentic, proactive_summarization) picks it up.

Additional Context

• Blocks: [FEATURE] Summarize through Anthropic native compaction #5401, [FEATURE] Include compaction calls in token usage #5404
• Design: team/designs/0022-server-side-compaction-support.md (docs(designs): server-side compaction support #5242)
• Reference (earlier prototype, not the spec): feat(context-manager): route retroactive summarization through Anthropic native compaction #5287