[FEATURE] Summarize through Anthropic native compaction #5401

Human Overview

Anthropic's preserved thinking models will return a 400 error when the prefix of a thinking block is changed. In order to keep thinking as a part of the conversation and not dropping thinking, summarization needs to happen through Claude's Native compact API. AnthropicModel overrides Model.compact() so that sessions with these models summarize through native compaction.

SDK Language

Both / Not language-specific

Problem Statement

Offload.summarize replaces older messages with a client-side summary. On preserved-thinking Claude models, that changes the history the kept turns' thinking is bound to, so the next request fails with 400 Invalid signature in thinking block, or drop_block discards the reasoning. This affects context_manager="auto", agentic, and proactive_summarization.

Proposed Solution

Part 1 of design 0022. Override AnthropicModel.compact() with on-demand compaction (compact-2026-09-04):

• Request settings:
  • compaction.instructions comes from instructions (SummarizeConfig.systemPrompt), capped at 16,384 characters.
  • Use the model's max_tokens. A stop_reason: "max_tokens" result leaves history unchanged.
  • Don't send stop_sequences, output_config.format, or a forced tool_choice (the API rejects them).
  • Don't use threshold compaction (compact_20260112).
• Gating on serverCompaction:
  • true: native compaction.
  • undefined: try native, and record a capability rejection as false.
  • false: fall back to super.compact(). If historyBound is true, warn and return undefined instead, since a client-side summary would break the thinking.
  • Timeouts, throttles and 5xx errors are never recorded. Warn and return undefined.
• Until the metadata work lands, the capability comes from a private list of preserved-thinking model IDs (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5) plus a temporary AnthropicModelConfig.serverCompaction override, mirroring use_native_token_count.
• Once a compaction block is in history, stream() and countTokens() use beta.messages with the compaction beta.
• The result is stored as a summary TextBlock with providerData ([FEATURE] Core: providerData on TextBlock for session restore and model switches #5402), with no [Summarized: …] header.

Done when:

• Python and TypeScript have unit tests for the gating table, the request settings, and the max_tokens stop case.
• An integ test with prefix_mismatch_behavior: "error" compacts repeatedly with no 400, including when the fix(models): retry with drop_block when a history edit invalidates replayed thinking #5171 retry fires (both betas sent).

Use Case

A long-running agent on Opus 5.5 with Offload.summarize("*").when(utilization=0.8, preserve_recent=4) summarizes and keeps the reasoning in its recent turns, instead of failing or losing it.

Additional Context

• Depends on: [FEATURE] Add an overridable Model.compact() and route summarization through it #5464, [FEATURE] Core: providerData on TextBlock for session restore and model switches #5402
• 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