Skip to content
Home › Context Management & Reliability

Server-side Compaction ⚠ verify

Context Management & Reliability

Definition

A beta capability where the API automatically compresses accumulated conversation context to stay within the window during long sessions, plus the related manual and pattern-based context-reduction tactics.

Key points

  • Compaction (beta) available for Fable 5, Mythos 5, Opus 4.8, Mythos Preview, Opus 4.7, Opus 4.6, Sonnet 5, Sonnet 4.6. ⚠ verify current model coverage (2026-07-08).
  • Related manual/pattern tactics: /compact (manual reduction in Claude Code), context editing (tool-result / thinking clearing), and the Scratchpad Pattern (file-based external memory).
  • Structured-summary approach (playbook): keep the full verbatim message history only for the active, unresolved issue; summarise earlier resolved turns into a narrative block. Layout = narrative summary + full verbatim active history + active issue at the end.
  • Complements Prompt Caching and isolated subagent contexts as a defence against Context Rot.

Why it matters for the exam

  • Long-session scenarios (e.g. a 48-turn support session approaching the window) test choosing compaction / structured summaries over dumping raw transcripts.

Common gotchas

  • Naive progressive summarisation compounds detail loss — preserve verbatim history for the active issue rather than re-summarising everything each turn.
  • Compaction is beta and model-gated; don't assume it's on for 200k models like Haiku 4.5 / Sonnet 4.5.

See also

Sources

Referenced by