CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents

A rule-based context-compaction method for long-running coding agents, by Trang Nguyen, Eulrang Cho and Tim Dettmers of Carnegie Mellon with Bingqing Chen of the Bosch Center for AI. The context grows normally until it crosses a token threshold; compaction then truncates or drops the parts that dominate its length, chiefly tool outputs and tool calls (84% of the tokens in a Terminal-Bench 2.0 run with GLM 5.1), keeping verbatim excerpts and never rephrasing or summarizing. Each compaction also discards the previous compacted block and works only from original content, so a compaction is never compacted again, and context length drops back to about the same level after every event (the "cliff"). The authors deliberately give up full-history recall for what they call high compaction precision: with summaries, an agent tends to assume it has everything and stops revisiting past state, so whatever the summary omitted turns into drift. The method needs no training and no auxiliary LLM calls, works with any model, and keeps the prompt cache reusable between compactions, cutting cache-read cost by up to 90%.

On Terminal-Bench 2.0 with the Terminus-2 harness, Kimi K2.6 resolved 61.4% of tasks at 32K- and 16K-token thresholds, against 59.2% with full context and 55.5% with Terminus-2's own LLM summarization at 16K, and GLM 5.1 rose from 49.8% to 54.3%. On Terminal-Bench 2.1 inside Claude Code, GLM 5.3 Flash reached 76.7%, against 71.0% for Claude Code's auto-compaction at the same budget of about 45K tokens and 73.0% at its default 200K. Because each rollout gets cheaper, parallel test-time scaling pays off: three compacted Kimi K2.6 rollouts gain 10.5 points for 1.9× the cost of one uncompacted run, and cost $58.01 for the whole of Terminal-Bench, less than one GPT-5.3 Codex run, while matching Claude Opus 4.7. On SWE-bench Verified it preserves full-context success for GLM 5.1 and Kimi K2.6 at 32K and 16K thresholds. For continual kernel optimization on KernelBench Level 3, with sessions exceeding a million tokens, Kimi K2.7 in OpenHands reached a 2.23× speedup after 200 steps and 3.58× after 400, against 1.30× without compaction, and GPT-5-mini reached 2.09× after 200 steps against 1.78× for AdaExplore, a specialized kernel agent using the same model. The open-source implementation is a scaffold-agnostic API proxy that works with Claude Code, Codex CLI and other harnesses (MIT license, pip package cliffcompaction, about 50 GitHub stars).