CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents
CliffCompaction cuts long-horizon coding-agent context cost by up to 50% without rewriting retained text.
CliffCompaction compacts long agent contexts by truncating or dropping original content only, never rewriting it or compacting a prior compaction. The authors report up to 50% lower cost under a bounded context while maintaining or improving Terminal-Bench results and improving test-time scaling efficiency. Under parallel scaling, Kimi K2.6 matches Opus 4.7 and exceeds Opus 4.6 and GPT-5.3 Codex at lower cost. On KernelBench it reports 2.23x CUDA kernel speedups after 200 steps and 3.58x after 400 steps, and they open-source a scaffold-agnostic API proxy for Claude Code, Codex, and other harnesses.
- Compaction only truncates or drops original content and never rewrites it.
- Prior compacted output is discarded so context drift does not accumulate.
- Reports up to 50% cost reduction while holding Terminal-Bench performance.
- Kimi K2.6 matches Opus 4.7 under parallel scaling at lower cost.
- Open-source API proxy works with Claude Code, Codex, and other harnesses.
Full article214 words · extracted from arxiv.org · click to collapse
Agents often work on complex problems that require millions of tokens of context, which necessitates compacting across sessions due to limited context windows. We develop CliffCompaction, an autocompaction technique that reduces cost by up to 50% under a bounded context while maintaining or improving performance on Terminal-Bench and achieving new levels of efficiency for test-time scaling and state-of-the-art results on KernelBench. The per-rollout savings of CliffCompaction make the performance--cost trade-off of test-time scaling more efficient, adding over 10 percentage points on Terminal-Bench for less than the cost of two full-context runs. Under parallel test-time scaling, CliffCompaction lets Kimi K2.6 match Opus 4.7, and exceed Opus 4.6 and GPT-5.3 Codex at lower cost. The key to CliffCompaction's effectiveness is that it keeps compacted information faithful by only truncating or dropping content, never rephrasing or rewriting it. We never compact a compaction---each pass operates only on original content, and prior compacted output is discarded, preventing context drift from accumulating. These properties sustain continual learning over sessions exceeding a million tokens: on KernelBench, CliffCompaction reaches CUDA kernel speedups of $2.23\times$ after 200 steps and $3.58\times$ after 400 steps, surpassing specialized search algorithms and trained agents despite being a general-purpose compaction technique. We open-source a scaffold-agnostic API-proxy implementation of CliffCompaction usable with Claude Code, Codex and other harnesses.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.26779