{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2609.26779v1",
  "slug": "2609-26779v1-1nezvu9",
  "url": "https://feed7.dev/p/2609-26779v1-1nezvu9",
  "title": "CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents",
  "why_included": "CliffCompaction cuts long-horizon agent costs by dropping or truncating original context instead of rewriting summaries, aiming to prevent drift across repeated compactions.",
  "summary": "CliffCompaction reports **up to 50% lower cost** under bounded context while maintaining or improving Terminal-Bench performance. It only truncates or drops original material, never rewrites it, and never compacts an earlier compaction.",
  "practical_implication": "For long-running coding agents, preserve original evidence and regenerate each compacted view from that source instead of recursively summarizing summaries. The open API proxy works across Claude Code, Codex, and other harnesses, making the approach testable without replacing the scaffold.",
  "agent_context": "CliffCompaction reports **up to 50% lower cost** under bounded context while maintaining or improving Terminal-Bench performance. It only truncates or drops original material, never rewrites it, and never compacts an earlier compaction.\n\nFor long-running coding agents, preserve original evidence and regenerate each compacted view from that source instead of recursively summarizing summaries. The open API proxy works across Claude Code, Codex, and other harnesses, making the approach testable without replacing the scaffold.\n\nResults come from the authors' benchmark setup: Terminal-Bench gains exceed **10 percentage points**, while KernelBench reaches **2.23× at 200 steps** and **3.58× at 400 steps**. Whether these trade-offs hold for a builder's repositories and context mix remains open.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2609.26779v1",
    "published_at": "2026-09-22T17:55:59.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "context",
  "domains": [
    "coding"
  ],
  "topics": [
    "context-engineering",
    "context-caching",
    "agent-reliability"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "Results come from the authors' benchmark setup: Terminal-Bench gains exceed **10 percentage points**, while KernelBench reaches **2.23× at 200 steps** and **3.58× at 400 steps**. Whether these trade-offs hold for a builder's repositories and context mix remains open."
  ],
  "connected_context": {
    "meaning": "This sharpens context compaction into a source-preservation rule: derive each bounded view from original material instead of recursively rewriting prior summaries. That complements reversible eviction and externalized tool history while offering cross-harness benchmark evidence of lower cost; it also confirms that compaction policies must be tested against each repository and context mix rather than adopted from headline savings.",
    "corpus_size": 856,
    "generated_at": "2026-09-23T09:06:29.105Z",
    "connections": [
      {
        "title": "Blast Radius",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.07440v1",
        "feed7_url": "https://feed7.dev/p/2608-07440v1-1gi5t78",
        "reason": "Both preserve recoverability instead of relying on lossy recursive summaries; Blast Radius archives predicted-unused history, whereas CliffCompaction regenerates compacted views from original material."
      },
      {
        "title": "mksglu/context-mode",
        "source_name": "GitHub",
        "source_url": "https://github.com/mksglu/context-mode",
        "feed7_url": "https://feed7.dev/p/context-mode-1olsb2y",
        "reason": "Context Mode supplies an external searchable store for bulky outputs and post-compaction recovery, complementing CliffCompaction’s rule that compacted views should remain grounded in original evidence."
      },
      {
        "title": "JuliusBrussee/caveman",
        "source_name": "GitHub",
        "source_url": "https://github.com/JuliusBrussee/caveman",
        "feed7_url": "https://feed7.dev/p/caveman-0yoqowc",
        "reason": "Caveman’s finding that compression overhead can outweigh savings reinforces CliffCompaction’s unresolved need for workload-specific cost and exactness trials."
      },
      {
        "title": "Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General Reasoning",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=2bvtay8wGYI",
        "feed7_url": "https://feed7.dev/p/scaling-to-long-horizons-ross-taylor-chengxi-taylor-general-reasoning-0jwtg4d",
        "reason": "CliffCompaction provides a concrete, benchmarked compaction policy for the long-horizon context pressure identified here, without resolving the separate reward and scheduling constraints."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-09-22T17:55:59.000Z",
  "modified_at": "2026-09-22T17:55:59.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2609-26779v1-1nezvu9",
    "json": "https://feed7.dev/p/2609-26779v1-1nezvu9.json",
    "markdown": "https://feed7.dev/p/2609-26779v1-1nezvu9.md"
  }
}