The Zhipu (Z.ai) team publicly released CompactionRL (2607.05378) on arXiv on July 6, rewriting "context compression" from an inference-time trick into a trainable RL primitive, directly serving GLM-5.2's (750B-A40B) training pipeline. The method uses cross-trajectory GAE + token-level loss normalization, jointly optimizing task execution and summary generation, letting the model both work and compress in the same policy. The effect reproduces neatly on open-source MoE: GLM-4.5-Air (106B-A30B) SWE-bench Verified Pass@1 pulled to 66.8% (+7.0), Terminal-Bench 2.0 24.5% (+3.1); GLM-4.7-Flash (30B-A3B) on the same two items gets +5.5, +6.8, running 56.0% / 20.2%. More noteworthy is the real problem it solves: once long-horizon agent trajectories are compressed, methods like GRPO based on group-level advantage will fail, because the sub-trajectory count and length of different rollouts for the same prompt no longer align; CompactionRL switches to critic-based PPO + token-level advantage estimation, changing the RL training assumption from "fixed-length trajectories" to "variable-length sub-trajectories". The engineering side also provides a base: training/rollout uses Zhipu's self-developed slime framework, supporting white/black-box rollout, sub-agent workflows, and FP8 KV-cache, perfectly carrying GLM-5.2's parallel OPD post-training workload. This means the RL pipeline has for the first time treated "context budget" as a learnable knob rather than just an inference-time fallback — for all teams still struggling with agent long-horizon planning, this is a reusable training recipe.