Subconscious + GLM-5.2 Makes "/compact" Obsolete

Large context windows for AI coding agents are slow, expensive, and unreliable due to the "lost-in-the-middle" problem where models forget critical information. A system called TIMRUN enables models like GLM-5.2 to perform model-driven self-compaction, dynamically pruning their own reasoning history during a task. This process utilizes a Subconscious Cache, which retains the KV cache states of remaining tokens to avoid latency penalties and preserve a "subconscious" memory of pruned information. This allows an agent to operate with the memory of millions of tokens while maintaining the speed and cost of a much smaller active context, eliminating context-limit errors.