LLM agents increasingly rely on externalized skill libraries, yet existing maintainers typically delegate skill editing to a prompted LLM, leaving the question of how to train a skill maintainer open. We present RepairGate, a 9B drop-in wrapper around the standard candidate patcher: a counterfactually-trained post-trial gate accepts, blocks, or overrides each candidate patch, with a structured repair protocol for field-level edits and overlap-aware regularization to control library growth. Across six base agents on SkillFlow, RepairGate delivers +7.73 pp over Vanilla and +9.13 pp over prompted +Iter, with the largest gains on the weakest base agents where prompted iteration regresses, at a standardized 0.033/trial local overhead on top of the shared candidate-patching pipeline. Code and data: https://github.com/xxy33/RepairGate

Full-state latent communication in LLM-based multi-agent systems offers richer semantics than text but suffers from memory overhead scaling linearly with collaboration rounds. We propose CondenseFlow, which introduces the Latent Thought Condenser (LTC)—a lightweight module using learnable semantic probes to compress KV caches into fixed-size representations, achieving 𝒪(1) communication complexity regardless of context length. We theoretically prove that compression error is bounded by attention concentration and accumulates controllably across rounds. On seven benchmarks spanning six models, CondenseFlow reduces KV cache memory by over 99% and inference latency by approximately 20% compared to dense transfer with negligible accuracy degradation, while outperforming text-based methods by 1.7 percentage points on average across all configurations. Code is available at https://github.com/xxy33/condenseflow.