Compact KV caches to share task-specific orchestrator state with workers.
npx skills add https://github.com/466852675/Skills-2026 --skill latent-briefing
Please help me install this Agent Skill. Skill: latent-briefing Source: https://github.com/466852675/Skills-2026/tree/main/latent-briefing Command: npx skills add https://github.com/466852675/Skills-2026 --skill latent-briefing
Replaying the full orchestrator reasoning trajectory to every worker inflates token usage and introduces latency, while summarization loses critical context needed for downstream tasks.
Ask the agent to compact the orchestrator’s KV cache for the upcoming worker task.
High-intent search queries and answers about installing and using this skill.
Share orchestrator state by compacting the KV cache into task-specific slices. This retains only the KV positions most relevant to the current worker task, avoiding full-text handoff and reducing token replay latency.
Task-guided attention scoring uses the worker's task prompt to score trajectory tokens and build a shared global mask. It applies median-plus-MAD thresholding to decide which KV entries to keep for downstream tasks.
Yes, compacting orchestrator state requires direct access to the worker model's KV tensors. You also need the ability to compute task-guided attention scores to selectively retain the necessary latent reasoning trajectory.
Use representation-level sharing to pass context in hierarchical multi-agent pipelines. This method transfers selective slices of the orchestrator's latent state directly, preventing the latency and context loss of summarization.
Summarizing orchestrator reasoning loses critical context because it drops intermediate reasoning steps. Compact KV cache sharing avoids this by retaining exact latent state positions needed for downstream tasks.
You should not use KV cache compaction if your system lacks access to worker model KV tensors. It is also unsuitable if your workload cannot leverage task-guided attention scoring for selective retention.
lucasFR2Compact orchestrator trajectories in worker KV caches using task-guided queries.
guanyangCompact orchestrator trajectories into worker KV caches via attention matching.
m31uk3Design AI workflows with a five-phase structure and human-in-the-loop checkpoints.
wangyouanShare relevant orchestrator trajectory parts with workers via Attention Matching.
