context-optimization

Apply observation masking to verbose tool outputs first, then trigger compaction when context utilization exceeds 70-80%. Stabilize your prompt prefix for KV-cache reuse, which alone can cut costs by 50% or more on cached tokens.

Observation masking replaces verbose tool outputs with compact references once their key points are extracted, storing full content externally for later retrieval. It targets 60-80% reduction in masked observations with under 2% quality impact.

Mask first to remove low-value bulk like old tool outputs, then compact what remains when utilization crosses 70-80%. Compaction is lossy summarization, so it works best after masking has already stripped redundant content.

Any change to the prompt prefix, even a single whitespace or a timestamp in the system prompt, invalidates the cached block downstream. Keep system prompts immutable and move dynamic metadata into user messages or tool results.

Partition when estimated task context exceeds 60% of the window limit and the task decomposes into at least three independent subtasks. Below that threshold, coordination overhead typically exceeds the token savings.

Over-aggressive compaction drops task goals, user constraints, and nuanced state, especially when the compacting model is itself above 85% utilization. Trigger compaction at 70-80% and re-validate summaries against the current task goal.