Read-only performance diagnostics for DeepSeek Harness: session load/restore timing, spill-hit counts, compaction count and trigger, context-injection volume (AGENTS.md/skills/tool-schema token share), and LLM cache hit rate — surfaced via /fast, persisted as reconstructable session events with async sampling off the model path.
Find the root-cause wall, apply the top intervention, prove it with a two-sided gate — or roll back. Engine- and hardware-neutral: no implicit defaults, spec-sheet values are never load-bearing.