OpenCode 2.0 vs OpenFox: local coding harness cache tests

This comparison tests how ten coding harnesses preserve a local LLM's prompt cache during development tasks. The speaker uses Cache Hunter to inspect requests during planning, implementation, skill and MCP additions, an AGENTS.md update, and forced conversation compaction. Each harness runs in an isolated container. The stated server setup uses two DGX Spark machines, DeepSeek V4 Flash and vLLM.

OpenCode 2.0.16 performs well in the speaker's tests, preserving the cache through most operations and making new instructions available immediately. Its compaction request replaces the last user message, which the speaker records as a minor invalidation. OpenFox, the speaker's own harness, receives a perfect score in this benchmark.

Codex handles the initial coding task and skill addition without cache invalidation, but its compaction request destroys the cached prefix. QwenCode's main conversation largely preserves the cache, while unexplained background requests interfere with performance. DeepSeek Harness Alpha also preserves the cache, but lacks skill management and immediate access to the newly added MCP tool.

Claude Code, Cline, Goose, Pi and Hermes encounter various cache or access problems. These results concern the tested configurations. The speaker's focus is the cost of processing a full context again, especially when compaction happens near the context limit.