dsh-sleev

After wiring up Sleev context-optimization routes for your model-inference provider, a common blind spot remains: the stream works, yet you have no lightweight way to see how many effective input tokens each call actually consumed or whether it was an agent step, a compaction pass, or a one-shot utility call. dsh-sleev fills that gap without touching the pipeline—it does not rewrite prompts, perform compaction, or proxy traffic; it simply records call metadata at the llm/stream boundary for matching routes.

Core capabilities include matching routes by exact alias or name prefix (default: sleev-), classifying each call as agent, compaction, session-title, or one-shot, yielding every stream chunk unchanged, recording provider usage and effective input token volume, retaining a bounded secret-free in-memory history, and emitting one structured completion record per observed call. All observer settings—matching rules, retention limit, and log level—are editable in the Settings > Plugins page and apply to the next matching call without a Host restart.

It suits DeepSeek Harness users who have already configured Sleev routes and want to quantify whether context compression actually reduced token overhead, as well as small teams that need to attribute agent, compaction, and one-shot call costs separately. It makes no token-savings promise; it simply makes the numbers visible.