dsh-jev-prune: A DeepSeek Harness Context Compression Plugin Based on Jev Structured Output

The `dsh-jev-prune` plugin provides a two-layer context compression mechanism for DeepSeek Harness (DSH) based on Jev structured output. It replaces the existing size-based truncation and model summarization logic via an interception point, using deterministic code to generate receipts instead of model inference and ensuring that the decision process follows the principle that “the model will not generate what it should not generate.” It includes two functional layers—result pruning and receipt compression—and uses strict gating mechanisms (such as a tool allowlist, freshness checks, and balance checks) to determine compression behavior, thereby freeing context while retaining key results, avoiding hallucinations, and restoring the original text.