dsh-token-saver v0.1.0

First public release of the dsh-token-saver plugin for DeepSeek Harness.

Highlights

conservative (default): lossless head/tail truncation + grep fold for every large tool result, routed through tools/post-execute. Saves 97.6 % of input tokens on the 8-scenario synthetic suite.
balanced: additionally calls ctx.compaction.compactIfNeeded() on agent/pre-step once token pressure crosses 50 %.
aggressive: replaces dsh-compaction-basic with the bundled LlmSummaryCompactionEngine subclass (gracefully degrades to the official path on LLM failure).
Prompt-cache utility that pins stable system-prompt sections for high cache hit-rate.
63 vitest cases + smoke + benchmark harness covering strategies, regression snapshots, and the full dsh event flow against a mock scope.

dsh plugin --profile web install dsh-token-saver

Already listed in awesome-dsh-plugin.

Follows the dsh-plugin GitHub topic convention
Declares dsh.bundle + cordis.patch.yml so it's installable via dsh plugin add
Official @deepseek-ai/* packages declared as peerDependencies with explicit prerelease branches
All reductions go through dsh event flow — the "model-visible = logged" invariant holds end-to-end