dsh-plugin-save-token

English | 简体中文

In one sentence: a DeepSeek Harness (dsh) dynamic plugin that cuts token cost without cutting model intelligence.

Codex user? The same token-optimization brain ships as a standard Codex CLI plugin (MCP stdio server) in codex-plugin-save-token/ — see its README.

It slims down oversized tool outputs at the entrance of every model request — reversibly and structure-aware. The full original text is always saved to disk; what the model sees is always a condensed version carrying a retrieval path. Every optimization obeys one red line: any replacement must be restorable in one step, and the estimated token count after compression must be strictly smaller than the original.

---

Why you need it

• The bulk of agentic-session cost is tool output that keeps re-entering the context: large JSON from APIs, CLI tables, logs. The same 40KB price table may be billed again on every conversation turn.
• But blunt truncation degrades intelligence: research shows that even with perfect retrieval, merely padding the context with irrelevant content drops accuracy by 13.9–85% (curated in "Context Length Alone Hurts"); conversely, over-compression fails too — in production randomized controlled trials, aggressive compression (keep ratio 0.2) actually made cost +1.8% worse, while moderate compression (0.5) delivered −27.9%.
• Conclusion: the right way to save tokens is structure-preserving slimming, not content hacking. All strategies in this plugin are designed within that boundary.

---

Core optimizations

1. Structure-aware compression: no more blind table slicing

Detects pipe-delimited high-density table rows (≥70% of lines sharing the same separator profile). When matched, instead of a blind head/tail window it does: keep the first 60 rows verbatim + stride-sample the middle section every N rows with original line numbers annotated (L61: ...) + keep the last 40 rows verbatim. The model receives a "map with coordinates" — any segment can be fetched precisely by line number.

Reference: rtk's never-worse guard and error-line retention policy; the stride sampling is this plugin's improvement over rtk's head/tail window.

2. TOON-style lossless encoding first

Uniform JSON arrays (e.g. API responses with 300 homogeneous objects) first go through deterministic tabular re-encoding: prices[300]{model,input,output}: — one schema header line + CSV data rows. Keys are written once, zero information loss, and the notice explicitly says "zero information loss". Lossy paths are only used when the lossless route is unavailable.

Reference: TOON — Token-Oriented Object Notation; measured savings of 30–60% tokens on uniform arrays.

…