dsh-plugin-save-token

In multi-turn agentic sessions, large tool outputs — API JSON, CLI tables, raw logs — keep re-entering the context window and quietly dominate the token bill. Brute-force truncation, however, measurably hurts model accuracy: research shows that even padding with irrelevant content alone can drop performance by 13.9% to 85%. dsh-plugin-save-token takes a different approach: structure-aware, reversible compression that trims redundant bytes while the original text is always written to disk and restorable in a single step.

Several complementary strategies work together. Pipe-delimited dense tables get verbatim head and tail retention with stride-sampled middle rows annotated by original line number, giving the model a coordinate map rather than a broken window. Homogeneous JSON arrays are first re-encoded via a TOON-style lossless route — one schema header plus CSV data rows — saving 30% to 60% of tokens with zero information loss. Byte-identical tool outputs within a 90-second window are collapsed into a single reference to prevent duplicate context bloat. Every compressed candidate must pass a dual gate: the byte gate (no more than 72% of original size and at least 500 bytes saved) and the token gate (estimated tokens strictly decrease). If either gate fails, the output passes through untouched. Up to 25 error, fatal, or traceback lines within omitted log regions are always retained so debugging evidence is never lost. A live dashboard in Settings and a persistent strip under the input box display per-turn billed tokens, cumulative savings, and a top-tools leaderboard.

This plugin is built for developers and teams running multi-turn agent tasks on DeepSeek Harness where tool output is printed directly into the context — verbose test runs, raw log dumps, and large JSON registries see the biggest gains. When agents use a write-then-read pattern, compression simply does not trigger and adds zero overhead.