dsh-tool-squeeze

Anyone building an agent will eventually hit the same wall: a single tool call spits back tens of thousands of low-signal lines-a Maven log, a nested JSON payload, a crawled HTML page-and blind truncation risks erasing the one failure buried in the middle, while keeping the full blob inline burns context tokens on every subsequent model request. dsh-tool-squeeze slots into the public tools/post-execute lifecycle hook of DeepSeek Harness and applies deterministic, local, evidence-aware compression to logs, JSON, HTML, and plain text. Error lines, failed tests, stack roots, non-zero exit codes, and representative structural samples are prioritized; every omission is disclosed with exact counts; and the complete formatted original is retained through the official DSH spill store for on-demand retrieval. No LLM call, no patch to Harness Core, and a fail-open guarantee means a plugin exception never turns a tool call into an error. On the committed high-repetition benchmark fixtures, a 22K-token Maven failure compresses to 1.7K (92.5%), a 61K Gradle log to 123 tokens (99.8%), and a 159K JSON API response to 352 tokens (99.8%), with all critical-evidence gates passing. The plugin is purpose-built for developers who use DeepSeek Harness to orchestrate multi-tool agents and routinely process large tool outputs-build logs, CI transcripts, API responses, scraped pages-where predictable latency, reproducible behavior, and context-window efficiency matter as much as the model reasoning quality itself.