AI Agent Hub
Back to plugins
🧠

dsh-squeeze-command

Memory Updated 2026.09.07

Run the following command in DeepSeek Harness:

dsh plugin install hardes11/dsh-squeeze-command

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install hardes11/dsh-squeeze-command in DeepSeek Harness to install this plugin; source is at https://github.com/hardes11/dsh-squeeze-command

About this plugin

Long conversations with frontier models carry a hidden tax: every single turn re-sends the entire history, and input tokens dominate the bill far more than reasoning tokens. Provider prompt-caching discounts soften the problem only while the prefix stays stable. dsh-squeeze-command ships a /squeeze slash command that lets you compress a session's context down to a budget you choose, with summaries written entirely by a cheap flash-tier route you configure. The expensive model merely marks which spans are compressible; the actual rewriting never touches the frontier price.

Nothing is lost in the process. Original messages stay in the append-only session log while the surface shows only checkpoint summaries, each capped with an explicit per-task completion status and a trusted open-work digest computed from the uncompressed tail so the model never re-enters work a summary already describes as done. The command is strictly manual, has no auto-trigger, and refuses to run while an agent turn or compaction is in flight.

The companion /squeeze map command renders a self-contained HTML timeline in your browser: clickable checkpoint bars per invocation, an event-density histogram, a per-checkpoint survival-ratio table, and a side-by-side view of original messages versus their summaries. It is built for power users who regularly push long, expensive-model sessions past hundreds of thousands of tokens and want to keep the full record while cutting the per-turn cost.

Use Cases

  • Frontier-model conversations past 300k tokens where per-turn input cost dominates and needs manual compression to a target budget
  • Long coding sessions with heavy accumulated history that should be trimmed for sending while the full log is kept for audit
  • Reviewing multiple compression passes via an interactive HTML timeline showing survival ratios, event density, and before-after message pairs

Best For

  • Power users who regularly push frontier-model sessions past hundreds of thousands of tokens and watch input cost climb
  • Dev teams that want cheap-model parallel summarization so compression never spends frontier-tier tokens
  • Technical users who need the full session log intact while capping per-turn context length