dsh-squeeze-command: On-Demand Context Compression for DeepSeek Harness

In DeepSeek Harness (DSH), every turn of the conversation resends the entire history. As the conversation grows longer, input token costs can rise quickly and become the dominant overhead. dsh-squeeze-command is designed to address this issue. It provides manual, budget-oriented context compression: the dialogue model selects the scope, and a cheap, fast model (flash-class) writes the summary.

What It Is¶

This is a DSH plugin for compressing session context on demand. It is maintained by hardes11. It reduces input token costs by allowing the dialogue model to select the scope and having a cheap, fast model write the summary, while retaining the original session log.

Core Features¶

• Cut resent content: Squeeze the conversation into the budget range with a single command, and generate the summary using the configured cheap routing rather than the expensive dialogue model.
• No dropped content: Each compression leaves a checkpoint on the surface, while original messages remain in the append-only session log; use /squeeze map to browse them.
• Resist post-compression confusion: Each checkpoint summary includes explicit status lines and calculated reliable summaries from uncompressed tails, preventing the model from re-entering work that is described in the summary but not actually completed.
• Triggers only when you ask: There are no thresholds, no automatic triggers, and it refuses to run while an agent turn is in progress.
• No agent scheduling required: Summary delegation is built in; you only need to run a command, and everything else happens within it.

Install and Enable¶

1. Clone the plugin into the DSH commands directory:

git clone https://github.com/hardes11/dsh-squeeze-command.git ~/.dsh/commands/dsh-squeeze-command

2. Mount and configure it in the preset agent.cordis.yml. You need to specify the target token budget and summary model:

- id: command-squeeze
  name: 'dsh-squeeze-command'
  config:
    contextBudgetTokens: 100000
    summarizerProvider: <your-provider>
    summarizerModel: <your-flash-model>

Command Usage¶

1. /squeeze: Compress to the configured contextBudgetTokens (preset configuration).
2. /squeeze 60k: Specify a one-time compression target (for example, 60000 tokens).
3. /squeeze status: View the current session’s token estimate, budget, and summarizer status.
4. /squeeze map: Generate an interactive HTML map to compare before and after compression and view retention ratio.
5. /squeeze help: View the full usage guide.

Notes¶

• Manual mode only: The plugin refuses to run while an agent turn is in progress.
• Cache invalidation: Compression rewrites the context prefix, invalidating the prompt cache; it is recommended to run it when the cache has expired or caching is not needed.
• Model requirements: The summary routing should be configured as a cheap, fast model (flash-class recommended), because summary tasks do not require high reasoning intensity.
• Required configuration: Configure contextBudgetTokens, summarizerProvider, and summarizerModel before running.

Summary¶

dsh-squeeze-command is a practical tool in DeepSeek Harness for controlling context costs. It reduces token consumption while maintaining context integrity through manual triggering and a flash-class summary model, making it suitable for long-running, input-cost-sensitive session scenarios.