dsh-auto-compact

DeepSeek Harness ships its automatic context compaction with an 80% trigger threshold. For models with two-hundred-thousand or even one-million token windows, that timing is often far too late, and long sessions quietly become sluggish before anyone notices why. dsh-auto-compact closes that gap by dropping the trigger to 75% and letting you fine-tune the policy per model, so history is tidied up while there is still comfortable headroom.

It injects the compaction policy directly into user-installed agent presets, preserves the most recent 20 percent of the window verbatim, and ships sensible per-model overrides such as a 70 percent trigger for GPT-5.6 (272K context) and a 65 percent trigger for DeepSeek V4 (1M declared context). The whole process is idempotent, auto-backs up every modified preset, rolls back with a single command, never touches read-only factory presets, and guarantees that a failed summary preserves the original history rather than silently dropping context.

If you run long multi-turn sessions on large-context models, manage several agent presets, or simply want a predictable compaction cadence across your stack, this plugin turns context management from reactive cleanup into proactive maintenance without rewriting a line of Harness source code.