dsh-context-assembler

DeepSeek Harness gives you exactly two context states: everything in the window, or whatever a threshold fires into a summary you can neither inspect nor undo. On a long task a four-thousand-token grep dump, a stack trace you already diagnosed, a sub-task that finished and shipped all keep occupying the window until that blind compaction folds them away. dsh-context-assembler turns context from a length into a set of decisions: you walk the session log as a checkable tree and, per span, choose to keep it verbatim, compress it to a one-line summary, or drop it entirely, so the model actually receives what you intend.

The core capability is an interactive three-state editor (full / key / off) layered on the harness surface-folding primitive, an N-to-1 span replacement. The panel renders turns, steps, tool calls, and subagent sessions as a navigable tree; untouched prefix nodes stay in place so the provider KV cache remains reusable. Tool-call and tool-result pairs are treated as atomic groups to avoid feeding malformed shapes to the provider. A single-event fold restores verbatim; a multi-event fold replays as one separator-delimited message and the panel labels this honestly instead of pretending to undo. On the agent side, a registered context_assembler tool (tree, set, preset, messages) lets the model itself fold finished history into a concise conclusion after each sub-task, rather than waiting for an opaque threshold to fire.

Built for power users running long, multi-tool, multi-subtask sessions in DeepSeek Harness who want explicit, per-span control over what the model sees each turn. If all-or-nothing context feels like a coin flip and you prefer every inclusion or exclusion to be a deliberate, inspectable choice, this plugin is for you.