dsh-plugin-context-manager

After dozens of rounds in a long conversation the context window fills up: token bills climb while the model quietly drops the constraints, architecture decisions, and edge-case notes locked in during early turns. Most plugins tack a side-channel summary onto the system prompt, but the message array the model actually receives stays untouched. dsh-context-manager works on the real message stream directly, so both injection and folding operate on the exact messages fed to the model on every request.

Three capabilities sit at the center. First, each conversation turn is auto-recorded as a summary that you reorder by drag-and-drop for priority; pinned records and custom text are prepended into the real message stream every round, so the model always carries your key points. Second, the browser management window lets you inspect every message the model sees (role, text, token estimate), select a range of history, and queue it for compaction into a summary, truly removing old text from context and cutting token usage. Third, all records, injection parameters, queued folding tasks, and the folding audit log are persisted to disk and survive restarts; injection counts, character budgets, and record limits are adjustable at runtime without editing config files.

Ideal for developers running long iterative sessions in DeepSeek Harness, especially those who need the model to carry critical constraints every round, prefer a controllable token budget over blind truncation, or want a visual audit of exactly what the model reads each turn. Four tabbed panels cover record management, real-conversation inspection, injection settings, and global pinning, with a quick-start guide on first open and zero-config usability right after install.