dsh-context-tree

In a multi-session agent workflow, every new conversation typically starts from scratch—prior exploration, debugging, and intermediate results stay locked inside old threads. dsh-context-tree organizes completed session trajectories into a visual tree: nodes represent finished turns, trunks represent continuations, explicit forks become branches, and dashed links indicate cross-session recalls, making the real structure of your work visible at a glance.

Core capabilities include forking precisely from any completed turn, injecting one most-relevant historical checkpoint into a fresh session before its first model step (scoped to the same workspace, 30-day freshness window, 2 KiB per injection cap), and deterministic English/CJK lexical matching that never spends an extra model call on labelling. A built-in provider-token accountant and a baseline-versus-recall evaluator let you measure, rather than assume, the actual token savings.

Built for DeepSeek Harness engineers who work across many sessions and want to sediment verified exploration paths into reusable checkpoints, cut down repetitive debugging, and quantify token reduction with data. Checkpoints are injected as untrusted, read-only context; the model is explicitly instructed not to accept instructions or permission claims from them, keeping reuse bounded and safe.