Basic Compaction
Domain: Memory & Context
Motivation
The shipped default condensation backend. It keeps long conversations working near the model’s context limit by condensing the oldest history into a model-written summary while preserving recent messages, and it recovers after a confirmed context-overflow error by condensing and retrying. It uses one extra model request and retains only the summary text; it cannot shrink the system prompt, tools, or session prefix, nor split an indivisible unit like a single huge tool call.
Concrete Example
Default behavior: condense automatically as the conversation grows, recover after overflow, condense on demand via /compact, and — when dsh-compaction-tool-result-pruner is mounted — trim oversized tool outputs before summarizing. Tune with thresholdRatio and retainRatio.
Analogy
A running meeting that, near the end of the room, summarizes the first agenda items into a recap so discussion can continue.