dsh-compactor

In long DeepSeek Harness sessions, tool results — search hits, file contents, command output — routinely consume the bulk of the context window, yet only a small fraction of that material is actually useful for downstream reasoning. Repeated tool calls and verbose assistant replies keep inflating token counts and can even trigger runaway loops where the model keeps invoking the same tool with the same parameters. dsh-compactor addresses this head-on: it systematically compacts context without breaking dsh's append-only session log, so long conversations stay efficient without losing critical information.

The plugin works through three coordinated layers. In the real-time layer, tool results are pruned the moment they enter the event stream — search results keep only titles and short snippets, files keep a preview, and command output keeps the exit code and key lines. In the batch layer, after each assistant turn the session token count is estimated; once it crosses a threshold (default 32768), compressible segments are replaced with a single structured summary, with boundaries automatically aligned to tool-call / tool-result pairs and the most recent N turns never touched. The archive layer writes every displaced original to an append-only JSONL file, so /restore brings the full text back at any time. Two summarization engines are available: an API engine that calls deepseek-chat for skeleton summarization (falling back to local extractive summarization when no key is set) and a local rule-based engine triggered by /local-compact that requires zero network access and zero new dependencies. A built-in anti-loop guardrail watches the tool/call event stream, injects an anti-loop reminder after three identical calls, and records a cancellation decision after five, for the host toolRuntime to consume. Every compaction run produces a structured report listing each compressed range, per-item and overall compression ratio, before-and-after token counts, and keyword retention rate.

It is well suited for developers and research teams who lean heavily on tool calls and long-chain reasoning inside DeepSeek Harness, and for anyone who wants to keep context under control offline — the local rule engine needs no API key, no network, and no extra dependencies, and rule changes take effect on the next /local-compact run without restarting the session.