dsh-context-guard

Once a long agent session reaches several hundred turns, the context window often gets clogged first by tool output: a single command spitting out thousands of log lines, a file read pulling in an entire codebase. Before the next request even starts, most of the token budget is already consumed. dsh-context-guard solves this without replacing built-in compaction and without issuing its own LLM call. It swaps oversized tool output for a compact preview of roughly 2.3K characters right before each model request, keeping the focus on the actual conversation.

Pruned output is never lost; it stays intact in DSH's append-only session log. The plugin offers a layered recovery path: first use the retained file path or a narrow filesystem search, and only fall back to recall for the smallest relevant fragment when that is insufficient. A recall result is preserved in full for exactly the next model step, after which it re-enters normal pruning, so the model can act on complete content without letting large payloads linger in context.

Built for developers running multi-turn agent workloads in DeepSeek Harness who need to control context bloat without discarding raw tool data. It is compatible with the built-in compaction tool-result pruner and takes effect immediately in a new session after a DSH restart.