dsh-context-window

Summarization-based context compaction has a structural weakness: it is implicitly lossy. The model does not know what was dropped, quality degrades across multiple rounds, and lost information is unrecoverable. dsh-context-window brings the alternative path that Codex introduced into DSH workflows. Instead of shrinking context into a single summary, it replaces the entire window with a fresh initial context. The model knows the old content still exists elsewhere and can retrieve it when needed, while also skipping the extra model call that each compaction round would otherwise require.

Three core capabilities: token-budget threshold notices (default at 25%, 50%, and 75% of window consumption, plus a final reset reminder near exhaustion), a model-callable new_context tool (sets a flag that flips the window on the next turn), and configurable budget accounting (effective window ratio, hard clamp, and prefix-billing toggle all aligned with Codex defaults). Resets land on step boundaries, never splitting an in-flight tool call, so tool-pairing balance is always preserved.

Ideal for long-horizon agent workflows where context integrity matters: multi-step code generation, cross-file refactoring sessions, and complex reasoning chains that repeatedly reference early conversation details. The plugin ships with enabled set to false, so installing it changes nothing until you opt in. Mount it on the host plane for notices and the tool only, or inside a compaction realm to fully take over the compaction backend.