dsh-plugin-context-trim

Running several dsh plugins at once means every request carries skills, tools, and prompt-section text that the current conversation does not need, needlessly consuming a precious token budget. dsh-plugin-context-trim lets you decide, per session, which plugins' skills, tools, and system-prompt paragraphs are visible to the model; suppressed items simply do not appear in that conversation's context, while every other session keeps full injection without any cross-talk.

The plugin uses a shadow-entry mechanism: cancelled items are overlaid with same-name entries at the agent-scope layer (modelInvocable:false for skills, tools.restrict deny-lists for tools/MCP, empty-text sections for prompt paragraphs), so the global registry is never mutated and everything unwinds cleanly when the agent is destroyed. A funnel button opens a plugin-grouped, three-state tree panel where you batch-toggle entries, see a live character-savings estimate at the bottom, and optionally save the selection as the workspace default so every new session inherits it. Once a session has produced at least one model step it locks automatically, and a confirm-guarded force-edit unlocks it with hot effect on the next step. Zero configuration, zero runtime dependencies, loopback-only HTTP.

Built for everyday users and plugin developers who stack multiple dsh plugins and want precise per-scene control over what the model sees, without touching the core package, registering any account, or exposing a public endpoint.