dsh-reasoning-summary

Closed-source providers encrypt their reasoning, so it never fully enters the DSH context. Each subsequent step forces the model to re-speculate what the previous step accomplished, and that speculation is immediately discarded in the next round, silently burning time and tokens. dsh-reasoning-summary closes this gap: it instructs a selected route to emit a concise, visible action summary naming the relevant files, commands, confirmed observations, and the very next action before each tool call, then persists that summary as ordinary session history that every later step reads directly without re-deriving it.

The plugin targets a single route matched by provider and model, leaving all other models untouched. It does not hide history, block tool calls, enforce permission policies, or alter what any model can see. Summaries survive mid-session model switches and toggling the setting off and on; they are removed only when the route list is cleared. For turns where the model produces reasoning but no visible action, the plugin requests a continuation up to three times per turn. On a suspected stall—two complete summary tags with no tool call—it releases the stream and injects its own prompt so the user can intervene without re-describing the problem. Legacy fields from earlier versions are normalized and discarded on load without breaking the runtime.

Ideal for DSH users who mix closed-source and open-source models, want to cut redundant re-inference costs, and need context continuity when switching providers mid-conversation. The plugin is off by default: tick the target route in the settings card and it takes effect immediately with no restart or page refresh required.