dsh-queue-merge

The official DeepSeek Harness agent loop consumes one queued prompt per turn. When you fire off four follow-up messages while the agent is mid-task, they become four separate turns — do a bit, get corrected, redo, get corrected again — and productivity tanks with every iteration. dsh-queue-merge adds the missing third queue-consumption policy: before the formal turn begins, a single LLM call rewrites all queued messages into one consolidated prompt so the agent acts on the full intent at once.

Consolidation is not simple concatenation. Corrections override earlier statements on the same point, while unrelated requests stay as distinct items. Any batch containing non-text attachments (images, files) or non-human sources skips the merge entirely and falls back to per-message processing. The consolidation call runs before the inbox is touched, so a failure leaves the queue exactly as it was. If the user edits, deletes, or reorders the queue during synthesis, the merge aborts rather than splicing out messages that were not part of the synthesis. A provenance badge follows every consolidated bubble so the original messages remain inspectable, and the merge never invents permissions the user did not state.

If your daily agent workflow leans heavily on queued follow-ups and you want them understood as a whole rather than replayed one by one, this plugin is built for you. It complements the official QueueDock and dsh-queue-plus: those manage queue order and membership, while this plugin governs how the queue is actually consumed.