Panda

I just shipped SpecsRelay for DeepSeek 0.10.0, focused on the steps between discussing an idea and handing it to a coding Agent.

SpecsRelay opens DeepSeek inside DSH Desktop and turns your conversation into requirements you can review and send to a local project. Here's what's new today:

Continue clarification in the original conversation. When SpecsRelay finds an unresolved question, you can now copy the questions back into DeepSeek, discuss them there, then recapture the updated conversation and organize it again. You can still answer directly in the panel. Copying never sends a message automatically.

Optional help with long conversations. The organizer plans what to look for goals, confirmed decisions, constraints, acceptance criteria and open questions then Jev selects original evidence before the organizer writes the requirements. It all happens within Organize current conversation, with no separate compression step. Short conversations use the full text, and planning or screening failures fall back to the full conversation.

Hi makers! Today's update (September 21, 2026): I've prepared the SpecsRelay 1.0.0 build, focused on the gap between discussing an idea with AI and giving a coding agent a clear task.

What's new:

Long-conversation assistance. Optional TypeSafe Jev support selects relevant older assistant replies before your chosen model organizes the requirements. All user messages and the newest turns stay intact. If Jev is unavailable, SpecsRelay uses the full conversation. When input is reduced, you see an estimated token reduction.

Model and reasoning advice. After the requirements are ready, an optional advisor suggests a model capability tier and reasoning effort. Supported agent families can also receive a concrete model suggestion. You keep control of the agent and model selection.