@mattpocockuk@trq212 Ive been experimenting with letting claude run a small http server in a background bash so it can generate interactive html and be alerted on those interactions. I had it add a feature where I can comment on any element. Still wip.
@ZachSDaniel1@chris_mccord Agree. I went through the phase of complicated systems just to come back to a stock harness with a few skills installed. Things really do work well out of the box for the most part.
@thdxr Sweet. In case its helpful, we're assembling a span tree in memory to get agent turns, subagents and their metadata, especially model, tool calls, turns, etc and doing some hacky stuff to pass trace data through to our MCP server.
@thdxr Also would be nice to filter to the things interesting to a caller like prompts, tool calls, etc. The internal spans would be too much noise for us.
@thdxr Any plans to expose the traces via the sdk? We're building traces based on the event stream, which is clunky and difficult to keep spans correlated. Would be great to be able to pass in trace headers on session start and a collector config.
@trq212@thdxr Skills suck. Sure they're useful but that's not helpful when they're impossible to govern and secure. There's no way to track their origin or their dependencies, especially across harnesses. I've haven't seen such a half baked idea go so far since react went mainstream.