Yyongxin12· perspective-bridge· 2 days agohands-on
Tried the newest model on the review persona: fewer invented findings, slower
Switched my local copy of the security persona to the latest model for a week of PRs. It stopped flagging the "unsafe deserialization" false positive on our YAML loader, which the older model raised on every PR. Latency roughly doubled on a 900-line diff. For an advisory comment that's fine; for anything blocking it wouldn't be.
Net: worth it for /review security, not for /fix-suggestion where the log tail matters more than the model.
CCHANSingYeungSunny· job-agent· 3 days agodiscussion
Should the gateway expose one merged tool list, or one tool per agent that routes internally?
Right now clients see job-agent__tool_run and erkang__tool_run side by side. With 30 agents that's a long list for the model to pick from. Alternative: a single fleet__ask tool with a router. I lean toward keeping it explicit; the slug prefix is the only thing that makes the roster legible.
Tplatform· 2026-09-14release
Fleet status page is live for admins; Plaza (public) is next
Weekly snapshot of every member's scaffold version, gate result, hosting health and review quota. Same data will feed a public project gallery with the internal columns removed.
Xxuefeiyu2015· SciComAgent· 5 days agohands-on
Migrating a uv project onto the template: what broke and what didn't
Adopt mode was additive as promised: nothing of mine was overwritten. What broke was the gate's install step, which only knew poetry. The fix landed in policies v0.0.9 the same week. If you're on uv, pin @v0.0.9 or newer and skip the pip install uv workaround.
Wweikangceng7-ai· erkang· 1 week agotrend
MCP is becoming the default seam between agents and tools. Are we exposing enough?
Three of the tools I use daily now ship an MCP server by default. Our fleet already speaks it end to end, but every member still exposes 1–3 tools. Worth a shared pattern for resources and prompts, not just tools.
