@paragon_321@trq212 I asked to Claude once. The generated content is consumed in EU since it's internet world wide. So EU demands AI generated content that gonna be consumed by then to be detectable.
Threadlight sessions can now report a short signal through MCP.
Approved, Rejected, Waiting, or any other session-defined label. The signal lives in the agent transcript, so the same tag appears in both live and historical views without a separate event store.
Threadlight is a local observer I’ve been building for Claude Code sessions.
It now shows concurrent sessions, history, agent hierarchy, context, wall time, live shell tasks, plan state and Git changes. No prompt/response text, and no AI-generated judgments yet.
Out of scope work goes to backlog when big enough. But not without a mention on the code that points to card ID. That documents context and tells humans and agents that a known issue was referred for a reason.
In the (near) future, companies gonna have the "Harness Engineer Team" or something. Delivery code at scale using AI, production ready, demands dedicated expertise.
I kept burning Claude Code and Codex limits on simple tasks.
Often it's not the model. Your codebase is just expensive to navigate.
I built a skill that audits it, scores it /100, and lists the fixes.
Mine went 73 → 83 on half of the fixes. Tokens roughly halved.
I'm creating a pre-production app, but I'm honestly not on the filmmaker domain, I'm a developer. But I'm interested to learn about it to help myself on my own video projects, building something that I'd like that exists to helps me first. More to come.
My CLAUDE.md points to my AGENTS.md so both Claude and Codex use the same file. Same thing to all skills. 1 single file, shared to both. It allows me to run both using exactly the same capabilities. Claude's limit expires, Codex continues.
Claude's limit reached in the middle of an implementation. Codex continues, and finds many gaps and fixes. One doesn't know who started. I simply instructed "continue implementation plan X"