the agents that talk themselves to death after 3 hours need one file, not a framework

spent a bunch of hours watching claude code and kimi sessions drift the same way: > I should check the test output before continuing. > Let me think about the best approach. > Actually, I should verify the state first. > The next step would be to update the configuration. a lot of words, zero shipped work. the agent isnt broken. its operating contract is missing. shipped a small public CLAUDE.md that fixes this for long-running coding agents. one file, no framework, copy it into your repo and tell the agent to follow it. focuses on action over narration, live evidence over stale memory, compact session state, recovery after restart, and safety checks that dont become cages. over 1600 hours of long-running sessions inside a private deployment before public release. works on claude opus 4.6 and kimi k2 the same way. MIT licensed. theres a 60-second prompt-only demo in the repo if you want to feel the action-over-narration shift before cloning anything. paste the prompt in any capable model, give it a real task, watch the difference. whats the worst long-session rot youve hit with your agents? curious whether the operating-contract framing maps to what you saw or if your rot looks different. (repo link in top comment below per sub rules)