sdlc-close-looplisted

# Closing the loop Maintenance is the stage that stays manual longest. An alert fires overnight and is missed. A ticket sits until someone has capacity. Post-mortem actions never land because the next incident arrives first. Every path needs a person to restart it, and people are the scarce thing. Closing the loop means a trigger — a threshold crossed, a ticket filed, a message posted, a schedule elapsing — starts the work **with nobody in the invocation path**. The agent investigates, acts only through routes you've approved, and writes what it found as a new intent. People triage and review; they no longer have to notice and begin. ## What has to exist first An intent format (`sdlc-intent`), because the loop needs a structured way to hand work back. A review gate (`sdlc-pr-review`) and hooks as an action boundary (`sdlc-hooks`), because this is where an agent runs unsupervised. And **a rollback path you have actually rehearsed** — the highest tier calls it, so it cannot be a runbook nobody has executed. ## Detection stays deterministic ⚠️ **No model participates in detection.** A script decides whether something is wrong; the agent is invoked only after that decision, and its tier governs what it may then do. This is not a performance concern. A trigger you can't explain is a trigger you can't trust at 4am, and the first false alarm that wakes someone will end the experiment. Pick **one** metric with a baseline stable enough to have a normal range — test failure rat