AyeOps: voice-authorized incident command

AyeOps is a voice incident commander. It pages you when production breaks, diagnoses the fault on its own, and then refuses to change anything until a human authorizes it out loud. When a service fails it queries health, tails error logs and finds the deploy that did it, down to the commit and the line that crashes. Then it stops. It proposes a fix, and the relay generates a one-time code that appears only on the operator's screen. The model is never given that code. To authorize, the operator says the action, the service and the code together: "Roll back auth-service, Hotel Alpha." A code on its own is refused, because reading two words off a screen is not the same as understanding the change. The relay checks the readback and runs the change. The model never does. The claim is checkable. The repository ships a checker that reads AssemblyAI's own record of every session, searching system prompts, tool schemas, tool calls and results and every agent turn for the code the operator spoke. Its first run found a real leak: a post-incident summary quoted a spent code. We fixed it and kept the failing test. Since then, 6 sessions and 100 turns searched, 0 occurrences. Built on the AssemblyAI Voice Agent API: speech, model and voice over one WebSocket, JSON-schema tool calling in hold mode, tools that do not exist in the model's schema until there is something to fix, barge-in, session resume, and keyterm prompting so code words transcribe correctly the first time. Nothing in the demo is simulated. Every session gets three real Docker containers that genuinely crash-loop and genuinely recover. Every authorization ships with a recording, a turn timeline and a postmortem: the audit evidence a security review needs before it will let an autonomous agent touch production. Try it with no microphone and no signup: https://ayeops.vexralabs.com