Kalari: Contained Rehearsals for API Upgrades

API upgrades break silently. A consumer can exit zero and print a neat total while 2,331 of 2,431 records never arrive. Kalari exists for that exact failure. It plugs into the coding agent you already use over MCP and rehearses the upgrade in a contained arena before the change ships. Kalari gives the agent five tools. What_changed lists the API changes with stable IDs. Blast_radius maps every caller and test each change touches. Recall_lessons returns past lessons as checks that must pass. Run_in_sandbox executes the migration in a throwaway microVM with time, output, and network caps. Record_lesson stores what a verified run proved, keyed by change kind. Safety is proved, not promised. A Lean 4 gate checks every sandbox call against four theorems. Workspace writes stay inside the project. Network stays within one allowlisted endpoint. Every run is capped and no secret travels in any argument. The aashaan supervisor enforces the caps inside each microVM and reports killed_timeout when a loop never ends. A hidden grader decides whether the upgrade works by comparing records, not by trusting exit codes. We measured Kalari against the reference setup on a frozen pair of real API behaviours. Both arms passed every session: 20 of 20 with Kalari, 40 of 40 without. The difference is cost, time, and containment. Kalari cost $0.019 per session against $0.082, finished in 26 seconds against 84, and recorded zero stuck seconds against 799. Production published zero lessons, so condition 5 matched condition 4: these runs show no benefit from learned memory yet. The numbers above come from that honest comparison. Kalari runs on Vultr end to end. A VX1 virtual machine hosts the service. Vultr Serverless Inference serves the coding agent and the decision models. Microsandbox microVMs contain every run. Try the live judges page, read the evidence behind each figure, and run the open-source benchmark yourself.