Hao AI Lab at UC San Diego has released vstack, an open-source toolkit that puts coding agents to work on formal verification while trying to stop them from cheating at it. The hao-ai-lab/vstack repository went live on GitHub on November 10, 2026 under the MIT licence, together with a blog post titled "Towards Verifiable LLM Infra" by Yuxuan Zhang, Minshen Zhang, Junda Su, Yi Sun, Junda Chen and Hao Zhang, and a companion prototype engine called VeriServe.

The lab's argument is that coding agents are getting good enough to write deep infrastructure such as GPU kernels, schedulers and serving engines, which moves the bottleneck from writing code to proving it correct. Benchmarks for latency and throughput do not catch rare concurrency bugs, and human review does not scale when one engineer is directing many agents. The team also observed that naive agents asked to use formal checkers tend to game them, for instance by weakening an invariant, dropping a timeout branch or adding an unverified assumption until the proof passes.

vstack answers with a five-stage workflow (specification, abstraction, validation, verification and continuous checking) enforced by mechanical gates. Safety properties and theorem signatures are pinned by SHA-256 hash so a worker agent cannot quietly change the goal. Every state variable and action in a TLA+ or Z3 model must map to an exact file, line and commit in the real codebase. Before a proof is trusted, the model has to fail on historical buggy code, pass on the fix, fail again when a guard is removed, and replay recorded execution traces. Coordinator, worker and red-team agents are kept separate, and a finding only counts as caught by formal checking if a faithful model violates a named property, the red team cannot refute it and it reproduces on a real engine next to a passing control. The tool installs as a plugin for Claude Code and Cursor and can also set itself up for Codex.

The team pointed vstack at the prefill-decode disaggregated serving code of SGLang and vLLM, two of the most widely used open inference engines. According to the post, this surfaced three classes of concurrency problems. In one, transfer slots or GPU KV pages were reused after a request aborted or timed out while one-sided RDMA transfers were still in flight, which on NVIDIA GB200 hardware let a late transfer silently read or overwrite another request's KV cache while the server still answered HTTP 200. In another, KV pages corrupted by such a race could enter the global prefix cache and affect later requests. The third involved unrenewed transfer leases that left decode requests waiting indefinitely while holding GPU memory. The lab says it submitted six proposed upstream pull requests and links a related vLLM fix by another contributor, which was still open when checked on November 16.

VeriServe, the companion engine built on mini-sglang and Mooncake GPUDirect RDMA, centres on a single versioned page state machine called PdLedger. The lab reports a TLAPS safety proof spanning 226 modules and 516,973 proof obligations, with a mean p50 time-per-output-token overhead of 0.18 percent on GB200. On November 13 the team added five vLLM bug case studies and said it is tracking more than 15 bugs in a public sheet.

The practical message is aimed at teams letting agents touch production systems code: generating a proof is easy to fake, and the checks around it matter as much as the model doing the work. vstack's gates are an attempt to make agent-produced verification auditable.

The evidence comes from the lab itself. The bug findings, reproduction results and proof sizes are its own reports, the upstream fixes were proposals rather than merged changes at the time of checking, and no independent evaluation of vstack was found. The repository had 59 GitHub stars on November 16.

References