ZeroRun is a conservative deterministic test-reuse system for AI coding agents and CI. It is designed for a narrow question: when an agent asks to rerun a deterministic test target after an edit, can a previously passing result be reused without pretending changed or uncertain work is safe?
The current v0.4 boundary is intentionally constrained to Linux/amd64, immutable pinned OCI runtimes, reviewed source closures, read-only execution, no network access, and result-only reuse. Unsupported or uncertain cases execute fresh.
ZeroRun v0.4 passed its internal historical qualification across 10 merged pytest patches and 20 merged SymPy patches. The accepted measurements include 78.415% pytest p95 reduction, 78.309% SymPy p95 reduction, 8.1657× pytest same-runner compute efficiency, and 5.4545× SymPy same-runner compute efficiency, with zero observed stale-success events, shadow mismatches, or cache conflicts in the accepted evidence.
What remains open
The next milestones are external: 2–5 real developers or teams, at least 100 genuine external edit-to-test requests, repeated design-partner reliability, installation/support burden, customer economics, retention, and willingness to pay. Internal qualification does not substitute for those.
Why narrow
The product is not arbitrary-command memoization. It trades breadth for a boundary that can be explained, reviewed, and handled conservatively. The current focus is proving whether that narrow boundary creates enough repeated value in real agent-driven development workflows to deserve expansion.
Public and private surfaces
The curated public repository at floxy-21/zerorun contains the product boundary, sanitized evidence, pilot path, roadmap, and security guidance. Runtime implementation, raw receipts, private build tooling, and experimental history remain in the private implementation repository.
Contact
For a private design-partner pilot, use the contact form or email kapoorjishan2@gmail.com.