Zachary Woods

Coding agents got fast; reading what they did didn't. A long agent turn arrives as a wall of terminal scrollback, and buried somewhere in it are the two decisions that actually needed you. Shelly makes visual output the default working mode: agent sessions — Claude Code or Codex — run inside embedded terminals, and each turn the agent writes a self-contained HTML page (a status card, a plan, a diff review) that Shelly floats in a panel over your work without stealing focus. You read the page instead of the scrollback, then answer in place — a ✓/✎/✗ ballot on the proposed next steps and a comment affordance on every block. Click the exact paragraph you're asking about and your reply is compiled into prose and injected straight into that agent's live session; you never type into a terminal. It's a desktop app rather than a web page because the panels have to sit over your terminal without taking focus and the agents have to actually run somewhere: Shelly registers an Objective-C NSPanel subclass at runtime from Rust and swaps framework-owned windows into it, runs each agent CLI in an owned PTY, and treats agent-authored HTML as untrusted — sandboxed opaque-origin iframes, a typed message protocol, and a hard gate that stops remotely-fetched pages from reaching a local shell. The Board unifies every session in one surface: artifact history, unread counts, live status, and a read-only diff of the files a session has touched. ~22k lines of Rust and TypeScript on Tauri v2, 23 releases over two months, MIT and public.

ONNX (Open Neural Network Exchange) is the Linux Foundation AI standard for ML model interoperability, used by PyTorch, TensorFlow, scikit-learn, and every major ML framework. Rewrote the ReferenceEvaluator Resize operator using separable 1-D interpolation, replacing per-output-pixel recursion with vectorized numpy operations. Achieved ~600,000x speedup (15 minutes to 1.5 ms on a typical (1,32,20,20) → (1,32,40,40) bilinear/half_pixel resize) while preserving bit-exact numerical compatibility (max diff <1e-10) across 294 test configurations covering every mode × antialias × coordinate_transformation_mode × shape combination, plus tf_crop_and_resize extrapolation and axes-restricted resizes. Resolved a long-standing performance issue (#6554) open since November 2024. Full 1798-test backend suite passes; ruff and mypy clean.