Noise-canceling headphones for your AI agent: OMNI v0.6.1

v0.6.1 is a positioning release with one code change in it, and the code change turned out to be wrong. Both halves are worth writing down.

Pain-first positioning

The README was rewritten around the two problems an agent actually has in a terminal: noise (the 300 lines of npm WARN between you and the one error) and context amnesia (the agent forgetting, three commands later, what it already read). The line we landed on, and still use, is “noise-canceling headphones for your AI agent”. It says what the tool does without inventing a product category for it.

The logo was redrawn to match: a neon “cute brain” wearing headphones, the noise-canceling half and the memory half in one mark.

Terminal UI polish

The [OMNI Active] status tag was overlapping the command output it was supposed to sit under. Distilled payloads now carry a trailing newline, so the tag lands on its own line.

Two housekeeping fixes went with it: a clippy::needless_splitn warning in the directory classification logic, so make ci stays at zero warnings, and the insta snapshots resynchronised against the new find output.

The find distiller, and why we retracted it

This release rebuilt the SystemOpsDistiller for find to strip out the raw file paths entirely and emit a key-value summary of directory distributions and extension counts instead. The changelog entry credited it with up to 99.8% savings, and at the time that read like the best number in the project.

It was the worst one. The paths were the answer to the question the agent asked; deleting them and counting the deleted bytes as a saving is the exact failure mode OMNI exists to prevent. Worse, that single number went on to poison the published figures: the 97.3% all-time reduction we advertised traced back to find . alone, at minus 12.6 million tokens of destroyed payload.

v0.6.2 replaced it with lossless prefix factoring, where the shared directory prefix is hoisted into a header and each path is stated once relative to it, cut at a separator so every path round-trips byte for byte. The honest saving on the same output is about 72%, with nothing lost. We would rather publish 72% and mean it.

If you are reading the archive in order, read the v0.6.2 note next. It is the release where we went back over every number we had published and replaced the ones we could not reproduce.