The release where we caught OMNI lying: v0.6.6

OMNI is a lossy tool. That is the deal: it drops noise from your terminal so your agent reads less. A lossy tool you can trust has one job it can never fail. It must not invent, and it must not delete the answer. v0.6.6 is the release where we audited OMNI against that job, found it failing in seven places, and fixed them.

We are writing this the way we’d want a tool we depend on to write it: with the wrong output quoted, not summarised.

Two failure shapes

A distillation can go wrong in two ways, and we found both.

It fabricated success. A python3 -c "…" or ruby -c … script that printed no error came back as Build: ok, a string that appears nowhere in the real output, invented in the success direction. A test-shaped command with no test results came back as Tests: 0 passed, 0 failed. On a real GitOps verification, Build: ok would have reported a passing check for a broken manifest. An agent acts on that and is wrong.

It deleted the answer. Run find, ls, or a verbose git log and every line is the answer. There is no noise to drop. OMNI was “compressing” them anyway: find returned 30 of 98 paths, ls -la replaced every filename with a count, git log -12 came back with 2 of 12 commits. High reduction on those commands was never a win. It was the bug.

What v0.6.6 fixes

• No more invented verdicts. python/ruby scripts and unparsed test output now pass through verbatim instead of Build: ok / Tests: 0 passed. A distiller that recognised no signal returns the raw output, and it never guesses in the success direction.
• Enumeration commands are left alone. ls, find, ps, wc, df, env and friends are lists of data, not noise, so OMNI now hands them back byte-for-byte instead of truncating them to a shorter, plausible, incomplete list.
• Verbose git log keeps every commit. It now distils to one hash subject line per commit, the --oneline view, so all of them survive while the Author/Date/body noise goes.
• git log --oneline keeps its subjects, and the JS/TS and TOML paths stop dropping lines or picking a filter by accident.

Seven fixes in all, and every one of them makes the reported saving smaller on the commands where the old number was a lie. We think that is the right direction for a number to move.

A benchmark you can run

The headline on our site, 58.9% fewer bytes across 1,810 real commands, with 63.6% of calls saving nothing at all, used to come from a run nobody kept. v0.6.6 ships the reproducer: a committed replay that reads real command traces, runs each through the current pipeline, and refuses to count a failed or empty replay as a saving. The number is now a fact about the shipped binary, not a screenshot.

We also wrote down, in the repo, exactly where OMNI wins and where it should stay out of the way: it earns its place on cargo test, terraform plan, git log, build logs, output that is a verdict wrapped in noise. On a file listing or a process table it now does nothing, on purpose.

The standard

Every compressor asks you to trust that what it cut didn’t matter. We hold OMNI to the harder version of that promise: the original is always recoverable, and the tool never lies to your agent. v0.6.6 is us checking our own work against that promise and shipping the corrections in the open: measured evidence, the wrong numbers we published, and the fix, for each one.

Upgrade with brew upgrade fajarhide/tap/omni, or omni update.