The work was done and thrown away: v0.7.7, OMNI

Most of this release is OMNI catching itself. Not catching bugs in other people’s tools, catching the places where it had been quietly telling you something that was not so, or doing nothing while reporting that it had done something.

The work was done and thrown away

The Pi plugin ran the hook on every tool result. It spawned the process, handed over the payload, waited for the answer, and then read hookSpecificOutput.updatedResponse out of the reply.

OMNI stopped emitting that key in 0.6.5. It emits updatedToolOutput.

So the value was always undefined, the handler always returned undefined, and Pi received the raw output every single time. Measured on a real payload rather than reasoned about: 29,099 bytes went in, and the answer that came back and was discarded carried 7,365. 74.7% removed, then dropped on the floor, on every tool result for the whole life of that plugin.

The plugin’s own type declaration named the wrong key too, which is why its TypeScript check passed for months. The contract was wrong in the one place that would otherwise have caught it.

This is the same mistake as an older one, and the older one’s epitaph is still in the source three lines above the fix. The Rust side has asserted since then that it writes updatedToolOutput. Nothing had ever asserted that anyone reads it, and that was always the boundary that mattered. A check now scans every plugin that spawns the hook and fails unless it names the key OMNI actually writes.

The report could not see its own biggest engine

OMNI removes bytes two ways. A distiller shortens one command’s output. A ledger folds away lines you have already been shown in this project, across turns, so the second Read of a file is mostly markers.

omni stats computed everything from the distiller’s table and never read the ledger’s. Over one week on the machine that found it, the distiller removed 309 KB and the ledger 501 KB. The report showed the first number and not the second, and showed it as 4.5%, because that percentage was the distiller’s saving divided by the 15,000 calls OMNI had deliberately declined to touch.

The default view is three lines now, one per engine, each percentage against its own base, and the calls OMNI declines are shown as a count with a zero beside them rather than averaged into a number that reads as failure.

Two ratios from two different populations are never added. Bytes may be summed; percentages may not.

A field whose whole job was to qualify a number

omni stats --json reports measurement_method next to every token figure, so a reader knows how far to trust it. It chose between actual and estimated by testing whether a token count was above zero. That column has exactly one writer, and the writer is an estimator, so the test was never false.

Every one of the 16,405 rows ever recorded said actual for bytes divided by a constant calibrated against a vocabulary that is not even Claude’s.

It says estimated now, and a check scans the reporting surfaces and fails on the literal.

Two that could have cost you an afternoon

A failing terraform validate came back rendered as a clean plan. The summariser counted a created resource for any line containing “will be created”, and terraform’s own error prose contains that phrase.

And twenty-three places cut text at a byte index. Any output with a multi-byte character in the wrong position could panic the hook.

The win

A fold can sit in the middle of a Read now rather than only at its edges, because the folded view keeps the file’s line count and the editor’s numbering stays true. On the corpus that measured it, that class went from 27.3% to 82.3%.

And omni init finally tells you what a host lets OMNI do. Configuring a host with no hook used to print the same green tick as a host with one, and you found out from an empty report days later.

brew upgrade omni