Test runs, builds and services print far more than anyone reads. Coding agents read it anyway, all of it, and pay for every token. Then they read it again after the next change. The common fix is to have a language model summarise the output, but a summary can drop the one line that mattered, and it won’t tell you what it left out.
Abbreviatory takes the opposite approach. It condenses large tool output, test runs, logs, JSON, diffs, HTTP responses and directory listings, into a short report. Failures, errors and warnings come first. Repeated lines fold into one. The report stays inside a byte budget. Anything left out can be brought back with one command. It uses rules, not a language model, and it never touches the network.
The Pitch in Four Numbers
- 5.6 MB to 3,075 bytes. A log of 63,369 lines reduced to a report 0.05% of its size.
- 1,500 of 1,500 simulated CI runs matched the known new, fixed, still-failing and flaky tests exactly.
- 10.9 MB peak memory while a 1,107 MB log of 12.5 million lines streamed through.
- 105 Go tests passing, with nothing found by the race detector.
How It Works
Pipe output through abbreviatory reduce. It detects the format automatically and streams the input, so even huge files need little memory. Lines that differ only in numbers, IDs, addresses or timestamps fold into one line with a count, like x486. Failures, errors, warnings and anything you mark with -keep go to the top. The report stays under 4 KB by default.
Every omitted line has a reference, something like [ref d2f86a61aa96:g10], and abbreviatory expand turns it back into the original bytes. Same input and options always give the same output. Nothing is guessed and nothing is lost for good.
It can also remember previous runs. Compare a test run against history and you get the new failures, the fixed tests, the ones still failing with a run count, and the flaky ones, plus new warnings if you ask. If nothing changed, the report is one line. abbreviatory run wraps a test command and keeps its exit status, and a GitHub Action does the same in one workflow step.
Built for Agents
abbreviatory mcp gives coding agents four Model Context Protocol tools: run a command and get a report, reduce existing output, expand a reference, and list run histories. abbreviatory serve does the same over HTTP, local-only unless a token is set. It reads go test, pytest and JUnit XML, application logs, JSON, git diffs, HTTP responses, directory listings from find, ls, tree and du, and build output from cargo, tsc, ESLint, Gradle, Maven, npm and pnpm. It’s one static Go binary for Linux, macOS and Windows.
Who It’s For
- Developers running coding agents who want lower token bills and faster loops.
- CI owners who want “what changed since last run” instead of a thousand-line log.
- On-call engineers reading service logs under pressure.
- Anyone who doesn’t trust an AI summary to keep the line that matters.
Limits
It isn’t a general text summariser. Text it doesn’t recognise is treated as a log: it can find error and warning lines and fold repeats, but it can’t judge which sentence of free prose is important. Stored logs are kept gzipped on local disk for a week by default, up to 500 MB, so references can be expanded. -no-store turns that off. The project calls itself early.
The Four-Question Screening
Every project featured here gets the same four questions, the way you’d screen a startup idea.
1. Where is the money in it?
Abbreviatory is free and has no paid tier. Its value is measured in tokens and time, and companies running coding agents at scale are spending real money on both. Paid paths include team features like shared run history and dashboards, enterprise support for more formats, and licensing the engine to agent platforms and IDE makers, who all face the same context problem. A tool that cuts context by orders of magnitude has a clear return.
2. What real problem does it solve?
Too much output, read too often. Agents waste context and money on lines that don’t matter, and humans scroll past the one that does. AI summaries help but can’t be trusted to keep everything important.
3. The technology in plain words
Think of a good assistant reading your mail. They put the urgent letters on top, bundle the four hundred identical newsletters into one note that says “x400,” and keep everything else in a labelled box so you can ask for any letter later. They follow fixed rules, so they never throw away the letter that mattered.
4. Moat and differentiation
The default answer to “too much output” is to ask a model to summarise it. Abbreviatory’s edge is that it’s deterministic, bounded and reversible: same output every time, never over budget, and every removal recoverable. That’s what makes it safe to put between an agent and its tools. The moat that can grow is format coverage and the run-history comparison, which turns noisy CI into a short list of what actually changed.
Why It Matters
Context is the scarcest resource an agent has. Spending it on repeated log lines is waste. A small, honest reducer that keeps every failure and loses nothing for good makes agents cheaper, faster and easier to trust.
Follow the Project
Read what context reduction is, try the demo, browse the case studies, start with the quick start, or get the code from the Abbreviatory repository on GitHub. Abbreviatory is listed with the other open-source projects on SoftDevStudio.com, and more featured projects are in the Projects section.