qex — Queued EXecutor

qex is a job queue for long tasks on one machine. It controls the number of cores and the quantity of memory that the jobs use together.

qex is for coding agents and for the people who work with them. Several agents on one machine each start work, and no agent sees the load of the others. The machine then runs out of memory. qex gives those agents one queue.

ID=$(qex submit --cpu guess --mem guess -- uv run train.py)
qex wait $ID
qex logs $ID

Agents: run qex help agents first. It is one page and it covers everything.

The three problems that qex solves

1. No agent sees the load of the others. Each agent finds free memory, starts a large task, and the out-of-memory killer selects a victim. qex holds a budget for the machine. A job starts when the machine has capacity for its claim, and it waits when the machine does not.

2. Every hand-rolled watcher waits on a proxy, and a proxy can go permanently false. An agent with no way to wait writes a monitor. That monitor watches a proxy for the work — a pattern in the process list, a line in a log, a file that should appear — and a proxy can stop being reachable without anything noticing. Four real watchers on one machine, in one day, slept for 95 hours combined on conditions that could never become true. Three of them:

while pgrep -f "solve.py"; do sleep 60; done      # matches its own command line
until grep -q "DONE" run.log; do sleep 60; done   # the writer was killed
until grep -q "READY" ~/other.log; do sleep 60; done  # that file never existed

A different user later found a fourth kind, on a machine two agents shared. It had slept for 63 hours waiting for ps -Ao args | grep -c solver to reach zero on that machine and on one reached over ssh. Its own work had finished two days earlier; the other agent's work held the count above zero. On a shared machine, "wait until nothing matches" is unsatisfiable by construction.

Only the first is the classic pgrep -f self-match. The others contain no pattern bug at all — they are careful commands whose evidence simply stopped arriving, or whose condition depended on somebody else. That is the general failure, and it is why the fix is not "write a better pattern".

qex waits on the process, not on a proxy for it. qex is the parent of your task and calls waitpid on that exact process: it exits or it does not, and there is no third outcome. qex wait therefore always returns — including 125 the moment somebody kills the job, where a log-watcher would still be sleeping.

3. There is no handle on a running task. qex gives each job a UUID. Use that id to read the state of the job, read its output, stop it, or remove it from the queue.

Install

Take the file for your machine from the latest release:

curl -fsSL "https://github.com/stephenc/qex/releases/latest/download/qex-$(uname -s)-$(uname -m).tar.gz" | tar xz
install -m 755 qex ~/.local/bin/qex

Or build it from the source:

cargo install --path .

qex needs Linux or macOS. It has no other requirement.

On Windows, use WSL2 — qex builds and runs there unchanged, and the jobs an agent starts (make, cargo, uv) usually live in WSL2 anyway. A native Windows build is a port, not a flag: qex holds a job with waitpid on one process id and stops a job tree with a process group, and Windows has neither. Building for Windows fails with one message that says this, rather than a page of errors. The first command starts the background coordinator for you, and you do not configure a service.

Five minutes

ID=$(qex submit --cpu guess --mem guess -- make test)   # gives the id at once
qex list                                                # what operates now
qex status $ID --wait                                   # wait, then the result
qex logs $ID --tail 50                                  # the output
qex kill $ID                                            # stop it

Put qex run -- make test in front of a command instead when you wait for it now: the output arrives as it happens, and the exit code is the exit code of the job.

A pipeline gives each stage its own log, its own exit code and its own claim:

BUILD=$(qex submit --name build -- make)
TEST=$(qex submit --name test --needs $BUILD -- make test)
qex wait $TEST      # 1 if the test failed, 126 if the build failed

Your session can stop, and the work continues

The job is not a child of your shell, and it is not a child of your agent. A supervisor holds it in its own session, and the record of the job is on the disk.

Somebody can therefore stop your agent, close the terminal, or replace the qex binary. The job continues and it still writes its result. Your wait is the only thing that stops, and any later session attaches to the same job with the id:

qex submit --id-file build.id -- make    # session 1
# a person stops the agent here. `make` continues.
qex status "$(cat build.id)" --wait      # session 2, and the result is there

This is why an agent that uses qex is safe to interrupt. A monitor script holds the answer in its own memory: stop the monitor, and the answer is gone.

Put the id file where it lasts longer than the session — the project directory or the home directory, and not a scratch directory that the harness owns or /tmp. qex gives a warning when the file goes to such a directory, because the file would go away at the moment that the handle becomes necessary.

What qex does not do

qex does not limit the CPU of a job. The cpu controller of cgroup v2 is not available to a user on a usual Linux system, and macOS has no equivalent. The queue controls the number of cores instead.

qex does not limit the memory of a job by default. A claim controls the queue only. Linux can apply a real limit with cgroup v2; set [enforce] mode. qex never reports a limit that it did not apply.

qex does not divide the machine between agents. The budget controls the total load, and the queue is the order of submission. A fan-out of 11 jobs from one agent takes the capacity, and the 2 jobs of a second agent wait for it. There is no share for each agent.

qex controls one machine. Two machines that farm work to each other are not coordinated.

A claim is a promise, and not a measurement. A job that claims 2GB and uses 20GB can still fill the machine. qex tests the free memory before each start, which limits the damage, but an accurate claim is better. qex measures each job and uses the measurement for the next job of the same command.

The documentation

The full documentation is at stephenc.github.io/qex.

Every page is also in the binary, so an agent needs no network: qex help agents, qex help resources, qex help config, qex schema status.

Development

cargo test -- --test-threads=2
cargo build --release

Each end-to-end test makes its own config and state directory, starts its own coordinator, and stops it at the end. The tests do not touch the coordinator of the user, and they turn the peer accounting off.

Use two test threads. Each end-to-end test starts real processes and waits for them. With more threads, the machine becomes busy, a job starts late, and a test reports a failure that the program does not have.

The documentation, the code comments, the help text and the error messages use Simplified Technical English (ASD-STE100).

See CONTRIBUTING.md. The first line of each commit message gives the next version number, so its form is part of the build.

License

Apache License 2.0. See LICENSE.