qex — Queued EXecutor

qex gives independent agents and harnesses one shared resource scheduler for one machine.

A coding agent sees how much CPU and memory the machine has free now. It does not see the jobs that another agent, another subagent or another harness is about to start. Ten agents can therefore each decide that the same free memory is theirs.

qex makes those decisions in one place. Claude Code, Codex, subagents, harnesses such as CI, and ordinary shells all submit work independently. They do not know each other, and they do not have to. qex admits only the set of jobs whose claims fit the machine together.

qex submit --wait --cpu guess --mem guess --id-file train.id -- uv run train.py
qex logs "$(cat train.id)" --grep 'ERROR|FAIL'

--wait holds the command until the job stops, gives you the exit code of the job, and ends with the record: the state, the code and the last lines of both streams. The output of the job stays in the log file. guess claims one half of the budget, which is 75% of the machine by default; qex info gives the number.

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

Three faults that one scheduler removes

1. The machine runs out of memory. Each agent measures correctly and each agent decides alone. The out-of-memory killer then selects a victim. One queue holds every claim. 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. The first command starts the background coordinator for you, and you configure no service.

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.

Five minutes

qex submit --wait --cpu guess --mem guess --id-file t.id -- make test

That command blocks until the job stops, and it ends with the record. From another shell, while it runs:

ID=$(cat t.id)
qex list                    # what operates now, and why anything waits
qex logs $ID --follow       # the output as it arrives
qex kill $ID                # stop it
qex status $ID --wait       # attach again, from any session

Put qex run -- make test in front of a command when the work is short and heavy and you wait for it now: it takes its turn in the queue, so everyone else on the machine keeps the capacity they claimed, the output arrives as it happens, and the exit code is the exit code of the job. When something stops the job, qex run gives 125 in place of it.

qex run ties the job to that command, but only for the stops that it can catch. Ctrl-C stops the job, and a SIGTERM on qex run stops the job. A SIGKILL, and the hangup that a terminal sends when it closes, do not reach the job: it continues, and qex list finds it. Use qex submit for anything you might come back to.

Driving many jobs at one time? Read one stream in place of a loop that asks about each job:

qex events --json     # one JSON object on one line for each change of state

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      # the code of the test, or 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. This is about qex submit; a job of qex run also stops when Ctrl-C or a SIGTERM stops the command that waits for it.

qex submit --wait --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.

A script that runs a second time starts no second job

An agent that loses its context runs its script again. Give the submission a key, and the second run starts nothing:

qex submit --wait --dedupe-key train:$(pwd) -- uv run train.py

While a job with that key waits or operates, a second submission with the same key gives that job's id and exits with the code 0. ID=$(qex submit ...) is thus correct in both cases, and the script needs no test of its own.

The coordinator makes the test and the submission one step, so two agents that run the same script in the same moment get one job and one id. A test that a script makes itself — read qex list, decide, submit — has a gap between the read and the submission, and both agents start a job in that gap.

The budget, and what a claim is

A claim is a number of cores and a quantity of memory. qex starts a job when the claims of the jobs that operate, plus this claim, stay inside the budget.

The budget is 75% of the cores and 75% of the memory of the machine. Set another value in ~/.config/qex.toml, as a percentage or as an exact size:

[budget]
cpu = "50%"
mem = "20GB"

qex info gives the budget and what the jobs hold now:

$ qex info
cores:           2 of 12 in use
memory:          4GB of 24GB in use

guess and half claim one half of the budget, so two such jobs operate together. full and max claim all of it, so the job operates alone. qex calculates the word at the submission, and the record then holds an exact value.

A job with no claim is not free. It gets 1 core and the memory of the machine divided by the number of cores, or the values in [defaults]. qex also learns: a command that ran before gets a claim from its own measurements. No claim is ever zero, because a claim of zero would let qex start jobs without end.

A claim larger than the budget is accepted, and the job then operates alone. qex gives a warning at the submission. It starts the job when no other job operates and the queue was quiet for 3 seconds. The field forced is true for such a job, and the job can swap or stop for memory. Set [queue] oversized to reject to refuse it at the submission, or to queue to hold it. A claim above a pool, such as a GPU, is always refused: an empty machine makes no fifth device.

Jobs start in the order of submission, and one large job does not hold the queue. A smaller job behind it starts while the capacity is free. After 2 such jobs, qex keeps the capacity for the job at the front and starts nothing else, so the large job always gets its turn. [queue] max_bypass sets that number, and 0 gives a strict order. --priority N puts a job before the jobs with a smaller number.

qex keeps no capacity when another user, or a program that qex never started, holds it. Such a wait has no known end, so the jobs behind continue to start.

qex tells the job what it claimed

Give both --cpu and --mem, and qex writes the claim into the environment of the job:

QEX_CPU=2  GOMAXPROCS=2  OMP_NUM_THREADS=2  QEX_MEM=4294967296
GOMEMLIMIT=4294967296  NODE_OPTIONS=--max-old-space-size=3072

A runtime then sizes its thread pool and its heap to the claim, and not to the machine. This is the one place where qex changes how your program runs, and it can decide whether a job succeeds.

The value is the claim, and not a measurement. qex writes the number that you asked for.

Both halves must come from you. --cpu 2 with no --mem writes nothing, and a claim that qex chose or learned writes nothing. A claim of 1 core that qex invented would otherwise make a build of 16 cores 16 times slower, with no message.

qex never replaces a value that you set. A program that asks the operating system directly still sees the whole machine.

--no-limit-env-hints turns this off for one job, and [claims] export_env = false turns it off for every job.

What qex does not do

qex does not measure your program before it starts. A claim is a promise, and qex trusts it. qex measures a job while the job runs, and it uses that measurement as the claim for the next job of the same command.

qex applies no limit by default, so nothing stops a job that goes above its claim. 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. 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 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 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.

One coordinator serves one user. Every agent that runs as you shares one queue. A second user on the machine gets a second coordinator, and each coordinator reads the claims of the other from /tmp/qex before it starts a job. That accounting is cooperative. It needs no administrator rights, and it trusts what the other coordinator writes.

When the kernel stops a job for memory

A training run that the kernel stops at hour four gets the state oom.

qex reports it, and it starts no new attempt. The record says what you can do. Give a larger --mem value and submit the work again:

qex submit --wait --mem 16GB --id-file train.id -- uv run train.py

qex cannot do more than report while it applies no limit, which is the default. A claim is then a promise, and the count that the kernel keeps also rises when the kernel stops a different program of the same user. The machine can be full while your claim is correct, and a larger claim is then the wrong answer.

[enforce] mode puts each job in a cgroup of its own and applies the claim as a real limit. The kernel counts a kill at that limit separately, and that count is proof that the claim was too small. With that proof, qex is designed to multiply the claim, start the job again with the same id, and keep the lesson for the next job of the same command.

That correction does not run in this build. The supervisor joins the cgroup of the job and then stops every process in it, so it stops itself before it reaches the code that raises the claim. Issue #88 holds the fault. While it is open, [enforce] mode gives you a real memory limit and the state oom, and no job runs a second time.

See the reference for the rules of the correction.

The documentation

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

Page	What it holds
Agents	The page for an agent. Start here.
Sandbox	What qex needs when a harness runs each command in a sandbox.
Reference	Each command, option and configuration field.
Design	The coordinator, the supervisor and the files.
Security	What qex writes, and who can read it.

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

Completions for your shell

qex completions bash > ~/.local/share/bash-completion/completions/qex
qex completions zsh  > ~/.zfunc/_qex
qex completions fish > ~/.config/fish/completions/qex.fish

The shell then offers each job by its id and by its name, and it offers the set that each command accepts: the jobs that operate after qex kill, and the jobs that wait after qex cancel.

A skill for your agent

skills/qex/SKILL.md teaches an agent to use qex instead of a background command and a polling loop:

mkdir -p ~/.claude/skills && cp -r skills/qex ~/.claude/skills/qex

See skills/README.md. The same material is in the binary as qex help agents, for an agent with no network.

Tell us what happened

qex helped you? Leave a comment. Give the number if you have one — hours, jobs, the size of the machine — and, if you are an agent, which model you are and which harness runs you.

A recipe worth sharing? The tips and tricks issue takes it — for example, qex has no timer because cron plus qex submit --lock is better, and that recipe is there.

Something did not work? Search first — gh issue list --repo stephenc/qex --state all --search "your words" — and comment on the issue that exists rather than opening a second one for the same fault. If nothing matches, open an issue. A small fault counts: a word that sent you the wrong way is worth a report, and most of what qex does today came from one.

Never paste the output of --show-env into a public page.

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 title of a pull request gives the next version number, so its form is part of the build.

Do not change the version in Cargo.toml. main holds 0.0.0-dev for ever; the number of a release lives on the tag, and build.rs puts the commit into the version of your build, such as 0.0.0-dev+g98513e2.

License

Apache License 2.0. See LICENSE.