Repository navigation
Releases
Releases · FIZL0/agent-runner
v0.43.0
- More keys in the runner's pane, listed in a ruled-off block when the run starts and on
?+ Enter:
s!stops now (ends the agent sessions, each queued to resume in its slot next run, then exits once the
merges under way finish),ppauses without exiting,+/-add or take away a slot, andz/xsuspend or
shut the computer down once the run stops (or goes idle). A line it does not know is echoed back with the list. stoptakes--force,--suspendand--shutdown, the same ass!,zandx.- A key is acted on and logged the moment it is typed:
stoppingno longer waits up to a loop pass. - A restart on an upgrade keeps what the keys set: the slot count, a pause, a suspend or shutdown, and a stop.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.43.0v0.42.1
- Agents are told to edit files with their editing tool, not
sed -i,perl -ior a script in a heredoc, and
to run one plain shell command at a time: a command Claude Code's safety check cannot read stops for a human,
and in an unattended session it waits, is refused, and the minutes are lost. Also: norm -rfon a variable
or glob, nogit reset --hard,git clean,git checkout -- .,git stashorgit push --force. The same
text is in the runner's prompt and in the guide skill (a new rule 7, which also keeps sessions out of the
runner's home and.runner/).
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.42.1v0.42.0
- One repair session per failure, not per task: a repaired merge rebuilt on commits the shared branch got after
the repair, which fails the gate again, is a new failure and gets a repair session of its own (logged as such).
The same failure on the same base still gives up after one. run --dry-runaccounts for every open task: one held by a file clash says which files and who holds them
(waiting for files: F171 (web/app/** held by B20)), and others say the running sibling, the task ahead in an
ordered section, running, finished and waiting to merge, claimed by another runner, tried this run, or next
once a slot is free. "Nothing to start" no longer leaves an open task out of every list.- A branch that does not merge is logged as
not merged, parked as wip/<ID> — why, with when the task starts
again by itself (the next run, or this one once its line is edited). "LEFT FOR A HUMAN" is kept for what truly
is: a crashed merge, or a branch that could not be parked.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.42.0v0.41.0
- Every command a person runs (
plan,--version,help, …) keeps theagent-runner guideskill installed, not
onlyrun: anyone who has typed agent-runner on a machine has it. Said once on stderr when written. - A run adds one line to the repo's rules file (
[run] rules, else AGENTS.md or CLAUDE.md), committed on its own,
pointing every session at the skill andagent-runner guide: the guidance reaches sessions on machines without
the skill, and ones that would not pick it. Never into a file with uncommitted edits; it never creates a rules
file;[run] rules_pointer = falseturns it off.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.41.0v0.40.0
- A repair session that finds a flaky test, not a clash, ends with
FLAKE: <test>. The runner then runs the gate
once more on the merge as it is before parking anything, and files the flaky test to be fixed either way. A
flake used to park a tested branch aswip/<ID>for a person. - A merge held by uncommitted or staged edits in the main checkout is a loud STOPPED: a red banner in the runner's
pane, the bell, a tmux message and a desktop notification (notify-send), raised again every 30 min while it
holds, and a line when it clears. It was one yellow log line, easy to miss for hours.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.40.0v0.39.0
- Runs claim tasks with the plan in the code repo too, not only with
[plan] repo: a claim is a ref on the
code's remote (refs/agent-runner/claims/<ID>), and the first push wins. Two runs, on two machines or in
two clones on one machine and account, never start the same task, and keep off each other's files. agent-runner guide: how an agent session a person starts in the repo should work with the runner (queue
work withplan addinstead of editing TODO.md or doing it on a branch, leavetask/*andwip/*alone,
how to unstick a task). Each run keeps it installed as a Claude Code skill in~/.claude/skills/agent-runner,
rewritten whenever its text changes;[run] install_skill = falseturns that off.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.39.0v0.38.1
v0.38.1
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.38.1v0.38.0
- A run restarts itself on a new version installed under it (
uv tool upgrade,pipx upgrade, a pull in a
clone), once no merge is going, and adopts its sessions, merges and gates. Before, the loop kept running
the old code while the sessions'doneand Stop hook ran the new one. Not on Windows, and not across a
major version: that one is logged once for a human to restart. - The run log names every stale key in
runner.toml/runner.local.tomlat the start of a run: a renamed one
(it still works under its old name, from a table of renames each release adds to), a removed one with what to
do instead, and an unknown one with the nearest real setting. An upgrade never needs a config edited first.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.38.0v0.37.3
- A task tried and failed in a run goes back in the pool once its line is edited (a
plan notewith what to
do differently, a corrected line), not only once it is blocked and unblocked. A failed task waited for the
next run even after it was given the fix: Cydian's I2 sat out a day-long run with its note saying what broke.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.37.3v0.37.2
- A merge turn that finds a merge already gated for other waiting tasks gives its turn up to them, once,
instead of dropping that merge. After a restart, the gate a run left is adopted inside whichever turn
comes first; P347's turn threw away R38's batch of 16 tasks, which had just passed a 42-minute gate.
Install latest (tracks main)
uv tool install git+https://github.com/FIZL0/agent-runnerUpgrade an existing installation
uv tool upgrade agent-runnerInstall this release (pinned)
uv tool install git+https://github.com/FIZL0/agent-runner@v0.37.2