Skip to content

Troubleshooting

Most of this CLI’s errors are written to be the fix, not just the symptom. This page is for the ones that need a sentence more.

Either nobody has signed in and minted one on this machine, or a token exists but was filed under a different URL than the daemon is looking under — the fix in both cases is the same:

Terminal window
cawdev

It signs you in through the browser and mints a runner:operate token for whatever this config already serves. If a token exists for a different instance, the message names it — worth checking whether that’s actually the same cawdev reached through a different door (development’s console vs. API ports are the common case), in which case the fix is pointing url in the config at the one the token is filed under, not minting a second one.

“session expired — /login to sign in again”

Section titled ““session expired — /login to sign in again””

The platform stopped honouring your browser session: it expired, you signed out in the console, or your account was disabled. The terminal drops to watching only. Type /login and approve the code in your browser. The polls pick up on their own afterwards.

The config’s projects map is empty. Add at least one slug → path, or run cawdev --setup to be walked through cloning one.

The daemon refuses to start a run on top of somebody’s uncommitted work by accident. Commit or stash it, or start the run again from the console and choose explicitly to build on top of what’s there — see The runner daemon.

Two daemons are answering on this machine — normal if you deliberately run more than one. Name which: cawdev --runner <name>.

“This CLI is below what <url> requires”

Section titled ““This CLI is below what <url> requires””

The platform has a minimum CLI version and this one is under it. The daemon refuses to start until you update; cawdev --version shows both numbers.

Nothing is wrong by itself — check the transcript for the idle note (“this session has said nothing for N minutes”). It’s not a failure state, just a flag that the checkout is being held by something that might be stuck. If it genuinely is stuck, x twice in the attached terminal cancels it.

It moves to USAGE_LIMITED, holding no process, and keeps its checkout. It is picked back up automatically once the window resets only if the machine has auto-resume on. Auto-resume is per machine and off by default, set under Your settings → Machines; without it, Carry on from the run page restarts it. A PAUSED run, by contrast, is never picked up by auto-resume: a person stopped it, and only a person starts it.

The run’s briefing says exactly where the .patch file landed and gives you the git apply --3way command to run it by hand. Check git status afterwards before trusting the tree — a three-way apply that partially succeeds leaves conflict markers, not a clean failure.

M (always, on this machine) only appears when the config has "acceptsRulesFromConsole": true. Pressing it anyway tells you what to add. See Permissions and the shield for what turning that on actually means before you do.

Check that the coding agent itself is signed in on this machine — claude, agy or codex, run once by hand. cawdev --setup asks this up front for exactly this reason: a runner whose agent isn’t signed in boots perfectly and fails every single run, which is a much worse way to find out than a question.

“ssh needed to ask you something” during a clone

Section titled ““ssh needed to ask you something” during a clone”

An ssh URL whose key has a passphrase, or a host being trusted for the first time, needs ssh to ask a question on the terminal directly — something the attached screen can’t hand over mid-draw, so it clones in batch mode instead and fails at once rather than hanging. Add the key to your agent once (macOS: ssh-add --apple-use-keychain ~/.ssh/id_ed25519) and try again, or run cawdev config add-project from a plain shell, which can ask.

The workspace folder is missing (spawn … ENOENT)

Section titled “The workspace folder is missing (spawn … ENOENT)”

A project’s checkout path in the config no longer exists on disk (moved, deleted, an unmounted drive). The run queues and says so rather than failing at spawn. Restore the folder, or point the project at a new path with cawdev config, and the run goes on. cawdev config remove-project <slug> stops serving it without touching the files.

The daemon was restarted (DAEMON_RESTARTED)

Section titled “The daemon was restarted (DAEMON_RESTARTED)”

A run failed with this because the daemon driving it was killed and started again: by systemd, by cawdev, by hand. The new daemon fails what the old one was driving at once and holds the checkout, so Carry on picks the run up. With auto-resume on it does so once by itself; a second restart on the same run waits for you.

On a small Linux box the usual cause is memory: a test suite grows until the kernel’s OOM killer picks the biggest process, which is sometimes the daemon. journalctl -k | grep -i "killed process" says so. Set sessionMemoryMax (for example "6G") in the config file so a runaway session is killed on its own and the daemon stays up, give the daemon OOMScoreAdjust=-800 in its systemd unit, and add swap.

A run of theirs that writes code is refused before it starts unless the machine accepts rules from the console (cawdev config rules on) and has Allow everything ticked under Your settings → Machines. The refusal says which of the two is missing. See Agents.

On a coding run, report done is refused while the checkout has uncommitted files or commits the remote hasn’t seen. The answer names them. The session commits what belongs to the work, discards what doesn’t, pushes, and reports again. If it can’t (a push the remote refuses, a file that looks like a credential), the run fails with the reason and the checkout held. Then Carry on, Commit or Discard decide.

The stage’s turn ended with a background shell or an Agent still running, or one that finished and was never read, so its result was lost. The failure names the commands. /continue in the terminal, or Carry on in the console, re-runs the stage, told to run them in the foreground.

Opening, finding and merging a pull or merge request happens on your machine with the host’s own tool, and a machine without it says so in one sentence. Install and sign in gh for GitHub or glab for GitLab. See Git hosts. A daemon too old to know a request the console makes (creating a branch, for one) is never handed it, and the refusal says to restart the daemon.

cawdev config show prints this machine’s config with the token redacted — the first thing worth pasting into a bug report. The daemon’s own log is one key away in the attached terminal: g or /log.