Bash debug answers: exit codes, arguments and linters

The -x option makes Bash print every command, with its expanded arguments, right before it runs, so you see exactly where the script went wrong. You start it as bash -x script.sh in WSL or on a Linux system you own, and set -eu stops the run at the first error or unset variable. In Pageferry, the Shell card holds this trace together with the real exit code.

A bash linter is a separate program that reads your script without running it and points out likely mistakes, such as an unquoted variable. You choose your own linter, install it in WSL or on your Linux system and ask the agent to run it with Shell. The card then shows its messages and its exit code. A linter does not replace a test run.

Arguments are the values after the script name, for example bash script.sh one two, and the script reads them as $1, $2 and so on. $# holds the count, and "$@" holds all of them as separate words. To add required flags and arguments, test $# at the start, print a usage line to standard error and exit 2. Stop passing secrets as command-line arguments: other users see them in the process list.

An exit code of 127 means that the command was not found, usually because of a typo, a program that is not installed or a PATH that does not include it. The shell prints a message, and Pageferry shows the same message and the number in the Shell card. A Bash script running apps with flags should check the exit code of every app it starts.

A code of 126 means the file was found but cannot be executed, usually because it lacks the execute permission or is not a program. On your own script, chmod +x script.sh adds the permission, after which ./script.sh can start it through the shebang line. If you are unsure, run bash script.sh first and read the exit code in the card.

Yes, codedguru.com is the official Pageferry website. The download button gives the original setup program of about 34 MB, and you can compare its SHA-256 checksum with the published manifest before you run it. Windows SmartScreen may warn because the installer is not code-signed yet. Questions about the app go through the Send feedback button inside it.

Yes, through WSL on your own PC or over SSH to a Linux system you own. Pageferry itself is an app for Windows 10 and 11, so you chat on Windows, file tools work in the chat's working folder, and Shell runs commands in the environment you choose. Use only machines you are allowed to administer, and test on a copy first.

No. The agent runs the commands and shows real output, but a script can still be wrong, and a model can misread the problem. Read every diff, run the script with test values and check the exit code yourself. Pageferry puts the facts into a card; the decision to accept a change stays with you, and a rejected edit changes nothing.

Ask first is a mode in which the agent can act but pauses before every action that changes something, such as writing a file, and for an edit you see the diff and accept or reject it. Plan only writes a plan and runs nothing: no commands and no file changes. You switch the mode to carry the plan out. Auto-run acts at once.

Yes, the task tool starts sub-agents in parallel, and inside a Bash script you can start concurrent jobs in the background with & and use wait. Exit code aggregation is then your job: wait for each process id and collect its status yourself, because $? holds only the status of the last command. Test such a script on your own machine.

Yes, with a cloud model the prompt, including the text of files the agent reads, goes to that provider. With a local model nothing leaves the machine. In both cases chats, working folders and API keys stay on the PC, and secrets in tool output are redacted before they reach the model. You choose the endpoint, so choose it with your script in mind.

No, the installer is not code-signed yet, so Windows SmartScreen may warn when you run it. Compare the SHA-256 checksum of the downloaded file with the published manifest, and only then choose Run anyway under More info. In Windows, certutil -hashfile with the path of the file and SHA256 prints the value to compare.

34 MB is the size of the setup program that the download button gives. A portable single .exe of about 25 MB also exists, but the button does not give it. Both are built for Windows 10 and 11, and the current version is 2.9.26. After installing, the app shows an Update button when a new version is ready.

No account is needed to run the app. Whether you need a network depends on the model: a cloud AI provider needs it, a local model does not, and web search, update checks and the control of chats from a phone browser are optional network features. The phone control uses one outbound connection and can be switched off.