Troubleshooting

Troubleshooting

Match the message you see to the check below. Fix that part and run the same command again.

pari-cli: command not found

The installer finished, and the shell has not picked up the new PATH entry.

  1. Open a new terminal.
  2. Confirm ~/.local/bin is on PATH for a curl install that used a virtualenv.
  3. Run pari-cli --version.

If you used pipx or uv, their own bin directory has to be on PATH. pipx ensurepath prints what to add.

Python 3.11 or newer is required

python3 --version

Install a current Python from python.org or your distribution, then run the installer again. On Windows, py -3 --version should report 3.11 or newer. The Windows installer skips the Microsoft Store stub under WindowsApps.

bubblewrap is not installed

On Linux, install the distro package and check the binary. PARI does not use a copy that lives inside the project.

sudo apt-get install bubblewrap
bwrap --version
pari-cli --dry-run

Sandbox support is coming soon

You are not on Linux, or you selected seatbelt, cloud, or none. PARI will not run that command on the host. Use Linux, or Ubuntu on WSL2, and leave the sandbox on auto.

No model selected

pari-cli login
pari-cli --dry-run

You can also export PARI_MODEL and the provider’s API key. Environment keys override the saved file. If the auth file is corrupt, the error names ~/.pari/auth.json. Move that file aside and sign in again.

Unknown provider or model

Provider ids are openai, deepseek, anthropic, grok, and openrouter. A model id has to match that provider’s prefix, such as claude- or gpt-. /model with a bad name leaves the current model in place.

File is outside the workspace

--add and /add only accept a file inside the directory PARI was started in. cd to the repository root, or pass --workspace with that root. A missing path is reported as file not found.

Dry run failed, or the session started and a command did not

--dry-run checks the sandbox before a turn. Starting pari-cli checks the model first, and checks Bubblewrap when a command actually runs. If the prompt opened and a later command failed, read that command’s error. The usual cause on Linux is a missing bwrap. The usual cause elsewhere is that the sandbox backend is not available yet.

Next

  • Install →

    Repeat the install and the dry run from a clean shell.

  • Models →

    Sign in again if the credential was removed.