cli-agents · labs · ai-assisted programming

CLI agents lab — troubleshooting

Start with the setup check in this folder. It names what is missing:

python check_setup.py

Installing

copilot: command not found after installing

npm put it somewhere your shell is not looking. npm prefix -g prints npm's global folder, and its bin folder must be on your PATH. In a Codespace, opening a new terminal is usually enough.

npm complains about the Node version

Both agents need Node 22 or newer; node --version shows yours. The Codespace already has it. On your own machine, install a current LTS release of Node.

On your own Windows machine

The lab works in PowerShell too. npm install -g @github/copilot is the same, and winget install GitHub.Copilot is an alternative. PowerShell does not understand \ at the end of a line, so type multi-line commands on one line.

Signing in

Copilot says you have no access

• Check that a plan is active at https://github.com/settings/copilot — Copilot Free, or Copilot Student through GitHub Education.

• In a Codespace, Copilot starts signed in with the Codespace's own GITHUB_TOKEN. If that is not the account with the plan, type /login to sign in as the right one. If it still uses the wrong account, clear the token for that terminal only, start the agent, and type /login:

  unset GITHUB_TOKEN GH_TOKEN
  copilot
  

/login shows a code and a web address

That is the sign-in. Open the address, type the code, approve. There is nothing to paste into a file.

Gemini's key box says "Paste your API key here"

Gemini did not find your key, so it is asking for it. Quit it, then check that the file is sample-app/.env and its name starts with a dot, that it has a line GEMINI_API_KEY= followed by your key, and that you started gemini from sample-app. Start it again. Sign in with Google is not the way in: Google's own sign-in stopped serving free personal accounts on 18 June 2026. In a script (gemini -p) the same problem shows as Please set an Auth method and exit code 41.

Where does a token or key go?

Never in a file git tracks — the module's safety audit rejects one, and your repository may be public. Copilot's /login stores what it needs in its own configuration, outside the repo. Gemini's key goes in sample-app/.env, which .gitignore covers, and Gemini also keeps the key you confirm in a file under ~/.gemini/, outside the repo. Never type a key into a command or paste it into code.

In a session

The agent ignores my AGENTS.md

• Instruction files are read when a session starts. Quit and start a new one.
• Check where you created it: sample-app/AGENTS.md, beside the code, in the folder you start the agent from.
• Gemini reads GEMINI.md unless context.fileName says otherwise (DIY 3, step 2), and ignores a folder's settings until you trust the folder. /memory show prints what it loaded.

My command or skill does not appear

• Check the folder: .github/skills/<name>/SKILL.md at the root of your repository for Copilot, .gemini/commands/<name>.toml in the folder you start Gemini in.
• A skill's folder name and its name: must match.
• Copilot: /skills lists what it found. Gemini: /commands reload.

It keeps asking permission — or has stopped asking

An approval you give during a session can last for the rest of that session. In Copilot, /reset-allowed-tools clears them; starting a new session clears them in both agents.

Gemini never answers

Look at the bottom right of its screen. It should read gemini-3.5-flash-lite. If it reads Auto, Gemini did not read the GEMINI_MODEL=gemini-3.5-flash-lite line from sample-app/.env, and on a free key its default model gave no answer when this lab was checked in October 2026. Fix the file, quit, and start it again.

Gemini says it is not running in a trusted directory

That is a script run (gemini -p) in a folder you have not trusted yet. Start gemini once in sample-app and trust the folder, or add --skip-trust to the command.

It says I have run out of requests

The free plans have limits: a monthly allowance on the Copilot student plan, shared with Copilot Chat (/usage shows what the current session has used), and per-minute and daily limits on Gemini's free API key. Wait, or switch to the other agent.

The policy checker

policy error: ... has no wildcards

A rule in this checker names the words a command starts with, so shell(git) already covers every git command. Write shell(git), not shell(git:*).

No module named 'pytest'

Run pip install -r requirements.txt from this folder.

The checker and my agent disagree

That is DIY 7's finding, not a bug. The checker is one model of matching; each agent has its own, and it can change between versions.