Letting the agent work

Enable tools

Start the chat with a tool set and a permission mode:

codel00p ...scope agent chat \
  --provider anthropic \
  --model claude-opus-4-8 \
  --tool-set all \
  --permission-mode ask

--tool-set chooses what the agent may use: read, edit, command, git, or all. --permission-mode ask confirms before sensitive actions, which is the right default while you build trust; allow runs without prompting and deny blocks tools entirely. Run /tools at any time to see what is available this turn.

A real turn

Ask for something that needs the repository, not just the model:

you> Run the test suite and tell me what is failing.

The agent calls the command tool. In ask mode you confirm it first; the output streams back and it summarizes the result. Because the conversation persists, you can follow straight up:

you> Fix the first failure, then re-run just that test.

It reads the relevant files, makes an edit, runs the single test, and reports back, all inside the same session. Every step is a workspace-safe tool call confined to your project directory.

Switch models mid-session

You do not need the strongest model for every turn. /model shows the current one; /model <id> switches it for later turns, so you can drop to a faster model for routine edits and come back up for tricky reasoning.

you> /model claude-haiku-4-5
Model set to claude-haiku-4-5.