Claude Code

Claude Code is the default ACA when it is installed, and the one wdx reads the most from: besides its state, model and context, the question it asks and its options. Its id is claude_code.

Install

wdx runs the claude program your login shell finds. Install it as Anthropic’s docs say, and log in once in a terminal of your own. wdx was last checked on 2.1.323.

wdx agent create --aca claude_code --model opus

What wdx gives it

• Models: fable, opus, sonnet and haiku, Claude Code’s aliases. Claude Code has no command that lists its models, so this list is wdx’s; any other name is passed as given.
• Permissions: full is --permission-mode bypassPermissions, with the sandbox off and its warning about that mode skipped; ask passes nothing, leaving your own settings.
• The launch note: wdx’s instructions to work through wdx, added to its system prompt with --append-system-prompt, at every start.
• A session of wdx’s choosing: a new agent starts with --session-id, and a start later resumes it with --resume, in the directory it last worked in.
• Its own screen settings: wdx turns off its alternate screen and its mouse, so that TIDE and agent attach keep its scrollback and the mouse.

What wdx reads

• Its state, from the input box and footer: working, idle, or waiting, and a draft in the box. With a draft in the box unreleased, the line above the box tells whether a turn runs, so an agent at rest no longer flashes working.
• The question it asks, only Claude Code’s so far: yes/no for a permission or trust question, decision for a choice, with the question, its detail and its options, in wdx agent get and TIDE’s info tab. And unreleased what each option takes, a choice, your own text, Chat about this or a checkbox list’s Submit, the questions of a decision that asks several, and its review screen, so that TIDE’s prompts tab and wdx prompt answer them without attaching.
• Its model, from its banner and /model, and from the status line’s file unreleased, so a resumed agent, whose screen shows no banner, shows its model too, and is marked limited when its own model’s limit runs out.
• How full its context is, from a status line wdx gives it, which writes a small file in the agent’s backup. Your own status line still runs, after wdx’s.
• Its conversation, copied from its transcripts as they grow, with /clear followed through a hook, so wdx agent transcript and --brief have it all, and a transcript Claude Code deleted is put back.

Messages

• A message is pasted, then Enter as a key of its own. During a turn, Claude Code takes it in at its next tool result, in the same turn.
• An instant message is sent at once with Ctrl+X Ctrl+S: a running command moves to the background rather than stopping, and an answer being written is cut there.
• A leading ! is typed as \!, as Claude Code would run it as a shell command without asking.

What differs

It is the only ACA whose questions wdx reads, so its rows say yes/no and decision where the others say waiting, and the only one whose questions the prompts tab and wdx prompt answer unreleased: wdx presses the keys in its terminal one at a time, checking each on its screen, and stops if the question changes.