Part six of the Handmade Claude Code campaign: the agent gets a face. Everything so far ran headless — a prompt in, an answer out. A coding agent you would sit in is a conversation: you type, it streams, it shows you the tools it is running and asks before the dangerous ones, it tells you what the session has cost, and when it goes wrong you press a key and it stops.

This session continues the agent you built in parts one to five, in the same repository. The same command without -p opens the terminal UI in the working directory:

<command> -C <dir> [--yes]

The contract

The checks drive your TUI inside tmux (a 120×32 pane), typing keys and reading the screen — so tmux must be installed on your machine, and the screen must say what happened in plain text:

• An input line that begins with > (whitespace before it is fine). Enter submits; Enter on an empty line does nothing.
• The answer is shown as it arrives: streamed fragments appear on screen before the final message does.
• A tool call is shown with the tool's name and what it targets — for bash the command, for read, write, edit, glob and grep the path or pattern, for skill the skill's name, for task the agent's name and the delegated prompt — and, once it finishes, its output (or the head of it). A call that failed is marked as such (error or failed on its line). Several calls in one reply show as several lines.
• Skills work as they do headless: /<skill> <args> at the prompt sends the skill's body, and the typed line stays in the transcript.
• Subagents are visible: everything the child does — its tool calls, its answer — is shown marked with the agent's name, so the user can tell the child's activity from the parent's.
• Without --yes, a call the rules do not permit becomes a question on screen, showing the tool name and its input. y allows it, n denies it (the model gets the usual permission denied result).
• Slash commands: /help lists the commands, /clear forgets the conversation (the next request carries only the new prompt), /model <provider>/<model> switches to another provider from the settings, /quit exits (so does Ctrl-D on an empty line).
• A status line shows the session's total input and output tokens, as plain integers.
• Esc interrupts the running turn: a running tool is killed, no further model call is made for that turn, and the input line is back.

The scripted model is the same one the headless checks use, so what streams, what asks and what gets interrupted is entirely under the check's control — and yours.

The judges

The quality panel — architecture, performance, code quality, test quality, technical governance and DX review — sits on the rungs where its subject is decided, one verdict per judge per rung, on top of the rung's own points. There is no closing review: what you build is judged as you build it, and a rung you never reach is a verdict you never get.

The ladder

• Set up and carry parts one to five forward (10)
• It starts and it quits (20)
• Ask and answer: typed, streamed, billed (30)
• Tool calls on screen (40)
• The question (40)
• Slash commands (20)
• Skills on screen (30)
• Subagents on screen (40)
• Esc means stop (30)