Mistral rebuilds internal CLI for coding agents

Mistral published an engineering post describing Spaces, an internal platform CLI its solutions team uses to scaffold projects, spin up dev environments, generate configs and deploy to staging. The post frames the tool's redesign around a second class of user: coding agents.

The trigger was concrete. Mistral built a TUI module picker for spaces init; when an agent tried to use it, it saw raw ANSI escape codes (\x1b[36m?\x1b[0m Select components), could not send arrow keys, and was locked out of the command. The fix generalized into a rule the post calls "every prompt is a flag in disguise" — decide what information a command needs first, then expose it through an interactive prompt, a flag, or a config file, all feeding one execution path. A -y flag is treated as a contract that every prompt resolves to a flag value or default, failing loudly rather than hanging on stdin.

Two other patterns: a plugin registry where each component declares its properties, rendering as a TUI picker for humans and JSON for agents; and two files generated on every init — context.json, a structured snapshot of modules, ports, commands and env vars, and AGENTS.md, imperative rules written for LLMs.

Mistral's one cited result is anecdotal: an agent wired a fresh repo into Spaces, generated config.yaml, connected Dockerfile and registry settings, set up GitHub Actions, and deployed via Koyeb in under 10 minutes.

Spaces is described as internal. The post announces no release, model or benchmark.