ensemblrlisted

# Working inside Ensemblr You are running inside **Ensemblr**, a macOS desktop coding-workspace app. It opens git repositories, cuts an isolated workspace per stream of work, runs coding agents against them, and reviews what they changed. You are one of those agents. Your session prompt already carries the **playbook**: the tool inventory, the etiquette, and the bookkeeping expected of you each turn. That playbook is the live contract — where it and this skill disagree, the playbook wins. This skill is the reference underneath it: the vocabulary, the file formats, and the behaviour a tool answer is describing. ## Where you are running Ensemblr drives two **agent runtimes** natively, in a chat tab with a structured timeline: **Pi** and **Claude Code**. It also launches third-party **harnesses** — Claude Code, OpenAI Codex, Mistral Vibe — as their own TUI inside a workspace terminal tab. Claude Code is both, and the two paths are deliberately different: a harness tab always runs skip-permissions and holds no chat-tab tools, while a native chat tab honours the workspace permission mode and holds the full set. Harnesses are an **opt-in Experimental feature**, absent unless the user has switched them on — so if nothing in your tool list launches one, that is the switch being off rather than a tool to hunt for. Either way you reach the app through tools named `ensemblr_*`, served over an MCP endpoint. **A harness may re-expose them under its own naming scheme** — an extra `ens