Locus 2.1.0
Locus 2.1.0 rebuilds the app's core on a composed architecture and ships everything staged since 2.0.0: Grill mode, the Notebook, the Sepolia Wallet Hub alpha, Codex-native ChatGPT chats, and a long list of transcript, runs, and selection fixes.
Every feature now has its own model under the hood
The application core that used to hold nearly all interface state in one 16,600-line object has been decomposed: provider accounts, agent teams, run history, the live team-run card, extensions and MCP, workspace knowledge, evaluations, schedules, the activity center, transcript search, background services, the AGENTS.md editor, the landing flow, and toasts each live in their own tested, observable model. Behavior is deliberately unchanged — what you gain today is faster, safer iteration on every one of those features, and it clears the runway for making each panel redraw only when its own feature changes. The full story is in Docs/Architecture.md.
Grill mode replaces GSD
The fourth composer mode is now Grill (⌥G, /grill): a relentless one-question-at-a-time interview that stress-tests a plan before anything gets built. Approving a plan implements it in Work mode. Everything saved under the old mode keeps working — stored workspaces and older mobile clients that still say "build" land on Work.
ChatGPT plans answer the way Codex does
A per-account Codex-native mode runs Work, Plan, and Build turns under the model's own Codex prompt and tools while every call still passes through Locus's permission prompts and deny lists. You also get a reasoning-effort picker and an optional web-search toggle per account.
The Notebook, and a Wallet Hub alpha
The sidebar gear menu, the Locus menu (⇧⌘9), and the command palette open a searchable list of every note Locus has kept. Direct-download builds can also enable the isolated, Sepolia-only Locus Vault as an in-app Wallet Hub for the private alpha — off by default, gated, and ignored by Mac App Store builds.
Fixed
The Add Account menu no longer renders lettermark providers as invisible white-on-white rows. A run's Overview no longer borrows numbers from a run that is still going. Selecting text in the transcript finally works like text — one selection spans the whole conversation and survives scrolling. Switching models no longer reports a model you have as "not installed". Files the agent produces open properly and appear in Outputs. See the changelog for the full list.
Requirements: Apple Silicon Mac, macOS 14 or newer.