Regente 0.2.0: orchestration, enforcement, and zero-config lanes.

The auto-handoff queue

A lock tells the second agent no and leaves it stuck. Regente queues it instead. The moment a second agent reaches for a claimed file, Regente blocks it pre-merge, queues it, and the instant the holder releases, hands the file to the next agent in line. Automatically, with a note, no human merging. The whole arc reads claim, queue, release, owned, and nobody waits idle.

Zero-config lanes

Territory without setup. Give an agent a role and it owns a set of path globs, or let Regente infer a lane from what each agent is already touching, so ownership is visible from the first edit. Presets cover frontend, backend, debugger, tests, reviewer, docs, and infra, or you pass your own globs.

regente lanes                       # who owns what, inferred or assigned
regente role --agent codex --role backend     # server/**, src/api/**, …

Symbol-level claims and semantic coupling

Finer than a whole file. An agent can claim a single function or class, so two agents can edit two functions of the same file without colliding. And Regente warns when two agents hold different files that import each other, catching the coupling a file-level view misses.

regente symbols api.js                          # claimable functions + line ranges
regente claim api.js --symbol login --agent alice
regente claim api.js --symbol logout --agent bob   # same file, separate symbol claims

More on this in Concepts.

Enforcement: off, warn, strict

Claims are cooperative by default. Flip enforcement on and a pre-edit hook blocks editing a file or function another agent holds, and a git pre-commit guard blocks committing someone else's files. Each agent carries a stable identity, so the guardrail knows who is who. It is a guardrail for a cooperating fleet, not a security boundary.

regente enforce strict      # off | warn | strict

The full model is in Enforcement & identity.

Orchestration: plan, assign, drive

New in 0.2.0, and deliberately assist-only. A goal becomes lane-aligned tasks, and one command walks the whole flow with an accept, edit, or reject gate at each step, like a plan-mode approval. Nothing is written until you accept.

regente go "Add a /health endpoint and a status badge"

Tasks carry dependencies and reserve their files all at once, all-or-nothing, so multi-file work cannot deadlock. The planner proposes, you approve and assign, and your agents do the work in their own sessions. It stays join, don't spawn: Regente never launches your agents or writes code for them. Planning is the only step that calls a model, and it uses your own key. See Orchestration.

The reasoning layer

Coordinate the why, not just the who. Share the decision behind a change so the rest of the fleet inherits the context, and see semantic conflicts surfaced right next to the file claims. The point is that the next agent picks up not only the file, but the reason it changed.

A hardened engine and the desktop command center

The coordination engine got a large reliability pass in this release: tighter state locking, enforcement and identity scoping, and dozens of brittleness fixes found by adversarial review. And the macOS desktop app, on Apple Silicon and Intel, shares one engine with the CLI, so humans, agents, and the dashboard all read the same locked source of truth: the live fleet, the collision band, the auto-handoff queue, and the lanes board to main.

Get 0.2.0

Grab the macOS app, or install the CLI with one line. Local-first, no account, free while in beta.