AgentOS v0.3.0 — Developer Preview 3

AgentOS v0.3.0 — Developer Preview 3

This page is the release notes for v0.3.0, and the source text for the GitHub
Release body of the same tag.

AgentOS is a local, single-operator control plane for handing scoped software
tasks to coding CLIs already installed on your own machine, and for keeping what
they did observable and durable: tasks, agents, repository and file grants,
isolated run records, provider event streams, human questions, review gates and
git delivery, in one workflow.

What this release is for. The same envelope as v0.1.0 and v0.2.0:
evaluating AgentOS on a machine you own, against repositories you are willing to
have an agent write to. It is not a production install.

What changed since v0.2.0

The headline is delivery. A chain that finishes its work now gets itself
merged, and the steps between review and merge no longer need an operator:

• An autonomous merge tail. The repair loop reads review severity, recovers
  parked work, hands a repair task the chain's context and a second attempt, and
  binds the gate proof to the regression head it was produced against.
• Serialized exact-head delivery. Merge runs against the exact head it was
  proved on, recovers pre-merge base drift by itself, and holds the merge lease
  only across the merge — the diagnosis and the narrowing are in
  docs/adr/0001-merge-lease-hold-window.md.
  Merge credentials are minted through a private GitHub App, and the self-hosted
  merge executor has a public provisioning path.
• Chains execute in layers. The schema expands with a legacy backfill and
  contracts onto the layered shape; the API exposes layers and progress, and the
  console renders them with blocked-on markers.
• Review isolation. Blind review and adjudication are separate roles with
  their own authority guard and physical isolation, and the review base is
  pinned in the schema rather than inferred.
• Canonical agents re-pinned. Chain prompts are aligned with the upstream
  mattpocock-skills baseline, implementation runs on native Luna subagents, and
  model routing is re-pinned across the fleet. A six-step direct engineer
  workflow template ships alongside the full-assurance chain.
• Operations. Quiet-window production upgrades, a revived offshore gate
  worker with slot dispatch and failover, and a docs-only gate profile.
• The repository command-line interface is removed. v0.1.0 and v0.2.0
  shipped a help-only interface with no operational command families behind it.
  This release retires it rather than carrying an interface that does nothing.
  Nothing in the documented workflow used it.

The full list, by area, is in CHANGELOG.md.

Release identity

Version	0.3.0
Tag	v0.3.0 (annotated)
Prerelease	Yes
Release commit	The commit the v0.3.0 tag points to, stated on the GitHub Release page.
GOAL5A0_MASTER_SHA	8d69ee8544196a3310b3d63caf8ce5ec9a0e023b
GOAL5A0_CONTROL_PLANE_A_SHA	29f8dd354cb99d671c2e2e4e9e23716fd8004f3d

Install and upgrade

There is no upgrade path between developer previews. This release adds ten
migrations, including an expand/backfill/contract sequence over chain layers, so
install fresh, following the quickstart in the repository
README.md — it checks out v0.3.0 and exports this
release's GOAL5A0_* values above. The host warnings and verification steps in
the v0.1.0 release notes apply unchanged to this
preview.

Existing installations that track main can bring their database prompts up to
date with npm run db:sync-canonical-prompts -w @agentos/db after pulling.

Support, security and licensing

The support matrix remains the current statement of what is
supported and on what evidence. This release changes no platform or provider
classification: macOS on Apple Silicon is still the only target, and the
merge-executor and gate-worker profiles are documented rather than verified. The
security, license and assets and
migration and recovery documents remain current.