v0.4.0 — Omarchy branding and compatibility

Build Omarchy Plugins v0.4.0

This release brings consistent Omarchy branding and compatibility guidance to
the twelve-skill provider-neutral bundle. It also includes the marketplace preflight and workflow
fixes prepared for the unpublished v0.3.2 draft.

What's changed

• Commit-pinned identity/category and compatibility badges, with the official
  Omarchy icon, plus a compact README row ordered CI, licence, identity and
  compatibility. The bundle's own README now uses the layout.
• Omarchy | 4.0.0+ means the installed Omarchy version reported by
  omarchy-version. Projects must declare and test their own supported range;
  the badge is not a marketplace approval or universal compatibility claim.
• Badge categories pair with consistent GitHub topics to help people find
  Omarchy projects. Canonical guidance and the OpenAI adapter stay synchronized.
• Privilege advisories catch sudo and pkexec in negated README prose as well
  as commands. Review guidance separates documentation matches from execution,
  compatibility checks, security review, approval and publication.
• The marketplace security policy is pinned, and workflow handoff cases cover
  completing agreed builds, candidate drift and marketplace feedback.

Verification and limits

Portable validation covers 55 unit tests, generated-plugin tooling, synchronized
adapters and deterministic archive builds. Required CI includes nine
OS/Python combinations across Linux, macOS, and Windows, workflow lint and
strict Claude packaging validation. Candidate results are recorded in the
acceptance record.

The privilege check is deliberately broader than the upstream matcher. A warning
is neither proof of execution nor a security finding. Passing portable checks
is not marketplace approval or a security audit.

Fresh Astra/Fable behavioral runs, the documented handoff evaluations and live
Omarchy desktop acceptance remain outstanding. This bookkeeping pass does not
claim a fresh-install or v0.3.1-to-v0.4.0 host upgrade smoke test.

The release workflow passed on merged commit a12e90568b7e8f28ca7ca92400009758da7eb43b: all 55 tests passed, two builds matched byte-for-byte, provenance was generated, and uploaded assets were downloaded and verified. Required CI passed across all nine OS/Python combinations, with workflow lint and strict Claude packaging validation.

Installation

The transactional installation and updates remain supported for Codex, Claude Code,
Cursor, Gemini CLI and OpenCode.

Choose the skills, portable Agent Plugin, OpenAI adapter or Claude plugin archive
for your host; verify it against SHA256SUMS. Use either the Claude plugin or
the Claude skills installer to avoid duplicate skill loading. The submission
archive contains reviewer materials. Source manifests and the SPDX 2.3 SBOM bind
the archives to the release source tree.

Existing installations can preview and apply the v0.4.0 update
from its checkout with install_agent_skills.py --update --diff and then
--update, preserving their existing target and scope. Locally modified managed
files require explicit resolution. No official OpenAI or Anthropic marketplace
listing is implied.

v0.3.1

Build Omarchy Plugins v0.3.1

This patch release turns recurring Omarchy marketplace review feedback into
targeted authoring guidance, and adds native Claude Code plugin packaging.
The same twelve canonical skills remain provider-neutral across Codex, Claude
Code, Cursor, Gemini CLI and OpenCode. Existing transactional installation and
update safeguards are retained.

What's changed

• Native Claude plugin and community-marketplace manifests, a dedicated Claude
  archive, and strict official Claude CLI validation in required CI.
• Guidance for limiting data before collection, rendering external QML text as
  plain text, preserving file and executable identity, cleaning up child
  processes, and keeping credentials out of arguments and logs.
• Advisory discovery checks for QML inputs and agent-control payloads, plus a
  reviewer-response table connecting findings to fixes, reproductions and the
  exact submitted commit. Static checks do not certify plugin security.
• Updated Omacom submission forms, including Kids/Education and current tag
  labels, with compatibility for existing CLI spellings.
• Installation and stewardship documentation, aligned version metadata, and a
  release workflow that can create a new annotated tag after verifying the
  expected main commit and its required CI result.

Research and verification

The marketplace findings report
draws on 6,531 issue/PR records and 30,036 comments, including 6,286 comments by
HANCORE-linux across 3,589 threads. Every collected record was indexed and
representative reviews were read in depth. This was not an independent source
audit of every plugin; overlapping topic counts include resolved findings.

Portable validation covers 54 unit tests, generated-plugin tooling, synchronized
adapters and deterministic archive builds. Required CI includes nine
OS/Python combinations across Linux, macOS, and Windows, workflow lint and
strict Claude packaging validation.
Release automation builds twice, checks byte equality, generates provenance and
verifies downloaded draft assets before owner publication. The workflow cannot publish
a release; publication remains owner-controlled.

Fresh Astra/Fable behavioral sessions and live Omarchy desktop acceptance remain
outstanding. Packaging validation does not establish model performance or live
desktop compatibility. See the
acceptance record.

Installation

Choose the skills, portable Agent Plugin, OpenAI adapter or Claude plugin archive
for your host; verify it against SHA256SUMS. Use either the Claude plugin or
the Claude skills installer to avoid duplicate skill loading. The submission
archive contains reviewer materials. Source manifests and the SPDX 2.3 SBOM bind
the archives to the release source tree.

Existing installations can preview and apply the update from a v0.3.1 checkout
with install_agent_skills.py --update --diff and then --update, preserving
their existing target and scope. Locally modified managed files require explicit
resolution. No official OpenAI or Anthropic marketplace listing is implied.

v0.3.0 — Skills for Astra and Fable

Build Omarchy Plugins v0.3.0

Twelve Omarchy skills, revised for Astra and Claude Fable: clearer instructions
for capable models, with the desktop contracts they still need to get right.

Sometimes the agent stopping halfway through is following exactly what the
instructions accidentally asked for. This release takes a closer look at those
instructions.

• Requested builds continue past the design note. Existing approval and settled
  decisions carry forward, while diagnosis-only requests stay read-only.
• Small fixes get focused edits and relevant checks. Required repository and
  release checks still apply; missing desktop access is reported honestly.
• Bundled helpers resolve from the loaded skill directory. Workbench
  registration is optional, and an example no longer dictates repository layout.
• Host-conformance text output no longer crashes. Codex discovery now checks
  parent directories and reports duplicate skill names without inventing a winner.
• Ten behavioral cases cover builds, repairs, scope, continuity and evidence.

The bundle stays provider-neutral. Use the same skills with Codex, Claude Code,
Cursor, Gemini CLI or OpenCode; choose the model in your agent. Existing
transactional updates, reproducible archives, checksums and SPDX 2.3 manifests
are retained. Release automation prepares a verified draft and cannot publish it.

All 52 tests pass, with CI across Linux, macOS, and Windows. The Astra and Fable
reviews use their providers' official guidance. Actual Fable model execution and
live Omarchy desktop testing remain outstanding; this is not a claim of measured
model performance gains.

If a skill still makes your agent stop unnecessarily, overbuild a small fix or
reopen a settled decision, please open an issue with the prompt and what happened.

v0.2.3

Release notes — 0.2.3

Hardens the provider-neutral toolkit and adds first-class OpenCode support. The
same twelve skills install into shared Agent Skills paths or native Codex,
Cursor, Gemini CLI, Claude Code, and OpenCode locations without calling a model
API.

This release adds transactional managed updates and removal, read-only host
discovery diagnostics, a deny-by-default OpenCode conformance probe,
deterministic exact-tree archives, source and release manifests, an SPDX 2.3
SBOM, SHA-256 coverage, stricter scaffold and QML trust-boundary checks, and
tag/asset provenance verification. CI now covers Linux, macOS, and Windows;
release automation can prepare a verified draft but cannot publish it. The
release path now validates every workflow with pinned actionlint, inspects
unpublished drafts through the supported GitHub CLI surface, and retries initial
asset downloads before comparing the uploaded release byte-for-byte.

Live Omarchy and model-provider compatibility remains explicitly evidence
bounded. Static validation and filesystem discovery are not described as a
security audit or a successful host invocation.