v2.20.0

Added

• Agent Plugins support — the repository is now a conformant Agent Plugins 1.0.0 package, so its skills load in any client implementing the standard rather than only in Claude Code. A root plugin.json manifest declares the plugin against the canonical 1.0.0 schema, and scripts/sync_skills.py mirrors it into .claude/ and scientific_writer/.claude/ so each bundled payload is a loadable plugin root of its own — a pip install scientific-writer is enough to point another client at the skills. .claude-plugin/marketplace.json is unchanged and still drives Claude Code. See docs/AGENT_PLUGINS.md.
• scripts/validate_agent_plugin.py — validates any plugin root against the specification: the closed plugin.json and mcp.json schemas, Agent Skills frontmatter for every discovered skill, plugin-root path containment, MCP transport and command rules, and ${PLUGIN_ROOT} / ${PLUGIN_DATA} placeholder expansion. It is dependency-free and offline — the canonical schemas are vendored under scripts/schemas/agent-plugins/1.0.0/, as §5.2 requires of clients — and supporting a future spec version means dropping in a new schema directory. CI runs it on every push, and it is the first thing that reports the pre-existing skills/document-skills bundle as undiscoverable (§7.1 forbids recursing past immediate children of skills/, so docx, pdf, pptx, and xlsx are invisible to every client, Claude Code included; the fix belongs upstream in the hash-locked skills repository).
• AGENTS.md alongside CLAUDE.md — the repository root now carries AGENTS.md as a generated, byte-identical mirror of CLAUDE.md, enforced by scripts/sync_skills.py --check so the two cannot drift. /claude-scientific-writer:scientific-writer-init writes both files into a user's project from templates/CLAUDE.scientific-writer.md and the new templates/AGENTS.scientific-writer.md.
• .agents/ alongside .claude/ — load_system_instructions() now searches .claude/ then .agents/ for WRITER.md, AGENTS.md, and CLAUDE.md, then falls back to root AGENTS.md and CLAUDE.md, so a project that already documents its agents in either layout is honored. setup_claude_skills() still installs into .claude/ (the Claude Agent SDK discovers skills there) and now also refreshes .agents/ when the project already has one, and it accepts a bundled payload shipped under either name.
• GitHub releases are created by the release workflow — release.yml now creates the GitHub release from the CHANGELOG.md section for the tag, after PyPI publishing succeeds. Previously the workflow only published to PyPI and release entries were written by hand, so the releases page drifted (v2.19.0 shipped to PyPI while the page still showed v2.18.0 as latest, and several earlier versions have no entry at all). The step is idempotent: re-running a tag refreshes the notes instead of failing.
• scripts/changelog_notes.py — extracts the release body for one version from CHANGELOG.md and appends a compare link to the previous release. Run it to preview what a tag will publish: uv run scripts/changelog_notes.py X.Y.Z.

Changed

• README section for non-Claude agents rewritten — "Use with Gemini CLI and Other Agents" is now "Use with Antigravity, Pi, and Other Agents", leads with the portable Agent Plugins path, and treats hand-loading a skill as the fallback. The worked example no longer writes to a vendor-specific config path; it appends the skill body to the project's AGENTS.md. It also strips YAML frontmatter correctly — the old tail -n +6 leaked metadata:, version, skill-author, and the closing --- into the prompt, because the frontmatter in skills/scientific-writing/SKILL.md closes at line 9.
• A changelog entry is now required to release — the workflow extracts the notes before running any checks, so a tag whose version has no ## [X.Y.Z] section (or an empty one) fails immediately, rather than after the package is already on PyPI. docs/RELEASING.md documents this, along with why the local scripts/publish.py path does not produce a GitHub release.

Full Changelog: v2.19.0...v2.20.0