[architdhamija]

What shipped

mex setup and mex skills sync now install mex-inbox, mex-relay, and mex-statechart for Cursor, GitHub Copilot, OpenCode, and Windsurf, with the same commands and approval rules as Claude Code and Codex.

There are no per-tool forks. All four tools' docs list .agents/skills/, so they share the copy Codex already uses, generated from the same packaged skills/ source.

Tool	--tool	Instruction file	Invocation
Cursor	cursor	.cursorrules	/mex-inbox
GitHub Copilot	copilot	.github/copilot-instructions.md	/mex-inbox
OpenCode	opencode	AGENTS.md	By name, through its skill tool
Windsurf	windsurf	.windsurfrules	/mex-inbox

Changes beyond the client list:

• Installer: it now plans a shared path once. Without this, a second tool using .agents/skills or AGENTS.md would see the first tool's write as a concurrent change and fail the apply.
• Hub setup commit: the review scope now includes each tool's instruction file, so OpenCode's AGENTS.md isn't left out.
• Commit checkpoint commands: each shared path is staged once.
• mex-statechart: a new delivery step for clients that may not render Mermaid: the text view plus the image paths.

Install

mex skills sync --dry-run --tool cursor
mex skills sync --tool cursor

Use copilot, opencode, or windsurf for the others, then start a new agent session. Per-tool steps and smoke checklists are in docs/agent-skills.md.

Upgrade behavior: after upgrading, a plain mex skills sync also syncs any of these four tools recorded in .mex/config.json. Previously they were skipped. Bytes outside the managed block are preserved, and unmanaged .agents/skills/mex-* directories are reported as conflicts rather than overwritten.

Still Claude Code or Codex only

• .claude-plugin/marketplace.json lists Inbox and Relay for Claude Code only.
• agents/openai.yaml is installed everywhere, but only Codex reads it.
• Inline statechart delivery is only verified on Claude Code's desktop app (PNG) and Codex CLI (Mermaid).
• Single-session population and the finalize repair turn still run only on Claude Code or Codex, because they run mex commands.

Blockers and open points

• Live clients not tested: discovery paths and invocation come from each vendor's docs, checked on 2026-11-04. Nobody has run the smoke checklists in Cursor, Copilot, OpenCode, or Windsurf yet.
• Windsurf: its docs now redirect to Devin Desktop, which reads .agents/skills/ and treats .windsurf/skills/ as legacy. Older Windsurf builds may need a manual copy to .windsurf/skills/, which MEX doesn't manage.
• OpenCode: it shares AGENTS.md with Codex, so the block names skills as $mex-inbox. OpenCode has no $ syntax and should find the skill by name. This is the first thing to check in its smoke test.
• README hero line: it still says Claude Code or Codex, because it sits beside the pinned mex-agent@0.8.5 install command. Update it in the release that ships this.

The optional nested AGENTS.md per subsystem work isn't included.

Testing

• Typecheck: npm run typecheck passes.
• Related tests: 150 of 152 pass on Windows with --testTimeout=60000. At the default 15s limit, tests in commit.test.ts timed out on this machine.
• The 2 failures: "rolls instruction replacement back portably and preserves its mode" and "detects an instruction mode change after preview". Both are Claude-only file mode checks, and Windows reads chmod 0o640 back as 666 on this machine. I didn't run them on main for comparison.
• Real CLI: each of the four tools was installed with the built CLI in a throwaway Git repo. The dry run wrote nothing, a second sync changed nothing, and installing all five .agents/skills tools together gave 7 actions with no conflicts and left a hand-written .cursorrules line intact.
• Not run: the full vitest suite was stopped before finishing, and npm run build:hub wasn't rerun after these changes, so this PR relies on CI for both.

🤖 Generated with Claude Code