feat!: emit Agent Plugins packages from codex and add an agent-plugins target

[steve-calvert-glean]

Second of three PRs adding Agent Plugins support. Stacked on #37 (the MCP module); review that first. This PR's base is feat/mcp-dialects.

Why

Agent Plugins 1.0 standardizes the plugin package: a closed root plugin.json, skills/, mcp.json, and reverse-domain extension namespaces. It deliberately leaves out marketplaces, hooks, agents, commands, and rules. OpenAI's packaging docs now prefer it for Codex and ChatGPT and keep .codex-plugin/plugin.json only as a compatibility fallback. Copilot/VS Code loads it too, and Cursor loads the skills and MCP parts.

What

• src/agent-plugins.ts, the package module. Builds the portable manifest, moves client fields under extensions[<namespace>], and validates a package against the spec: closed manifest, plugin name (§5.5), mcp.json (via feat: render MCP config into each target's MCP dialect #37's agent-plugins dialect), and each skill's name per the Agent Skills spec, which must also match its directory. Marketplaces stay on each target, as the spec intends.
• codex emits packages by default. It writes root plugin.json with OpenAI settings under extensions.com.openai, plus mcp.json, and keeps .agents/plugins/marketplace.json. Existing configs need no changes: authored interface/apps/hooks move under the namespace automatically. format: "legacy" keeps the .codex-plugin layout. validate detects each plugin's layout on disk, the way Codex does.
• New agent-plugins target. Portable packages only, under plugins/<name>/, with no marketplace (the spec defines none) and no client profile. Including agents/commands/rules/hooks is a config error, and a non-portable manifest field is a build error. install-info reports no install command, with a citation.
• Target interface. Two optional members: forConfig (picks a definition per target config) and finalizeManifest (runs after the author's manifest override is merged).
• Conformance. Vendors the official 1.0.0 plugin.schema.json and mcp.schema.json (commit SHA and SHA-256 in tests/fixtures/agent-plugins/SOURCE.md). Codex and agent-plugins output must validate against them via ajv's 2020-12 build, and pluginpack validate must pass on the same output.
• Docs. README (targets table, an Agent Plugins section with upgrade notes, MCP table, config reference), CONFORMANCE.md, CLAUDE.md, and the authoring skill. ADR 0001 records why there's no dual layout.

Breaking change

The codex target's default output changes from .codex-plugin/plugin.json + .mcp.json to plugin.json + mcp.json. To keep the old layout, set format: "legacy" on the codex target.

Verification

• npm run check passes (249 tests; 9 new in tests/agent-plugins.test.ts, plus new conformance tests for codex AP, codex legacy, and agent-plugins).
• Built gleanwork/agent-plugins from a temp copy:
  • claude and cursor output is byte-identical to 0.11.0.
  • codex changes only in layout: .codex-plugin/ and .mcp.json are replaced by plugin.json and mcp.json. interface lands under extensions.com.openai, and the shared Claude-style MCP config converts to valid Agent Plugins mcp.json.
  • Two upgrade blockers surface on purpose, both handled in the follow-up (c): the hand-written overrides/codex/glean/mcp/config.json (cwd: ".") fails the build, and once it's removed, validate rejects the glean_run skill name, which Agent Plugins clients would skip.

Next

(c) in gleanwork/agent-plugins: delete the codex MCP override, decide on the glean_run rename, upgrade to 0.12, and regenerate codex-plugins.