Add Codex CLI plugin support

[wayyoungboy]

Summary

• add a Codex plugin manifest for memory-powermem while keeping Claude Code behavior intact
• rename the shared plugin root to apps/agent-plugin so Claude Code and Codex CLI use a neutral directory
• make installed-plugin init/status/stop/reset skills work for Claude Code and Codex CLI by resolving a shared plugin root
• update Codex setup docs to use codex plugin plus explicit codex mcp add, and refresh marketplace descriptions
• merge the latest main, preserving the no-LLM hook regression, noop provider tests, and package targets under apps/agent-plugin

Closes #1027.

Validation

• Codex plugin manifest validation for apps/agent-plugin
• JSON validation for marketplace, plugin manifest, MCP, and hook config files
• go test ./... from apps/agent-plugin
• python tests/regression/test_claude_hook_no_llm.py using the project virtual environment
• POWERMEM_HOOK_BIN=apps/agent-plugin/hooks/bin/powermem-hook-linux-amd64 python tests/regression/test_claude_hook_no_llm.py using the project virtual environment
• make test-claude-hook-docker
• python -m pytest tests/unit/test_noop_embedding.py tests/unit/test_noop_llm.py tests/integration/test_noop_embedding_mode.py tests/integration/test_noop_llm_mode.py -q using the project virtual environment
• python -m py_compile tests/regression/test_claude_hook_no_llm.py using the project virtual environment
• sh -n for apps/agent-plugin shell launchers and scripts
• make -n package-agent-plugin, make -n package-claude-plugin, and make -n test-claude-hook-docker
• git diff --check
• two independent read-only reviews of the merge-resolution diff, no blocking findings

[wayyoungboy]

Merged the latest main into this branch to clear the stale conflict state.

Validation:

• Using the project virtual environment, ran python -m pytest tests/unit/test_claude_plugin_uv_install.py tests/unit/test_noop_llm.py tests/integration/test_noop_llm_mode.py.
• Ran cd apps/agent-plugin && go test ./....
• Ran bash apps/agent-plugin/scripts/package-plugin.sh and kept generated package artifacts out of the PR.

[wayyoungboy]

Sweep update: this branch is dirty again after the recent Claude hook/no-LLM regression changes on main.

I checked the merge shape. This is not a small textual conflict: the PR renames apps/claude-code-plugin to apps/agent-plugin, while main added hook regression, Docker, and Makefile changes under the original Claude plugin path. Resolving it safely needs a larger respin that maps those new hook changes into the shared agent-plugin layout, rather than a quick merge commit.

I am leaving this PR open for that dedicated rebase/respin instead of pushing a risky conflict resolution in this sweep.

[wayyoungboy]

Updated this PR with the latest main and resolved the merge conflict.

What changed in the merge resolution:

• Preserved package-agent-plugin and the compatibility package-claude-plugin alias.
• Preserved the upstream test-claude-hook-docker target.
• Moved the new hook regression workflow, Dockerfile, and regression test paths from the old Claude-specific plugin directory to apps/agent-plugin.
• Moved the new uv-init-flow.svg asset into apps/agent-plugin.
• Rebuilt packaged hook binaries after merging the latest hook changes.

Validation:

• go test ./... from apps/agent-plugin
• python tests/regression/test_claude_hook_no_llm.py using the project virtual environment
• POWERMEM_HOOK_BIN=apps/agent-plugin/hooks/bin/powermem-hook-linux-amd64 python tests/regression/test_claude_hook_no_llm.py using the project virtual environment
• make test-claude-hook-docker
• python -m pytest tests/unit/test_noop_embedding.py tests/unit/test_noop_llm.py tests/integration/test_noop_embedding_mode.py tests/integration/test_noop_llm_mode.py -q using the project virtual environment
• JSON validation for plugin and hook config files
• shell syntax checks for agent plugin scripts
• make -n package-agent-plugin, make -n package-claude-plugin, and make -n test-claude-hook-docker
• git diff --check
• two independent read-only reviews of the merge-resolution diff, no blocking findings

[lightzt99]

Heads up on the interaction between this PR and #1031 (SQLite default for coding agent).

#1031 modifies files exclusively under apps/claude-code-plugin/ — scripts/init.sh, README.md, SETUP.md, skills/init/SKILL.md, tests/test_env_generation.py, plus tests/unit/test_claude_plugin_uv_install.py and tests/unit/storage/test_sqlite_init.py at the repo root. Since #1025 renames that entire directory to apps/agent-plugin/, whichever PR merges second will need a non-trivial rebase:

• If Add Codex CLI plugin support #1025 merges first, feat: sqlite default to coding agent #1031's file paths all move. The static-text assertions in test_claude_plugin_uv_install.py (e.g. INIT_SH = ROOT / "apps" / "claude-code-plugin" / "scripts" / "init.sh") break until updated, and every doc/skill reference to apps/claude-code-plugin/... needs to become apps/agent-plugin/....
• If feat: sqlite default to coding agent #1031 merges first, your next rebase of Add Codex CLI plugin support #1025 needs to carry feat: sqlite default to coding agent #1031's init.sh changes through the rename: the db_provider-aware PACKAGE selection (sqlite → powermem[server,extras], oceanbase → powermem[server,seekdb]), the --with pysqlite3-binary uvx injection, the removal of HF_HUB_OFFLINE=1 writing, and the POWERMEM_INIT_PRELOAD_MODEL deprecation no-op.

The second path is lighter — #1031's changes are all internal to init.sh and its docs, so they should survive a directory rename cleanly. The first path requires touching test files that hard-code the old path.

Happy to rebase #1031 onto #1025 once #1025 lands if you'd prefer that order; just let me know which way you want to go.

[wayyoungboy]

Review status for the current head: GitHub currently reports this branch as conflicting with main, so I cannot complete final review or mark it ready. This PR still needs a rebase/respin before the diff can be reviewed safely, especially around the plugin directory rename and the newer Claude hook/no-LLM regression changes that landed on main.

No LGTM: the PR is currently conflicting and is authored from the wayyoungboy branch.