setup

Set up the installed MemHub plugin's host integration. This skill configures the plugin install; `/memhub:onboard` separately connects the current repo to its agent brain. Resolve `ROOT` as the ancestor of this skill file containing `scripts/`. Codex sets `PLUGIN_ROOT`; Claude Code sets `CLAUDE_PLUGIN_ROOT`. Prefer those values, then derive it from the skill path if needed. Arguments: `$ARGUMENTS` ## Codex Codex hook support depends on the host version. Current desktop builds can load bundled plugin hooks; older CLI releases need the user-level compatibility bridge. This build, from the Claude plugin directory, ships no Codex manifest; Codex installs the `XTraceAI/agent-plugins` build, whose manifest points at `hooks/codex-hooks.json`. Claude compatibility handlers must skip Codex payloads. When the user bridge is installed, bundled Codex handlers defer to it for matching events; the user handlers still need review and trust. Do not install both production and staging plugins. A projectless task may receive shell hook payloads containing the session directory instead of the command's explicit `workdir`. Use a shell-quoted absolute `cd <repo> && ...` prefix for repository shell calls, or start the task in the repository. The plugin cannot reconstruct an omitted working directory. This workaround enables command rules; session-start rules still require starting in the repository. Install the bridge: ```bash python3 "$ROOT/scripts/setup_codex_hooks.py" install ``` The installer is idempotent. It merges four MemHub handlers into `$CODEX_HOME/hooks.json` (default `~/.codex/hooks.json`), preserves unrelated events and handlers, backs up a changed existing file, and installs a stable trampoline that follows plugin version upgrades. It enables: - `SessionStart`: rulebook posture rules and any plugin-upgrade notice, the repo brain brief, and a capture-health warning — the same three scripts Claude Code runs at session start; - `PreToolUse`: the rulebook hook (which can deny the call); - `PostToolUse`: the rulebook hook, artifact-link reminders, and PR-link recording after GitHub-touching shell calls (GitHub MCP calls only through the bundled plugin hooks — the bridge's matcher stays narrow so upgrading it needs no re-trust); - `PostToolUse` + `Stop`: incremental session capture. After installation, report installation and trust as separate states. Codex deliberately does not let a plugin approve command hooks, and the setup script cannot inspect or change that approval. Do not claim capture or team rules are active until the user confirms the review is done. Give these precise review instructions: 1. Restart Codex, then choose **Review hooks** at startup or open `/hooks`. 2. Review the single MemHub handler under each of `SessionStart`, `PreToolUse`, `PostToolUse`, and `Stop`. 3. Trust only handlers whose source is `User config - ~/.codex/hooks.json` and whose command contains `memhub_hook_bridge.py`. If Codex reports more than four handlers awaiting review, the extras are not from this installer. A user who trusted the earlier three-handler bridge is asked once more after upgrading: the new `SessionStart` handler is the only addition. Warn the user not to choose **Trust all** in that case. Never describe this manual approval as automatable by MemHub. For `$ARGUMENTS == --status`, run this and make no changes: ```bash python3 "$ROOT/scripts/setup_codex_hooks.py" status ``` For `$ARGUMENTS == --remove`, run this and stop: ```bash python3 "$ROOT/scripts/setup_codex_hooks.py" remove ``` ## Authentication and health After an install, check the hook credential without opening a browser: ```bash CLAUDE_PLUGIN_ROOT="$ROOT" python3 "$ROOT/scripts/login.py" --status ``` If it reports `NOT LOGGED IN`, explain that the hook credential is separate from the MCP connector login and run `/memhub:login` before calling setup complete. Do not silently start a browser login. Then run the local health check. On Codex: ```bash echo '{}' | python3 "$ROOT/scripts/capture_health.py" --host codex --plugin-root "$ROOT" ``` On Claude Code and Cursor: ```bash echo '{}' | CLAUDE_PLUGIN_ROOT="$ROOT" python3 "$ROOT/scripts/capture_health.py" ``` (It reads a hook payload from stdin; without the pipe it blocks on the tty. Without `--host codex` it judges Claude Code's capture state, which on Codex reports the wrong host's health — the same flags the bridge passes.) No output is healthy. Relay any warning exactly enough that the user knows the remedy. Finish with a compact status: hook bridge installed/current, hook trust unverified or user-confirmed, plugin credential healthy or missing, and whether the current repo still needs MemHub onboarding. ## Other hosts Claude Code and Cursor load their bundled MemHub hooks natively, so never write Codex's bridge into their config. On those hosts, only run the authentication and health checks above. Cursor Teams/Enterprise policy can block unofficial marketplaces before this skill is available; that is an admin installation policy, not something this plugin should attempt to bypass.