b-open-io/codex-agent-setup

Codex Agent Setup

Installs the plugin's generated Codex custom-agent adapters as regular files into a Codex agents directory. Explicit invocation only — do not run unless the user asked to install, update, check, or uninstall these agents.

Safety rules (non-negotiable)

• Never edit ~/.codex/config.toml or any global Codex config.
• Never symlink into the plugin cache; the installer copies regular files.
• Never delete unrelated user/project agents.
• Default scope is the current project (.codex/agents/), not user-global.
• Prefer --check first when the user asks what would change.

Commands

Run via the bundled launcher (resolves checkout or installed plugin cache):

bash "${SKILL_DIR}/scripts/setup.sh" [options]

If SKILL_DIR is unknown, resolve from this skill's path or set BOPEN_PLUGIN_ROOT to the plugin root and run:

python3 "${BOPEN_PLUGIN_ROOT}/scripts/codex-agents/install.py" [options]

Common invocations

| User intent | Command | |---|---| | Install curated roster into project .codex/agents/ | bash scripts/setup.sh | | Install curated roster into user ${CODEX_HOME:-~/.codex}/agents/ | bash scripts/setup.sh --user | | Install all generated agents (project) | bash scripts/setup.sh --all | | Install all into user scope | bash scripts/setup.sh --user --all | | Dry-run / drift check | bash scripts/setup.sh --check (add --user / --all as needed) | | Uninstall managed agents | bash scripts/setup.sh --uninstall (add --user if installed user-scope) | | Overwrite an unmanaged name collision | bash scripts/setup.sh --force |

Default curated roster

front-desk, agent-builder, prompt-engineer, researcher, tester, documentation-writer, architecture-reviewer, code-auditor.

Each file installs as bopen-<name>.toml (for example bopen-front-desk.toml). Its Codex runtime agent name uses the identifier-safe bopen_<name_with_underscores> form, such as bopen_front_desk.

Full roster

--all installs every generated adapter under codex/agents/. Only installable plugin personas live in agents/; app-specific deployments and user-created agents are not source inputs for this installer.

What the installer does

1. Reads committed generated adapters from <plugin-root>/codex/agents/.
2. Copies selected TOML files atomically into the target agents directory.
3. Maintains an ownership manifest: .bopen-tools-agents.json in the target.
4. Preserves unrelated agents always.
5. Refuses unmanaged filename collisions unless --force. With --force, quarantines the colliding unmanaged file first; if quarantine fails, aborts that file without overwriting.
6. Quarantines only stale unmodified managed files — those whose source adapter no longer exists in the full current plugin roster. A later curated install after --all does not drop still-valid non-curated agents.
7. Stores quarantined files under a target-local hidden directory: <target>/.bopen-tools-trash/quarantine/... (nested non-TOML path so Codex agent discovery never loads them). No temp-directory fallback.
8. Leaves modified managed files in place and prints a warning.
9. Prints installed / updated / quarantined / skipped counts.
10. Reminds the user that a new Codex session is required to load agents.

How Codex invokes an installed agent

Codex has no codex exec --agent flag. Installed adapters are subagents: the Codex orchestrator spawns them by name when the multi_agent feature is on (a thread_spawn under the hood). The runtime name is the identifier-safe form (bopen_researcher, bopen_front_desk), not the bopen-<name>.toml filename.

multi_agent is a Codex feature flag. It is often already enabled globally under [features] in ~/.codex/config.toml; if not, enable it per invocation with -c 'features.multi_agent=true' rather than editing global config (safety rule).

To use an agent, name it in the prompt and tell the orchestrator to delegate:

codex exec -c 'features.multi_agent=true' \
  "Delegate this to the bopen_researcher subagent and return its answer verbatim: <your task>"

Verify (one-shot, after install)

Prove an installed adapter's persona is actually active. Run from a new Codex session (agents are read at session start):

codex exec --sandbox read-only --skip-git-repo-check \
  -c 'features.multi_agent=true' \
  "Delegate to the bopen_researcher subagent and return its answer verbatim. \
Task: State your role, operating constraints, your Pre-Task Contract fields, \
and which agents you delegate code analysis and architecture review to. \
Do not answer yourself — spawn the bopen_researcher agent."

Pass criteria — the returned answer carries the source persona (agents/researcher.md), not a generic assistant. For bopen_researcher expect: identifies as Parker / Research Analyst, states it is read-only, lists the Pre-Task Contract fields (Scope / Sources / Deliverable), and delegates code analysis to code-auditor and architecture review to architecture-reviewer. If the reply is generic or ignores those constraints, the agent was not spawned — confirm you started a new session and that the adapter exists in the target agents dir.

Swap bopen_researcher and the pass criteria for whichever agent you installed.

Maintainer: regenerate adapters

When agents/*.md changes, regenerate committed artifacts from the plugin root:

python3 scripts/codex-agents/generate.py
python3 scripts/codex-agents/generate.py --check

Or: bash skills/codex-agent-setup/scripts/generate.sh.

After install

Tell the user:

1. Summary counts from the installer output.
2. Target directory used (project vs user).
3. They must start a new Codex session before spawning bopen_* agents.
4. Global Codex config was not modified.
5. How to invoke: agents are Codex subagents spawned by name (bopen_researcher, not the .toml filename) when multi_agent is on — see "How Codex invokes an installed agent" above.
6. Offer to run the Verify one-shot to prove the persona is active.