Agents

Agent	Tested version	Status
Claude Code	2.1.318	✅
Codex	0.164.0	✅
OpenCode	2.0.18	✅
Pi	0.86.0	✅
Command Code	1.88.3	✅

Each was tested end to end in the v1 release check.

Note

Any other agent that reads AGENTS.md and .agents/skills/ can follow the workflow, but without a hook nothing stops it before a gate is approved; the check still catches it at commit and in CI.

Wiring an agent

lyceum init detects an agent when its binary is on your PATH or its project folder exists (.claude/, .codex/, .opencode/, .pi/, .commandcode/), and asks about each one it detects. To wire one without being asked:

lyceum init --harness claude

--harness is repeatable and accepts claude, codex, commandcode, opencode and pi.

For each agent it wires, init installs skills, slash commands where supported, and a hook. Skills go to .agents/skills/lyceum-*/, and to .claude/skills/lyceum-*/ when Claude Code is wired. Wiring Claude Code also adds a one-line pointer to AGENTS.md in CLAUDE.md.

Agent	Hook
Claude Code	PreToolUse in .claude/settings.json
Codex	PreToolUse in .codex/hooks.json
OpenCode	A plugin in .opencode/plugins/
Pi	An extension in .pi/extensions/
Command Code	PreToolUse in .commandcode/settings.json

Existing settings files are merged: every existing key and value is kept.

Caution

Codex and Pi run project hooks only after you trust the project in that agent. Until you do, the hook silently doesn’t run. init prints the step for each.

Starting a skill

Agent	Start a skill
Claude Code	/lyceum-feat
Codex	$lyceum-feat
Pi	/skill:lyceum-feat
OpenCode, Command Code	Ask for the skill by name

The skills are lyceum-feat, lyceum-bug, lyceum-doc, lyceum-test, lyceum-refactor, lyceum-chore, lyceum-init (the new-project interview), lyceum-status (run lyceum status and follow the next step), and the inspect skills lyceum-understand, lyceum-trace and lyceum-audit.

What the hook does

While the unit’s code-opening gate is not cleared, the hook blocks writes to code. Test files are allowed at the gates that allow them. The plan’s files: list is never a block, only a check warning. The hook also blocks any shell command that runs lyceum approve or lyceum eject. When it denies, the reason names the gate and the next step, and tells the agent not to retry.

The hook does not parse shell writes such as printf hi > f; those are left to check.

In Claude Code, OpenCode and Pi, the hook can instead ask you to confirm an approval in the agent’s own prompt. See Approvals.

Codex sandbox

Wiring Codex adds the repo’s .git folder as a writable root in .codex/config.toml, so the agent can commit under workspace-write. That sandbox also keeps the network off, so a suite that binds ports can fail there and Codex asks to escalate on each run. You can set network_access = true under [sandbox_workspace_write] in .codex/config.toml, which opens outbound network to the agent too, or run such tests outside the sandbox.

OpenCode server

When a re-run of init changes an existing OpenCode plugin, restart OpenCode’s server (opencode serve): a running server keeps the plugin version it first loaded.

Removing Lyceum

lyceum eject removes skills, slash commands, hooks, the CI workflow, the pre-commit hook and the appended Lyceum sections in AGENTS.md and CLAUDE.md. It leaves lyceum/ and the root docs in place as plain markdown. Like approve, it runs only in your own terminal and is confirmed by typing eject.