More than one CLI

⚙ Settings → Which tools archeus shows, and the tab strip on New session.

archeus began with one coding CLI behind it. It now reads three, and the distinction that makes that work is worth stating once:

• a provider is an endpoint. Model providers repoint ANTHROPIC_BASE_URL and the same claude binary talks to a different model.
• a harness is the binary itself, plus the conventions it keeps on disk — where its transcripts live, what its instructions file is called, what a session id even is.

The two are orthogonal. OmniRoute is a provider; Codex is a harness. The launch picker shows both in one strip because they are both answers to "what runs this session", but they are not the same kind of answer.

What archeus finds on its own

Nothing to configure. archeus looks for each CLI's binary and its home directory, and a CLI it cannot find is simply not offered — a home with no binary is a leftover, not an installation.

CLI	Home	Instructions file	Found via
Claude Code	~/.claude (and every other account)	CLAUDE.md	CLAUDE_CONFIG_DIR, ~/.local/bin, PATH
OpenAI Codex	~/.codex	AGENTS.md	CODEX_HOME, its hashed install directory, PATH
pi	~/.pi/agent	AGENTS.md	PI_CODING_AGENT_DIR, npm's global bin, PATH

Everything that follows works off that: the project list, the sessions list, previews, turn counts, models, token spend, search, usage and the dashboard all merge across every CLI you have.

One memory graph, three CLIs

This is the part that is not a convenience.

A project's memory graph is the project's — what it is, how it is laid out, what has been learned about it. It is not Claude Code's memory of the project, and it would be absurd for it to be rebuilt once per tool. So archeus keeps one graph and delivers the same digest into every instructions file an installed CLI reads: CLAUDE.md for Claude Code, AGENTS.md for Codex, and both for pi, which reads either.

A machine with no Codex never grows an AGENTS.md, and a CLI you switch off stops getting one.

The agent routing table deliberately does not follow. Delegating to a subagent is a Claude Code capability; writing "delegate with the Agent tool" into AGENTS.md would instruct Codex to use a tool it does not have.

Skills do follow, because SKILL.md is the Agent Skills standard and all three read the same file. Installing one personally puts it in every account and every CLI; installing one into a project writes .claude/skills and .agents/skills — two directories, not three, because .agents/skills is the cross-harness convention that Codex and pi share.

Choosing one at launch

New session opens on a tab strip: your installed CLIs, then each configured backend by name. The default is Claude Code and you can change it in ⚙ Settings → Defaults → Starts on.

The tab is above the rest of the form because it decides what the rest of the form even means. A Codex session has no worktree and no name-at-launch; a pi session has no permission mode. Those fields are not greyed out — they are gone, and the strip says why:

Codex names a session afterwards, with thread name set, not at launch.

A resume has no strip at all. A session already on disk belongs to the CLI that wrote it, and a Codex thread cannot be resumed by pi.

What each CLI cannot do here

archeus never pretends. A screen a CLI has no equivalent for is shown greyed with the reason, and a button on a session row works the same way.

	Claude Code	Codex	pi
Sessions, resume, fork	✅	✅	✅
Token spend, by day and project	✅	✅	✅
Plan usage & reset windows	✅ OAuth usage endpoint	✅ read from its own rollout	— bills per provider, so there is no single plan window
Project memory	✅	✅	✅
Skills	✅	✅	✅
Path-scoped rule files	✅	— reads the digest in AGENTS.md instead	— reads the digest in AGENTS.md instead
Archive a session	✅	✅	— no archive; a session is kept or deleted
MCP servers	✅	✅ codex mcp	— no MCP client
Plugins & marketplaces	✅	✅ no per-plugin update	— pi install packages, not marketplaces
Install & update the CLI	✅	✅ codex update	✅ pi update
Checkpoints (/rewind)	✅	— SQLite, not a file-history store	— branches inside one session file
Hooks	✅	— gated on an undocumented trusted_hash	— TypeScript extensions instead
Subagents	✅	— .agents, with no CLI to manage them	— none
Output styles	✅	—	—
More than one login	✅ one CLAUDE_CONFIG_DIR each	✅ one CODEX_HOME each	✅ one PI_CODING_AGENT_DIR each
Status line	✅	—	— draws its own

Every row is checked against the installed binary, not assumed from the name — and four of them changed when that was actually done. codex mcp list/get/add/remove is a full MCP surface, codex plugin list reads every marketplace, and codex doctor reports the installed version and whether a newer one exists; all three run offline and without a login. Token spend was never Claude Code's alone either: every CLI records its usage in its own transcripts.

The plugins row changed a second time, and it is a vocabulary difference rather than a capability one: Codex calls the same operations by different names — plugin add where Claude Code says install, plugin remove where it says uninstall, and plugin marketplace upgrade where it says update. archeus keeps one table of those verbs, so the Plugins page adds, installs and removes under either CLI. The one thing Codex genuinely has no command for is upgrading a single installed plugin, so that button is simply not drawn there.

Two of those rows changed again, for the same reason. The plan window is not Anthropic's alone: Codex records its own windows in its rollout, and codex doctor — which this page used to name as the source — prints no limit at all, as its own sections show. And a login is a home for all three CLIs, selected by one environment variable each, so "more than one account" needed no new mechanism in any of them.

Hooks is the one that survived the check, and it is worth saying why it stays off. Codex's hook contract is byte-identical to Claude Code's — the binary carries the same session_id, transcript_path, hook_event_name and stop_hook_active payload fields — so writing one would be easy. But every handler in its hooks.json is gated on an undocumented trusted_hash, and a wrong hash is refused silently. A hook archeus installs that never fires and never says so is worse than no hook at all.

Where a gap is archeus reading Claude Code's file for something the other CLI keeps elsewhere, the reason says so on the page itself — so you can tell "this tool does not have it" from "archeus has not got to it yet".

Where each CLI's own screens live

The sidebar carries what more than one CLI has. Everything that is one CLI's own lives behind Harnesses, under a tab for that CLI: Setup — whether it is installed, where, which version, whether an update is waiting and what it can do — plus, for Claude Code, Accounts, its own client state, output styles, subagents and hooks.

Five screens moved there rather than staying in the sidebar greyed out for the two CLIs that do not have them. A greyed row teaches you the app has the feature; five of them teach you the app is about one tool.

Models and quick starts

Each CLI gets its own model list and its own effort scale, because they are not translatable: --effort max and ultracode are Claude Code's and Codex rejects both, pi's scale runs from off to max, and a priced Anthropic model card means nothing next to gpt-5.5.

The model field is free text with two sets of suggestions behind it — run here, what that CLI has actually been pointed at on this machine, and ships with this CLI, read out of the catalogue it installs. Nothing is typed into archeus: the catalogue versions with the CLI, so pi update updates it.

That catalogue also carries prices, which is what makes a gpt-5.5 or Gemini session cost something rather than reading n/a. A model nothing publishes a price for stays unpriced — archeus does not guess, because an Opus-tier fallback on a gpt-5.6-luna session would be wrong by a factor of fifty.

Turning one off

⚙ Settings → Which tools archeus shows.

Off means everywhere — the launch tabs, the project list, the sessions list, the usage table, and the instructions files that get a memory block. Hiding it from one surface and leaving it in the others is what makes a setting feel broken.

Nothing is uninstalled and nothing on disk is touched. Switch it back on and its projects come straight back.

Claude Code is listed and locked. It is not only a CLI archeus reads: it is the binary archeus itself runs for memory extraction, lessons, CLAUDE.md generation and every other AI feature, so switching it off would hide the app's own engine while it kept running.

What archeus will not guess

Two deliberate refusals, both the same discipline checkpoints already follows:

• A Codex thread's transcript is named by its index, not derived from its id. archeus reads the path out of state_*.sqlite rather than reconstructing it.
• A pi session directory's name is never decoded. pi maps every path separator and the drive colon to -, which is lossy; the first line of every session file carries the real working directory, and that is what archeus reads.

Both stores are opened read-only. archeus never writes to another CLI's session state.