Build Codex and Hermes commands from their own interfaces

[Alexgodoroja]

What this changes

The new-session form carried a warning on GPT Codex and Hermes Agent: "These flags come from the published interface for … rather than from its own help output, so check the command below before starting."

The warning was accurate. codex --full-auto is not a flag in current Codex:

$ codex --full-auto --help
error: unexpected argument '--full-auto' found

So the one non-default option the form offered for Codex built a command that could only fail on the machine it was sent to, after the round trip through POST /api/commands and the agent poll.

Rather than keep the warning, I installed both tools, read their real interfaces, and rebuilt the two kinds so there is nothing left to warn about. Codex and Hermes now sit at the same bar as Claude Code and OpenClaw.

Verified against

tool	version	install
codex	0.153.4	npm i -g @openai/codex
hermes	0.19.0	uv tool install hermes-agent

Every option offered by the two builders parses against the installed binary. All 17 command shapes the builders can produce were generated from the shipped build() functions and run against the real CLIs: 17/17 accepted.

Codex

Dropped --full-auto. Added, all read from codex --help and codex resume --help:

• resume by id, or --last
• --sandbox — read-only / workspace-write / danger-full-access
• --ask-for-approval — on-request / never
• --search

Options follow resume <id>, which is where the subcommand accepts them, so one order works for a fresh and a resumed run.

Hermes

Replaced the free-text argument box with real controls: a subcommand picker, --resume, --continue, --model, --worktree, --yolo.

Two ordering rules the parser imposes are encoded and tested:

1. Every option belongs to the top-level parser, so all of them precede the subcommand. hermes sessions --yolo is error: unrecognized arguments: --yolo; hermes --yolo sessions is fine.
2. A bare --continue takes an optional value, so it is greedy — in hermes --continue sessions the subcommand is read as the session name to resume. It is emitted last, and only when no subcommand and no explicit session could be eaten by it.

Discovery

Unchanged, and already correct: codex and hermes were present in cmd/shell/harness.go, in KNOWN_HARNESSES in app/server/app.ts, and in DETECTED_HARNESSES in app/src/lib/harnesses.ts. Confirmed end to end on a machine with all four installed — installedHarnesses() returns [claude-code codex hermes openclaw], and the "not found on this machine" note still renders for a machine that reports without them.

Tests

session-kinds.test.ts covers both builders: resume precedence, empty selects meaning "no flag", option ordering, and the --continue guard. The catalogue test that asserted which kinds were built from a published interface is replaced by one asserting no builder can emit --full-auto.

• npm run check (root) — pass
• npm test in app/ — 616 passed, 3 skipped
• npm run typecheck, oxlint — clean, no new warnings
• go test -race ./..., go vet ./..., gofmt -l — clean
• npm run build:web — builds

The modal was also driven in a browser to confirm the live command preview matches the builders and the warning is gone.

Note: npm run build in app/ fails at check-bundle for want of VITE_FIREBASE_*. That reproduces on main on a machine with no .env.local and is unrelated to this change.