Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.
Codex agent status doesn't work (RUNNING/NEEDS-YOU never light up)
Symptom. A Codex session shows no live status on the canvas or on the phone — no
"RUNNING" badge, no walking-pet mascot, no NEEDS-YOU, no completion notification. Claude and
Gemini on the same host report fine, and Codex works when run from a local desktop
session. It only breaks on a host where Codex was installed as a snap (Ubuntu's
snap install codex, /snap/bin/codex).
Root cause — snap confinement, not a nodeterm bug. nodeterm installs Codex's managed
hooks correctly: ~/.codex/hooks.json (six events → the managed codex.sh) plus the matching
[hooks.state."…"] trusted_hash = "sha256:…" trust entries in ~/.codex/config.toml. Verified
in the field: the files are present, the script is executable, the trust hashes match. But the
snap build of Codex never reads them, for two stacked reasons:
$HOME/CODEX_HOMEremap. A snap runs withHOME(and thereforeCODEX_HOME) pointed at its own sandbox dir — e.g./root/snap/codex/<rev>/— sosnapCodex looks for its config under~/snap/codex/<rev>/.codex/, not the standard~/.codex/where nodeterm (and every other tool) writes. The hooks are installed in a directory the snap Codex simply never consults.- Exec confinement. Even if the hook config were found, the hook command runs
~/.nodeterm/agent-hooks/codex.sh. A snap's AppArmor profile only allows its own$HOMEsandbox; a script under/root/.nodeterm/…(outside the snap home) can't be executed. So the hook chain is structurally unreachable under snap confinement.
Claude/Gemini are unaffected because they aren't distributed as snaps here.
Fix — install Codex outside snap confinement
snap remove codex
npm i -g @openai/codex # now CODEX_HOME resolves to ~/.codex (no remap)
codex login # the fresh ~/.codex has no auth.json yet — sign in once
Ensure the npm binary wins on PATH (it normally does — /usr/bin before /snap/bin):
command -v codex should print /usr/bin/codex, and codex --version a recent (≥ 0.145)
build. Then reconnect the SSH project (or restart the headless Server Edition) so nodeterm
reinstalls the hooks against the now-visible ~/.codex, and open a fresh Codex node —
RUNNING/NEEDS-YOU and the walking pet should light up.
Notes
- Alternatives that also work: run Codex as a non-snap install under any user whose
~/.codexis the real path, or pointCODEX_HOMEat a directory nodeterm can install into. The npm install is the simplest. - The managed hook/trust install itself is correct and version-checked (
codex-trust.tsgolden-locks the hashes). If a much newer Codex ever changes its hook-hash serialization, the trust would stop matching — that's a separate, tracked concern, not this snap issue. - Codex ≥ 0.145 still uses
config.tomltrust hashes (the--dangerously-bypass-hook-trustflag confirms the trust gate is intact).