nc://command-tray

Command Tray: the shell commands this agent ran, with their exit codes, as two tools and /commands. A Nodecode rewrite of aydnOktay/hermes-command-tray by aydnOktay.

Every (sh ...) the agent runs through eval lands in the tray as it finishes: the command, ok or fail with the exit code, the directory it ran in, the last non-empty line of its output, when it last ran and how many times. A command that runs twice is one row with a run count, the list keeps the newest 40, and token-shaped strings are redacted before anything is written.

It runs nothing itself, calls no service and needs no key. The rows are a JSON file under the home (command-tray/state.json), so a restart does not lose them.

Get it

Tell the agent "add the command-tray cell", or run nodecode add nc://command-tray.

Use it

Ask the agent what it ran, and it calls command_tray_list:

1. pytest -q — ok (0) ×2
   3 passed
   in /work/project
2. make deploy — fail (2)
   ERROR: no such target

• command_tray_list — the rows; markdown: false answers the same rows as JSON instead.
• command_tray_clear — empty the tray, for a fresh list on a new topic.
• /commands in any shell — the same list; /commands clear empties it.
• (help :command-tray) — the manual the model reads.
• (command-tray:runs) and (command-tray:clear) — the same two verbs through eval, for a layer or a snippet.

Settings

Nothing is required; the tray fills itself. In /settings, the page's Settings tab, or the command-tray section of config.jsonc:

"command-tray": { "max_items": 40, "preview_chars": 160, "redact": true }

max_items bounds the list, preview_chars how much of an output line a row keeps (0 for none), and redact masks token-shaped strings — GitHub and OpenAI-shaped keys, AWS key ids, key = value assignments and bearer headers — in both the command and the kept output line.

How it is wired

One piece of advice around NODECODE.EVOLVED:SH: the command runs through the chain first, its own answer comes back untouched, and the row is recorded from what the advice saw. A tray that cannot write its file never fails the command it was watching. The rows are written atomically — a temporary file and one rename — and read once at start; a file that is missing, or is not the shape this cell writes, is an empty tray and never a refused start.

Every row assertion in the tests runs against a stubbed shell, so the suite spawns nothing and dials nothing: run .github/test-cell.sh tools/nodecode-command-tray from the cells repository.