Putting your repos on it

Run the wall

macOS. Homebrew is the way in, and it brings its own Node:

brew install orochi235/tap/transom
transom install    # starts the wall under launchd

launchd keeps it running across logins and restarts; transom stat says whether it answers, and transom cycle restarts it.

To hack on it, run it from a clone instead. That needs Node 22.18 or newer:

git clone https://github.com/orochi235/transom
cd transom
npm install
npm run dev

The client is at localhost:7750 and the daemon at localhost:8787, watching ~/transom/inbox/. Put it on the monitor you do not type on.

Wire it to your agents

Once per machine, and again after you change anything here — it is idempotent. From a clone it is bin/transom wire.

transom wire    # --dry says what it would change
                # --off takes it all back out

It puts three things where an agent will meet them:

• The skill, symlinked into every ~/.claude*/skills/. It tells an agent how to send, how to diagnose a render that never arrived, and how to except a repo.
• The wall-nudge hook, a PostToolUse entry in each settings.json. It fires when an agent reads or creates an image that never reached the wall and hands the model a one-line correction, which is what stops renders dying in a scratch directory.
• transom on PATH, symlinked into ~/.local/bin. Without it an agent that half-remembers the rule types transom post chart.png, gets command not found, and goes back to reporting a file path.

A session that was already running keeps the hooks it started with, so a live one is not evidence that wiring failed. transom wire also prints a rule for your CLAUDE.md and leaves you to paste it: that file is hand-written prose, and a tool that rewrites preferences fights their author.

A repo needs no setup

A zone is one pile on the wall, and the zone is the directory name of the repository you ran in. Nothing has to register a project: the first send creates its zone, and bin/transom records which directory the renders came from so the daemon can read that project’s .hued file and paint the pile in its color.

transom post render.png              # the zone is this repo
transom post --zone icons sheet.png  # or say where it goes
some-generator | transom post        # a pipe is a send too

What a send can carry

Every flag is optional, and the file alone is the whole protocol. These are the ones worth knowing on the first day:

transom post --ttl 5m draft.png           # how long it lives
transom post --caption "after the fix" a.png
transom post --attention soon broken.png  # or look, urgent, problem
transom ask "which crop reads better?" \
     --choice left --choice right crop.png

transom ask puts the question on the card, waits, and prints the answer, so an agent can stop and ask you something about a picture instead of guessing. Groups of pictures share one card with --group; video, self-contained HTML pages and .glb/.stl meshes land as cards whose lightbox plays, runs or orbits the real thing.

Settings for one repo

A .transom.yaml at the repo root changes what a send from that repo does. It is committed with the repo, and every key is optional; a flag on transom post still beats it. Inside the repo:

# write a starter, or check the file
transom wire --repo
# remove a starter nobody edited
transom wire --repo --off

zone: icons          # not the directory name
show: preview        # open locally, skip the wall
defaults:
  ttl: 30m
  apps: [LDView]     # “Open in LDView”
attention:
  loudest: soon      # louder sends are lowered
  sound: false

The file’s first line points your editor at the schema, which completes the keys and flags a wrong one. transom post checks it too: a file with a mistake in it stops the send and names each problem rather than being quietly ignored.