Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

4.1 KB
Raw Blame

DSH tools

Store categories and README visuals

browse-categories.json maps precise manifest domains into the Store's browsing categories. Coding comes first; the other categories group domain harnesses. The metadata-only generator reads the same listed catalog as the registry and updates the desktop category map, README inventory, and overview graphic together:

node store/tools/presentation.mjs
node store/tools/presentation.mjs --check

The README slideshow uses the screenshots selected in readme-showcase.json and their complete original prompts from each package's store.json. With the release-pinned Flutter SDK and FFmpeg on PATH:

node store/tools/showcase.mjs

Set FLUTTER_BIN and FFMPEG_BIN if needed. The script renders captions with Flutter and encodes an infinitely looping GIF. It also writes a still preview and readable prompt links into the README. Sources and visual-review notes are in the asset guide.

Daemon checks

Daemon-level checks over the same loopback WebSocket the desktop uses (ws://127.0.0.1:18473/api/local-ws). They need a running daemon with the harness installed on this machine, and cli/node_modules (for ws). Run them as the OS user Harness runs as: they reach the daemon over its 0600 socket in ADAPTER_DATA_DIR (default ~/.harness/cli/data, or ~/.harness/s/<hash> for a folder too long for a socket), local-daemon.mjs, and nothing goes to its TCP port, which any account can hold. On Windows they reach it over its named pipe, which the daemon's record daemon-<port>.json names: each request proves the daemon's token (daemon-token) without sending it, and an answer counts only with the daemon's proof. With no live pipe they say Harness is not running; nothing goes to the port, which the Team account's daemon can hold from boot (audit W1). Only a home too long to hold a socket uses the port, with hook-credential as X-Harness-Local (local-credential.mjs). The Store packages share the same transport, a copy of local-daemon.mjs in each (lib/localDaemon.mjs).

node store/tools/dsh-e2e.mjs <machineId> <dshId> <engine> <workspace> [--keep] [--restart] [--bypass] [--no-verdict]
    # create → materialize → viewer up → HARNESS_DSH on the pane → (verdict) → delete
node store/tools/dsh-discover.mjs ...   # a pane started by hand with HARNESS_DSH is discovered as its harness
node store/tools/dsh-delete.mjs <agentId> <machineId>

<machineId> is this machine's id from harness status / ~/.harness/cli/data/machines.json.

runtimes.sh — what a package runs on

A new machine has Apple's Python 3.9, no Homebrew and usually no Node on PATH, so a package's setup never asks the person to install an interpreter: it sources runtimes.sh and gets one. harness_node falls back to the Node Harness itself runs on, harness_venv makes a venv on a pinned CPython that uv downloads (uv itself is fetched, pinned and checksummed, when absent), and harness_conda_env covers native libraries PyPI has no wheel for. Everything lands in the package directory or ~/.harness/runtime.

A package installs alone, so each one carries a copy (toolchain/runtimes.sh in an agent, runtimes.sh in a viewer). Edit only store/tools/runtimes.sh, then:

node store/tools/sync-runtimes.mjs           # rewrite every copy
node store/tools/sync-runtimes.mjs --check   # exit 1 on a copy that drifted or a package missing one

Specialist studio runtime copies

studio_runtime.py validates project controls, runs one named action, verifies its artifacts, and writes immutable results plus the Harness verdict. studio_fetch.py fetches reviewed commit pins without overwriting local source changes. Every specialist package carries its own copy for sparse installs. After editing the canonical files, run:

node store/tools/sync-studios.mjs
node store/tools/sync-studios.mjs --check

The shared Studio Viewer test guide covers the runtime, pin fetching, domain workflows, browser controls, and actual Harness lifecycle.