@alisio/plugin-laya

A local Laya decision provider for Alisio. Features that ask Alisio for a select, boolean or ordinal decision get an answer from a small model running on your machine, without an LLM generation call and without sending the data anywhere.

This is an independent plugin. It is not affiliated with the Laya project. Laya is optional: Alisio works without it, and every feature keeps its own fallback when no provider answers.

Requirements

• Alisio core >=0.3.0 (the plugin only registers its commands on older cores).
• Python 3.10 or newer (64-bit) with the venv module (on Debian or Ubuntu: apt install python3-venv).
• Disk: the managed runtime takes about 6 GB when a CUDA-capable machine installs PyTorch's default wheels (measured on 2026-10-04; the setup consent text says "about 6.1 GB"), and about 1 GB on a CPU-only machine. The multilingual model checkpoint download is 678 MB. Running setup again replaces the previous runtime, so a single environment remains.
• Memory: peak resident memory of the server measured at about 2.3 GB with multilingual and about 2.8 GB with english (CPU, one checkpoint loaded); plan for 3 GB free.
• Network only during /laya:setup. Afterwards the server runs offline.

Install

alisio install npm:@alisio/plugin-laya

Then, inside Alisio:

1. /laya:setup — shows exactly what will be downloaded and where from, and asks for explicit consent before anything is fetched or written. It starts a background job and returns at once. The first setup took about 3 minutes with warm pip and Hugging Face caches; a cold one depends on your connection.
2. /laya:status — follow progress, then check the provider. /laya:cancel aborts a running setup.
3. Activation: when the setup finishes successfully it asks once whether to activate the provider and, if you accept, Alisio itself writes decisions.provider to the global configuration (core 0.4.2 or newer; "Yes" is pre-selected as the recommended answer on core 0.4.3 or newer). It never overrides a different provider you configured. If you decline, run /laya:activate to retry. Only on older cores, set decisions.provider = "laya" in the global Alisio configuration by hand. The plugin itself never edits your config.json.

Installing the package alone downloads nothing and starts nothing.

Commands

| Command | What it does | | --- | --- | | /laya:setup [--device auto\|cpu\|cuda\|mps] [--model multilingual\|english\|typed-decisions\|auto] [--yes] [--repair] [--uninstall] | Installs (or repairs, or removes) the managed runtime. Without an interactive UI it refuses unless --yes is passed. | | /laya:status | Host support, setup job, installed pins, effective config with its source, server state, restarts, last error, observed latency. | | /laya:activate | Retries making Laya the active decision provider (the host asks for confirmation, with "Yes" recommended on core 0.4.3 or newer; needs core 0.4.2 or newer). | | /laya:cancel | Cancels a running setup job; the previous runtime stays untouched. |

A setup job started by another Alisio process is shown as running there and is left untouched; it can only be cancelled from the process that started it.

Configuration

| Key | Values | Default | | --- | --- | --- | | device | auto, cpu, cuda, mps | auto | | model | multilingual, english, typed-decisions, auto | multilingual | | preload | true, false | true |

Precedence, highest first: ALISIO_LAYA_DEVICE / ALISIO_LAYA_MODEL / ALISIO_LAYA_PRELOAD environment variables, the plugin options in your Alisio configuration (pluginOverrides.laya.options, user level only), config.json in the plugin config directory (written by /laya:setup --device ... --model ...), then the defaults. ALISIO_LAYA_PYTHON selects the interpreter used by setup; ALISIO_LAYA_HOME moves the runtime directory.

preload: true starts the server and warms the model when Alisio activates the provider, so the first decisions are not lost to a cold start. Set it to false to save memory; the server then starts on the first decision, which falls back while it warms up.

How it behaves

• The server is bound to 127.0.0.1 on a free port, protected by a random per-start token, and runs with a minimal environment. It stops when the provider is deactivated and when Alisio closes.
• Cold start takes seconds. Until the model is warm, decisions fall back to the feature's default immediately instead of waiting.
• Quality is modest and uneven. On 375 labelled dashboard decisions (half Spanish), multilingual was right 62 % of the time: about 90 % for the dashboard purpose (a select), but close to chance for yes/no toggles and for choosing a column, which is why Smart Dashboard asks Laya only for the dashboard purpose. Treat answers as hints that the feature can override, not as facts.
• Laya's confidence is not calibrated: treat it as a weak signal and do not make sensitive decisions depend on it. Alisio's confidence threshold is a heuristic filter.
• On a slow CPU, requests with many decisions can exceed Alisio's 1500 ms decision timeout; use smaller packs or raise decisions.timeoutMs. Measured on a busy CPU with 4 threads: about 0.3 s for one decision and about 2 s (p95 3 s) for a request of seven with multilingual; english is about three times slower.

Privacy and security

• Decisions never leave your machine. After setup there is no network use beyond loopback.
• Setup is the only network phase. It contacts PyPI, PyPI's file host, the PyTorch wheel index (CPU installs) and Hugging Face, and pins the Laya version, the model revision and the SHA-256 of every model file.
• A subprocess is not a sandbox: the server runs with your user's privileges. The controls reduce exposure; they do not contain a compromised dependency.
• Nothing is written to your workspace. Files live only in the plugin's own config, state and cache directories. Secrets are never stored by this plugin.

Troubleshooting

• /laya:status says the runtime is outdated: run /laya:setup --repair.
• Setup fails at "creating environment": install the Python venv module (see Requirements).
• The server keeps failing: /laya:status shows the last error; /laya:setup --repair rebuilds it.
• Installing the plugin fails with a peer @alisio/sdk conflict (ERESOLVE): older plugins are still installed. Run alisio install --update (core 0.4.1 or newer), which updates all installed plugins at once.
• Right after a new version is published to npm, the first install can fetch the previous version for a few minutes (npm metadata cache): alisio install --update fixes it.
• Remove everything with /laya:setup --uninstall.

License

MIT