dsh-laya

Laya typed decisions — noul (yes/no),
choice, score — as a first-class Cordis service and model-visible tools for
DeepSeek Harness.

English · 简体中文 · Español · Português · हिन्दी

Status: 0.1.4, work in progress. The plugin mounts and activates in the
Harness, and the sidecar contract it speaks is verified end-to-end on an RTX
5060 (load 9.8 s, 411 ms for three questions on CUDA). It has not yet been
exercised through a live model turn in CI.

Install

Two pieces, because Laya is PyTorch and therefore cannot live inside a Node
plugin.

1. The sidecar — this is what actually holds the model:

pip install "laya-mcp[mcp]"
laya-mcp serve                      # loads once, listens on 127.0.0.1:8787

2. This plugin:

dsh plugin --profile <profile> add dsh-laya

Then confirm the row is active, not failed, and read the startup line — it
says in as many words whether state stays on this machine.

Why the sidecar is separate

A plugin that shelled out to pip install and then downloaded a 650 MB
checkpoint behind your back would be hostile, however convenient. So this plugin
installs nothing and downloads nothing. It is a client of a process you
start, and when that process is not running it says so rather than failing
obscurely at the first tool call.

The split also buys the warm model: Laya's cold build costs seconds to tens of
seconds, and its default lazy router rebuilds a checkpoint on every language
switch. laya-mcp serve pays that once.

What it contributes

One service, ctx.laya — so Host code and other plugins can ask for a
judgment directly, without a model round-trip:

const laya = ctx.get('laya')
const result = await laya.ask({ state, questions })

It exposes ask, plan, health, capabilities, sidecarUrl, and loopback
— the last being whether state stays on this machine, as a fact rather than a
policy. plan is the preflight: the same arithmetic /ask reports, with no
forward pass.

Two tools — laya_ask for a batch of typed questions, and laya_plan to
check the token budget before spending a forward pass.

What it is honest about

Laya truncates silently and its confidence number is widely misread, so the tool
description and the response both say so:

• An oversized state is cut from the END, and the answer is then about the
  surviving prefix. The response reports it under truncated, and laya_plan
  tells you before you ask. Pass strict: true to refuse instead — verified
  against the live sidecar, which returns HTTP 400 with state_truncated and a
  hint naming max_len.
• confidence is not accuracy. It is a concentration statistic: low whenever
  probability is spread across options even when the top option is right, and
  high on a confident wrong answer. A noul also carries a no / uncertain /
  yes band, because a calibrated probability is not a decision — a real
  measured example from this stack returned 0.5457 with the band uncertain.
• A silent CPU demotion is reported. If Laya falls back to CPU after a device
  error it never returns to the accelerator, and nothing in its own output admits
  it. The health surface carries degraded, and the tool card appends
  CPU (degraded).

Configuration

- insert:
    - id: laya
      name: 'dsh-laya'
      config:
        sidecarUrl: 'http://127.0.0.1:8787'
        requestTimeoutMs: 120000
        lifecycle: never      # 'attach' to log a reachability check,
                              # 'spawn' to start the sidecar yourself
        spawnCommand: null    # required by 'spawn', e.g. ['laya-mcp', 'serve']
        spawnTimeoutMs: 120000
        logLevel: info

lifecycle decides what happens when nothing is answering at sidecarUrl:

spawn exists because the alternative was worse. The plugin still installs
nothing and downloads nothing — a tool that ran pip install and then fetched a
650 MB checkpoint behind your back would be hostile, and that has not changed.
But launching a sidecar you already installed is a different act, and without it
every session began by starting a Python process in a terminal by hand, and began
failing again every time that process went away.

Two rules make it safe to leave on:

• A sidecar that is already answering is attached to and never touched, so two
  harness sessions cannot put two models on one port.
• A sidecar this plugin started is stopped when the plugin unmounts; one it did
  not start is left exactly as it was found.

There is still no option that installs anything or fetches a model, and
spawnCommand has no default: this plugin will not guess at an interpreter.

Pointing sidecarUrl somewhere that is not loopback is allowed and warns once at
startup, naming the destination: that is the moment the privacy story changes,
and it should not be discoverable only by reading a config file.

Upstream limits worth knowing before you rely on this

Repeated from upstream's own measurements, because an integration that implies
otherwise is lying to you. The base checkpoints are near chance zero-shot on
typed decisions (0.362 for English against a 0.461 majority-class baseline);
score is the weakest primitive (35% vs 70% for Jev in independent measurement);
raw calibration error is 0.466 before temperature fitting; and accuracy falls off
above roughly 20 options.

Calibration makes a probability honest. It cannot make a model right.

Applicable DSH version: verified against dsh-v0.1.7-alpha.1 (the host release this build targets); requires >=0.1.7-alpha.1 <0.3.0.

PerryLink DSH Plugin Family

This project is one of the 45 DeepSeek Harness plugins maintained by PerryLink. If this one helps you, the others likely will too:

Licence

Apache-2.0. Laya is Apache-2.0 by Convai Innovations. This is an independent
integration and is not affiliated with or endorsed by that project.