PerryLink/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:

never	Start nothing; assume something else manages the sidecar. The default.
attach	Also check /health once at load and log what it found — useful when the harness and the sidecar race at startup.
spawn	Also run spawnCommand if nothing answers, then wait for it to come up.

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:

Plugin	One-liner
dsh-auto-review	Second-model auto-review on the approval chain, fail-closed by default
dsh-autotier	Automatic strong/cheap model-tier routing with deterministic risk guards and a /tier command
dsh-background-agents	Durable background child agents with a Web UI sidebar, messaging and interrupt
dsh-budget	Cost governance for DeepSeek Harness: budgets, carbon, and latency in one panel.
dsh-catalog	DSH Desktop Market standard catalog source for the PerryLink family
dsh-cert-mcp	Read-only MCP server exposing the certification registry: grades, snapshots and five-dimension evidence
dsh-checkpoint-rewind	Claude Code /rewind-equivalent: snapshots, session forks, one-shot restore
dsh-claude-move	Migrate Claude Code sessions, memory, skills and CLAUDE.md into DSH
dsh-click	Cross-platform native desktop control for DeepSeek Harness — Windows first.
dsh-composer-history	Terminal-style input history for the web composer: arrows, Ctrl+R search
dsh-data-quality	Dataset quality checks and citation cross-checks (the optional numeric bridge consumed here)
dsh-defend	Prompt-injection, jailbreak, and secret-leak defense for DeepSeek Harness.
dsh-doublecheck	Engineering-discipline guard: requirements grill, test gates, adversary review
dsh-draw	Unified static-image generation routing for DeepSeek Harness.
dsh-fast	Read-only performance diagnostics for DeepSeek Harness.
dsh-fund-research	Deterministic research reports for Chinese public mutual funds
dsh-github	GitHub PR/issues integration for DSH, every write gated by approval
dsh-industry-research	Industry research orchestration that seals its deliverables through this plugin's ctx.researchReport.assemble
dsh-laya	Laya typed decisions (noul/choice/score) as a first-class Cordis service and model-visible tools
dsh-library	Local document knowledge base for DeepSeek Harness.
dsh-local-ai	Local-model (Ollama) integration for DeepSeek Harness.
dsh-lsp-actions	LSP diagnostics, formatting, completion, code actions and rename over language servers
dsh-mask	PII masking middleware: anonymize at the model boundary, restore at the display layer
dsh-mcp-panel	Read-only MCP runtime panel: /mcp command + Settings tab with status, tools and errors
dsh-memento	Approval-gated cross-session memory: ctx.memory seam + SQLite + memory tool
dsh-observe	OpenTelemetry and Langfuse observability exporter for DeepSeek Harness.
dsh-output-styles	Claude Code outputStyles-equivalent runtime style switching
dsh-permission-rules	Claude Code-style declarative allow/deny/ask permission rules with audit
dsh-plugin-certification	Community certification registry with repro-checkable grades and badges
dsh-plugin-doctor	Zero-dependency static + sandbox smoke detector for DSH plugins
dsh-plugin-guide	Plugin-development knowledge base as an on-demand agent skill
dsh-plugin-kit	Shared zero-runtime-dependency toolkit for the PerryLink DSH plugins
dsh-plugin-upgrade	One-package, one-corridor-index plugin upgrade skill: routes a repository to the matching closed corridor card
dsh-plugin-upgrade-015	Merged 0.1.3-alpha.1 → 0.1.5-rc.1 upgrade corridor card plus a zero-dependency seam scanner
dsh-reach	Multi-channel approval/question bridge: WeChat/Telegram/Feishu, session console
dsh-research-report	Verifiable research-report engine: content-addressed evidence ledger and sealed versions
dsh-score	Multi-dimensional quality scoring for DeepSeek Harness plugins.
dsh-session-pin	Pin sessions in the Web sidebar with durable ordering
dsh-session-sync	Cross-device session sync for DeepSeek Harness — a dedicated git mirror of your session store.
dsh-skill-pack-security	Security-audit skill pack: secret scan, dependency and supply-chain review
dsh-talk	Voice-first session loop for DeepSeek Harness: talk to it, hear it answer.
dsh-team-rooms	Cross-session team rooms: shared message bus, task board and timeline
dsh-test-drive	Isolated install-and-smoke test drives for DeepSeek Harness plugins.
dsh-ticktick	TickTick/Dida365 task bridge: session-header panel + 11 tools
dsh-translate	Vendor parameter translation and deterministic JSON repair for DeepSeek Harness.

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.