dsh-plugin-model-capability

dsh plugin --profile web add dsh-plugin-model-capability

dsh-plugin-model-capability






Model Capability Manager — manage the llm-pi-ai provider routes of DeepSeek Harness (DSH Web) from a dedicated Model Capability page in the in-app settings: per-model thinking levels, context window, output cap, input modalities, per-route defaults, gateway compatibility fields, one-click presets, and an EN/中文 switchable UI.

[简体中文说明](./docs/README-zh.md) · Report Bug · Request Feature

---

Table of Contents

• [Why this plugin exists](#why-this-plugin-exists)
• [Screenshots](#screenshots)
• [Features](#features)
• [Installation](#installation)
• [Quick Start](#quick-start)
• [Uninstall](#uninstall)
• [FAQ / Troubleshooting](#faq--troubleshooting)
• [How it works](#how-it-works)
• [Development](#development)
• [Publishing](#publishing)
• [Contributing](#contributing)
• [License](#license)

---

Why this plugin exists

DSH stores provider configuration in the active profile's llm-pi-ai.providers entry. Editing profile patches by hand is error-prone, and two classes of problems bite people often:

1. Gateway incompatibility — not every vendor accepts the same protocol dialect. For example Alibaba Cloud (DashScope) in compatible-mode, Moonshot/Kimi, Zhipu/BigModel, MiniMax, Volcengine Ark, SiliconFlow, Baidu Qianfan and other gateways may reject developer role messages or reasoning_effort echoes the way the OpenAI/Anthropic dialects expect. Turning compat.supportsDeveloperRole on against such a gateway produces 400-style errors.

2. Thinking-level wiring — the 7 levels (off / minimal / low / medium / high / xhigh / max) each need a wire value the upstream provider understands (e.g. low → "low" for one vendor, "h3" for another). Max-thinking configs and per-model reasoningEfforts are tedious to author by hand.

This plugin gives you a GUI for all of it, plus one-click presets that bake in dialect-safe configurations (see [Presets](#presets)).

Screenshots

Settings entry · Section overview (EN)

 · 

Model editor (EN) · Gateway compatibility fold (ZH) · Section overview (ZH)

 ·  · 

Features

• Per-model editor for every route:
• name, contextWindow, maxTokens — capacity fields accept plain numbers or K/M suffixes (262144, 256K, 1M).
• Supports both explicit models lists and DSH 0.1.7 catalog-preserving modelOverrides entries.
• input modalities — text / image checkboxes with de-duplication.
• Thinking toggle — switch the whole model between reasoning off and the full 7-level matrix (off/minimal/low/medium/high/xhigh/max), each level with its own wire value. Empty non-off levels are prevented (the Host rejects them), and a one-click fill all levels with the same value button is included.
• Apply field to all models of the route (name / contextWindow / maxTokens / input / reasoningEfforts).
• Per-model compat editor (folded away by default).
• Per-route editor:
• displayName, baseURL, api (openai-completions / openai-responses / anthropic-messages).
• Defaults: defaultContextWindow, defaultMaxTokens, defaultInput, reasoning, thinkingBudgets (minimal/low/medium/high), cacheRetention, transport.
• Route-level compat editor, including DSH 0.1.7 finish-reason, thinking-budget, max_output_tokens, Baseten and vLLM controls, plus an Advanced fold for timeouts, image limits, headers, and raw JSON.
• Controlled models.dev synchronization:
• Reads the cache or fetches api.json only when you click Load catalog; Fetch latest catalog explicitly refreshes it. Each local route can be mapped to a models.dev provider or excluded.
• Shows a dry-run count before writing. Only existing models whose IDs match (case-insensitively) are updated; no route or model is added or removed.
• name, context limit, output limit, and supported text/image inputs are independently selectable. Reasoning is opt-in because provider wire semantics vary: explicit effort lists are mapped, reasoning: false disables reasoning, and toggle-only/unknown formats preserve the manual value.
• A validated, compact copy of the last successful catalog is cached in browser localStorage for 24 hours. A failed refresh can use that stale last-known-good catalog; without a valid cache, the operation stops without touching settings.
• Applying a preview is one revision-fenced settings mutation. Unmanaged and manually edited fields are preserved.
• One-click presets — 7 built-in recipes plus your own saved presets:

Preset · What it does

Safe gateway · compat.supportsDeveloperRole=false, supportsReasoningEffort=true — for gateways that reject developer role messages (DashScope compatible-mode, Kimi/Moonshot, Zhipu, Mini

…(truncated)