JuneLearn/dsh-reasoning-settings
Add Low, Medium, High, and other reasoning levels to DeepSeek Harness's third-party API to support low, medium, and high reasoning levels and model selection for each sub-agent call
DSH Reasoning Settings adds an independent reasoning-effort (thinking intensity) settings page for custom llm-pi-ai providers in DeepSeek Harness — the official UI only exposes effort selection for first-party DeepSeek models, so third-party/relay APIs get a dedicated control. It auto-reads the custom providers and models you already added on the official Models page, declares a per-model effort level (off/minimal/low/medium/high/xhigh/max), customizes the actual wire value sent to the API for each level, sets a provider-wide default effort restricted to the levels supported by every model in that provider, and configures the reasoning wire format for openai-completions providers — all written back through the official settings.mutate API with API keys never touched. It also fixes subagent behavior: in-process subagents inherit the parent's live third-party provider/model selection and an explicit parent reasoning effort, model-only child overrides resolve against your configured third-party providers, and the subagent/subagent_fork tools gain per-call provider/model/reasoning_effort fields without replacing their official lifecycle.
Install
dsh plugin --profile web add github:JuneLearn/dsh-reasoning-settingsGitHub source install per README (npm package not found on the registry as of 2026-08-24): dsh plugin --profile web add github:JuneLearn/dsh-reasoning-settings, or from a DSH source checkout: pnpm dsh plugin --profile web add github:JuneLearn/dsh-reasoning-settings. Prerequisites: Node.js (22.19.x or 24+), Git, and pnpm (dsh plugin invokes it in the profile directory). The README also notes a third-party relay recommendation with a referral link (WPIronman API) — the plugin itself is independent of that service and does not require any particular relay.
Compatibility
DeepSeek Harness web profile with custom llm-pi-ai providers. Adds an independent reasoning-effort (thinking intensity) page for custom/third-party providers — the official UI only exposes effort selection for first-party DeepSeek models. Features: auto-reads custom providers/models from the official Models page; per-model effort level off/minimal/low/medium/high/xhigh/max with customized wire values; provider-wide default effort restricted to levels every model supports; reasoning wire format config for openai-completions providers; writes back through the official settings.mutate API without touching API keys; fixes in-process subagents inheriting the Agent's create-time official model instead of the parent's live third-party selection; makes subagents inherit an explicit parent reasoning effort; resolves model-only child overrides (e.g. grok-4.5) against user-configured third-party providers; extends the ordinary subagent/subagent_fork tools with per-call provider/model/reasoning_effort fields without replacing their official lifecycle.
Details
- Repo: JuneLearn/dsh-reasoning-settings
- Category: Agent Capabilities
- Stars: 6
- Version: GitHub source install github:JuneLearn/dsh-reasoning-settings (README-documented; npm 404 as of 2026-08-24)
- Last push: 2026-08-16
- First seen: 2026-08-14
Recent updates
The current English README documents: features (per-model effort levels, wire value customization, provider-wide defaults, reasoning wire format, settings.mutate write-back, subagent provider/model/effort inheritance), the live demo, install (github: source, prerequisites), and the third-party relay recommendation note (referral link disclosed).
FAQ
- Why do I need this if DSH already has reasoning settings?
- The official UI only exposes reasoning-effort selection for first-party DeepSeek models. This plugin makes it work for custom llm-pi-ai providers and relay APIs too.
- Does it touch my API keys?
- No — everything is written back through the official settings.mutate API and API keys are never touched.
- How do subagents inherit settings?
- In-process subagents inherit the parent's live third-party provider/model selection and an explicit parent reasoning effort; the subagent/subagent_fork tools gain per-call provider/model/reasoning_effort fields.
Alternatives
MEMZ-JZY/DSH-Claude-Style-Reasoning-Slider · qjcnmd/dsh-reasoning-slider · vvvspec/better-reasoning-slider