dsh-effort-config

A DeepSeek Harness (dsh) plugin that lets you configure reasoning-effort levels for third-party models from the settings page, writing to the llm-pi-ai settings namespace. Level selection reuses the native model picker's Effort panel (composer model seat / /model command) — the selector is not reimplemented; it appears automatically once a model declares levels.

Demo Screenshots

Features

• Settings-page entry: a new "思考努力度" (Reasoning Effort) page in the settings panel (settings.section), showing only configured providers, newest additions first
• Add provider from the header (settings.action button): the candidate list puts already-configured (your own) providers first, hand-written routes before catalog routes; clicking an unconfigured built-in catalog route enables it (writes an empty profile; models and protocol come from the pi-ai catalog)
• Level editing: each model can declare the 7 levels off / minimal / low / medium / high / xhigh / max plus a wire spelling (defaults to the level name)
  • nothing checked = stays undeclared (field not written)
  • only off checked = no reasoning (reasoningEfforts: false)
  • non-off levels require a non-empty wire value; empty wire values or unknown level names are rejected loudly and never persisted
• Route default level (reasoning) and Anthropic token budgets (thinkingBudgets, shown only when the protocol is anthropic-messages)
• Delete a provider (together with all of its reasoning-level configuration), with two-step confirmation
• Persistent: writes land in settings.yaml under llm-pi-ai and survive restarts
• Live sync: the model picker refreshes right after saving (via the native llm/adapters-updated and settings/document-updated events)
• Bilingual UI: copy follows the Harness interface language automatically (中文 / English)

How it works

All data flows through existing wire remotes — no custom Host Remote (no typert codegen toolchain needed):

• connection.api.settings.describe / replace — read/write the llm-pi-ai user layer (optimistic concurrency: expectedRevision plus replay-and-retry on conflict)
• connection.api.llm.providers / models — the configurable-provider directory (displayName / declared / active) and the host model catalog (model enumeration for catalog routes)
• remote.$on('llm/adapters-updated' | 'settings/document-updated') — automatic refresh

Level semantics follow the dsh-llm-pi-ai reasoningEfforts contract (keys fixed to the 7 levels; values are the provider-side wire spellings; off may be left null = send nothing). Writes are additionally guarded by the settings seam's schema + validate (e.g. "offers no level beyond off" is rejected at write time), so invalid values never persist.

Ordering semantics

• Main list: reverse of the providers key order in settings.yaml (newest first)
• Add-provider candidate list: configured (your own) first, hand-written routes before catalog routes

Differences from the dynamic Cordis plugin

This package runs in the browser realm and is unaffected by the dynamic host's node:vm prototype-identity issue (the settings seam's isPlainObject checks against the host Object.prototype, so objects created inside the sandbox realm are rejected). The write path is therefore simpler: deep-copy the user layer → apply operations → settings.replace.

Installation

# Using the dsh CLI (recommended): forwards to pnpm in the web profile
dsh plugin --profile web add /home/benzhoupo/ds-plugins/dsh-effort-config

# Or manually:
# cd ~/.dsh/profiles/web && pnpm add file:/home/benzhoupo/ds-plugins/dsh-effort-config

Then restart dsh web (the patch row and the browser bundle need to be recomposed). To uninstall:

dsh plugin --profile web remove dsh-effort-config

Usage

1. Open Settings → "思考努力度" in the left nav (between "模型" and "插件")
2. If you have no provider yet: click "添加提供商" in the header → pick one from the candidate list (configured providers come first and are marked "已配置"; they cannot be re-added)
3. On a provider card:
   • For each model, click "编辑档位" → check levels (wire defaults to the level name; edit as needed) → Save
   • Use the "路由默认档位" dropdown to set the request default level
   • Anthropic-compatible providers (protocol anthropic-messages) show "Anthropic token 预算"
4. Back in the composer: the model seat shows "model name · level" and the Effort panel lists all declared levels — pick one directly
5. The "删除" button (two-step confirmation) removes the provider and all of its reasoning configuration

settings.yaml example

llm-pi-ai:
  providers:
    bearlab:
      displayName: BearLab
      apiKeyEnv: BEARLAB_API_KEY
      api: openai-responses
      baseURL: https://bearlab.ai/v1
      models:
        - id: gpt-5.6-sol
          name: GPT 5.6 Sol
          contextWindow: 256000
          reasoningEfforts:        # hand-written route: written on the models[] entry
            xhigh: xhigh           # wire value = provider-side spelling (openai-responses → reasoning.effort)
            off: null              # off left empty = send no parameter
      reasoning: xhigh             # route default level
    anthropic:                     # catalog route: goes through modelOverrides
      modelOverrides:
        claude-sonnet-4-5:
          reasoningEfforts:
            high: high
      thinkingBudgets:             # meaningful only for the anthropic protocol (budget_tokens)
        minimal: 1024
        low: 2048
        medium: 4096
        high: 8192

Wire mapping cheat sheet

Troubleshooting

• "思考努力度" missing from settings: make sure dsh web was restarted and the package is installed (dsh plugin --profile web why dsh-effort-config)
• Page says "llm-pi-ai 命名空间未注册": @deepseek-ai/dsh-llm-pi-ai is not loaded or was removed
• Model list empty right after adding a provider: the route was just registered — click "刷新" (catalog enumeration depends on adapter registration completing)
• Default model points at a model with no declared levels: requests fail with UNSUPPORTED_REASONING_EFFORT before network I/O (native pi-ai behavior); declare levels for that model or clear the default level

License

CC BY 4.0 (Creative Commons Attribution 4.0 International), attribution: benzhoupo. You are free to share and adapt the work, provided you give appropriate credit.