xingzhen199186/dsh-advisor-group

A DeepSeek Harness (DSH) plugin that lets the main model consult multiple expert advisor models in a retro chat-group card — for professional, long-tail world-knowledge, high-risk, or uncertain questions.

📖 中文文档 / Chinese: README.zh.md

---

✨ Features

• Auto-deepen consultation pipeline — one ask_advisors call runs up to maxRounds rounds automatically: each round is a driver deep-question → advisor A → advisor B (sees A) → advisor C (sees A+B) → … sequential relay, closed by a driver-generated synthesis conclusion.
• Zero-config driver model — the driver formulation of deep follow-ups reuses the current agent's provider/model, so no extra API key or model setup is needed; it falls back to discussion.driverModel when the session header is unavailable.
• Three ways to activate — @顾问群 mention (force-start), same question repeated 3 times without resolution, or main-model self-assessed confidence below the threshold.
• Optional Jev semantic pre-classification (off by default) — when trigger.jev.enabled is on, a configured Jev model judges non-force-started consultations first (escalation / high-risk / web-search fit, thresholds under trigger.jev.*); if Jev is unavailable the local rules-based classifier still decides. With trigger.jev.useEnglishState, the judge reads the caller-supplied English gist (questionEn) instead of the Chinese question — display and session records stay Chinese.
• Retro CRT chat cards — green/amber/blue CRT themes, scanlines, LIVE/DONE headers, auto-expanded thinking panel with auto-scroll; advisor Markdown rendered with a link-protocol whitelist (headings, lists, code, quotes, links, tables).
• Provider presets (11 platforms · 26 presets) — DeepSeek, Moonshot Kimi, Kimi Code, Aliyun Bailian, Zhipu AI, OpenAI, Claude, Gemini, SiliconFlow, AIHubMix, OpenRouter (OpenAI/Anthropic-compatible variants included).
• Security-minded by design — API keys use official SecretField semantics (never returned to the browser; apiKeysByProvider key history is server-side only), SSRF-guarded diagnostics (https-only / loopback, no IP literals, no redirects), per-boot token auth on /advisor-group/* routes, and an atomic configurable daily consultation cap (default 50, can be disabled) persisted across restarts.
• Runtime toggle — toggle_advisor_group enables/disables the plugin and persists the flag to settings.

✅ Compatibility

Surface	Status
Harness	DeepSeek Harness 0.1.7-rc.2 → 0.2.x (incl. 0.2.0-rc.1)
Node	^22.19.0 || >=24.0.0
Platforms	DSH Web and the desktop app (same client bundle) + headless host logic

📦 Installation

# 1. Install the plugin (one package, everything included)
dsh plugin --profile web add dsh-advisor-group

# 2. Restart DSH web
npx @deepseek-ai/dsh web

From source (development)

npm install --legacy-peer-deps --no-audit --no-fund
npm run build
dsh plugin --profile web add ./dsh-advisor-group-<version>.tgz   # after npm pack (version from package.json)

Desktop app: the desktop app owns its own profile — dsh plugin --profile desktop … is refused by design, not by permissions. Install from inside the app instead: Plugins in the sidebar → Add plugin → enter the package name or the absolute path of a local tarball.

🚀 Quick start

1. Restart DSH web and hard-refresh the browser (Ctrl+Shift+R).
2. Go to Settings → Plugins → Advisor Group: configure your advisors (provider route + model; use 获取模型列表 to pull the authoritative model list) and tune maxRounds, thresholds, and the UI theme.
3. Just start a conversation:
   • type @顾问群 in your question to force a consultation, or
   • ask a professional/uncertain question — the plugin escalates automatically when appropriate.

ask_advisors tools available to the model: ask_advisors, toggle_advisor_group.

⚙️ Configuration

Key	Type	Default	Description
enabled	boolean	true	Enable advisor group
discussion.maxRounds	number	2	Total rounds of the auto-deepen pipeline (each round = driver question + all advisors in relay)
discussion.maxAdvisorsPerCall	number	3	Max advisors per call (1–10)
discussion.autoDeepen	boolean	true	Run the auto-deepen pipeline (driver follow-ups + final synthesis)
discussion.driverModel	object	–	Fallback driver model {provider, model} when the session header cannot be read
discussion.advisorTimeoutMs	number	600000	Per-advisor call timeout (ms, 1000–600000), applied to both channels
discussion.driverTimeoutMs	number	600000	Driver generation timeout (deep-question / conclusion; ms, 1000–1200000)
discussion.advisorTools	string	'readonly'	Global default advisor tool scope when an advisor sets no own tools: readonly / all (every session-visible tool incl. writable — elevated risk, see Security) / off. Tool calling needs the direct-http (OpenAI/Anthropic) channel.
quota.enabled	boolean	true	Enable the daily new-consultation cap (cost safety valve)
quota.maxPerDay	number	50	Max new consultations per UTC day (1–100000); ignored when quota.enabled is false
trigger.requireClassifier	boolean	true	Run the pre-classifier before starting
trigger.allowWebFallback	boolean	true	Allow classifier to recommend web search
trigger.confidenceThreshold	number	0.6	Escalate when main-model confidence is below this
trigger.jev.enabled	boolean	false	Use the external Jev model for semantic pre-classification of non-forced consultations; falls back to the local rules-based classifier when Jev is unavailable
trigger.jev.provider / trigger.jev.model	string	'typesafe' / 'jev-latest'	Jev route (typesafe or openrouter) and model; key via trigger.jev.apiKey (secret) or trigger.jev.apiKeyEnv; optional baseURL, timeoutMs (default 10000)
trigger.jev.confidenceThreshold	number	0.6	Escalate when Jev answers needsAdvisor yes (or its 0–1 score reaches this value)
trigger.jev.highRiskThreshold	number	follows jev.confidenceThreshold	Flag high-risk when Jev answers high_risk yes (or its 0–1 score reaches this value); unset = follows trigger.jev.confidenceThreshold
trigger.jev.useEnglishState	boolean	false	Judge using the caller-supplied English gist (questionEn) instead of the Chinese question; display and session records stay Chinese
ui.theme	string	retro-green	Chat card theme (retro-green / retro-amber / retro-blue)
ui.showTimestamps	boolean	true	Show timestamps
ui.autoExpand	boolean	true	Auto-expand card
advisors[].tools	string	–	Per-advisor tool-scope override (readonly/all/off; unset follows the global default)
advisors	array	[]	Advisor list (provider/model/baseURL/apiKey/apiKeyEnv/protocol…, keys are role('secret'))

discussion.parallel and discussion.stopOnConsensus are deprecated leftovers kept only for stored-config compatibility.

🧮 Daily quota

The plugin ships a cost safety valve: it counts new consultations per UTC calendar day, default cap 50, changeable or switchable on the settings page. Only new consultations count — resuming (「▶ continue」) and follow-ups never consume quota. The check and the increment happen inside one synchronous block, so two consultations cannot slip through the same window; the counter is written atomically to $DSH_HOME/storages/advisor-group/daily-guard.json, survives restarts, and resets at the UTC day boundary. When the cap is reached ask_advisors does not start and returns a one-line explanation; the settings page shows "today's remaining consultations X / Y", and with the cap off the plugin still counts for display without blocking.

💸 Usage and cost

One consultation is not one call but a chain of them. With the defaults (maxRounds = 2, at most 3 advisors) a new consultation costs roughly 9 model calls: 2 rounds × (1 driver follow-up + 1 per advisor) + 1 driver synthesis. Advisors that use tools add one more model call per tool round (at most 4 rounds per advisor plus one forced text-only round), so the real count can be noticeably higher than 9.

Billing follows each channel: advisors bill on the route and model you gave them, while the driver bills on the current session's model. To keep costs down, lower maxRounds, configure fewer advisors, set tools: 'off' for advisors, and pick a daily cap you are comfortable with. Resuming after an interruption ("▶ continue") only asks the advisors that had not finished, so it costs less than starting over.

🔒 Security & privacy

• Direct API keys can be stored in the DSH settings.yaml (marked secret, never returned to the browser by describe); prefer apiKeyEnv (env-var mode) if you don't want keys on disk.
• /advisor-group/* routes use a per-boot shared token for local single-user use — not multi-user auth. Add a reverse-proxy auth layer before LAN exposure.
• Diagnostics endpoints resolve the real key server-side and validate the target URL against an https-only / loopback SSRF guard; DNS-rebinding protection is a documented out-of-scope limitation.
• The daily consultation cap (configurable via quota.*; default 50; can be disabled) is persisted to $DSH_HOME/storages/advisor-group/daily-guard.json (UTC day key), so restarts don't reset it.
• Classifier shadow mode appends one observation sample per non-forced classification (read-only /advisor-group/shadow), used for threshold tuning only — never influences behavior.
• advisorTools: 'all' is an elevated-risk scope. It exposes every session-visible tool to the advisor models, including writable/execution ones (pwsh, bash, write, config/SSH tools), executed through the official guarded pipeline. Only enable it for advisors you trust (e.g., your own local models), keep them on the direct-http channel, and note that every non-read-only invocation is logged with console.warn for audit. Prefer the default readonly (read/grep/glob/web_search/web_fetch/scan_discover/list_imported_sessions) or off.

⚠️ Known limitations

• A hard crash of the DSH host can leave an already-open card showing LIVE until the page is refreshed (a disconnect notice now appears); refresh restores the true state from the session log.
• Risk notes are fact-based, never text-based: a consultation reports truncation (an advisor stream cut off by timeout or network) and cancellation/stop. Users see them on stopped or completed cards and in the session log, and the main model receives them in the ask_advisors result. Advisor prose is not scanned for keywords — the earlier text heuristic was removed in 0.1.1 because it also fired on plain negations such as 「没有任何风险」.

🛠️ Development

npm install --legacy-peer-deps --no-audit --no-fund
npm run typecheck
npm test        # vitest suite, incl. provider streaming contracts against a local fake LLM server
npm run build   # tsdown; client bundle must NOT be built with minify: true

📄 License

Apache License 2.0 © 2026 dsh-advisor-group contributors.