Wei Cross Research Skill#

Overview#

Use wei-cross-research when you need a reliable answer — not just one model's opinion.

This skill queries multiple LLMs in parallel and uses a judge model to synthesize their responses into a single cross-validated answer. When models agree, confidence is high. When they disagree, the disagreement is surfaced — not silently resolved.

Best for:

• High-stakes questions where a wrong answer has real consequences
• Topics where a single model may have blind spots or biases
• Analysis that benefits from multiple independent viewpoints
• Reducing hallucination via cross-model comparison

Cost note: This skill queries 2–3 models per request. Expect approximately 2–3x the token usage of a single-model query. Use it when answer quality justifies the cost; avoid it for simple or low-stakes questions.

---

Requirements#

• Bun >= 1.0

Install Bun#

bash

curl -fsSL https://bun.sh/install | bash

Or on macOS with Homebrew:

bash

brew install oven-sh/bun/bun

Environment Variables#

bun supports accessing environment variables from .env file in the project root.

Note: If you don't have an OpenRouter API key, you can modify config.json to use other providers. Change the provider field from "openrouter" to "bailian" or "openai_compliant" and update the api_key_env to point to your available API key.

Configuration Files#

遇到模型访问问题？ 请参考 README.md 了解如何根据你的网络环境选择和配置 config.json。

Install Dependencies#

bash

bun install

Usage#

bash

bun run scripts/index.ts "your question"
bun run scripts/index.ts -t financial "美联储2026年会降息吗？"

Domain-Specific Judges#

When queryType is set to "financial", the judge step uses a finance-specialized prompt that produces:

• Base Case Analysis — probabilistic scenario with data-driven reasoning
• Bull Case — arguments for upside scenario
• Bear Case — arguments for downside scenario
• Key Variables / Risks — macro events, earnings, policy changes, market sentiment

This avoids deterministic predictions and enforces probability ranges (e.g., 60–70% likelihood). Use it for investment, macroeconomic, and market analysis questions.

Example:

json

{
  "query": "美联储2026年会降息吗？",
  "queryType": "financial"
}

---

Supported Models#

All models are accessed via OpenRouter or other configured providers. Answering models may use live retrieval depending on the provider configuration.

The system selects 2–3 answering models in parallel (based on roles) and uses a judge model to synthesize the final response.

Model Roles#

Each model in config.json is tagged with one or more roles indicating its capabilities:

RoleDescriptionTypical UseretrievalHas web/live data accessCurrent events, real-time infocodingStrong programming capabilityTechnical implementation, debuggingsocialSocial media data accessX/Twitter sentiment, trendingreasoningDeep analytical capabilityComplex analysis, synthesiscreativeCreative writing strengthStorytelling, open-ended taskslongcontextLarge context windowDocument analysis, long inputsgeneralBroad balanced capabilityFallback, ambiguous queriesjudgeAnswer synthesisFinal synthesis (judge models only)

Note: Specific model names and their roles are defined in config.json → models. Refer to that file for the current model roster.

Judge Models#

Judge models synthesize answers already in context and normally do not require retrieval.

They are configured in config.json with role "judge" and selected via the judge_model config key.

Judge models are independent of answering models and may synthesize outputs from any answering pool.

---

Model Selection#

Model selection is controlled via config.json using a roles-based routing system. Instead of hard-coding model names, you select models by the capabilities (roles) they provide.

How to Select Models#

As the calling model, follow this process:

1. Classify the query — Match keywords to determine the queryType
2. Pass queryType — The skill will look up the routing.xxx.models in config.json
3. (Optional) Pass explicit models — Use the models parameter to bypass auto-selection

Query Types (Domain)#

queryTypeDescriptionfinancialMarkets, investing, macroeconomicstechnicalProgramming, systems, engineeringsocialPublic opinion, social media sentimentcurrent_eventsRecent news and real-time informationscientificObjective knowledge, definitions, theoriescreativeWriting, design, ideationgeneralDefault fallback

Intent (Task Type)#

In addition to queryType, queries may include an optional intent field. queryType defines the domain (what the question is about), while intent defines the task (what to do with the question). If intent is not provided, the system defaults to analysis for complex queries and lookup for simple factual queries.

intentDescriptionlookupRetrieve factual informationanalysisDeep reasoning and explanationcomparisonCompare multiple entitiespredictionForecast future outcomes (used in financial)generationCreate content (text, ideas, design)

Example:

{ "query": "美联储2026年会降息吗？", "queryType": "financial", "intent": "prediction" }

Selection Algorithm#

text

1. Analyze query → match keywords → determine queryType
2. Pass queryType to skill → skill looks up `routing.<queryType>.models` in config.json
3. Skill selects top 2–3 models from the routing config
4. If queryType === 'financial', skill uses judge_financial.txt for synthesis

Examples#

Example 1: Financial Query#

Query: "美联储2026年会降息吗？"

Selection process:

1. Keywords: 美联储, 降息 → queryType: financial
2. Pass to skill: { "query": "...", "queryType": "financial" }
3. Skill looks up: config.json → routing.financial.models
4. Skill selects: First 2 models from the routing config
5. Judge: Uses judge_financial.txt (Bull/Bear/Base Case analysis)

Example 2: Technical Query#

Query: "How do I implement a distributed transaction?"

Selection process:

1. Keywords: implement, distributed → queryType: technical
2. Pass to skill: { "query": "...", "queryType": "technical" }
3. Skill looks up: config.json → routing.technical.models
4. Skill selects: Models configured for technical queries

Example 3: Social Query#

Query: "What are people saying about SpaceX on Twitter?"

Selection process:

1. Keywords: Twitter, saying → queryType: social
2. Pass to skill: { "query": "...", "queryType": "social" }
3. Skill looks up: config.json → routing.social.models
   • Note: grok-4.1 has social, sentiment, trending roles + X data access

When to Reference Specific Models#

Only hard-code model names when:

1. Special data access — e.g., grok-4.1 for X/Twitter data, kimi-k2.5 for 200K context
2. Known strengths — e.g., qwen3.5 for coding tasks based on benchmarks
3. Avoiding specific models — e.g., excluding models known to underperform for certain tasks

In these cases, document why that specific model is needed, not just its name.

---

When To Use This Skill#

Use this skill when:

• The user asks a complex research question
• The question requires high confidence or cross-validation
• The topic has multiple competing viewpoints
• A factual error would have significant consequences

Do NOT use this skill for:

• Simple factual lookups
• Quick definitions or summaries
• Trivial tasks a single model can answer reliably
• Time-sensitive queries where 8–15s latency is unacceptable

---

Skill Parameters#

ParameterTypeDescriptionquerystringThe research questionqueryTypestringDomain classification (financial, technical, etc.)intentstringTask type (analysis, prediction, etc.)modelsarrayOverride model selectionmaxModelsnumberMax modelsdepthstringsimple / treejudgeModelstringOverride judge

Note: The domain parameter has been deprecated. Use queryType: 'financial' instead for financial queries.

Example:

json

{
  "query": "What are the economic impacts of AI agents?",
  "queryType": "general",
  "intent": "analysis"
}

json

{
  "query": "美联储2026年会降息吗？",
  "queryType": "financial",
  "maxModels": 2
}

Depth Modes#

ModeBehaviorUse Whensimple (default)Single-pass: each model answers the query once, judge synthesizesMost research questionstreeMulti-pass: follow-up sub-queries are generated and answered before synthesisComplex topics requiring decomposition (adds ~10–20s latency)

---

Output Format#

Success (all models respond):

json

{
  "query": "user question",
  "models_used": ["glm-5", "kimi-k2.5"],
  "answers": [
    { "model": "glm-5", "answer": "..." },
    { "model": "kimi-k2.5", "answer": "..." }
  ],
  "final_answer": "...",
  "confidence": 0.85
}

Partial failure (one model timed out or errored):

json

{
  "query": "user question",
  "models_used": ["glm-5"],
  "models_failed": [
    { "model": "kimi-k2.5", "reason": "timeout" }
  ],
  "answers": [
    { "model": "glm-5", "answer": "..." }
  ],
  "final_answer": "...",
  "confidence": 0.61,
  "warning": "Synthesis based on partial responses. Confidence may be reduced."
}

Full failure:

json

{
  "query": "user question",
  "models_used": [],
  "models_failed": [
    { "model": "glm-5", "reason": "timeout" },
    { "model": "kimi-k2.5", "reason": "api_error" }
  ],
  "final_answer": null,
  "error": "All models failed. Please retry."
}

Confidence scale: All confidence values use a 0–1 scale (e.g., 0.85 = 85% confidence). This applies consistently across normalizer outputs and judge outputs.

---

Result Files#

Each run produces files identified by a shared timestamp in YYYY-MM-DDTHH-MM-SS format (ISO 8601, colons replaced with hyphens).

The timestamp is logged at the start of execution:

text

[ResearchAgent] Timestamp: 2026-03-19T14-30-05

File Locations#

FilePathContentReportreports/report-{timestamp}.txtFinal synthesized answer from judgeModel responsesintermediate/{model}-{timestamp}.txtRaw response from each answering modelJudge rawintermediate/{judge}-{timestamp}.txtRaw judge synthesis output

Example#

For a run at 2026-03-19T14:30:05 with models kimi-k2.5 and gpt-5.4, judge glm-5:

text

reports/report-2026-03-19T14-30-05.txt        ← final answer
intermediate/kimi-k2.5-2026-03-19T14-30-05.txt
intermediate/gpt-5.4-2026-03-19T14-30-05.txt
intermediate/glm-5-judge-raw-2026-03-19T14-30-05.txt

Use the timestamp from console output to locate all files from a specific run.

---

Performance Characteristics#

StageTypical LatencyRouter~1s (skipped when models passed directly)Model inference (parallel)20–100sJudge synthesis20-60sTotal****40–120s

Timeout per model: 60-120 seconds Retries per model: 1

---

Failure Handling#

The skill tolerates partial failures:

• If a model times out or errors, the skill continues with remaining responses
• The judge synthesizes available answers and notes missing models in output
• If all models fail, a structured error is returned (see Output Format above)
• The router has a default fallback pair ( glm-5 + kimi-k2.5) if routing fails

---

Security Notes#

• User-supplied query values are included in prompts sent to external model APIs. Avoid passing unsanitized inputs from untrusted sources.
• The skill does not validate or filter query content — callers are responsible for input sanitization upstream.
• Do not include secrets, PII, or confidential data in queries unless the target model APIs are approved for that data classification.

---

Quality Evaluation#

A synthesized answer is considered high quality when:

• Consensus points across models are clearly identified
• Disagreements are surfaced (not silently resolved)
• Confidence ≥ 0.75
• The judge does not fabricate citations or sources

For ongoing quality tracking, log confidence, models_used, and models_failed per request.

---

Best Practices#

Recommended Model Combinations#

intentRole CombinationExamplelookupretrieval + generalQuick factual lookup + balanced fallbackanalysisreasoning + retrievalDeep analysis + live data contextpredictionreasoning + synthesisForecast with multi-source synthesiscomparisonreasoning + structuredEvaluate options systematicallygenerationcreative + synthesisCreate + refine outputqueryTypeRecommended RolesWhyfinancialretrieval + researchLive data + analysistechnicalcoding + generalTechnical + broader contextsocialsocial + retrievalSentiment + current contextcreativecreative + synthesisGenerate + refine

Why Role Diversity Matters#

Combining models with different roles improves reliability:

• retrieval + reasoning: Up-to-date facts + deep analysis
• coding + general: Technical accuracy + broader context
• social + retrieval: Platform-specific sentiment + general web context

Benefits:

• Higher reliability through capability diversity
• Reduced hallucination via cross-validation
• Improved reasoning quality on ambiguous topics

---

Example Usage#

text

use cross-research

query="What are the major AI breakthroughs in the past 12 months?"
queryType="current_events"

Selection process:

1. Keywords: "past 12 months" → implies current_events
2. Pass queryType: "current_events" to skill
3. Skill looks up config.json → routing.current_events.models
4. Judge synthesizes responses

Example result:

text

Final Answer:
AI breakthroughs in the last year include...

Consensus:
- Agent frameworks matured significantly
- Multimodal models expanded in capability
- Inference costs decreased substantially

Confidence: 0.87

---

Changelog#

VersionChanges1.0.3updated metadata1.0.2model updates; bug fixes1.0.0Initial release