Multi-Agent Trading Debate

Multi-agent trading debate framework for collective market decision-making. Use when a trading signal is detected or a position decision is needed. Triggers...

Security Scan

The skill's description (multi-agent debate with Feishu messaging and order execution) is plausible, but the instructions and included files have mismatches and missing pieces (notably no Feishu credentials or integration code), so the package is internally inconsistent and needs clarification before trust.

Purpose & Capability

The skill's purpose is to run a debate and notify/coordinate via Feishu and produce execution decisions. SKILL.md and references explicitly instruct sending messages to a Feishu group and posting execution commands to agent handles, but the registry metadata lists no required credentials or API tokens (no FEISHU_TOKEN/APP_ID/APP_SECRET). That is inconsistent: sending messages to Feishu normally requires credentials. Also the SKILL.md key-file paths (regime/detector.py, risk/position_sizer.py) do not match the actual file locations (scripts/regime_detector.py, scripts/position_sizer.py), which suggests sloppy packaging or incomplete integration.

Instruction Scope

Runtime instructions ask the agent to: send debate batches to a Feishu trading group, collect analyst responses, synthesize verdicts, execute orders (via @trading-execution), and write prediction logs (data/predictions.jsonl). The instructions do not show how Feishu or execution APIs are authorized, and they reference specific paths and files that don't match the provided scripts. Instructions imply network communication and potential triggering of executions but no code in the package performs API calls — leaving ambiguity about how messages/orders would be sent and who/what has authority to execute them.

Install Mechanism

This is instruction-only with two small Python scripts included; there is no install spec, no external downloads, and nothing is written to disk by an installer. That reduces supply-chain risk. The scripts optionally import numpy but do not automatically install it.

Credentials

The skill declares no required environment variables or primary credential, yet its workflow depends on an external Feishu group ID (hardcoded in references/feishu_format.md) and on sending messages and execution commands. The lack of declared tokens/credentials is disproportionate to the stated capability. Also, the hardcoded Feishu group ID is present in the package (may be sensitive), and the scripts' behavior can change based on availability of numpy (no install specified).

Persistence & Privilege

The skill does not request always:true or other elevated persistence. It writes/reads local prediction and TCA log files per SKILL.md; otherwise it does not modify other skills or global agent settings. Autonomous invocation is allowed by default but not combined here with other high privileges.

What to consider before installing

This package looks like a draft integration for team debate and trading coordination, but there are important inconsistencies you should resolve before using it with real funds or credentials: - Clarify Feishu integration: ask the author how Feishu messages and execution commands are actually sent. If the skill will post to Feishu or call an execution API, it must declare the required credentials (app id/secret or bot token) and show the code that uses them. - Do not provide any API keys or exchange credentials until you confirm who runs the network calls and whether human approval is required for execution. The SKILL.md implies order execution via an @trading-execution handle; confirm whether that is manual or automated. - Fix file/path mismatches: SKILL.md references regime/detector.py and risk/position_sizer.py but the package has scripts/regime_detector.py and scripts/position_sizer.py. Ensure the runtime instructions point to the actual code you reviewed. - Review hardcoded identifiers: the Feishu group ID is embedded in references/feishu_format.md. Confirm that ID belongs to your organization and that posting to it is acceptable. - Dependency behavior: the regime detector falls back to a simulated mode if numpy is unavailable; decide whether simulation is acceptable or whether numpy should be installed in a controlled environment. - Run the included scripts in a sandbox (no network, test data) to verify they only perform local computations and to confirm there are no hidden network calls. If you plan to enable real messaging/execution, require explicit opt-in and least-privilege credentials, and add logging/auditing and manual approval steps. Given these gaps, treat the skill as untrusted until the author provides a clear explanation of how external messaging and order execution are authorized and implemented, and until the packaging/path issues are fixed.

A structured multi-agent debate system for financial market decisions. Agents represent distinct roles: regime detection, technical analysis, on-chain data, sentiment, bull perspective, bear perspective, and final judgment.

Workflow

Signal Detected
     ↓
Regime Detection (trend/range/volatility)
     ↓
Analysts Report (technical + onchain + sentiment)
     ↓
Bull/Bear Arguments (independent positions)
     ↓
Trading Judge Verdict (confidence + position size)
     ↓
Execution → Prediction Log → Reflection

Quick Start

1. Send debate batch to Feishu trading group
2. Wait for analyst reports (timeout: 25 minutes)
3. Collect all responses
4. Synthesize verdict with confidence score
5. Execute if signal confirmed, otherwise hold

Debate Message Format

📊 【批次 #[YYYYMMDD-NNN] 自动辩论触发】

🔍 Regime: [REGIME_RESULT]

当前持仓：[DIRECTION] [SIZE] @ [ENTRY] | 当前：[PRICE]
浮亏：-[AMOUNT] USDT

请各位分析师给出独立判断：
- technical-analyst: K线形态分析
- onchain-analyst: 链上大户动向
- sentiment-analyst: 市场情绪短期
- trading-bull: 看多理由（具体点位）
- trading-bear: 看空理由（具体点位）

@trading-judge 最终裁决，截止 [TIME]

Judge Verdict Template

📋 【交易裁决 #YYYYMMDD-NNN】

Regime: [REGIME]
Confidence: [0-100]%
Signal: [BUY/SELL/HOLD]
Position: [SIZE] @ [PRICE]

Bull case: [summary]
Bear case: [summary]

Decision: [ACTION] [REASON]
Risk: [RISK_LEVEL]
Next review: [TIME]

Position Sizing

Confidence	Max Position	Stop Loss
>80%	100%	Tight (1.5%)
60-80%	75%	Standard (3%)
40-60%	50%	Wide (5%)
<40%	25% or hold	No entry

Regime Detection Rules

Regime	Description	Preferred Strategy
TRENDING	ADX > 25, clear direction	Momentum follow
RANGING	ADX < 20, no clear trend	Mean reversion
VOLATILE	High variance, uncertain	Reduced size, wider SL

Prediction Log Entry

After each decision, record in predictions.jsonl:

{"timestamp":"ISO8601","regime":"TRENDING","signal":"HOLD","confidence":62,"reason":"No clear entry, ranging market","actual_outcome":"pending","pnl":0}

Key Files

• Regime detector: regime/detector.py (ADX + trend analysis)
• Position sizer: risk/position_sizer.py (Kelly + risk budget)
• Prediction log: data/predictions.jsonl
• TCA log: data/tca_log.jsonl

When to Trigger

✅ Trigger debate:

• Cron window: 08:00, 12:00, 16:00, 20:00 CST daily
• Price deviation > 5% from entry
• Major news event detected
• Manual request

🚫 Skip debate (hold current position):

• Confidence < 40%
• Regime unclear
• Risk/reward < 1.5