fund-manager - SKILL.md Agent Skill

name: fund-manager description: TradingAgents-inspired decision engine — synthesizes all skill outputs into actionable portfolio-wide trading decisions with human-in-the-loop. Triggers on "fund manager run", "trading decisions", "portfolio review", "bull bear debate", "collect signals", or "what should I trade".

Fund Manager

Core question: "Given everything I know, what should I buy, sell, hold, trim, or add — and why?"

Adapted from the TradingAgents multi-agent framework. Scripts compute quantitative signals into structured artifacts; the AI agent (or human) fills in qualitative reasoning through the review workflow. No auto-execution — human approves every trade.

Workflow

Step 1: Collect Signals (Script)

Gather quantitative signals from all skill artifacts and the REST API.

uv run python skills/fund-manager/scripts/fund_cli.py collect

What the script does:

Calls GET /api/portfolio/positions — all open positions
Per ticker, collects 6 signal sources:
- Market: GET /api/analysis/signals/{ticker} — momentum, SMA crossovers, RSI, MACD, Bollinger, ATR, VWMA
- Fundamentals: GET /api/analysis/signals/{ticker} + GET /api/financials/{ticker}/metrics — P/E, margins, growth, ROE
- Thesis: data/artifacts/{ticker}/thesis/thesis.json — health score, catalysts, position type
- Risk: GET /api/analysis/risk/{ticker} + data/artifacts/_portfolio/risk/ — beta, VaR, drawdown, alerts
- Earnings: data/artifacts/{ticker}/earnings/ — surprises, upcoming previews
- Model: data/artifacts/{ticker}/model/projections.json — target price, upside/downside
Calls POST /api/analysis/risk/portfolio — portfolio-level risk metrics

Output: data/artifacts/_portfolio/decisions/{date}_signals.json

Step 2: Bull/Bear Debate (Script)

Classify signals and structure the investment debate.

uv run python skills/fund-manager/scripts/fund_cli.py debate
uv run python skills/fund-manager/scripts/fund_cli.py debate --ticker NVDA

What the script does per ticker:

Classify each signal as bull / bear / neutral
Build investment debate scaffold:
- Bull case: bull signals + thesis alignment
- Bear case: bear signals + risk factors
Build risk assessment with three perspectives (from TradingAgents' 3-way risk debate):
- Growth view (aggressive): argues for maintaining/increasing exposure
- Risk view (conservative): advocates capital preservation
- Balanced view (neutral): weighs both, proposes moderate adjustments

Output: data/artifacts/_portfolio/decisions/{date}_debates.json

Step 3: Trading Decision (Script)

Produce preliminary action recommendations from debate scores.

uv run python skills/fund-manager/scripts/fund_cli.py decide
uv run python skills/fund-manager/scripts/fund_cli.py decide --persist

What the script does per ticker:

Quantitative scoring: net bull count minus bear count
Map score → preliminary action: BUY / SELL / HOLD / TRIM / ADD
Generate decision template with placeholders for agent review
Produce markdown report with portfolio context

Output: data/artifacts/_portfolio/decisions/{date}_decisions.json + {date}_decisions.md

Step 4: Human Review (Interactive)

Review each position's debate and set final actions. No trade executes without this step.

uv run python skills/fund-manager/scripts/fund_cli.py review
uv run python skills/fund-manager/scripts/fund_cli.py review --ticker NVDA

The agent (or human) fills in:

final_action — the actual recommended action after qualitative review
reasoning — why this action (quantitative signals + qualitative judgment)
position_sizing — specific sizing recommendation
risk_adjusted — whether the risk debate changed the decision
time_horizon — immediate / this_week / this_month
catalyst_trigger — what would change this decision

Status transitions: preliminary → reviewed

Artifacts

Output goes to data/artifacts/_portfolio/decisions/:

File	Contents
`{date}_signals.json`	Raw signal bundle from all 6 sources per ticker
`{date}_debates.json`	Bull/bear signal classification + debate scaffolds
`{date}_decisions.json`	Preliminary + final action decisions
`{date}_decisions.md`	Human-readable decision report

REST API Endpoints Used

Endpoint	What we read
`GET /api/portfolio/`	Portfolio summary
`GET /api/portfolio/positions`	All open positions with weights and P&L
`GET /api/analysis/signals/{ticker}`	Aggregated quant signals (technicals, momentum, fundamentals)
`GET /api/analysis/signals/portfolio/summary`	All-position signal aggregation
`GET /api/analysis/risk/{ticker}`	Per-ticker risk contribution
`POST /api/analysis/risk/portfolio`	Portfolio-level risk (beta, VaR, Sharpe, Sortino, drawdown)
`GET /api/financials/{ticker}/income-statements`	Revenue/earnings growth
`POST /api/analysis/reports`	Persist decision report to DB

Cross-Skill Reads

data/artifacts/{ticker}/thesis/thesis.json — thesis health score, catalysts, position type
data/artifacts/{ticker}/thesis/health_checks.json — latest health check scores
data/artifacts/{ticker}/earnings/ — earnings surprises and previews
data/artifacts/{ticker}/model/projections.json — target price, upside/downside
data/artifacts/_portfolio/risk/risk_report.json — latest portfolio risk metrics
data/artifacts/_portfolio/risk/alerts.json — active risk alerts

Important Notes

No auto-execution — every trade decision requires human review (hard rule)
Decision history artifacts enable learning from past decisions — review with fund_cli.py history
Run the full pipeline with fund_cli.py run for the recommended sequential workflow
Stale signals degrade debate quality — collect fresh signals before each decision cycle
The markdown report is auto-generated for Streamlit rendering