icy-replit

byKishan Sharma

# Replit Agent Prompt: Indian Intraday AI Paper-Trading System — Trial Version (PAPER TRADING ONLY) Build this complete system in one pass — no phases, no check-ins mid-build. Give one final report at the end: what was built, what was tested, what needs my attention. ## Safety rules (absolute) - PAPER TRADING ONLY. No real orders, no broker integration, no order-placement code anywhere, ever. - All trades are simulated: writing a row to our own `trades` table, nothing else. - No live broker account, no OAuth, no credentials to manage — this version has no broker connection at all. ## Data sources (two separate things — do not conflate them) 1. **Visual chart on the dashboard**: embed TradingView's free widget (their standard embeddable chart widget) showing the live candlestick chart for whatever stock is currently the active candidate or open position. This is for display only. 2. **Numeric data for the agents to actually analyze** (price, OHLC, volume — TradingView's widget does NOT expose this to our code): pull this from a free public source — use the `yfinance` Python library (Yahoo Finance data) as the default, with NSE symbols (`.NS` suffix). Structure this behind a simple `DataProvider` interface so the source can be swapped later without touching agent logic. ## What it does Once started, runs autonomously through NSE hours (09:15–15:30 IST, configurable) with a manual STOP/HALT control: scan → rank → analyze → risk-check → decide → paper-execute → monitor → exit → learn → repeat. ## Parameters (config file, not hardcoded) - Starting capital ₹15,000; kill switch at ₹5,000 equity - Max 1 open position; max risk 5% of equity per trade; min net R:R 1.5 - Eligible stocks: ≤ ₹1,500/share; top 5 candidates (long + short) - Opportunity Scout re-scans every 15 min (fixed); an open trade is monitored continuously and is never interrupted by a new scan - Flat cost: ₹25/transaction → ₹50 per round-trip trade; net_profit = gross − 50; R:R check uses net, not gross ## Agents — only these six 1. **Opportunity Scout** — deterministic ranking from the DataProvider, filtered ≤₹1,500/share, shortlists top 5 (long + short) every 15 min. LLM only writes a short rationale, never picks candidates. 2. **Technical Agent** — LLM-based; given one candidate, reads chart structure (support/resistance, trend, momentum, volume) and proposes PROCEED/WAIT with a full plan: entry, target, stop-loss. 3. **Regime Agent** — LLM-based; classifies overall market context (trending/choppy/high-vol/low-vol) so a setup isn't judged in isolation from the broader market. 4. **Risk Manager** — 100% deterministic code, no LLM, hard veto power. Enforces max risk %, max 1 position, kill switch, price cap. Cannot be overridden by anything else. 5. **Main AI** — sees Technical + Regime output + Risk Manager's clearance; makes the sole final PROCEED/WAIT decision using **net** R:R (after ₹50 cost) against the 1.5 minimum. Low/fixed temperature, strict schema output, cannot override the Risk Manager. 6. **Learning Engine** — logs every closed trade's full context (setup, reasoning, outcome) to a trade journal; can retrieve similar past cases as context for Technical/Regime agents. Passive only — no automatic changes to parameters or prompts without my explicit approval. ## Non-agent backend logic (still needed, but plain deterministic code, not an "AI agent") - **Trade execution**: once Main AI says PROCEED, write the simulated trade (entry, target, stop-loss, size) to the DB. No LLM involved. - **Position monitoring**: continuously check the open position's live price (from the DataProvider) against its stop-loss/target/time-based exit, and close it when hit. Runs independently of Scout's 15-min scan cycle — a live position is never interrupted by a new scan. ## Dashboard (separate pages, not one long page) - **Live Trading**: TradingView widget for the active/open symbol, current position with entry/target/stop-loss, equity vs. ₹15,000 and ₹5,000 kill line, STOP/HALT control. - **Trade History/Journal**: closed trades log from the Learning Engine. - **Settings**: editable config parameters (Section above). ## Do not add - Any real broker connection or order placement, under any framing. - Any agent beyond the six listed above (no News Agent, no Critic, no separate Market Data Agent). - Auto-tuning of parameters/prompts by the Learning Engine without my approval. - Hardcoded universe/thresholds/costs outside the config file. I need only dashboard and no other pages. Don't ask for api key. Don't write essay on the dashboard I know the details. NO need to write agents descriptions.