| name | indian-algo-trading |
| description | Write production-quality Python automated trading strategies for the Indian stock market. Covers the full lifecycle: backtesting, optimization, paper trading, and live deployment across equity, F&O, currency derivatives, and MCX commodities. Bakes in best practices for risk management, position sizing, order handling, and Indian market-specific rules. Rupeezy/Vortex is the primary broker with a template for community broker adapters.
MANDATORY TRIGGERS: Use this skill whenever the user mentions algo trading, automated trading, trading bot, trading strategy, backtest, backtesting, strategy code, quant strategy, systematic trading, or any task involving writing Python code to trade on Indian stock exchanges (NSE, BSE, MCX). Also trigger when the user mentions Rupeezy, Vortex API, vortex-api, or asks about F&O strategy automation, options selling bot, intraday strategy, or positional strategy. Even if the user just says "write a strategy" or "help me automate my trading" — use this skill.
|
| version | 1.1.14 |
Indian Algo Trading — Strategy Writing Skill
Write production-quality Python trading strategies for Indian markets. Every strategy must be safe enough to run with real money.
Before Writing Any Code
Step 1: Understand the User's Intent
Ask (skip any already answered):
- Asset class — equity, F&O, currency, commodities?
- Live or backtest?
- Broker — Rupeezy/Vortex (default; see
references/brokers/rupeezy-vortex.md) or other (check references/brokers/)?
- Deployment — self-hosted (user runs
python main.py themselves; MUST ship login.py+auth.py, see Rule 8) or Rupeezy container (zip uploaded via MCP; zero-arg VortexAPI(), no login files)?
- Risk tolerance — max loss per trade / day / drawdown. Defaults if unknown: 1% / 3% / 10%.
Step 2: Discuss Strategy Design
Before any code: entry logic, exit logic (stop-loss is mandatory), position sizing, scheduling, hedging (warn against naked options selling).
Step 3: Route to the Right References
Load only what's relevant. Core references — read when topic comes up:
| Need | File |
|---|
| New strategy | references/strategy-patterns.md |
| Risk / position sizing | references/risk-management.md |
| Backtesting | references/backtesting.md |
| Indian market rules (expiry, timings, margins) | references/indian-market.md |
| Error handling | references/error-handling.md |
| Code quality / testing | references/code-quality.md |
| Rupeezy/Vortex SDK | references/brokers/rupeezy-vortex.md |
Advanced (suggest proactively when context fits): options-greeks.md, regime-detection.md, india-data-edge.md, execution-alpha.md, robustness-testing.md, portfolio-construction.md, psychological-guardrails.md, tax-optimization.md, python-performance.md. Don't wait for the user to ask — for MA crossover suggest regime detection; for 40% CAGR backtest demand robustness testing.
Code Architecture Rules
Every strategy MUST follow this structure. No exceptions.
Separation of Concerns
main.py → entry point, initialization, scheduling
strategy.py → signal generation ONLY (no order placement)
execution.py → order placement, fill tracking (no signal logic)
risk_manager.py → position sizing, exposure checks, drawdown limits
guardrails.py → daily loss limits, cooldowns
config.py → all configurable parameters
Self-hosted only, add login.py (loopback OAuth server) + auth.py (get_client() helper). Container packages must not include these. See Rule 8.
Signal generation and execution are always in separate modules so signals can be tested independently and execution can be swapped between backtest and live without touching signal code.
Configuration Externalized
Every tunable parameter lives in config.py or env vars: symbols, quantities, thresholds, indicator periods, risk caps, schedules. Broker credentials always env-var, never in code.
Risk Manager as Gatekeeper
Every order goes through risk_manager.approve(signal) before submission. The manager checks position size, daily loss, drawdown, margin, and concentration. Full implementation in references/risk-management.md.
def place_order(signal):
if not risk_manager.approve(signal):
logger.warning(f"Risk manager rejected: {signal.reason}")
return None
return execution.submit_order(signal)
Structured Logging
Use logging, not print(). Log every decision with timestamp, symbol, action, reason, price, qty, current P&L.
Graceful Shutdown
Handle SIGTERM/SIGINT: cancel pending orders, optionally square off, log final state. Critical for container deployments where the platform can stop the strategy any time.
Critical Rules — Violations Cause Real Money Loss
These are non-negotiable. Every strategy must follow them.
⚠ Silent-until-live failure modes: Rules 1, 11, 12, 13, 15. These pass py_compile, pytest, and often a WebSocket smoke test — they blow up only on the first real broker interaction (or on a later trading day when something rolls). Walk these five explicitly whenever you scaffold or modify a strategy. Each rule's body names the specific failure.
1. NEVER hardcode instrument tokens
Tokens change daily. Pass ticker="<EXCHANGE>:<SYMBOL>" to every Vortex API (place_order, historical_candles, get_order_margin, client.quotes, wire.subscribe) on vortex-api >= 2.1.8. Never build a ticker from exchange:symbol (breaks on F&O) — read inst.ticker from the master. Feed ticks are tagged with tick.get("ticker") (best-effort enrichment; exchange/token always present).
client.place_order(ticker="NSE:RELIANCE", ...)
inst = client.instruments.get_by_ticker("NSE:RELIANCE")
Ticker shapes: equities "NSE:RELIANCE"; indices append IDX ("NSE:NIFTYIDX") but the F&O symbol field stays bare (so option-chain filter is symbol == "NIFTY"); F&O contracts each have their own ticker — use client.instruments.all_by_underlying("NSE_FO", "NIFTY") to enumerate. For other brokers, use whatever symbolic identifier they expose — never raw numeric tokens.
2. NEVER hardcode lot sizes
Lot sizes change with corporate actions and SEBI directives. Read them off the instrument object: client.instruments.get_by_ticker("NSE:NIFTYIDX").lot_size. For other brokers, look them up from the instrument master.
3. ALWAYS use stop-losses
Every strategy ships with one. If the user explicitly refuses, warn and add a wide-buffer stop anyway; document the risk.
4. ALWAYS check margin before placing orders
Call get_order_margin() first. If insufficient, log and skip — don't crash.
5. ALWAYS handle order rejections
Every place_order wrapped in try/except. Rejections happen routinely (margin, price-out-of-range, exchange down).
6. NEVER ignore partial fills
A "buy 100" can fill 60 now + 40 later, or 60 then cancel. Track fill state precisely.
7. ALWAYS set IST timezone explicitly
import pytz; IST = pytz.timezone("Asia/Kolkata")
Never rely on system timezone.
8. Self-hosted strategies MUST ship login.py + auth.py
For self-hosted (anywhere the user runs python main.py themselves, including headless boxes), generate login.py and auth.py. No exceptions. Never generate input("auth code: "), VORTEX_ACCESS_TOKEN in .env, or a broker.py that exposes a manual auth_code. Run scripts/scaffold_strategy.py --deployment self-hosted or replicate its output; the full pattern is in references/brokers/rupeezy-vortex.md.
Why: auth_token (the ?auth=... redirect param) and access_token (the long-lived bearer) are different. Users confuse them. The loopback HTTP server on 127.0.0.1:8765/callback does client.exchange_token() automatically so the user never sees either token.
Does not apply to the container platform — there VORTEX_ACCESS_TOKEN is injected at runtime, main.py uses zero-arg VortexAPI(), and login.py/auth.py would be dead code that breaks at runtime.
9. Order updates have split authority across surfaces. Never sleep-poll.
VortexFeed.on_order_update is the only way to learn an order changed in real time. Connect the feed before any place_order so the first push isn't lost. The callback fires for five envelope types — order, trade, sl_trigger, gtt_order, position_conversion — all signal that state changed for some order_id. See references/brokers/rupeezy-vortex.md for the payload table and per-type meaning.
Authority is split between surfaces because REST orderbook can lag postbacks by tens of seconds on the same order:
- Postback
data.status is authoritative for status transitions — believe terminal postbacks immediately, don't wait for REST.
client.orders() / client.trades() are authoritative for fill quantities, traded_price, avg fill — refresh from these on every postback (coalesced 500 ms debounce so a many-fill order doesn't trip the REST rate limit).- Downgrade guard: once terminal is recorded for an order_id, a later non-terminal REST row is stale — merge fill fields but DO NOT overwrite the status.
Do NOT poll client.orders() / client.order_history() / client.trades() on a sleep-loop. Refresh-on-postback is event-driven — REST is called only when a postback indicates a change.
Order outcomes are events, not return values. Live strategies wire tracker.on_terminal = strategy.on_order_terminal to receive (order_id, status) callbacks for every terminal transition. Do NOT call tracker.wait() inside the strategy main loop — wait() is the script / test / cancel_and_wait() primitive only; its timeout is required (pass float("inf") for forever). Call tracker.initialize() once at startup before placing orders so historical fills don't fire spurious callbacks.
Terminal-status detection is substring-based to catch broker-decorated forms (REJECTED_BY_RMS, AMO_CANCELLED); canonical tokens are COMPLETED / EXECUTED (same state — postback uses COMPLETED, orderbook uses EXECUTED), REJECTED, CANCELLED. Rejection-reason field name varies by surface (status_message / error_reason / message) — use tracker.rejection_reason(order_id).
The reference OrderTracker in references/code-quality.md (and the scaffolded order_tracker.py) implements all of this — wire it, don't re-implement.
10. NEVER short sell illiquid equities intraday
Auction risk: stock hits upper circuit → you can't exit → 20%+ penalty above your sell price. Check volume + circuit band before shorting; prefer F&O for shorts. Details in references/indian-market.md.
11. ALWAYS respect tick sizes
Round prices to the instrument's tick (from client.instruments.get_by_ticker(...).tick). Mis-tick prices get rejected.
def round_to_tick(price, tick): return round(round(price / tick) * tick, 2)
12. ALWAYS respect Daily Price Range (DPR)
Exchanges set a daily circuit-limit band per instrument. Orders outside it are rejected by the broker's OMS before reaching the exchange — common cause of failed deep stop-losses and ambitious targets. Read DPR from the broker's quote/market data and clamp price / trigger_price accordingly.
13. Account for calendar spread margin removal on expiry day
On expiry day, calendar-spread margin benefits are removed and required margin can jump 5-10x (₹26K → ₹2.6L per lot). Check if any spread leg expires today and pre-flight margin. See references/risk-management.md.
14. NSE has no public data API
All market data (quotes, historical candles, OI, FII/DII, depth) must come through the broker API or third-party providers. Never write code that scrapes or calls NSE endpoints directly.
15. Pass SDK enum instances, not their string values
The Vortex SDK runtime-typechecks every transaction_type, product, variety, validity, mode, exchange, resolution argument via isinstance(value, EnumClass). Passing the string value ("INTRADAY") where the SDK expects the enum (Vc.ProductTypes.INTRADAY) raises TypeError: product must be of type ProductTypes even though the enum's .value IS "INTRADAY" — the check is class-based, not value-based. Silent until the first live place_order / get_order_margin; py_compile and unit tests don't catch it.
from vortex_api import Constants as Vc
client.place_order(product="INTRADAY", ...)
client.place_order(product=Vc.ProductTypes.INTRADAY, ...)
If you store these in config.py, store the enum directly (the scaffolded config.py does). Name ≠ value for many Vc.* enums (Vc.VarietyTypes.REGULAR_LIMIT_ORDER.value == "RL", Vc.ExchangeTypes.NSE_EQUITY.value == "NSE_EQ"), so a stringly-typed config with getattr(Vc.X, name) lookup also works but only if you store the member name, not the value.
Backtesting Standards
Every backtest must include realistic friction. Zero-cost backtests produce fantasy returns.
- Costs: STT (eq 0.1%, fut 0.05%, opt 0.1%) + brokerage + exchange.
commission=0.001 minimum in backtesting.py (raise for options). Rates in references/indian-market.md.
- Slippage: ≥0.05% liquid, 0.1-0.2% illiquid; double near F&O expiry.
If CAGR > 30%, flag and require robustness testing (walk-forward, Monte Carlo, OOS). If parameters are tunable, run grid optimization with heatmap. See references/robustness-testing.md.
Strategy Output Format
Both modes: main.py, strategy.py, risk_manager.py, config.py, requirements.txt, README.md.
Self-hosted only, additionally: login.py, auth.py, .env.example (with VORTEX_API_KEY + VORTEX_APPLICATION_ID only — never VORTEX_ACCESS_TOKEN). requirements.txt includes python-dotenv.
Container only: requirements.txt does NOT include python-dotenv; no .env.example for credentials (platform injects them).
Backtest-only strategies can be a single file as long as risk management, realistic costs, and parameters are present. If a backtest touches VortexAPI, the self-hosted layout applies.
Proactive Suggestions
After generating a strategy, offer the relevant ones:
- Regime detection when the strategy assumes a single regime (trending OR sideways).
- Robustness testing (walk-forward, Monte Carlo) when going live or when CAGR > 30%.
- Psychological guardrails (daily loss caps, consecutive-loss pause) on any live strategy.
- Tax optimization when holding-period tweaks could move trades from STCG (20%) to LTCG (12.5%).
- VWAP execution for orders > 5% of ADV.
- Vectorization when you see Python loops over price data.
