# Autonomous Hyperliquid operations mode

Session learning for evolving a paper/testnet Hyperliquid strategy into a later autonomous but strictly risk-bounded live trader.

## User-facing operating model

- Do not force the user to manually approve every future micro-trade once live autonomy is explicitly granted.
- Before that grant, keep Mainnet live orders blocked and run paper/testnet only.
- For live-autonomous mode, require hard limits rather than per-trade chat approval:
  - platform: Hyperliquid only
  - strategy: v76_hl_confirmed_squeeze_hybrid strict/live-candidate
  - start long-only
  - no Martingale, no averaging down, no blind dip-buying
  - no risk increase without explicit approval
  - no new live strategy or live coin without prior shadow/paper evidence and user approval
  - no new entries during API/data degraded mode
  - if any position lacks a confirmed exchange-level reduceOnly stop, set/keep `block_new_entries=true` and alert critically

## Reporting contract

Avoid chat spam after every small trade. Default user-facing cadence:

- One daily digest at 19:00 Switzerland time.
- Immediate alerts only for critical events:
  - position without confirmed stop
  - kill-switch activation
  - API/data degraded for a sustained period
  - unexpected open position/order or Mainnet reconcile mismatch
  - daily loss limit reached
  - bot/process crashed and could not be restored safely

Daily report should include concise trading facts, not developer logs:

- account/equity/free collateral
- daily net PnL and number of trades
- open positions and open orders
- best/worst trades or coins
- most important blocked signal reasons
- market regime: risk_on, risk_off, choppy, squeeze_building, breakout_confirmed, failed_breakout, or no_trade_environment
- API/data-quality status
- what worked, what failed, learnings, observations, proposed changes
- whether to continue unchanged, pause, or run more paper tests
- WLD share and PnL ex-WLD when WLD is in the universe

## Operational safety pattern

1. Mainnet read-only reconcile first; never use a private key for read-only checks.
2. Compare exchange positions/orders against local journal/protection state.
3. Validate per open position:
   - stop exists
   - stop is reduceOnly
   - stop size matches position
   - stop direction closes the position
   - position has local journal entry
4. Detect local journal entries without exchange positions and exchange positions without local journal entries.
5. If reconcile is not clean:
   - no new entries
   - critical alert
   - repair before continuing
6. Supervisor/autostart may restart paper/live runtime only into a safety-checking state. Autostart must never mean blind trading after VM reboot.
7. On startup/restart, before any new trade:
   - read account state
   - read open positions
   - read open orders
   - verify stops
   - verify API health/data quality
   - verify kill-switch inactive
   - restore persistent runtime/nonce/position state
8. For paper operations, a cron-backed supervisor is acceptable. For final live operations, prefer a systemd user service or equivalent durable supervisor with `Restart=on-failure`, but still gate trading on safety preflight.

## Statistical learning boundaries

Per-trade journals should capture rich data internally: coin, setup type, entry/exit reason, entry/exit, stop, size, fees/slippage/funding, net PnL, R-multiple, holding time, MFE/MAE, market regime, and rule compliance.

Learning is allowed, but live mutation is constrained:

- small filter experiments may run in paper/shadow autonomously
- new live parameters require report + explicit user approval
- risk increases are never autonomous
- new live coins require paper/shadow observation, liquidity checks, and user approval

## Coin policy

Start focus: BTC, ETH, SOL, LINK.
Secondary if setup and data quality are strong: SUI, ENA, BCH.
WLD is not permanently banned, but it must not become a default focus or sole PnL driver. Require clean setup, strong liquidity/spread/depth/funding/trend, non-dominant recent share, and daily visibility of WLD share plus PnL ex-WLD.

For new coins: observe in shadow/paper first, evaluate signal quality, liquidity, spread/depth/funding/volume, gather a small paper history, then propose live inclusion in the daily report.

## Communication pitfall

When the user asks for trading/autonomy decisions, do not answer with long developer logs, commit lists, or implementation archaeology. Give the trading decision, current readiness, remaining blockers, and explicit recommendation concisely. Keep implementation details in journals/reports unless asked.