Where agents
[and humans]
win on Hyperliquid

Two phases — Phase 1 ("Let it breathe") uses a wide 3% retrace with 3 breach checks to avoid early shakeouts. Phase 2 ("Lock the bag") triggers on the first profit tier, tightens to 1.5% retrace (or per-tier retrace in v4), and ratchets a profit floor upward that can never go back down. The script closes positions directly via mcporter with retry — no agent intervention needed for the critical path.

• Works for both LONG and SHORT positions
• ROE-based tier ratcheting (leveraged return on margin, not raw price)
• Per-tier retrace: tighten trailing stops as profit grows
• Breach decay: soft or hard mode for volatile assets
• Error handling with close retry and pendingClose recovery
• Self-contained: one Python script + one JSON state file

Same core as DSL v4 (all thresholds are ROE-based, leverage-adjusted). The difference is philosophy: DSL-Tight ships with aggressive defaults and fewer knobs. Phase 1 sets a 5% ROE trailing floor. Phase 2 uses 4 tiers that lock 50→65→75→85% of the high-water move, with breach counts that tighten as profit grows (3→2→2→1). Auto-calculates all price floors from entry and leverage — no manual math. Includes a stagnation take-profit that auto-closes if ROE ≥ 8% but the high-water mark hasn't improved for 1 hour.

• ROE-native: auto-calculated price floors from leverage
• 4 tiers with per-tier breach counts (3→2→2→1)
• Stagnation take-profit for stalled winners
• Fewer knobs than DSL v4 — opinionated preset

4-stage funnel that burns near-zero LLM tokens. Stage 0 checks BTC macro context. Stage 1 bulk-screens all perps via a single API call (volume filter). Stage 2 overlays smart money flow + freshness signals. Stage 3 deep-dives the top 15 with multi-timeframe candle analysis (4h trend, 1h setup, 15m entry) with parallel fetches. Final score: 0–400 across 4 equal pillars, gated by hourly trend alignment.

• Smart money flow (pnl contribution, trader count, acceleration)
• Market structure (4h EMA crossover, hourly trend, S/R levels, candlestick patterns)
• Technicals (multi-TF RSI convergence, volume ratio, momentum)
• Funding rate (contrarian edge when funding opposes setup direction)

The agent creates a strategy wallet, configures risk rules based on your profile, then enters a scan → evaluate → trade → protect loop. It uses the Opportunity Scanner and Emerging Movers to find setups and DSL to protect every position. You get Telegram alerts for every action.

• 3 risk profiles: conservative, moderate, aggressive
• Race condition prevention — no phantom closes across crons
• Conviction collapse = instant cut (220→24 traders = close now)
• Hourly trend alignment — hard requirement for all entries and flips
• Cross-margin buffer math — concentration beats diversification
• Speed filter — best moves happen fast, slow movers are suspects
• Daily loss limits and total drawdown hard stops

Tracks the SM profit concentration leaderboard over time using a single API call (leaderboard_get_markets). Scans top 50 markets every 60 seconds. Detects acceleration patterns — assets climbing the ranks with increasing profit contribution. v3.1 adds quality filters to prevent false IMMEDIATE signals from erratic rank histories, low velocity, thin trader pools, and low-leverage instruments.

• IMMEDIATE_MOVER — 10+ rank jump from #25+ in one scan
• NEW_ENTRY_DEEP — appears in top 20 from nowhere
• CONTRIB_EXPLOSION — 3x+ contribution increase in one scan
• DEEP_CLIMBER — 5+ rank jump from #25+
• v3.1 quality filters: erratic rank, velocity gate, trader count floor, max leverage check
• Self-contained: one Python script + one JSON history file. Near-zero LLM tokens

Mechanical exits, discretionary entries. DSL v4 handles all exit logic — human/AI judgment on entry signals only. Runs 2 position slots with 7 cron jobs: Emerging Movers every 60s, Scanner every 15min, DSL per position, SM flip detection every 5min, and health checks. Favors aggressive rotation into higher-conviction setups.

• Fixed 2-slot max — concentration beats diversification
• IMMEDIATE_MOVER is the primary entry trigger
• 7-cron architecture with race condition prevention
• Conviction collapse and dead weight = instant cut
• Aggressive rotation into higher-conviction setups
• Proven session results with trade-by-trade data

A daily cron spawns an isolated sub-agent that gathers trade memory files, DSL state, and Senpi trade history. It analyzes each closed trade (PnL, ROE, signal quality, DSL tier reached, close trigger), computes metrics (win rate, profit factor, slot utilization, dead weight duration), identifies patterns distinguishing winners from losers, and produces a full report with recommended config changes. The report is saved to memory and a concise summary is sent via Telegram.

• Automated nightly retrospective via cron
• Per-trade analysis: signal quality correlation, DSL tier distribution
• Missed opportunity detection (top movers you didn't trade)
• Data-backed improvement suggestions (high/medium/low confidence)
• Cumulative trend tracking — today vs all-time stats
• Brutally honest — no sugarcoating losses

Scans the top 50 Discovery traders, hard-filters by consistency and risk labels matching the user's risk level, then scores on PnL rank (35%), win rate (25%), consistency (20%), hold time (10%), and drawdown (10%). Creates 2–5 mirror strategies with overlap checks to prevent correlated blowups. Daily cron monitors trader health — a trader must be on WATCH for 2+ consecutive days AND a replacement must score 15% higher before any swap happens.

• 3 risk profiles with trader label filters and leverage caps
• Score-weighted allocation with 35% cap per slot
• Strategy-level stop losses (-10% to -25% by risk)
• Position overlap check across selected traders
• 2-day watch period prevents churn