cheetah-strategy（猎豹策略）

🐆 CHEETAH v5.1.1 APEX — The Arena Sniper

v5.1.1 changelog (fleet-fix batch 4)

• Leverage safety fix. v5.1 fired MON LONG at 7x but MON's Hyperliquid max is 5x. Order rejected with CREATE_INVALID_LEVERAGE — but the MCP wrapper returned outer success=true, so the scanner logged a phantom ENTRY. v5.1.1 adds get_safe_leverage() which queries strategy_get_asset_trading_limits and clamps requested leverage to the asset's Hyperliquid max before submission.
• Inner-order success validation. After create_position returns, inspect data.orders[0].success and surface INNER_FAILURE when the per-order status is false even if the outer envelope claims success. Prevents phantom ENTRY logs from corrupting the daily counter.

SM commits. Quality traders commit. Price confirms. Volume commits. All at once. Cheetah pounces once.

Why v5.0 is a complete rewrite

Cheetah iterated through four theses, all on HYPE, all on Wolverine's turf:

The pattern: Cheetah kept fighting for Wolverine's HYPE territory. Fleet analysis showed Wolverine (HYPE momentum), Pangolin (funding fader), and Owl (crowding fader) already cover every HYPE thesis Cheetah was trying. Cheetah was redundant.

v5.0 does something no predator in the fleet does: it refuses to trade unless every major signal aligns simultaneously. MIN_SCORE is 14 out of 15 — the highest gate in the fleet by a wide margin.

Arena thesis

The Senpi Arena ranks by ROE %, not absolute PnL, with a $25k weekly volume minimum. Looking at the current week's leaderboard top 5:

• #1 pr0br000: 67% ROE, 19 trades — sniper pattern (high conviction, low frequency)
• #2 0xschelling: 34% ROE, 57 trades
• #3 dih: 30% ROE, 191 trades — scalper pattern (high frequency, tiny per-trade)
• #4 ysr: 21% ROE, 133 trades
• #5 Magi300a: 17% ROE, 190 trades

Two winning shapes exist. The scalper path is fee-drag death for a fleet agent — our fleet is 37 red / 2 green precisely because Hyperliquid fees eat high-frequency strategies. The sniper path (pr0br000) works because each trade has maximum edge, and 19 trades × ~3.5% ROE per trade = 67% ROE total.

APEX is purpose-built to replicate the sniper shape.

Signal pipeline

Every scan iterates through all non-XYZ assets from leaderboard_get_markets(limit=100) and scores each one. An entry fires only when score ≥ 14 AND all hard gates pass.

Hard gates (any failure = reject)

1. SM consensus — pct_of_top_traders_gain ≥ 10% AND trader_count ≥ 25
2. Velocity gate — contribution_pct_change_15m ≥ 1.0 OR contribution_pct_change_1h ≥ 3.0
3. Not currently held — APEX has no existing position on this coin
4. Not in cooldown — the asset passed the per-asset cooldown window
5. Not XYZ DEX — XYZ equities/indices banned

Scoring (max = 15, threshold = 14)

Score = 14 requires all signals except either the rank climb OR dropping one of the velocity subcomponents. Score = 15 requires everything. Partial signal stacks (score ≤13) never fire.

Position sizing

Leverage tiers

• Score 14: 8x leverage
• Score 15: 10x leverage (perfect setup)

Dynamic daily cap (rebased to $648)

if pnl_pct >= 5:     return 8   # Hot hand — more shots
elif pnl_pct >= 0:   return 5   # Target rate for Arena volume floor
elif pnl_pct >= -5:  return 3   # Careful
elif pnl_pct >= -15: return 2   # Defensive
elif pnl_pct >= -25: return 1   # Preserve
else:                return 0   # HARD STOP

Starting at 0% PnL baseline = 5 entries/day cap = matches target weekly volume.

DSL configuration (aggressive ratcheting)

Designed for the Arena's ROE% optimization:

The 5% max loss is the single most important setting. A 5% ROE loss is survivable in a 7-day Arena window; a 20% loss ends your week.

Fleet-standard guardrails (all present)

• Self-executing — scanner calls create_position via mcporter directly (Wolverine pattern)
• Dynamic P&L-aware daily cap (rebased to $648, fleet PR #176 pattern)
• Auto-cancel stale resting orders — non-reduceOnly orders older than 10 min get cancelled (fleet PR #177 pattern)
• Persistent entry log — state/entry-log.jsonl records every ENTRY/EXIT event, survives openclaw sessions clear --current (Wolverine v2.3 pattern)
• Per-asset cooldown — 240 min standard, 120 min extended after a loss
• Stale-date bug fix — load_trade_counter checks date on every call (fleet PR #177)

Runtime Setup

# Set wallet and chat ID
sed -i 's/${WALLET_ADDRESS}/<WALLET>/' /data/workspace/skills/cheetah-strategy/runtime.yaml
sed -i 's/${TELEGRAM_CHAT_ID}/<CHAT_ID>/' /data/workspace/skills/cheetah-strategy/runtime.yaml

# Install runtime
openclaw senpi runtime create --path /data/workspace/skills/cheetah-strategy/runtime.yaml
openclaw senpi runtime list && openclaw senpi status

# Pull latest scanner via curl
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/cheetah/scripts/cheetah-scanner.py -o /data/workspace/skills/cheetah-strategy/scripts/cheetah-scanner.py
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/cheetah/scripts/cheetah_config.py -o /data/workspace/skills/cheetah-strategy/scripts/cheetah_config.py
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/cheetah/config/cheetah-config.json -o /data/workspace/skills/cheetah-strategy/config/cheetah-config.json

# Verify with a manual run
python3 /data/workspace/skills/cheetah-strategy/scripts/cheetah-scanner.py

Expected first-run output: no exceptions, _cheetah_version: "5.1.1-APEX" in the JSON, and either note: "no score >= 14 candidates" (most common) or a rare action: "ENTRY" if a confluence setup exists right now.

Cron configuration

Use the Turbine pattern — detached bash loop, zero LLM wake:

nohup bash -c 'while true; do python3 /data/workspace/skills/cheetah-strategy/scripts/cheetah-scanner.py >> /tmp/cheetah-loop.log 2>&1; sleep 180; done' > /tmp/cheetah-nohup.log 2>&1 &

3-minute cadence. Zero LLM wake cost. Scanner decides everything in Python.

Arena week expectations

Expected value of week 1 prize: ~$530. That's the cost-weighted average across outcomes. The sniper design means APEX either wins significantly or wins nothing — it does not grind out mediocre returns.

Why this will work as one of the 10 consistent winners

APEX isn't just an Arena one-shot. The thesis is a permanent edge:

1. Low trade frequency = low fee drag. The fleet's biggest hidden cost (7.6-8.4 bps fee/vol) barely affects APEX at 5 trades/week.
2. Hyperfeed data refreshes every 15 min — the 3-min scan cadence catches the early window of any genuine confluence setup.
3. Confluence setups don't get arbitraged away — markets always have rare moments when everything aligns, and those moments always pay.
4. Tight -5% floor is self-correcting. Bad calibration means APEX simply doesn't trade — it won't bleed like over-eager agents. We get a clean signal if the score thresholds need tuning.
5. Letting 99% of signals pass is the feature, not a bug. Selectivity IS the edge.

License

MIT — Built by Senpi (https://senpi.ai).

Skill Attribution

When creating a strategy, include skill_name and skill_version in the call. See references/skill-attribution.md for details.