SuperAPI AI Directory
Partner中文AI newsFavorites
Skillswolverine-strategy

wolverine-strategy

Agents & models

View repository

Updated 10-16 02:25

About this skill

🦡 WOLVERINE v6.1.0 — HYPE Patient-Conviction

v6.1.0 (2026-06-01) — Phase 1 retrace leash widened for HYPE volatility

One isolated DSL change. NO scoring, gate, leverage, or Phase 2 change. exit.dsl_preset.phase1.retrace_threshold 8 → 15.

Why. During HYPE's epic run Wolverine caught a +16.4% ROE winner but left the bulk of the move on the table, and its recent window ran a 25% win rate with losses clustered at −5% to −9% ROE. The provable structural cause: at 5x leverage an 8-ROE-point retrace is a 1.6% HYPE price leash (8 ÷ 5). HYPE routinely wicks 3-5% inside a single 15m candle, so this single-breach (consecutive_breaches_required: 1) retrace fired on ordinary noise — exiting positions at small losses before they could become the BIG trend the wide Phase 2 "let winners run" ladder exists to ride. The −5% to −9% loss cluster is consistent with a trade peaking +2-3% ROE, pulling back 1.6% price, and tripping the 8-point retrace.

The fix. 15 ROE pts = ~3.0% price leash at 5x — roughly double the buffer, still well below the 20% max_loss_pct hard stop so it remains a distinct trailing-from-peak rule, not a collapse into the hard stop (which is why 15 and not the 20 that would equal max_loss).

Why ONLY this. The recent window is n=12 (score buckets n=2-4) — statistically indistinguishable from a 50% win rate, far too thin to justify retuning Phase 2 locks, the entry score floor, or MIN_4H_STRUCTURE. This retrace fix is the one change defensible on stop-width-vs-volatility math alone, independent of sample size. Everything else waits for data.

Decision rule. Re-measure expectancy after 40-50 trades. A let-winners-run trend agent is supposed to have a low win rate; it only fails if the wins can't clear the loss size. Break-even at a 25% win rate needs a 3:1 avg-win/ avg-loss ratio (recent window was 1.56:1 — wins truncated). If, after this fix and a real sample, expectancy is still negative, the trend-follower archetype is wrong for HYPE and a rewrite is earned. Until then it's a mis-calibrated strategy, not a broken one.

v6.0.0 (2026-05-18) — Patient-Conviction adoption (Bison v3.0.1 port)

NO scoring change. NO thesis change. NO gate change. Behavioral profile shifts from "4 entries/day on score ≥9 with tight T0 lock=35" to "2 entries/day on score ≥10 with wide T0 lock=0 multi-day holds."

Why. Bison v1.0 (BTC/ETH/SOL conviction holder) shipped this exact pattern on 2026-05-12 and 4 days later was sitting on a SOL SHORT at +93.5% ROE that had been held 4 days through chop, V-recoveries, and an intraday paper drawdown to -8%. The SOL position rode SOL $92.30 → $83.67 (-9.4% spot) — a move that every fast-rotation agent in the fleet (Cheetah, Scorpion, Turbine) flipped out of within minutes of entry. Bison held through the noise because its DSL gave the position room to breathe past +10% retrace before locking anything. That T0 lock=0 first tier is the single highest-leverage knob in the Bison ladder.

Wolverine has been red since the +12.8% HYPE-day post-mortem (PR #281, 2026-05-15) tightened gates to reject the obvious chase. The gates now correctly identify clean trends. What was missing was the back-end discipline to let those trends run. v6.0.0 ports the Bison back-end without touching the Wolverine front-end.

What changed (5 small edits):

File Change
runtime.yaml risk.guard_rails.max_entries_per_day 4 → 2
runtime.yaml risk.guard_rails.per_asset_cooldown_minutes 120 → 240
runtime.yaml exit.interval_seconds 30 → 60 (REDUCE_ONLY spam mitigation)
runtime.yaml exit.dsl_preset.phase2.tiers wide Bison ladder (T0 lock 0 through T5 lock 85)
config/wolverine-config.json minScore 9 → 10
scripts/wolverine-producer.py + wolverine_config.py recent-signals race-window dedup (Bison v3.0.1 pattern)

What did NOT change.

  • Six-gate entry validation (4h trend, structure ≥0.65, 1h-4h alignment, 15m momentum ≥0.15, base-tech floor, 4h magnitude ≥1.0%) preserved verbatim.
  • SM hard block on opposing direction preserved.
  • RSI hard gates (74 LONG / 26 SHORT) preserved.
  • Multi-factor scoring (~17 max) preserved.
  • Conviction-tiered leverage (3x standard / 5x apex) preserved. Bison uses 7-10x on majors; HYPE's volatility doesn't yet warrant that — we get the patient-conviction benefits from the DSL ladder, not from leverage.
  • FP-001 quiet hours (00-04 UTC unless apex score 11+) preserved.
  • All time-cuts remain DISABLED (hard_timeout, weak_peak_cut, dead_weight_cut from v3.0.4 sweep).

Tradeoffs to monitor.

  • Wider T0 lock means more giveback on weak winners (a +12% mover that retraces to +0% will close at breakeven instead of v5.x's locked +4.2%). Net positive ONLY if the captured BIG trends exceed the lost small winners. Bison's track record validates this on majors; HYPE-specific data needed for 30+ days post-deploy.
  • Cutting entries 4 → 2 means more idle time when HYPE chops. Acceptable — Wolverine's red period under v5.x was driven by too-frequent entries on borderline setups, not too-rare ones.
  • The 4h cooldown stacks with the recent-signals dedup — Wolverine should fire HYPE at most every 4h regardless of signal cadence.

Rollback condition. If v6.0.0 produces zero or one entry per week for 2 consecutive weeks, the floor is too high. Drop minScore back to 9 in wolverine-config.json (the producer floor remains 9; only the operator's selectivity is config- driven). Other deltas should be preserved.

v5.0.0 (2026-05-08) → v5.0.1 (preserved)

v3 → v4 architectural rewrite. v3.x was a full-agency scanner. v4.0 flips to the standard senpi-trading-runtime plugin pattern: producer emits signals, runtime owns execution + state.

What changed structurally:

  • wolverine-producer.py (NEW) replaces wolverine-scanner.py (DELETED)
  • v2-runtime-native: external_scanner + LLM-pass-through gate + native risk.guard_rails
  • DSL exits via FEE_OPTIMIZED_LIMIT (saves ~0.020-0.030% per maker-filled close)
  • Trade chain DB emits per-trade telemetry — chain DB visibility on Wolverine for the first time
  • Python-state-crash class of bug (load_tc / set_cooldown / has_resting_orders) is structurally impossible in v4.0

What's preserved from v3.0.3/v3.0.4 EXACTLY:

  • HYPE single-asset thesis
  • Six-gate entry validation:
    1. 4h trend != NEUTRAL
    2. 4h structural strength ≥ 0.75
    3. 1h direction matches 4h
    4. 15m momentum ≥ MIN_MOM_15M (0.15) in direction
    5. Base-tech floor (strong_15m OR aligned_5m)
    6. 4h MAGNITUDE ≥ 1.5% — the v3.0.3 fix that rejects dead-flat chop. Wolverine's own 2026-04-23 self-diagnostic: all 6 Week 5 trades died because trend_strength_4h is structural (lower-highs count) and passes even when 4h price magnitude is nearly flat.
  • SM HARD BLOCK if direction opposes
  • RSI hard gates (74 LONG / 26 SHORT)
  • Multi-factor scoring (~17 max points): base-tech + SM concentration + SM velocity + funding alignment + funding regime + funding persistence + volume + OI velocity + BTC correlation + RSI room + 4h momentum bonus + move-exhaustion penalty
  • MIN_SCORE = 9 (config-overridable)
  • Conviction-tiered leverage: 5x apex (score ≥11) / 3x standard (≥9)
  • DSL preset preserved: time-cuts ALL DISABLED, Phase 1 max_loss 20% / retrace 8 (v6.1.0: → 15) / 3 breaches, Phase 2 ladder (10/15, 20/35, 35/55, 55/70, 80/85)

v3.0.1/3.0.2/3.0.4 v1-DSL fixes preserved:

  • dead_weight_cut: DISABLED (single-asset has no rotation cost)
  • hard_timeout: DISABLED (v1 DSL fired this in Phase 2 incorrectly)
  • weak_peak_cut: DISABLED (v3.0.4 — completes time-cuts sweep)
  • All exits now 100% price-action

⛔ Hard Rules (Fleet Patches)

RULE FP-002: User-conversation Claude sessions MUST NOT trade

Hard rule, not a heuristic. When responding to a user message, the Claude Code session MUST NOT call any of:

  • create_position
  • close_position
  • edit_position
  • ratchet_stop_add / ratchet_stop_edit / ratchet_stop_delete
  • cancel_order
  • strategy_close / strategy_close_positions

These tools are reserved for the producer cron (wolverine-producer.py) and the DSL ratchet engine. The cron is the only entry path. The DSL is the only exit path. User-conversation sessions are read-only.

RULE FP-001: Quiet hours for low-liquidity windows

Producer skips emission during 00:00-04:00 UTC by default. Apex setups (score >= quietHours.apexBypassScore, default 11) bypass.

Configurable via quietHours.{startUtc,endUtc,apexBypassScore} in wolverine-config.json. Set startUtc == endUtc to disable.

🦡 WOLVERINE — Original Thesis

Single asset. One thesis. Smart money commits, Wolverine pounces, DSL trails the trend.

What it does

Wolverine scans HYPE on Hyperliquid every 3 minutes for a confluence of momentum signals:

  • Smart Money consensus — pct of top traders gain ≥ minimum, trader count above threshold
  • Velocity acceleration — 15m and 1h contribution change positive AND building (15m > 1h)
  • Price confirmation — 4H and 1H both moving in the signal direction
  • Volume confirmation — bonus when volume spikes above 6h average
  • 15m freshness gate — penalizes stale signals (15m velocity < 0)

When the score reaches 8+, Wolverine fires a HYPE entry at conviction-scaled leverage (7x or 10x) with FEE_OPTIMIZED_LIMIT order type. Position management is delegated entirely to the DSL exit engine — Wolverine never closes a position itself, only opens.

v2.3 — chop-hardened

On 2026-04-14 Wolverine v2.2 took 5 consecutive losing HYPE trades over ~3 hours of chop for -$113 total. Every entry was a "fresh signal" by v2.2's criteria, but the scanner had no concept of "we just lost N times on this same coin." v2.3 adds three protections:

1. Chop-detection lockout

After 2 losses on HYPE within 3 hours, Wolverine refuses any new entry on HYPE for 6 hours from the last loss. The scanner emits CHOP_LOCKED and waits out the chop. Constants:

  • CHOP_WINDOW_HOURS = 3
  • CHOP_MAX_LOSSES = 2
  • CHOP_LOCKOUT_HOURS = 6

2. Direction-flip hard gate

If the new signal is OPPOSITE direction to the most recent trade on HYPE within 2 hours AND that trade was a loss, Wolverine hard-gates the new entry. Catches the LONG → SHORT → LONG whipsaw pattern even before the chop-lockout threshold fires. Winning flips are still allowed (legit reversal catch); losing flips are blocked (chop chasing).

3. Persistent entry log

Every ENTRY, EXIT, CHOP_LOCKOUT, and FLIP_BLOCKED event writes a JSON line to state/entry-log.jsonl. This log survives openclaw sessions clear --current — the data lives on disk, not in LLM context. When you ask "what scored that trade?" after a session reset, you can answer by tailing the log:

tail -20 /data/workspace/skills/wolverine-strategy/state/entry-log.jsonl | jq

4. Exit tracking hook

sync_closed_positions(wallet, state) runs at the start of every scan, compares current positions to the previous scan's snapshot, and appends EXIT events linked to prior ENTRY metadata. Closes the loop on chop detection so the scanner sees its own realized PnL per trade.

Fleet-standard guardrails (all present)

  • Self-executing via create_position (no external action layer needed)
  • Dynamic P&L-aware daily entry cap (PR #176)
  • has_resting_orders() auto-cancels stale maker orders >10 min (PR #177)
  • Stale-date bug fix in load_trade_counter() (PR #177)
  • Per-asset cooldown (180 min default)
  • Conviction-scaled leverage capped at 10x (fleet H12 audit)

Key settings

Setting Value Why
Asset HYPE Single-asset focus, no parallel bets
Max positions 1 Concentration
Margin per trade 50% High conviction commits high capital
Max leverage 10x Fleet cap (H12 audit found >10x destroys edge via fees)
Min score 8 Above this, conviction is real
Per-asset cooldown 180 min Patience between trades
DSL hard timeout 240 min HYPE winners often need 2-4 hours to develop
DSL Phase 1 max loss 25% ROE Standard fleet floor
DSL Phase 2 tier 1 +8% / 25% lock Earlier than fleet standard

⛔ Critical agent rules

  1. Install path is /data/workspace/skills/wolverine-strategy/
  2. THE SCANNER DOES NOT EXIT POSITIONS — DSL handles all exits
  3. MAX 1 POSITION — HYPE only
  4. Do not modify scanner constants without testing — fleet-wide thresholds were tuned by audit
  5. Do not rebase the daemon to a faster tick than 3 min — 3 min is the right cadence for HYPE momentum signals

Setup

Step 1 — Install path

The skill must live at /data/workspace/skills/wolverine-strategy/. The package contains:

wolverine/
├── README.md                     # User-facing summary
├── SKILL.md                      # This file (LLM-facing)
├── runtime.yaml                  # OpenClaw runtime config + DSL preset
├── config/
│   └── wolverine-config.json     # Wallet, strategy ID, chat ID
└── scripts/
    ├── wolverine-scanner.py      # Main scanner (v2.3)
    └── wolverine_config.py       # Helper module (atomic write, MCP, state I/O)

Pull from the senpi-skills GitHub repo:

mkdir -p /data/workspace/skills/wolverine-strategy/{config,scripts,state}

curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/wolverine/runtime.yaml \
  -o /data/workspace/skills/wolverine-strategy/runtime.yaml

curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/wolverine/SKILL.md \
  -o /data/workspace/skills/wolverine-strategy/SKILL.md

curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/wolverine/config/wolverine-config.json \
  -o /data/workspace/skills/wolverine-strategy/config/wolverine-config.json

curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/wolverine/scripts/wolverine-scanner.py \
  -o /data/workspace/skills/wolverine-strategy/scripts/wolverine-scanner.py

curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/wolverine/scripts/wolverine_config.py \
  -o /data/workspace/skills/wolverine-strategy/scripts/wolverine_config.py

Step 2 — Set wallet and chat ID

Set your strategy wallet address in runtime.yaml:

sed -i 's/${WALLET_ADDRESS}/<YOUR_STRATEGY_WALLET>/' /data/workspace/skills/wolverine-strategy/runtime.yaml
sed -i 's/${TELEGRAM_CHAT_ID}/<YOUR_TELEGRAM_CHAT_ID>/' /data/workspace/skills/wolverine-strategy/runtime.yaml

OR set them in config/wolverine-config.json directly (the scanner will fall back to the JSON values if env vars / runtime placeholders aren't set):

{
  "strategyId": "your-strategy-id",
  "wallet": "0xYourStrategyWallet",
  "chatId": "your-telegram-chat-id"
}

The scanner also supports environment variables: WOLVERINE_WALLET and WOLVERINE_STRATEGY_ID.

Step 3 — Install the runtime in OpenClaw

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

Step 4 — Verify with a single manual scan

python3 /data/workspace/skills/wolverine-strategy/scripts/wolverine-scanner.py

Expected: clean exit, JSON output includes "_wolverine_version": "2.3". Most likely first run shows a heartbeat (no signal) — that's normal. Wolverine fires on HYPE momentum confluence, which is rare.

Step 5 — Run the scanner on a recurring schedule

The recommended pattern is a detached bash loop with zero LLM wake cost (matches Turbine pattern):

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

# Verify the loop is running
ps aux | grep wolverine-scanner | grep -v grep
tail -5 /tmp/wolverine-loop.log

Alternative: configure an OpenClaw cron with sessionTarget: isolated. Avoid sessionTarget: main — that pattern is a known fee-burn vector when the cron spawns LLM-driven agent turns instead of running the producer script directly.

Operational notes

Tailing the entry log

After Wolverine takes its first trade, you can read the structured entry log to see exact scores, reasons, leverage, and outcomes:

tail -10 /data/workspace/skills/wolverine-strategy/state/entry-log.jsonl | jq

The log persists across session clears, so you can answer "what was the last entry score?" even after openclaw sessions clear --current.

Checking chop-lockout state

If Wolverine outputs CHOP_LOCKED notes, it has detected 2+ losses on HYPE within 3 hours and is sitting out for 6 hours from the last loss. To see when it unlocks:

grep CHOP_LOCKOUT /data/workspace/skills/wolverine-strategy/state/entry-log.jsonl | tail -1 | jq '.unlock_ts'

This is intentional behavior, not a bug. Wolverine is protecting capital during chop.

Verifying the resting-order auto-cancel

Wolverine cancels stale FEE_OPTIMIZED_LIMIT maker orders older than 10 minutes. If you see Auto-cancel attempted on stale order ... in the log, that's the safety mechanism working.

Best for

  • Operators who want a single-asset HYPE momentum specialist
  • Trading environments where 2-4 trades per day is the right cadence (not high-frequency)
  • Accounts where one bad chop day could be expensive without circuit breakers
  • Operators who want the agent to be self-managing — no manual position closing required

Not for

  • Multi-asset diversification (use Phoenix, Condor, or Polar instead)
  • High-frequency scalping
  • Anyone who wants the scanner to make exit decisions (DSL handles all exits)
  • XYZ DEX equities (Wolverine is HYPE-only)

License

MIT — Copyright 2026 Senpi (https://senpi.ai).