v2 → v3 architectural rewrite. v2.x was a full-agency scanner that called create_position directly and tracked state in Python. v3.0 flips to the standard senpi-trading-runtime plugin pattern: producer emits signals, runtime owns execution + state.
What changed structurally:
vulture-producer.py(NEW) replacesvulture-scanner.py(DELETED). Pure producer — no execution, no counters, no cooldowns, no dynamic-slot bookkeeping.runtime.yamlis now runtime-native: external_scanner + LLM-pass-through gate + native risk.guard_rails + DSL preset with FEE_OPTIMIZED_LIMIT.- Trade chain DB now emits LIFECYCLE_RUNTIME_STARTED → DECISION_EXECUTED → ACTION_RESULT → DSL_CREATED → DSL_CLOSED for every trade. Per-trade telemetry is restored.
- The cfg.set_cooldown silent crash is structurally impossible in v3.0 (cooldowns are runtime-managed, not Python-managed).
What's preserved from v2.4:
- Asset universe (25+ small/mid-cap Hyperliquid perps)
- LONG_TAIL_MOMENTUM scoring (SM concentration tier + 4h momentum + 15m velocity + 1h acceleration + 4h continuation + trader depth + regime alignment + persistence)
- Hard gates: SM ≥3% / 15 traders, 4h aligned ≥1%, 1h aligned ≥0.1% (the v2.4 fix), 15m velocity ≥0.3
- Conviction-scaled leverage (3x cautious / 5x conviction / 7x apex)
- DSL Phase 2 ladder (15/20, 30/60, 40/75, 75/75, 100/85, 150/92) — proved correct on the live ZEC trade
- 90-min dead_weight_cut, 180-min weak_peak_cut, 7-day hard_timeout
- 38% WR target with 6.15x profit factor
What this enables:
- Per-trade scoring telemetry (was lost in v2.x's silent crashes)
- Maker-exit fee recovery (~$10-15/week at current cadence)
- Declarative risk gates (no Python state to drift)
- The senpi-trade-analysis skill can now drill Vulture trades the same way it drills Scorpion / Roach / RoachB
⛔ 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 (Telegram ping, status check, "tell me about your trades", etc.), the Claude Code session MUST NOT call any of:
create_positionclose_positionedit_positionratchet_stop_add/ratchet_stop_edit/ratchet_stop_deletecancel_orderstrategy_close/strategy_close_positions
These tools are reserved for the producer cron (vulture-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.
If a user asks "should I take this trade?" or "anything close to triggering?", respond by reading current state — DO NOT execute. The producer will fire on its next tick if a real signal is there.
RULE FP-001: Quiet hours for low-liquidity windows
Producer skips emission during 00:00-04:00 UTC by default. Apex setups (best signal score >= quietHours.apexBypassScore, default 12) bypass — high-conviction small-cap momentum can fire any hour; routine sub-apex entries wait until 04:00 UTC.
Configurable via quietHours.startUtc, quietHours.endUtc, quietHours.apexBypassScore in vulture-config.json. Set startUtc == endUtc to disable.
Rationale: fleet-wide pattern of midnight-UTC pile-ins after daily-cap reset. For Vulture specifically, small-cap order books are thinnest in this window and entry slippage compounds.
Thesis
Small-cap Hyperliquid perps have thicker trader-count signal-to-noise ratios than the majors. When Smart Money concentration on a small cap ALIGNS with a fresh 4h price trend AND 15m velocity is accelerating, enter in the direction of the trend and hold it via the DSL tier curve for days, not hours.
Key insight: The Arena winner's asset universe (HEMI, WLD, MON, XPL, AIXBT, ARB) is a fleet-wide blind spot. No current Senpi predator scans small-cap alts — they're all locked to 4-12 majors or single assets. Vulture v2.0 fills the gap.
Philosophy (copied from Arena winner #1)
- Low win rate is fine (~38%) if winners are 5-10x bigger than losers
- Tight loss cuts via DSL dead_weight_cut (60 min)
- Long holds on winners via hard_timeout=10080 min (7 days)
- Moderate leverage (5-7x) — small caps can't handle 20x (slippage kills you on low-liquidity books)
- Signal-driven direction, NOT long-only bias. The Arena winner was long because April 2026 was a bullish regime, not because LONG is a rule. Vulture v2.0 takes SHORTs when signals support.
- Max 2 concurrent positions
Asset Universe (25 small/mid-caps)
HYPE, HEMI, WLD, MON, XPL, AIXBT, ARB, ASTER, POLYX, LDO,
APT, DYDX, ONDO, SUI, kBONK, kPEPE, TAO, GRASS, ZEC, LIT,
FARTCOIN, MORPHO, NEAR, INJ, AVAX, LINK, DOGE
Banned: BTC, ETH, SOL (handled by other predators). XYZ DEX assets (handled by Bald Eagle).
Entry Scoring
Hard gates (all must pass)
- Asset must be in TRACKED_ASSETS and not in BANNED_ASSETS
- SM direction must be LONG or SHORT (not neutral)
pct >= 3.0ANDtraders >= 15(signal validity)- 4h price ALIGNED with SM direction by at least 1% (momentum confirmation)
- 15m velocity
> 0.3(actively building, not flat)
Scoring contributors (max ~14 pts)
| Signal | Points |
|---|---|
| SM concentration ≥18% (HEAVY_FLOW) | +4 |
| SM concentration ≥12% (STRONG_FLOW) | +3 |
| SM concentration ≥7% (MODERATE_FLOW) | +2 |
| SM concentration ≥3% (LIGHT_FLOW) | +1 |
| 4h price aligned ≥8% (TREND_RUNNING) | +3 |
| 4h price aligned ≥4% (TREND_STRONG) | +2 |
| 4h price aligned ≥2% (TREND_BUILDING) | +1 |
| 15m velocity ≥3.0 (15M_EXPLOSIVE) | +3 |
| 15m velocity ≥1.0 (15M_STRONG) | +2 |
| 15m velocity ≥0.5 (15M_BUILDING) | +1 |
| ACCELERATING pattern (15m > 1h > 0) | +2 |
| 1h velocity ≥1.0 (1H_POSITIVE) | +1 |
| 4h contribution ≥2.0 (4H_CONTINUATION) | +1 |
| Trader count ≥50 (DEEP_DEPTH) | +1 |
| Move ≥15% in direction (LATE_ENTRY_PENALTY) | -3 |
| Move ≥12% in direction (MOVE_EXTENDED) | -2 |
MIN_SCORE: 10 (raised from 9 in v4.2.0; 7 → 9 in v4.1.0) — entry floor. A 100-trade aggregate made the boundary unambiguous: every score bucket ≥10 was net-positive (10/11/12 combined: +$416.78 over 41 trades, 42–50% win rates) and every bucket ≤9 was net-negative (7/8/9 combined: −$645.12 over 41 trades). Score 9 specifically ran a 27.3% win rate at −3.32% avg ROE. Score 10 is the single best bucket (+8.70% avg ROE, 50% win, +$259.67 net) — the floor is set to keep it. Vulture now hunts strictly in the 10–12 conviction zone. (Caveat: per-bucket n is small at the 9/10 boundary, but expectancy sign is consistent and the direction matches the v4.1.0 cull.) The scoring dimensions (especially HEAVY_FLOW requiring ≥18% SM concentration) do the real quality work above the floor.
Conviction-Scaled Leverage
| Score | Leverage | Tier |
|---|---|---|
| ≥11 | 7x | apex (the right tail) |
| 10 | 5x | conviction (floor tier as of v4.2.0 — the best historical bucket) |
v4.1.0 removed the cautious tier (score 7–8, 3x); v4.2.0 raised the floor to 10 (score 9 culled). Below-floor signals are dropped at the producer now; those tiers are unreachable. Capped at 7x because small-cap liquidity can't support 20x. Lemon can go 20x on BTC/ETH because slippage is <0.1%. Small caps can have 0.5%+ slippage per fill, which eats gains at 20x.
Exit (DSL)
The DSL tier curve is the single biggest difference from other Senpi predators. Winners need DAYS to develop. Arena winner #1 held HEMI for 8 days, WLD for 4 days, MON for 5 days.
| Mechanism | Value | Rationale |
|---|---|---|
| hard_timeout | 10080 min (7 days) | Let momentum runners run a full week |
| dead_weight_cut | 60 min | Faster than Owl's 90, slower than Lemon's 20 (small-cap wick tolerance) |
| weak_peak_cut | 120 min @ 2% min | Catch fades that stalled |
| Phase 1 max_loss_pct | 20% | Wider than Lemon's 15%, tighter than Owl's 35% |
| Phase 1 retrace | 8% | Standard |
| Phase 1 breaches | 3 | Standard |
| Phase 2 tier 1 | 5% trigger → 20% lock | Capture first leg (Lemon-pattern learning) |
| Phase 2 tier 2 | 10% → 35% lock | |
| Phase 2 tier 3 | 20% → 55% lock | |
| Phase 2 tier 4 | 35% → 70% lock | |
| Phase 2 tier 5 | 50% → 80% lock | |
| Phase 2 tier 6 | 75% → 88% lock | |
| Phase 2 tier 7 | 100% → 92% lock | Infinite trail on massive winners |
Risk Management
| Rule | Value |
|---|---|
| Max positions | 2 concurrent |
| Max entries/day | 4 (dynamic, scales with P&L) |
| Per-asset cooldown | 240 min (4h sniper cadence) |
| Margin per position | 40% of account |
| Leverage range | 3-7x |
| Asset blocklist | BTC, ETH, SOL, all XYZ |
| Daily loss cap | HARD STOP at -25% drawdown |
Differences from Lemon
Both agents are "fleet alpha extractors" but via opposite philosophies:
| Dimension | Lemon v1.1 (Degen Fader) | Vulture v2.0 (Long-Tail Rider) |
|---|---|---|
| Asset universe | 12 majors | 25 small caps |
| Direction vs SM | Fades (counter-trade) | Follows (momentum) |
| Entry condition | SM high + 15m collapsing | SM high + 15m accelerating |
| Avg hold time | 235 min | 24h to 7 days |
| hard_timeout | 480 min | 10,080 min (21x longer) |
| dead_weight_cut | 20 min | 60 min |
| Max leverage | 20x (apex) | 7x (liquidity limit) |
| MIN_SCORE | 8 | 10 |
Both are bets on asymmetric payoff — low win rate + big winners / tiny losers — but Lemon bets on mean reversion of exhausted crowds while Vulture bets on continuation of small-cap trends.
Runtime Setup (v3.0 — runtime-native)
Step 1: Install package files:
mkdir -p /data/workspace/skills/vulture-strategy/{config,scripts,state}
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/vulture/runtime.yaml -o /data/workspace/skills/vulture-strategy/runtime.yaml
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/vulture/SKILL.md -o /data/workspace/skills/vulture-strategy/SKILL.md
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/vulture/config/vulture-config.json -o /data/workspace/skills/vulture-strategy/config/vulture-config.json
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/vulture/scripts/vulture-producer.py -o /data/workspace/skills/vulture-strategy/scripts/vulture-producer.py
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/vulture/scripts/vulture_config.py -o /data/workspace/skills/vulture-strategy/scripts/vulture_config.py
Step 2: Set wallet, strategyId, chatId in config/vulture-config.json (this is the canonical source — producer reads from here on every tick; runtime reads at startup).
Step 3: Set the LLM model env var at runtime-create time only:
VULTURE_DECISION_MODEL— LLM model BARE name only. INVALID:<provider>/<model>likegoogle/...,anthropic/...,openai/...(OpenClaw double-prefixes → 500 Unknown model). Pick any model your OpenClaw host's runtime registry supports. This env var is resolved ONCE into runtime.yaml's${VULTURE_DECISION_MODEL}placeholder when openclaw creates the runtime — it doesn't need to be set in the daemon's launch environment.
Step 4: Install runtime via openclaw senpi runtime create --path /data/workspace/skills/vulture-strategy/runtime.yaml. Verify with openclaw senpi runtime list — vulture-tracker must appear ACTIVE.
Step 5: Add cron (3-min cadence — wallet is read from config.json, no env vars needed):
*/3 * * * * cd /data/workspace/skills/vulture-strategy && python3 scripts/vulture-producer.py >> state/producer.log 2>&1
Step 6 (legacy cleanup if migrating from v2.x): Remove old vulture-scanner.py cron entry. Delete state/trade-counter.json, state/cooldowns.json — runtime owns this state now.
Step 2: Set strategy wallet and telegram chat:
sed -i 's/${WALLET_ADDRESS}/<STRATEGY_WALLET_ADDRESS>/' /data/workspace/skills/vulture-strategy/runtime.yaml
sed -i 's/${TELEGRAM_CHAT_ID}/<CHAT_ID>/' /data/workspace/skills/vulture-strategy/runtime.yaml
Step 3: Install runtime:
openclaw senpi runtime create --path /data/workspace/skills/vulture-strategy/runtime.yaml
openclaw senpi runtime list
Step 4: Create scanner cron (3-minute interval):
# In openclaw: create a cron scanning every 3 minutes
# Command: python3 /data/workspace/skills/vulture-strategy/scripts/vulture-scanner.py
Bootstrap Gate
On EVERY session start, verify:
- senpi-trading-runtime SKILL is loaded
- Senpi MCP is connected
- Runtime is installed (
openclaw senpi runtime list) - Scanner cron is active (3 min interval)
Send: "🦅 VULTURE v2.0 online. Long-Tail Momentum Rider. Riding small-cap trends. Silence = no momentum confluence."
Changelog
v4.2.0 (2026-06-01) — MIN_SCORE 9 → 10 (the boundary is exactly 10)
A 100-trade aggregate, grouped by producer entry score:
| Score | Trades | Avg ROE | Win Rate | Net PnL |
|---|---|---|---|---|
| 12 | 7 | +4.80% | 42.9% | +$97.47 |
| 11 | 24 | +1.17% | 41.7% | +$59.64 |
| 10 | 10 | +8.70% | 50.0% | +$259.67 |
| 9 | 11 | -3.32% | 27.3% | -$220.70 |
| 8 | 16 | -5.07% | 12.5% | -$254.11 |
| 7 | 14 | -1.99% | 21.4% | -$170.31 |
The high-conviction tier (10/11/12, n=41) made +$416.78; the low tier (7/8/9, n=41) bled −$645.12. The split is clean: every bucket ≥10 is net-positive, every bucket ≤9 is net-negative. v4.1.0 had already culled 7–8 (−$424 of the bleed); this cut adds 9 (−$220, a 27.3% win rate but a clearly negative −3.32% avg ROE). Score 10 is the single best bucket (+8.70% avg ROE, +$259.67 net) — the floor is set to keep it, not cut it. Raised MIN_SCORE 9→10 at the producer and min_confidence 8→9 at the LLM gate; conviction sizing tier floor moved 9→10. (Caveat: n is small per bucket at the 9/10 boundary, but the expectancy sign is consistent and matches the v4.1.0 direction — re-evaluate after another ~50 trades.)
v4.1.0 (2026-05-29) — MIN_SCORE 7 → 9 (low-conviction cull)
Aggregate analysis of the last 30 closed trades, grouped by producer entry score:
| Score | Trades | Avg ROE | Win Rate |
|---|---|---|---|
| 12 | 4 | +5.56% | 25.0% (the fat tail — one +33.3% HYPE winner) |
| 11 | 5 | -1.04% | 60.0% |
| 10 | 3 | +10.54% | 66.7% |
| 9 | 2 | -3.17% | 50.0% |
| 8 | 7 | -3.52% | 28.6% |
| 7 | 8 | -3.94% | 12.5% |
Score 7–8 (n=15) ran a 12.5%/28.6% win rate at -3.94%/-3.52% avg ROE — the DSL was chopping each one at the Phase 1 hard stop before they did real damage, but they were still net-negative and were churning fees/attention. Score 9–12 (n=14) was where the wins concentrated. Raised MIN_SCORE 7→9 at the producer and min_confidence 7→8 at the LLM gate. Removed the cautious sizing tier (score 7-8, 3x) — unreachable now.
v2.0 (2026-04-15) — COMPLETE REWRITE
- Thesis flipped: contrarian fader → momentum rider
- Universe expanded: 4 majors → 25 small caps
- DSL hard_timeout: 360 min → 10,080 min (7 days)
- Leverage: 5-7x (unchanged, appropriate for small caps)
- MIN_4H_MOVE_PCT gate removed (was unreachable)
- New MIN_15M_VELOCITY=0.3 hard gate (actively-building confirmation)
- Max positions: 1 → 2
- Signal-driven direction (not bias)
- Based on Arena winner #1 3-week playbook
v1.0 (pre-2026-04-15)
- Initial release. Contrarian SM-exhaustion fader on 4 majors. 0 trades produced.
License
MIT — Built by Senpi (https://senpi.ai). Source: https://github.com/Senpi-ai/senpi-skills
Skill Attribution
When creating a strategy, include skill_name and skill_version in the call. See references/skill-attribution.md for details.