vulture-strategy

What this skill does

# 🦅 VULTURE v4.0.0 — Long-Tail Momentum Rider (senpi_runtime_helpers)

**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) replaces `vulture-scanner.py` (DELETED). Pure producer — no execution, no counters, no cooldowns, no dynamic-slot bookkeeping.
- `runtime.yaml` is 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_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** (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 >= 3.0` AND `traders >= 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