🦅 KESTREL v2.0 — XYZ Macro Breakout Rider

The first Senpi agent to trade Hyperliquid XYZ DEX (commodities/metals/indices/equities) on a v2 producer architecture.

What v2.0 fixes from v1.x

v1.1 was a v1 full-agency Python scanner that:

• Called create_position directly (no LLM gate, no chain telemetry)
• Used ensureExecutionAsTaker: false on entries (resting-on-book bug)
• Had over-strict scoring math: MOVE_EXHAUSTION penalty cancelled out 4H alignment, leaving most breakouts below MIN_SCORE 6

v2.0 fixes all three:

Architecture: v2 producer + LLM gate + DSL

• kestrel-producer.py emits scored breakout signals via SenpiClient.push_signal() (direct HTTP POST). NO execution code.
• Runtime LLM gate is regime-aware — applies macro vetoes (BTC drawdown propagating to risk-off equities, vol expansion spikes) the producer can't see.
• risk.guard_rails declarative — runtime owns daily caps, drawdown halt, cooldowns.
• Trade chain DB emits LIFECYCLE / DECISION_EXECUTED / ACTION_RESULT / DSL_CREATED / DSL_CLOSED — first time Kestrel has telemetry.

Execution: maker-first, taker-fallback

• Both entries AND exits use FEE_OPTIMIZED_LIMIT with ensure_execution_as_taker: true.
• After 60s ALO timeout, falls back to taker — no more "resting on book" with phantom realized PnL.
• Saves ~0.020-0.030% per maker-filled close vs v1.x MARKET orders.

Calibration: scoring relaxed to actually trade

v1.1 only fired 2 trades / $60 volume in its lifetime. v2.0 calibration:

• 1H base scores +1: 1.5% triggers +3 (was +2), 2% triggers +4 (was +3), 3% triggers +5 (was +4). The 1H breakout is the PRIMARY signal — weight it heavier.
• MOVE_EXHAUSTION threshold 4%→6%, MOVE_TIRING 2.5%→4%. v1.1's penalty kicked in too early, cancelling 4H alignment bonus and yielding net 0-1 contribution.
• MIN_SCORE 6→5 modest relaxation.
• Spread gate 0.2%→0.35% — XYZ tickers have wider natural spreads than crypto.

Expected impact: 2-3 trades/month → 5-15 trades/month.

Thesis (preserved from v1.1.1)

When a macro asset moves >=1.5% in an hour with volume confirmation, the move usually continues for 1-3 hours. Ride the trend with wide DSL.

12-asset universe:

• Commodities: CL (oil), BRENTOIL
• Precious metals: GOLD (SILVER removed v1.1.1 — HL doesn't support)
• Indices: SP500, XYZ100
• High-volume equities: AAPL, NVDA, GOOGL, TSLA, AMZN, META, MSFT

24/7 trading. No market-hours gate — Hyperliquid XYZ trades all hours including weekends and overnight (Iran/oil scenarios are the whole point).

Scoring components (v2.0)

Component	Range	Source
1H breakout magnitude	+3 to +5	hard gate >=1.5%
4H trend alignment	0 to +2	matches breakout direction
MOVE_EXHAUSTION penalty	0 to -2	when 4H >=6% (was >=4% in v1.1)
MOVE_TIRING penalty	0 to -1	when 4H >=4% (was >=2.5%)
Volume surge	0 to +2	vs 4-bar 1H average
SM confirmation	0 to +2	XYZ-filtered SM leaderboard
Funding alignment	0 to +1	favorable for breakout direction

Score-tiered leverage:

Score	Leverage
9+	5x
5-8	3x

DSL preset (v2.0)

Phase	Component	Setting
Phase 1	max_loss_pct	18%
Phase 1	retrace_threshold	8
Phase 1	consecutive_breaches	3
Phase 2 T0	trigger 5% / lock 35%	fleet-standard
Phase 2 T1	trigger 10% / lock 50%
Phase 2 T2	trigger 20% / lock 65%
Phase 2 T3	trigger 35% / lock 80%
Phase 2 T4	trigger 50% / lock 90%	apex
hard_timeout	480 min (8h)	enabled
weak_peak_cut	60 min, min 2.0	enabled
dead_weight_cut	45 min	enabled (catches false breakouts)

Risk gates (runtime.yaml risk.guard_rails)

Gate	Setting
daily_loss_limit_pct	10%
max_entries_per_day	12 (producer dynamic-cap below)
max_consecutive_losses	3
cooldown_minutes (post-loss)	30
drawdown_halt_pct	25%
drawdown_reset_on_day_rollover	false
per_asset_cooldown_minutes	180 (3h)

Producer-side dynamic daily cap (P&L-aware):

• +5% PnL → 12 trades/day
• ≥0% → 8
• -5% → 5
• -15% → 3
• -25% → 1
• <-25% → 0 (HARD STOP)

Fleet patches incorporated

• ✓ Held-asset dedup (3-layer: producer pre-filter, LLM gate, runtime per_asset_cooldown)
• ✓ Post-close cooldown (180min; producer-side runtime-cooldown backstop)
• ✓ Reentrancy guard (producer_daemon scanner_lock with stale-PID auto-recovery)
• ✓ Wallet-from-config (no hardcoding)
• ✓ drawdown_reset_on_day_rollover: false
• ✓ FEE_OPTIMIZED_LIMIT entries AND exits with ensure_execution_as_taker: true
• ✓ Fleet-standard T0/T1 ladder
• ✗ FP-001 quiet hours — deferred fleet-wide

Hard rule for user-conversation Claude sessions

User-conversation Claude sessions 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 daemon (entry path) and
the DSL ratchet engine (exit path). User-conversation sessions
are read-only.

License

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