mantis-strategy

🦎 MANTIS v6.0.0 — Slipstream

The mantis is patience plus speed. Stillness, then strike.

v6.0.0 (2026-05-12) — plumbing-only migration

NO thesis change. v5.0 entry filters, confidence-tier sizing, dynamic hard_timeout, and leader-reversal veto logic all preserved verbatim. Six-layer plumbing flip:

1. MCP transport: mcporter subprocess (2.5-5s cold-start) → SenpiClient.mcp_call() in-process HTTPS (~280ms).
2. Signal emit: scanner called create_position directly; producer now emits via push_signal() to runtime /signals. Runtime LLM-gated mantis_entry action opens via FEE_OPTIMIZED_LIMIT, passing leverage + marginUsd through from signal data for tier-based sizing.
3. Leader-reversal veto NOW WORKS. v5.0 emitted VETO_CLOSE stdout JSON that no runtime action consumed (silent no-op). v6.0 wires it as a direct producer-authoritative action — cfg.mcp_call("close_position", ...) fires when the leader reverses >1% from entry value. The runtime cannot express "close this position if a separate asset's 4h move reverses by X%"; that thesis lives in the producer.
4. Scheduler: openclaw cron (60s) → producer_daemon(interval_seconds=60). Long-lived process; scanner_lock with stale-PID auto-recovery.
5. Risk gates: Python MAX_DAILY_ENTRIES + asset-cooldowns.json → declarative risk.guard_rails block. state/asset-cooldowns.json deleted (runtime per_asset_cooldown_minutes: 240 is authoritative).
6. Exit fee: DSL exits switched from MARKET (taker, 0.045%) to FEE_OPTIMIZED_LIMIT (maker-first, 0.015%, 60s ALO timeout, taker fallback).

State files retained (producer-authoritative, NOT runtime-leaky):

• state/position-metadata.json — leader_pct_at_entry per asset (required for veto pass)
• state/entry-log.jsonl — observability only (no longer used for daily-cap counting)

What Mantis does

Mantis monitors market_get_cross_asset_flows every 60 seconds. When a leader (currently BTC) makes a significant 4h move, the tool surfaces correlated alts that historically follow that leader but haven't yet responded. Mantis filters those laggards by strict criteria, picks the highest-confidence one, and emits a signal to enter the laggard in the direction the leader moved — sized by confidence, with a hard timeout calibrated to the alt's typical lag distribution.

The thesis

Pure statistical edge with quantified time bounds:

• Edge: alts that follow a leader 85%+ of the time, with low timing variance, are likely to catch up after a leader move
• Timing: the alt's historical avg_lag_minutes defines the trade's window
• Exit: the catchup either happens within avg_lag × 1.5 minutes — or it doesn't happen at all
• Veto: if the leader reverses mid-position, the entire thesis is invalidated and Mantis closes immediately (producer-side close_position call)

Entry criteria (all must pass — preserved from v5.0)

Filter	Threshold	What it ensures
follow_rate	≥ 0.85	Alt historically follows leader 85%+ of moves
confidence	≥ 0.75	Composite of correlation + follow + lag + SM alignment
gap_pct	≥ 1.5% (abs)	Catchup opportunity is meaningful
sm_starting_to_rotate	true	Smart money is confirming the catchup
lag_stddev_minutes	≤ 90	Timing is predictable enough to trade
Asset	not already open	No doubling up on same asset
Asset	not in per_asset_cooldown_minutes	runtime-enforced (240min)

Sizing — conviction tiers (preserved from v5.0)

Sizing scales directly off the confidence score. Producer computes the tier-based marginUsd and leverage, embeds in signal data; LLM passes both through to the runtime's OPEN_POSITION payload.

Confidence	Margin %	Leverage
≥ 0.92	75%	8x
≥ 0.85	50%	7x
≥ 0.75	25%	5x

Hard cap: 8x leverage, 75% notional of available margin.

Direction

Match the leader: positive leader 4h move → LONG the laggard, negative → SHORT. Mantis never goes against the leader's direction; the entire setup is "the alt is going to catch up to where the leader is."

Exit logic — three layers in priority order

1. Leader-reversal veto (producer-side, v6.0 wired correctly) — every 60s tick, Mantis checks the current 4h leader move vs. the leader's move at entry. If the leader has reversed by >1.0% from entry, Mantis calls close_position directly with orderType: FEE_OPTIMIZED_LIMIT + ensureExecutionAsTaker: true, 30s ALO timeout. The thesis was the leader's move; if it dies, the trade dies — but maker-first with a guaranteed taker fallback saves 0.030% per maker-filled close. Half the DSL exit timeout (30s vs 60s) because vetos are slightly more time-sensitive than retrace-based DSL exits.

2. Dynamic hard timeout (producer-side, via signal data) — avg_lag_minutes × 1.5, clamped to [30, 240] minutes, passed in signal data.hardTimeoutMinutes. If the alt hasn't caught up within that window, the catchup probably isn't happening — DSL hard_timeout: 240 ceiling backstops.

3. DSL trail — standard Phase 1 / Phase 2 trailing logic for in-window winners:

   • Phase 1 max loss: 12%
   • Phase 2 trail tiers: 5/30, 10/60, 15/75, 25/85
   • Weak peak cut: 20 min, 1.5% min value
   • Exits via FEE_OPTIMIZED_LIMIT (v6.0 win)

Risk controls

Control	v6.0 source	Value
Max concurrent positions	strategy.slots	2
Per-asset cooldown	risk.guard_rails.per_asset_cooldown_minutes	240 min
Daily entry cap	risk.guard_rails.max_entries_per_day	6
Max position notional	producer-side cap	75% of available margin
Daily loss limit	risk.guard_rails.daily_loss_limit_pct	10%
Max consecutive losses	risk.guard_rails.max_consecutive_losses	3
Cooldown after loss	risk.guard_rails.cooldown_minutes	60 min
Drawdown halt	risk.guard_rails.drawdown_halt_pct	20%
Reset on day rollover	risk.guard_rails.drawdown_reset_on_day_rollover	false

MCP tools used

Producer side (in-process via SenpiClient.mcp_call):

• market_get_cross_asset_flows — primary signal source
• strategy_get_clearinghouse_state — current positions + account value
• close_position — direct call when leader-reversal veto fires

Runtime side (called by senpi-trading-runtime in response to the signal):

• create_position (FEE_OPTIMIZED_LIMIT entry via mantis_entry action)
• DSL trailing exits via FEE_OPTIMIZED_LIMIT close

Cron cadence

Producer daemon runs every 60 seconds. The cross-asset flow data itself is refreshed daily at 00:00 UTC by the hyperliquid-cross-asset-flow-cron job; Mantis consumes the live output.

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 → runtime action path (entry), the DSL ratchet engine (DSL exits), and the producer's leader-reversal veto pass (the only producer-direct close_position call site, scoped strictly to assets in position-metadata.json whose tracked leader has reversed). User-conversation sessions are read-only.

Operator install

See README.md for fresh-install + migration commands from v5.0.

Status

v6.0.0 — helpers-native migration. Currently BTC-only as the leader. Adding ETH/SOL/HYPE as leaders is a one-line config change in mantis_config.py once pre-computed lag data for those assets ships.

License

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