spider-strategy

name: spider-strategy description: >- SPIDER v5.1.1 — Two autonomous style legs on two wallets, one producer. NOT a copy-trader: each leg scores its own universe to a STYLE and pushes signals; the runtime owns the LLM gate (pass-through), DSL exits, and all risk.guard_rails. SWING leg = Tech & AI multi-day momentum on a DYNAMIC XYZ-equity universe (curated semis/AI/space include-set + auto-caught fresh listings like CBRS/Cerebras and SPCX/SpaceX) plus static crypto alts (SUI/ONDO/HYPE/NIL/GRASS/ZEC), LONG only, 4h+1h trend structure + 24h relative-strength, conviction leverage clamped 10x, wide let-winners-run DSL. SCALP leg = Macro & majors fast mean-reversion (BTC/ETH/SOL/HYPE + xyz:BRENTOIL/xyz:CL), BOTH directions, short-TF stretch + RSI extreme with a 1h trend filter, strict 5x, tight fast-capture DSL. SPIDER_LEG env selects the leg. license: Apache-2.0 metadata: author: jason-goldberg version: "5.1.1" platform: senpi exchange: hyperliquid requires: - senpi-trading-runtime>=1.1.0 - senpi_runtime_helpers

🕷️ SPIDER v5.0 — Two-Persona Style Hunter

Spider runs two concurrent strategy wallets, each a distinct trading style. One producer script (spider-producer.py) serves both; the SPIDER_LEG env var selects which leg a given daemon is. Each leg binds to its own wallet, runtime YAML, DSL, and risk envelope.

Not a copy-trader. Earlier Spider versions chased other traders' books. v5.0 hunts the style: each leg scores its own universe with its own technicals and emits its own signals. The runtime LLM gate is pass-through; the runtime owns execution + DSL + risk.

What changed from v4.0

v4.0 was a single-leg "patient anchor sniper" — one 7-day LONG scored on arena-leader alignment + SM consensus. v5.0 is a full thesis replacement: two style legs, multi-signal baskets, both directions on the scalp leg, and a leg-parameterized producer. The old single-leg runtime.yaml / spider-config.json are removed.

v5.1 made the swing XYZ-equity universe dynamic: instead of a fixed allowedAssets list, the leg rebuilds its equity pool each tick from the live instrument board and auto-catches freshly listed Pre-IPO Perpetuals / AI IPOs (the edge behind Spider's CBRS/Cerebras win — a name that did not exist as a tradeable ticker when a static list would have been written). The scalp universe stays static (majors + energy). See the SWING universe section above.

SWING leg — Tech & AI multi-day momentum (LONG only)

Multi-day trend rider. Holds a small basket of the strongest names in a semiconductor/AI + high-beta-alt universe, sits through drawdowns while the multi-timeframe trend holds.

Universe (config/spider-swing-config.json) — dynamic, rebuilt every tick by build_universe():

• Static crypto alts (cryptoAlts): SUI / ONDO / HYPE / NIL / GRASS / ZEC.
• Dynamic XYZ equities: derived from the live market_list_instruments board (the same call already made for venue-leverage caps — zero extra cost). An XYZ name is eligible if it is liquid (dayNtlVlm ≥ xyzVolFloorUsd, default $5M) and either (a) its bare ticker is in the curated tech/AI/space xyzIncludeSet (NVDA AMD INTC MRVL MU TSM ASML ARM SMSN SKHX DRAM SNDK DELL LITE CRWV PLTR ORCL GOOGL META MSFT AMZN AAPL NFLX IBM COIN MSTR CRCL HOOD SPCX RKLB CBRS), or (b) it was first seen < xyzFreshDays (default 21d) ago and is not in the commodity/FX/ index xyzExcludeSet. Branch (b) auto-catches new Pre-IPO Perpetuals / AI IPOs (Spider's CBRS/Cerebras and SPCX/SpaceX wins) with no code edit. Qualifiers are capped to the top xyzMaxNames (default
  20. by 24h volume to bound per-tick candle fetches; the score gate below does the final quality filtering.

First-seen timestamps live in state/xyz-first-seen-swing.json. On first run every current name is back-dated as already-old, so the auto-catch fires only on names that appear after deploy. A freshly listed perp needs ~24h of candles (≥6 4h bars) before it can score. If the instrument board is unavailable, the leg falls back to a static allowedAssets list.

Scoring (raw integer; minScore 5)

Direction is always LONG — a bearish 4h structure drives the score below minScore, so the leg simply waits. Requires 4h and 1h bullish alignment to clear the floor.

Execution & exit

• slots 3, margin_pct 28% (3 × 28 = 84% max committed), tick 300s
• conviction leverage clamped 10x, then to each asset's venue max
• DSL wide let-winners-run: phase1 max_loss 22% / retrace 12 / 1 breach; all time-cuts OFF; hard_timeout 7d staleness backstop; phase2 ladder 15%→lock0 / 30%→45 / 60%→68 / 100%→80 / 150%→90
• low turnover: max_entries_per_day 12 (raised from 4 — don't throttle the winning style by trade count; the loss-based gates own risk), per_asset_cooldown 4h, bypass_max_entries_per_day_on_profit true

SCALP leg — Macro & majors fast mean-reversion (BOTH directions)

Fades short-timeframe stretch and rides the snap-back. Long-biased but actively shorts when overbought. Strict 5x, minutes-to-hour holds, high win rate, small per-trade size.

Universe (config/spider-scalp-config.json → allowedAssets): majors BTC / ETH / SOL / HYPE + energy xyz:BRENTOIL / xyz:CL.

Scoring (raw integer; minScore 4)

Picks the more-extreme side (oversold → LONG, overbought → SHORT), then:

The −2 knife guard keeps the leg from longing a falling knife or shorting a ripping uptrend — it only fires on genuine intraday extremes.

Execution & exit

• slots 4, margin_pct 15% (small scalp size), tick 60s
• strict 5x leverage clamp (the style's defining discipline), then venue max
• DSL tight fast-capture: phase1 max_loss 5% (tightened from 7% to lose less) / retrace 3 / 1 breach; weak_peak_cut ON (30m @ 1.5), dead_weight_cut ON (25m), hard_timeout 2h; phase2 ladder 3%→lock35 / 6%→55 / 12%→70 / 25%→85
• turnover: max_entries_per_day 18 (trimmed from 30) — fee-sensitive; lower further if net-of-fees underperforms

Leverage clamping (both legs)

Desired leverage = the leg cap (swingMaxLeverage 10 / scalpMaxLeverage 5). The producer then clamps to each asset's Hyperliquid venue max, read from market_list_instruments → instruments[].max_leverage (one call per tick, whole universe). This prevents emitting an unfillable order — e.g. GRASS and NIL cap at 3x, so a swing 10x desire is clamped to 3x for those names. The runtime decision gate also rejects any leverage above the leg cap as a clamp-breach defense.

XYZ (HIP-3) asset handling

XYZ equities/energy require dex="xyz" on market_get_asset_data (an empty dex errors with Asset prefix "xyz" must match dex). Main-DEX assets pass dex="". The main and xyz sub-DEX sections are two VIEWS of ONE cross-margined wallet (not two collateral silos), so both report the same accountValue. get_positions() takes accountValue ONCE via max() across the two views — never sums it. Summing double-counts the balance and sizes every position 2x too large (margin_usd = account_value * margin_pct). assetPositions ARE per-sub-DEX, so those are enumerated across both. Candle fields are HL-native (o/h/l/c/v).

Race-window dedup

Each push_signal(coin) is recorded in state/recent-signals-<leg>.json (per-leg). The producer skips any coin seen within 180s before scoring — covering the gap between a push returning OK and the position appearing in the next-tick clearinghouse pull. On-chain held-asset filtering (get_positions) is the safety floor underneath.

Risk gates (risk.guard_rails)

Entries and exits both use FEE_OPTIMIZED_LIMIT (ensure_execution_as_taker true; swing 60s maker-first window, scalp 20s).

Files

Operator install

See README.md — the two legs are two daemons (SPIDER_LEG=swing and SPIDER_LEG=scalp) on two wallets, each with its own runtime YAML.

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. Each producer daemon handles real signals on its next tick.

License

Apache-2.0 — Copyright 2026 Senpi (https://senpi.ai)