owl-strategy

name: owl-strategy description: >- OWL v8.0 — Pure contrarian crowding-unwind hunter (senpi_runtime_helpers). Wait for the crowd to overcommit. Wait for them to exhaust. Then eat their liquidations. Producer pushes signals to the runtime via direct HTTP POST; runtime LLM gates them (pass-through), executes via FEE_OPTIMIZED_LIMIT (maker-first), and manages DSL exits autonomously. Conviction-scaled leverage (7/8/10 by score) per the fleet Polar v2.4 / Bald Eagle v3.0 pattern. Entry direction is OPPOSITE of crowd direction — Owl is the only Senpi agent that fades crowding. license: Apache-2.0 metadata: author: jason-goldberg version: "8.0.1" platform: senpi exchange: hyperliquid requires: - senpi-trading-runtime

🦉 OWL v8.0 — Pure Contrarian. Helpers-Native. Maker Exits.

Wait for the crowd to overcommit. Wait for them to exhaust. Then eat their liquidations.

One thesis: the crowd is wrong. Every other skill in the fleet enters WITH momentum, WITH the trend, WITH smart money. OWL is the only skill that enters AGAINST the crowd. The edge: crowded trades unwind violently and predictably.

What changed in v8.0.0

Plumbing-only migration. NO thesis change. v7.1's crowding/exhaustion scoring, persistence gates, MACRO_TREND_GATE, conviction leverage, DSL preset are all preserved verbatim.

Six-layer plumbing flip on @senpi-ai/runtime 1.1.0 (same pattern as Wolverine/Lemon/Cheetah v3.0+ migrations):

Why v8.0: keeps Owl on the same plumbing as the rest of the fleet. Eliminates per-tick subprocess overhead (~30-60s → ~1-3s ticks). No thesis change — v7.1 was the most recent thesis-level work.

Thesis (preserved from v7.1 / v6.2)

• Crowding score: funding extremity (0-4) + SM tilt (0-3, +1 if confirms funding) + OI concentration (0-2). Floor 6.
• Persistence: 1+ hour above floor, with 2-tick tolerance for noise (v5.3)
• Exhaustion score: volume declining + price stalling + volume spike no follow-through + RSI divergence. ≥ 2 distinct signals, score ≥ 5.
• Combined score ≥ 12 to fire.
• Macro trend gate: block crypto fades when |BTC 4h move| > 3% (v7.1).
• 6h post-loss per-asset cooldown.
• Universe: ALL crypto perps with OI > $3M (v6.1 expansion, no top-30 truncation).
• XYZ banned (different unwind dynamics; needs separate calibration).

Files

Producer behavior

Runs every 15 minutes (long-lived daemon, NOT openclaw cron). On each tick:

1. Resolve wallet (OWL_WALLET env var, falls back to config.wallet); fail loud if both empty.
2. Read account value via strategy_get_clearinghouse_state for sizing + dynamic cap.
3. Apply dynamic daily cap (defense-in-depth alongside runtime guard_rails).
4. Fetch universe: market_list_instruments (~80-100 crypto perps with OI > $3M).
5. Fetch SM positioning map (one MCP call shared across all assets).
6. MACRO_TREND_GATE: if |BTC 4h move| > 3%, skip tick.
7. Score crowding per asset + update persistence history.
8. Filter to persisted candidates (≥ 1h above crowding floor).
9. Detect exhaustion for persisted candidates only (saves MCP calls).
10. Filter combined score ≥ 12 + per-asset cooldown.
11. Emit top contrarian candidate via cfg._wrapper_client.push_signal(..., signal_type="OWL_CONTRARIAN_FADE", ...) with conviction-scaled leverage (7/8/10 by score).
12. Persist crowding history + cooldown state under state/<wallet-hash>/.

NO execution code. NO position-tracking. NO DSL state. The runtime owns all of that.

Direction is OPPOSITE of crowd direction. This is the entire edge — Owl is the only fleet agent that fades crowding.

Entry flow

Producer daemon tick (every 900s)
  ↓ Crowding score >= 6, persisted >= 1h
  ↓ Exhaustion score >= 5 with >= 2 signals
  ↓ Combined score >= 12
  ↓ SenpiClient.push_signal(scanner="owl_signals", signal_type="OWL_CONTRARIAN_FADE", ...)
Runtime (127.0.0.1:8787)
  ↓ Schema-validates fields against runtime.yaml
  ↓ LLM gate (decision_model = ${OWL_DECISION_MODEL})
  ↓ Pass-through unless malformed; rejects if direction == crowdDirection
  ↓ OPEN_POSITION via FEE_OPTIMIZED_LIMIT (maker-only, 60s, NO taker fallback)
DSL (runtime-managed)
  ↓ Phase 1 max_loss 35% / single-breach (wide for contrarian retrace)
  ↓ Phase 2 6-tier ladder starting at 5% trigger (Lemon-pattern first leg)
  ↓ hard_timeout 480m, weak_peak 120m@2%, dead_weight 30m
  ↓ Exit via FEE_OPTIMIZED_LIMIT (maker-first, 60s, taker fallback)

Required env vars

The runtime YAML uses these substitutions:

The producer reads:

Risk envelope (declarative, runtime-enforced)

Dynamic daily cap (producer-side, v6 carryover):

Expected behavior

Silence is correct. Owl waits for both the crowd state AND the exhaustion trigger. Most days, one or both is missing. Don't improvise.

v8.0.0 changelog

• Plumbing-only. Six-layer flip per Wolverine v5.0 / Lemon v2.0 template. Thesis preserved verbatim from v7.1.
• owl_config.py: mcporter_call() now routes through SenpiClient.mcp_call() (direct HTTPS). Exposes _wrapper_client proxy for push_signal().
• owl-producer.py: drops fcntl lockfile + subprocess.run([...ingest]). Uses producer_daemon(fn=main, interval_seconds=900, ...) and cfg._wrapper_client.push_signal(signal_type="OWL_CONTRARIAN_FADE", ...).
• runtime.yaml: schema version bump 1.7.1 → 1.8.0. Description rewritten. DSL preset, scanner owl_signals, action owl_entry, consecutive_breaches_required: 1 unchanged.
• State files unchanged. crowding-history.json / asset-cooldowns.json / trade-counter.json continue to live under state/<wallet-hash>/.

License

Apache-2.0 — Built by Senpi (https://senpi.ai). Attribution required for derivative works. 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.