Grizzly Strategy

🐻 GRIZZLY v7.0.0 — BTC Alpha Hunter (senpiruntimehelpers)

Trend-continuation on BTC, v2-runtime-native. Same Kodiak-family architecture as Wolverine / Polar / Kodiak. Producer emits signals, runtime executes via LLM gate, DSL manages exits via maker-first FEEOPTIMIZEDLIMIT.

Trades WITH the trend and WITH the smart money. Not contrarian. SM-opposes is a hard block.

---

⛔ CRITICAL AGENT RULES

RULE 1: Install path is /data/workspace/skills/grizzly-strategy/

RULE 2: THE SCANNER DOES NOT EXIT POSITIONS — DSL only.

RULE 3: MAX 1 POSITION (BTC only)

RULE 4: Verify runtime on every session start

RULE 5: Never modify parameters without family-wide review

RULE 6: HARD_STOP circuit breaker triggers at -25% drawdown

RULE 7 (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 daemon (grizzly-producer.py) and the DSL ratchet engine. The daemon is the only entry path. The DSL is the only exit path. User-conversation sessions are read-only.

If the user asks an action-implying question ("anything close to triggering?", "should I take this trade?"), respond by reading current state — DO NOT execute. The producer will fire on its next tick if a real signal is there. Pattern observed across the fleet: agents have been opening positions immediately after user pings because their conversation sessions had MCP write access. This rule closes that gap.

RULE 8 (FP-001): Quiet hours for low-liquidity windows

Producer skips emission during 00:00-04:00 UTC by default. Apex setups (score >= quietHoursApexBypassScore, default 14) bypass — high-conviction BTC trends can fire any hour; routine sub-apex entries wait until 04:00 UTC.

Configurable via quietHoursStartUtc, quietHoursEndUtc, quietHoursApexBypassScore in grizzly-config.json. Set start == end to disable.

Rationale: fleet-wide pattern of 00:00 UTC pile-ins after daily-cap reset. For Grizzly specifically, BTC overnight liquidity is thinnest in this window and entry slippage can eat a low-conviction edge.

---

Thesis

Pure trend continuation on BTC, never counter-trend.

From Kodiak's lifetime top-3 SOL winners (+$133 / +$87 / +$78):

> "The absolute highest predictor of a massive directional swing is when > the 4H, 1H, 15m, and 5m price momentum are perfectly unified in a > single direction, AND the Smart Money leaderboard is heavily lopsided > (>65% directional consensus) in that exact same direction."

Grizzly v5.0 applies that exact thesis to BTC. BTC is slower and heavier than SOL, so thresholds are tightened — but the pattern is identical.

---

v5.0 — COMPLETE REWRITE

v4.x was a contrarian SM-fader. The April 10 inversion test read "if we flipped direction we'd win 81.8%" as evidence that SM was broken on BTC. The correct read was: the scanner was late, entering after the move exhausted. Flipping direction made those same late entries profitable temporarily, but it also set Grizzly up to fight every real trend.

v4.x lost $245. Grizzly was the only family member running a direction opposite to the family intent.

v5.0 returns to the family pattern:

• REMOVED: CONTRARIAN_FLIP — Grizzly now trades WITH SM consensus
• REMOVED: CONTRARIANEXHAUSTIONGATE — gating contrarian behavior
• ADDED: 3-mode state machine (HUNTING / RIDING / STALKING)
• ADDED: 4-timeframe alignment (5m + 15m + 1h + 4h)
• ADDED: 4TF_aligned bonus scoring
• ADDED: SM strongly tilted bonus (>65% consensus)
• ADDED: SM opposes HARD BLOCK
• ADDED: Move-exhaustion + tiring penalties
• ADDED: RSI filter (70/30 BTC-tuned)
• ADDED: Volume trend + OI-proxy growth signal
• ADDED: STALKING mode — reload on dip after DSL exit
• KEPT: Dynamic daily cap (P&L-aware circuit breaker)
• KEPT: hasrestingorders with auto-cancel stale
• KEPT: 10x MAX_LEVERAGE cap

---

Hard Gates (all must pass)

1. 4h trend structure ≠ NEUTRAL (BULLISH or BEARISH, >60% higher-lows or lower-highs)
2. 1h trend agrees with 4h direction
3. 15m momentum confirms direction (>0.05% in direction)
4. SM opposes = HARD BLOCK (leaderboardgetmarkets)
5. RSI filter — LONG requires RSI ≤ 70, SHORT requires RSI ≥ 30

---

BTC-Specific Tuning vs Kodiak (SOL)

BTC moves ~0.6% in 4h; SOL moves 2–4%. Thresholds tightened ~3x:

| Parameter | Kodiak (SOL) | Grizzly (BTC) | Why | |---|---|---|---| | minmom15m | 0.10% | 0.05% | BTC slower | | rsimaxlong | 74 | 70 | BTC rarely pushes 74 | | rsiminshort | 26 | 30 | BTC rarely pushes 26 | | moveexhaustion | 4.0% | 2.5% | BTC exhaustion at lower % | | movetiring | 2.5% | 1.5% | | | 4hstrong | 4.0% | 2.0% | | | volratiomin | 1.2 | 1.1 | BTC volume always deep | | fundingextreme | 0.005 | 0.003 | BTC lower-magnitude |

MIN_SCORE = 10 (family standard). Dynamic daily cap same as family.

---

Scoring (max ~18 pts, MIN_SCORE = 10)

| Signal | Points | |---|---:| | 4h trend structure | +3 | | 1h trend agrees | +2 | | 15m momentum strength | +1 | | 4TFaligned (5m agrees) | +1 | | SM aligned | +2 | | SM strongly tilted (>65%) | +1 | | 15m velocity fresh (>0.5) | +1 | | 15MSTALEPENALTY (cc ≤ 0) | -3 | | Funding pays direction | +2 | | Funding crowded against | -1 | | Volume ratio ≥ 1.1x | +1 | | Volume rising >15% | +1 | | OI growing >10% | +1 | | RSI room | +1 | | 4h strong (>2%) | +1 | | MOVEEXHAUSTION (>2.5%) | -2 | | MOVE_TIRING (>1.5%) | -1 |

---

Position Sizing

| Score | Leverage | Margin % | |---|---:|---:| | 10-11 | 10x (clamped to HL max) | 30% | | 12-13 | 10x | 37.5% | | 14+ | 10x | 45% |

Leverage auto-clamped to BTC's Hyperliquid max via strategy_get_asset_trading_limits.

---

Exit (DSL) — BTC-Tuned per RatchetStop Timing Guide

| Mechanism | Value | Why | |---|---|---| | hardtimeout | 360 min (6h) | BTC trends take time | | weakpeakcut | 120 min @ 2% min | BTC peaks consolidate slowly | | deadweightcut | 90 min | BTC can sit quietly within a trend | | Phase 1 maxloss | 15% | Same as Kodiak | | Phase 1 retrace | 8% | | | Phase 2 tier 1 | +5% → 25% lock | First profit lock | | Phase 2 tier 2 | +10% → 45% | | | Phase 2 tier 3 | +15% → 65% | | | Phase 2 tier 4 | +20% → 80% | | | Phase 2 tier 5 | +30% → 90% | Apex lock | | Phase 2 tier 6 | +50% → 94% | Monster winner trail |

---

Lifecycle (v6.0 — runtime-owned)

producer.py  → [all gates pass, score >= MIN_SCORE] → client.push_signal()
runtime LLM  → [pass-through gate, min_confidence 7] → OPEN_POSITION
runtime DSL  → [tracks position, ratchets through Phase 2 ladder] → exit
runtime risk → [enforces daily caps + drawdown halt + cooldowns] → block next entry

Producer (HUNTING): scanner evaluates BTC every 3 minutes. All hard gates must pass. NO Python state — emits a signal and exits.

Runtime (RIDING): position_tracker scanner detects open via POSITION_OPENED, fires ON_POSITION_OPENED, DSL engine starts trailing. Producer's held-asset check prevents duplicate emission while position is open.

Runtime (POST-EXIT): when DSL closes, per_asset_cooldown_minutes gate blocks new BTC entries for 60 min. v5.x's STALKING mode and reload-on-dip logic are dropped — DSL owns position lifecycle and the producer does NOT add to existing positions.

v5.x Python state machine (HUNTING/RIDING/STALKING) is gone. Runtime owns all state. No Python state files to corrupt.

---

Family Relationship

All four single-asset specialists run the same architecture:

| Agent | Asset | MIN_SCORE | DSL Profile | |---|---|---:|---| | Kodiak | SOL | 10 | Mid-beta (240/60/90min) | | Wolverine | HYPE | 9 | High-beta, time-cuts disabled (v4.0+) | | Polar | ETH | 10 | Mid-beta (240/60/90min) | | Grizzly | BTC | 12 | Low-beta, time-cuts disabled (v6.0) |

Bug fixes and signal improvements propagate across the family. If one member discovers an edge, all four get it. If one has a bleed, all four get audited.

---

Install

Prerequisite: plugin must be the published @senpi-ai/runtime >= 1.1.0. The senpi-trading-runtime skill must also be installed on this host — it ships the Python Producer SDK (senpi_runtime_helpers) that this producer imports. See README install steps.

mkdir -p /data/workspace/skills/grizzly-strategy/{config,scripts,state}

curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/grizzly/runtime.yaml -o /data/workspace/skills/grizzly-strategy/runtime.yaml
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/grizzly/SKILL.md -o /data/workspace/skills/grizzly-strategy/SKILL.md
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/grizzly/config/grizzly-config.json -o /data/workspace/skills/grizzly-strategy/config/grizzly-config.json
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/grizzly/scripts/grizzly-producer.py -o /data/workspace/skills/grizzly-strategy/scripts/grizzly-producer.py
curl -s https://raw.githubusercontent.com/Senpi-ai/senpi-skills/main/grizzly/scripts/grizzly_config.py -o /data/workspace/skills/grizzly-strategy/scripts/grizzly_config.py

Configure

Edit config/grizzly-config.json with wallet, startingBudget, optional minScore / quietHours* overrides. Runtime resolves ${WALLET_ADDRESS}, ${TELEGRAM_CHAT_ID}, ${GRIZZLY_DECISION_MODEL} from environment.

Install runtime + launch producer daemon

openclaw senpi runtime create --path /data/workspace/skills/grizzly-strategy/runtime.yaml
openclaw senpi runtime list

SENPI_AUTH_TOKEN= \
GRIZZLY_WALLET=0x... \
  nohup python3 -u /data/workspace/skills/grizzly-strategy/scripts/grizzly-producer.py \
  > /tmp/grizzly-producer.log 2>&1 &

senpi-helpers list                                       # daemon visible with recent LAST_TICK
senpi-helpers health grizzly-             # exit 0 = healthy

` = first 8 hex chars after 0x of the strategy wallet (matches the daemon's LOCK_NAME`).

---

Changelog

v6.0.0 (2026-05-06) — v2-RUNTIME-NATIVE REWRITE

• Architecture: v1 full-agency Python scanner (1073 lines) → v2 producer + LLM gate + native risk + DSL maker exits
• Producer: grizzly-producer.py emits via SenpiClient.push_signal() (direct HTTP POST). NO execution code.
• LLM gate: pass-through (min_confidence 7); honors producer signals unless structurally broken
• risk.guard_rails: declarative daily caps / drawdown halt / consecutive losses / per-asset cooldown — replaces Python get_dynamic_daily_cap, set_cooldown, etc.
• DSL: FEE_OPTIMIZED_LIMIT on entries AND exits (~0.020-0.030% fee recovery per maker close)
• Trade chain DB: LIFECYCLE_RUNTIME_STARTED → DECISION_EXECUTED → ACTION_RESULT → DSL_CREATED → DSL_CLOSED per trade
• DROPPED: 3-mode state machine (HUNTING/RIDING/STALKING), evaluate_reload, has_resting_orders, create_position, get_safe_leverage, get_dynamic_daily_cap
• PRESERVED VERBATIM: all 6 entry gates (4h trend != NEUTRAL, 4h structure ≥ 0.75, 1h matches 4h, 15m momentum, base-tech floor, v5.5 macro V-recovery), SM hard block, RSI 70/30 hard gates, multi-factor scoring (~17 max), MIN_SCORE 12, BTC-tuned thresholds, leverage tiers, DSL preset
• Observability fix: forensic-logging pattern — INGEST_FAILED / INGEST_REJECTED / INGEST_EXCEPTION all log rc + stderr + stdout + payload

v5.0 (2026-04-16) — COMPLETE REWRITE

• Direction logic flipped back from v4.x contrarian to trend-following
• Ported Kodiak's full 3-mode state machine (HUNTING/RIDING/STALKING)
• Added 4-timeframe alignment with 4TF_aligned bonus scoring
• Added SM-opposes HARD BLOCK
• Added move-exhaustion + tiring penalties
• Added RSI filter (70/30 BTC-tuned)
• Added volume trend + OI-proxy growth signals
• Added STALKING reload-on-dip mode
• DSL profile retuned per RatchetStop timing guide for BTC (low-beta)
• MIN_SCORE raised 8 → 10 (family standard)
• Margin tiers rebuilt with conviction scaling (30/37.5/45%)
• Inner-order success validation in execute_entry
• getsafeleverage() clamps to BTC's HL max

v4.x (deprecated) — contrarian fader

• Direction-inverted from family intent
• Lost $245 fighting real trends
• Replaced by v5.0

v3.2 — trend-follower (pre-contrarian)

---

Skill Attribution

When creating a strategy, include skill_name and skill_version in the call. See references/skill-attribution.md for details.

License

MIT — Built by Senpi (https://senpi.ai). Source: https://github.com/Senpi-ai/senpi-skills

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Senpi-ai
• Source: Senpi-ai/senpi-skills
• License: MIT

Install and usage instructions live in the source repository linked above.