🐆 JAGUAR v4.0.2 — Striker (helpers-native)

Plumbing-only migration from v3.7. NO thesis change. v3.x striker
scoring + risk.guard_rails preserved verbatim. Producer
flips to in-process SenpiClient, daemon replaces cron, runtime owns
execution. v4.0.2 relaxes one DSL time-cut (dead_weight_cut 12 → 30min);
scoring and gates are untouched.

v4.0.2 changelog (2026-05-29) — dead_weight_cut 12 → 30min

Symptom: a GRASS LONG was cut at exactly 12:00 with price essentially
flat ($0.4925 → $0.4924) — dead_weight_cut fired on a position that
hadn't failed, it just hadn't started moving yet. Entry + exit fees paid
for zero edge.

Diagnosis: the v3.7 dead_weight_cut: 12min was too eager. A Striker
rank-jump can take longer than 12min to accelerate; cutting that early
turns "not moving yet" into "guaranteed fee loss." Per fleet HYPE-run
lessons (fees are the biggest killer; momentum needs room before it
accelerates), the 12min floor was chopping winners-in-waiting.

Fix: dead_weight_cut.interval_in_minutes 12 → 30. At 30min,
weak_peak_cut (25min @ 3% min) becomes the effective backstop for flat
trades, and dead_weight_cut now only catches positions that DID peak
but then went stagnant. hard_timeout (45min) and all scoring/gates
unchanged. Hypothesis test — re-evaluate after ~50 more trades.

v4.0.1 changelog (2026-05-29) — order-of-operations bug fix

Bug: v4.0 appended current_scan to history["scans"] BEFORE calling
detect_striker_signals(). Inside the detection function, prev_scans[-1]
then returned the same scan that was just appended — so every asset's
rank_jump computed as current_rank − current_rank = 0. No asset ever
met STRIKER_MIN_RANK_JUMP. The scanner went completely silent.

Impact: A historical-replay test on the bug missed 25 valid signals over
2 days, including spikes on VVV, XMR, and GRASS. The agent diagnosed the
bug from its own diagnostics on 2026-05-29 ("score 11 GRASS LONG —
FIRST_JUMP #34→#15" fired within minutes of the hot-patch) and the fix
is now ferried into the repo.

Fix: Reordered main():

1. Build current_scan
2. Load history
3. If history is empty, seed it and return early (first-ever scan)
4. Call detect_striker_signals(current_scan, history) FIRST — so
   prev_scans[-1] returns the actual previous scan
5. Append current_scan to history AFTER detection
6. Save history

The detect-then-append sequence is now the contract; future
jaguar-producer.py edits must preserve it.

v4.0.0 changelog

• jaguar-producer.py (NEW) replaces jaguar-scanner.py (DELETED).
  Pure producer — no create_position calls, no trade counters, no
  cooldown state, no held+pending dedup, no daily-cap state file. The
  runtime's risk.guard_rails and per_asset_cooldown_minutes own all
  of that.
• MCP calls flip from mcporter subprocess to in-process
  SenpiClient.mcp_call() via jaguar_config.mcporter_call shim.
• Cron → long-lived daemon via producer_daemon (180s tick).
• fcntl reentrancy guard removed — producer_daemon owns the per-tick
  scanner_lock with stale-PID auto-recovery.
• runtime.yaml now declares the jaguar_signals external_scanner,
  an LLM-pass-through jaguar_entry action with FEE_OPTIMIZED_LIMIT
  entries, and the v3.7 DSL preset + risk.guard_rails verbatim.
• Phase 1 consecutive_breaches_required set to 1 (v2 DSL single-breach).
• Trade chain DB now emits LIFECYCLE_RUNTIME_STARTED →
  DECISION_EXECUTED → ACTION_RESULT → DSL_CREATED → DSL_CLOSED for
  every trade. Per-trade telemetry restored.

What's preserved from v3.7 verbatim

• v3.2 rank-jump threshold floor (STRIKER_MIN_RANK_JUMP = 10)
• v3.3 prev-rank floor (STRIKER_MIN_PREV_RANK = 20)
• v3.3 reason-count floor (STRIKER_MIN_REASONS = 3)
• v3.4 silent-None fix: $3M day-notional absolute liquidity floor
  (replaces unreachable vol_ratio hard gate)
• v3.0 Striker-only thesis (no Stalker, no Hunter, no Pyramiding)
• Conviction-scaled leverage (score≥10 → 10x apex, score≥9 → 7x conviction)
• 15m contribution velocity hard gate (contrib_15m > 0)
• 4h direction-alignment gate
• XYZ ban
• v3.7 risk.guard_rails: max_entries_per_day=3, bypass_on_profit=true,
  daily_loss_limit_pct=10, drawdown_halt_pct=25,
  max_consecutive_losses=3, cooldown_minutes=60,
  per_asset_cooldown_minutes=120
• DSL preset: hard_timeout=45min, weak_peak_cut=25min@3%,
  dead_weight_cut=12min, phase1 max_loss=15%, phase2 ladder
  (7/40, 12/55, 15/75, 20/85)
• Per-asset HL leverage clamp via strategy_get_asset_trading_limits
  (XMR-class small caps with 5x venue ceilings)

---

⛔ Hard Rules

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

RULE 2: The producer does NOT exit positions

DSL ratchet engine owns all exits via runtime.yaml's exit: block.

RULE 3: MAX 2 concurrent positions

Enforced via strategy.slots: 2 in runtime.yaml.

RULE 4: User-conversation Claude sessions MUST NOT trade

Hard rule. When responding to a user message (Telegram ping, status
check, 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 and the DSL
ratchet engine. The daemon is the only entry path. The DSL is the
only exit path. User sessions are read-only.

RULE 5: Verify runtime is installed on every session start

Run openclaw senpi runtime list. jaguar-tracker must be listed
ACTIVE.

RULE 6: 120-minute per-asset cooldown (runtime-enforced)

per_asset_cooldown_minutes: 120 in risk.guard_rails. Producer no
longer tracks this in Python.

---

Thesis

Violent rank-jump explosions only. When an asset rockets from
rank 20+ into the top 10 with a ≥10 rank jump AND its 15m
contribution velocity is actively building AND 4h price is aligned
with the SM direction, that's a Striker. Rare but high-conviction.

Striker = "one amazing trade per day." The producer is the filter;
the LLM gate and risk.guard_rails are sanity checks, not
second-guessing layers.

Entry Scoring (preserved verbatim from v3.7)

Hard gates (all must pass)

• Current rank > 10 (we only want assets climbing into the top 10)
• 4h price ALIGNED with SM direction
• Previous-scan match (must have been on the leaderboard prior tick)
• Rank jump ≥ 10
• Previous rank ≥ 20 (deep-from)
• 15m contribution velocity > 0 (actively building)
• Day-notional volume ≥ $3M (liquidity floor)
• XYZ DEX banned
• Reason count ≥ 3

Scoring contributors (max ~14 pts)

Signal	Points
FIRST_JUMP (was below top 50 OR prev_rank ≥ 30)	+3
IMMEDIATE_MOVER (rank_jump ≥ 10 from rank ≥ 20)	+2
HIGH_VELOCITY (contrib velocity > 10)	+2
DEEP_CLIMBER (prev_rank ≥ 40)	+1
STRONG_4H (price chg 4h > 3%)	+1
DEEP_SM (≥30 traders)	+1
15M_STRONG_SPIKE (contrib_15m > 2.0)	+3
15M_SPIKE (contrib_15m > 0.5)	+2
15M_BUILDING (contrib_15m > 0.1)	+1
1H_ACCEL (contrib_1h > 1.0)	+1
ACCEL_PATTERN (15m > 1h > 0)	+1

MIN_SCORE: 9 — entry floor.

---

Conviction-Scaled Leverage

Score	Leverage
≥10 (apex)	10x
≥9 (conviction)	7x

Producer pre-clamps via strategy_get_asset_trading_limits so the
LLM gate's leverage echo is venue-safe (XMR=5x, etc.).

---

Exit (DSL — preserved verbatim from v3.7)

Mechanism	Value
hard_timeout	45 min
weak_peak_cut	25 min @ 3% min
dead_weight_cut	30 min (v4.0.2: was 12)
Phase 1 max_loss_pct	15.0
Phase 1 retrace_threshold	8
Phase 1 consecutive_breaches_required	1
Phase 2 T0	7% trigger → 40% lock
Phase 2 T1	12% trigger → 55% lock
Phase 2 T2	15% trigger → 75% lock
Phase 2 T3	20% trigger → 85% lock

---

Risk Management (runtime.guard_rails — preserved verbatim from v3.7)

Rule	Value
Max positions	2 concurrent
Max entries/day (when RED)	3
Max entries/day (when GREEN)	unlimited (bypass on profit)
Per-asset cooldown	120 min
Daily loss cap	10%
Drawdown halt	25%
Max consecutive losses	3 → 60 min cooldown
Asset blocklist	all XYZ

Policy: cap losers, ride winners. When today's PnL > 0 →
max_entries_per_day bypass → unlimited entries. When today's PnL ≤ 0
→ enforce 3/day → top 3 by score, then stop.

---

License

MIT — Built by Senpi (https://senpi.ai).
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.