Lynx Strategy

🐯 LYNX v1.0.0 — Adaptive MIN_SCORE Self-Tuner

The Vulture v4.1 story productized. Lynx is a momentum agent that audits its own closed-trade history every 6 hours and raises its own MIN_SCORE when low-conviction buckets bleed.

Why this strategy exists

When Vulture's autonomous agent ran the 30-trade audit that led to v4.1, it manually identified that Score 7–8 trades had been net-negative and raised MIN_SCORE 7→9 by hand. The fix worked. The question Lynx asks: why is that operation manual?

Lynx bakes the same operation into a scheduled cron. Every auditEverySec (default 6h), it:

1. Pulls its own closed-trade history via audit_query(user_ids=[senpiUserId], action_type="close", limit=auditLimit)
2. Parses the entry score from each trade's ai_reasoning field
3. Buckets trades by score and computes {n, avg_roe_pct, win_rate_pct} per bucket
4. Identifies the highest-scoring bucket at or above the current floor that has n >= minBucketN AND avg_roe_pct = 4% → +3, >= 2% → +2, >= 1% → +1 | up to 3 |

| 1h confirmation: 1h move aligned with 4h direction, |move| >= 0.5% | +2 | | Smart Money aligned: SM tilts in direction, >= smTiltMinPct (55%) | +2 | | Volume rising: 12h vol vs 36h baseline ratio >= 1.3 | +1 |

Initial MIN_SCORE: 4 (permissive — gives the audit data to work with). Auto-tunes up to maxMinScore: 7. Operators see the current MIN_SCORE in every tick's output (currentMinScore field) — so they can watch Lynx learn live.

DSL preset (let_winners_run — momentum-class)

Standard momentum-class let-winners-run: maxloss 20%, retrace 12, 7d hardtimeout, no weakpeakcut, Phase 2 ladder +10/0 → +20/25 → +35/50 → +60/70 → +100/85. The Lynx innovation is in entry, not exit.

Scanner pattern

NEW archetype #15: Self-tuning / adaptive-threshold agent — see senpi-trading-runtime/references/producer-patterns.md. Primary MCP calls: market_get_asset_data per whitelisted asset (the producer signature), leaderboard_get_markets (SM gate), audit_query (the self-tuning audit — every 6h). The pure functions (parse_score_from_reasoning, compute_bucket_stats, recommend_min_score, should_update_threshold, pct_move, trend_direction, lynx_score) are unit-tested in tests/test_signal.py.

Operator install

See [README.md](README.md).

Changelog

v1.0.0 (2026-05-29) — initial release

First fleet agent to modify its own behavior based on its own trade history. Productizes the Vulture v4.1 manual cull as a scheduled audit. Introduces archetype #15 (self-tuning agent). 19/19 unit tests covering both the self-tuning logic (parsescore, computebucketstats, recommendminscore, shouldupdatethreshold) and the scoring logic (pctmove, trenddirection, lynxscore). Taker-true entry, disown-safe launch.

License

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

Skill Attribution

When creating a strategy, include skill_name and skill_version. See references/skill-attribution.md.

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.