bug-scan

bug-scan

Single source of truth for the classifier-anomaly rules the swing-scan cron has been hunting since 2026-06-21. Designed for cron + LLM-agent surfacing so morning briefs and external scanners stop missing the same anomalies that swing-scan catches.

When to use

• After running run-all-l2 or run-all-l3 — pipe the envelope through --from-json to surface bugs the brief would otherwise miss.
• From a cron tick (no network): --from-state reads the existing swing-scan state tracker and translates its open_findings into the bug-scan envelope.
• Standalone for a fresh scan on ad-hoc tickers (does the fetch + run + detect all in one call).

When NOT to use

• For full L2/L3 analysis — use run-all-l2 / run-all-l3 directly. This skill is a detector on top of their output, not a replacement.
• For live trade signals — bug-scan is purely diagnostic. Never pair its output with execution.

Detection rules

Shape	Tag	Trigger	Severity
pattern_b_1 (absent-with-subs)	[BUG]	l2_fired=False AND ≥2 sub-signals present AND wsum > 0.30	medium / high (wsum ≥ 0.5)
pattern_b_2 (silent)	[BUG]	pattern.present=True AND classification=None AND any sub-signal present	high
pattern_b_3 (ghost)	[BUG]	pattern.present=False AND classification populated	high
weight_drift	[DRIFT]	sub-signal weights sum outside 1.0 ± 0.05	medium
l3_calibration_skew	[INFO]	≥6 ideas, zero with conviction ≥ 4	low (regime signal)
cross_tf_contradiction	[INFO]	same ticker + L2 skill, healthy vs weakening across TFs	medium

The first three shapes are the recurring Pattern B family: classifier anomaly + L1 absence + non-null classification (ghost-classifier shape). Weight drift catches the market-exhaustion 0.900 regression. The L3 and cross-TF checks detect calibration skew and inter-timeframe contradictions.

Usage

# Fresh fetch — positional tickers, comma-separated interval/period
uv run skills/bug-scan/scripts/run.py HYPEUSD SOLUSD AEROUSD \
    --interval=1h,4h --period=1mo,6mo --json

# Read the swing-scan state tracker (no network)
uv run skills/bug-scan/scripts/run.py --from-state --json

# Pipe a pre-fetched run-all-l2 envelope
uv run skills/bug-scan/scripts/run.py --from-json /path/to/l2_envelope.json --json

The --interval / --period flags accept comma-separated lists paired 1:1. A single --period value is reused for every interval.

Output schema

{
  "ok": true,
  "findings": [
    {
      "tag": "[BUG]",
      "shape": "pattern_b_1",
      "ticker": "HYPEUSD",
      "tf": "1h",
      "skill": "market-trend-quality",
      "summary": "Pattern B Shape #1: 3 subs (w=0.60) but pattern absent",
      "wsum": 0.60,
      "present_sub_signals": ["ema_alignment", "pullback_depth", "volume_confirmation"],
      "severity": "medium"
    }
  ]
}

Findings are sorted high → medium → low severity for terminal output. Use --json to get the full list unsorted.

State-resolved paths

• --from-state (bare) reads from $XDG_DATA_HOME/market-skills/swing_scan_state.json. Pass --from-state=PATH to override. The env var must be set — the runner does not fall back to a host-specific default.
• The fresh-fetch mode uses the same data provider stack as run-all-l2 / run-all-l3 (Hyperliquid → CCXT → Kraken → YFinance).

Layer rules

• Sits between L2/L3 and the LLM narrative layer. Reads L2/L3 output, emits diagnostic findings. Does not gate trades.
• Detection rules use analysis.contracts.l2_fired and l2_classification (the single source of truth for "did the L2 actually fire?") — never raw pattern.get("classification").