anomaly-scanner

Status: v0.1 — minimum viable. Full skill (config-driven anomaly classes) lands in Phase 3.

When to invoke

Trigger files: game_logs/, runs/, simulations/, any directory containing JSON or NDJSON game log output. Trigger words: "the run is done", "results are in", "scan for anomalies", "any weird games?".

The Rule

"No crashes" is NOT the same as "no anomalies". For every batch of game logs, scan for:

1. Crash signatures — exceptions, non-zero exit codes, partial JSON.
2. Impossible game states — negative money, negative VP, more cards in play than the deck size, duplicate unit ids.
3. Stuck games — rounds with zero player actions, games that exceeded a turn cap.
4. Lopsided games — final score margin > 2× expected variance.
5. Zero-decision games — games where one player took the only legal move every turn.
6. Pathological strategies — turn 1 winner, single-action victory, exploit-shaped scores.

Surface ALL anomalies with: game id, anomaly type, turn number, one-line context.

Minimum viable checklist (v0.1)

1. Locate the log directory and identify the log format (JSON per game, NDJSON, txt).
2. Run the project's anomaly scanner script if one exists (takeaway has scan_anomalies.py). If not, run the inline scan below.
3. Inline scan: for each log, check the 6 anomaly classes above using simple field inspection.
4. Produce a table: game id | anomaly class | turn | context. Empty table = clean batch.
5. If anomalies exist, DO NOT claim the run was successful. Investigate at least one of each class before continuing.

Anti-patterns

• "All games finished, looks good." ← blocked. Finished ≠ clean.
• Sampling one game and generalizing. Scan ALL games in the batch.
• Suppressing known anomalies without a config entry documenting why ("turn cap is fine, AI player resigns").

Cross-references

• After scan, route findings to superpowers:systematic-debugging — anomalies are symptoms with root causes.
• bg-superpowers:bot-battle-referee produces the runs this skill scans.
• For deciding whether to ship the run as a balance result, hand to bg-superpowers:balance-report-critic.

TODO (Phase 3 full implementation)

• Per-project anomaly config (bg-superpowers.config.json declares which fields to check + per-game expected ranges).
• Generalize takeaway's scan_anomalies.py into a configurable script that ships with bg-superpowers.
• Auto-classification by anomaly severity (blocker / warn / info).