Eve library

Kuan Yin — Troubleshooting

Guide

Revision 1

On this page

Technical doc 32.16.1.10.

Build & test#

Symptom Cause / fix
NX Failed to parse pnpm lockfile … reading 'forEach' A stray pnpm install rewrote the lockfile. git checkout -- pnpm-lock.yaml; do not re-run install.
Nx reports duplicate projects Worktrees expose duplicate project names. Bypass Nx: npx tsc --noEmit and npx vitest run from the lib dir.
Cross-lib test import fails to resolve @kuanyin/* The test lib's vitest.config.ts needs resolve.alias mapping each @kuanyin/* to ../<lib>/src/index.ts (see @kuanyin/evaluation).
Pre-commit "stub-indicator" hit on a benign comment Reword to avoid the token, or annotate // stub:legitimate <reason> / // random:legitimate <reason> on the same line.

Runtime#

Symptom Cause / fix
Auth returns authenticated: false Credential keys are camelCase (clientId, codeVerifier, …); MFA requires an mfaToken in credentials.
Trust composite is unexpectedly low / reliabilityLabel: new Trust factor scores are on a 0–100 scale, not 0–1.
predictCascadeRisk returns 0 currentHealthScore must be 0–1; assessCommunityHealth().overallScore is 0–100 — divide by 100.
Rate limiter never throttles The token bucket refills by elapsed time; with currentTokens > 0 or a stale lastRefillTimestamp, requests are admitted.
Cache reads miss right after write TTL elapsed (setWithTTL) or the key was invalidated by tag/prefix. Check getStats().
Irreversible action silently not applied Working as intended: requireHumanReviewForBan gates bans/terminations for human review.

Analysis quality#

  • Harm score seems low for an obviously hostile message — keyword matches drive the base score; caps/exclamation add boosts. A message with no seeded keywords still scores via boosts but lower. Tune via the harm-keyword seeds / community config rather than hard-coding.
  • Sarcasm/coded language slips through — enable detectSarcasm and detectCodedLanguage in the precognition pipeline and the advanced-nlp dog-whistle/obfuscation detectors; single-message classifiers miss context, so add the threat-detection contextual conversation analysis.

Escalation#

If a problem affects safety (a crisis not routed, a shield not raised, a minor exposed), treat it as a P1 and follow 15-incident-response.md — do not wait on a normal fix cycle.