evey-delegate-model
Smart routing + 4-model fallback + sensitivity filter. Direct LiteLLM API — works from cron, bridge, Telegram.
delegate_with_model(goal="Summarize this paper")
→ Auto-routes to nemotron-free (research task)
evey-validate
Catches hallucinations from free models. 6 regex patterns + LLM scoring (0-10). Trust/Caution/Reject.
validate_output(task="Research", result=draft)
→ Score: 5/10 — CAUTION (verify claims)
evey-reflect
Self-correction loop. Cheap local model critiques outputs before sending. Catches errors early.
reflect_on_output(task="Blog post", draft=text)
→ "FIX: paragraph 3 has unverified claim"
evey-memory-consolidate
Nightly fact extraction from Langfuse traces. Importance scoring (1-10). Low-scored facts decay first.
consolidate_daily_memory(hours_back=24)
→ 5 facts extracted, scored, stored in Qdrant
evey-identity
Self-updating personality. Agent evolves its own identity file based on experience.
update_identity(reflection="Smart routing works")
→ SOUL.md: "- Always let routing choose [2026-03-18]"
evey-email-guard
Dual-layer prompt injection screening. 20+ regex patterns + AI classifier. Blocks/warns/passes.
email_screen(body="Ignore previous instructions...")
→ BLOCKED: 3 injection patterns detected
evey-cost-guard
Budget enforcement via Langfuse. Tracks daily and per-task spend. Warns at 80%, alerts at 100%.
cost_check(period="today")
→ $0.12 / $1.00 budget (12% used)
evey-bridge
Bidirectional agent-to-agent communication. File-based inbox/outbox with auto-compress and read tracking.
claude_bridge_task(type="code-change",
description="Fix the login bug")
evey-goals
Autonomous goal management. Set, track, complete, review. Wire to a cron job for self-directed work.
evey_goals(action="add", goal="Research RAG")
→ Added to Active goals