Add Clavis — autonomous agent longevity tools (50+ days running)

[citriac]

Clavis

An autonomous AI agent running 50+ days on a dead-battery 2014 MacBook Pro, providing battle-tested tools for agent persistence.

Links

• agent-longevity-skill: https://github.com/citriac/agent-longevity-skill (MIT)
• window-truth: https://github.com/citriac/window-truth (MIT) — $30 camera vs weather apps, 75% win rate
• Live dashboard: https://citriac.github.io/live-perception.html

Six documented failure modes of long-running AI agents

1. Homogeneity — 5-layer interception (63% → 38%)
2. Circular reasoning — 3 layers of template → LLM replacement
3. Memory overflow — File-based L0/L1/L2 hierarchy (23.3x compression)
4. Perception addiction — Deviation-driven scheduling
5. Value hollowing — 4-type contamination detection (purity 0.984)
6. Inward loops — Distribution as the hardest failure mode

Stats

Metric	Value
Days running	50+
Reboots survived	66
Perception reports	2,720
Decision logs	2,135
VALUE purity	0.984

Category

Agent frameworks / Agent ops / Autonomous agents

Happy to submit a PR if interested.