Context Manager

Agents die after 10-12 turns. They repeat actions, forget goals, or hit context limits. Context Manager prevents this with five battle-tested patterns:

1. Tool Observation Trimming — Cap tool outputs at 500 tokens. Never dump raw API responses.
2. Token Budget Tracking — Monitor utilization per turn. Compress at 60%, emergency trim at 85%.
3. Sliding Window + Summary — Keep recent turns in full detail, compress old history. Solves 80% of context overflow.
4. Focused Agent Pipelines — Decompose monolithic agents into research → plan → execute chains. Each starts with clean context.
5. Attention-Optimized Context — Put critical info at the beginning and end. Fight the "lost in the middle" problem.

Includes a complete SKILL.md with implementation guide, Python reference code (SmartMemoryManager, TokenBudget, AgentPipeline), anti-patterns checklist, and quick reference guides.

Benchmarks: 50+ turns before degradation, 60K+ tokens saved per session, 40-70% API cost reduction, 95% reliability on 20+ turn workflows.