Context Death Spiral Prevention — Claude Code Compaction Primer

Context death spirals do not announce themselves. One session your agent is sharp — the next it is looping, contradicting past decisions, or refusing tasks it handled cleanly last week. This primer explains the structural cause so you can stop blaming the model.

---

⚡ What's Inside

• The compaction sequence What actually happens at the token boundary. Not "the agent forgets" — the real ordered sequence of events that determines what survives and what gets dropped

• The three failure modes Premature trigger, over-aggressive truncation, and the freshTailCount trap that silently destroys your most recent context without any indication it happened

• The threshold logic Why autocompact at 167k/200k is not arbitrary and what happens when you cross each boundary without the right gates in place

---

🏭 Built From Production

After watching the same session failure pattern repeat at the same token boundary across dozens of runs, the cause became undeniable: agents were not degrading randomly. They were hitting an architectural limit with no guard in place.

---

💡 What This Is Not

Not a fix — a diagnosis. If you want the full production compaction architecture with every threshold, gate, and 3-strike circuit breaker, that is the paid SKILL.md. This primer makes sure you understand exactly what problem that file solves before you buy it.

---

✅ Core Capabilities

• ✅ Complete context compaction model explained in plain operational terms
• ✅ Three compaction failure modes with recognition patterns
• ✅ Threshold reference: autocompact 167k, warning 160k, hard block 177k
• ✅ freshTailCount explained — the parameter that controls what survives compaction
• ✅ Entry point to the Agent Compaction Architecture SKILL.md ($19)