Claude Code context window full: what happens and what to do
A full context window does not end a Claude Code session, it compacts on its own. What loads before you type, when compaction runs, what survives it, and how to keep a long session cheap.

A full context window does not end a Claude Code session, it compacts on its own. What loads before you type, when compaction runs, what survives it, and how to keep a long session cheap.

What an MCP server adds to Claude Code context at session start, how tool search keeps it small, when tool definitions load upfront anyway, and what to do about large tool results.

Which memory files Claude Code loads at session start, what they cost on every request, how to measure them per file, and how to trim them without losing instructions.

Claude Code hooks block by default and run on every matching event. Which hooks slow a session down, how to measure each one, and how to fix a slow hook without losing it.

How to see where Claude Code tokens go on Windows and Linux: the built-in commands, ccusage, Claude-Code-Usage-Monitor, status lines and the SkillKeeper mod that runs inside Claude Code.

What disable-model-invocation does in a Claude Code skill, how it differs from user-invocable, when to use it, how to set it without editing SKILL.md, and why it also frees up the skill listing budget.
How the Claude Code 5-hour session limit and the weekly limit work, why they run out faster than expected, how to check what is left, and how to see which skills, agents and sessions used it up.
Five ways to track Claude Code token usage: /usage, /context, the status line, ccusage and a menu bar monitor. What each one shows, and which answers where your tokens actually went.
Claude Code token usage climbs for a handful of reasons: long context, cache misses, MCP results, memory files, subagents and loops. How to measure which one it is for you before you cut anything.
The skill description limit is 1,536 characters in Claude Code and 1,024 in the Agent Skills API. What gets truncated, how when_to_use shares the budget, and how to check your skills.
A checklist for Claude Code skills that never fire: discovery, disabled auto-invocation, a description that reads like documentation, truncation, and overlapping skills.
Four ways to count Claude Code skill usage: the built-in /skill-doctor, OpenTelemetry, a 30-line transcript script, and a menu bar app. What each shows, and when to use which.