Context Windows

Draft · Version 0.3.0 · 7 pages · PDF · Word

What is the context window?

The context window is everything the model can see on one call. It holds the instructions and tool descriptions the harness loads, your conversation, and every file and command output the agent has read. The window has a fixed size, so a long session fills it.

An Accidental Guide to: Agents, Harnesses, Skills, and Tools covers the basics and lists what the window holds. These four ideas explain how the window fills up and how you keep room for the work.

Term	Meaning
Token	A short piece of text used to count the window’s size and usage.
Startup context	What the harness loads before your first message.
Compaction	Replacing older conversation with a summary when the window is nearly full.
New session	An empty conversation that starts with only the startup context.

What takes up room?

The startup context loads before you type your first message and stays in the window for the whole session. The conversation grows with every message you exchange and every file and command output the agent reads. Compaction replaces the conversation with a short summary. A new session has only the startup context. The sizes in the picture are examples.

Tip

A crowded window can give worse answers before it’s full, because the model has more to read and the details you need are harder to find. To see how full yours is, run /context in Claude Code or GitHub Copilot CLI, or /status in Codex, as of October 2026.

01 Token

Field	Text
Identify by	A window size, usage count or price stated in tokens.
What	A token is a short piece of text. Window sizes, usage and prices are all counted in tokens.
So what	The whole window is sent on every call. A large file or log costs its tokens again on every later call.

Field marks

Field	Short answer
Controlled by	The model you pick
Cannot	Be counted from words

Confused with: Word

A short, common word is often one token. A long or rare word can take several.

02 Startup context

Field	Text
Identify by	Content in the window before your first message.
What	Startup context is what the harness loads first. That’s the system prompt, instructions files, skill and sub-agent descriptions, and tool definitions.
So what	Startup context is sent on every call. Long instructions files and many MCP servers take room before you start.

Field marks

Field	Short answer
Controlled by	You and the harness
Cannot	Hold every procedure cheaply

Confused with: Memory

Harness memory is a file of notes loaded with the startup context. The model keeps nothing between calls.

03 Compaction

Field	Text
Identify by	A message saying the conversation was compacted.
What	Compaction swaps older conversation for a summary. GitHub Copilot CLI does it at about 80 percent full, as of October 2026.
So what	Compaction can lose early instructions and exact output. Rules that must last belong in an instructions file.

Field marks

Field	Short answer
Controlled by	The harness and you
Cannot	Bring lost detail back

Confused with: Trimming

Trimming drops old content without a summary. Some harnesses trim before they compact.

04 New session

Field	Text
Identify by	An empty conversation after /clear, /new or opening a session.
What	A new session starts with only the startup context. Only content saved in files carries over.
So what	A fresh window has room for a new task. Have the agent save its plan and progress to a file first.

Field marks

Field	Short answer
Controlled by	You
Cannot	Recall the old conversation

Confused with: Compaction

Compaction keeps a summary in the same session. A new session keeps only files.

What keeps the window useful?

A few habits keep the window small and on topic, so the model can find what matters.

Habit	Why it helps
Start fresh for unrelated work	A new session gives the next task an empty conversation.
Save the plan and progress	A file carries what matters into the next session.
Send long work to sub-agents	Searches, test runs and log reading stay in separate windows. Only summaries return.
Read only what you need	Ask for one function or only the failing tests.
Keep instructions files short	Startup context is sent on every call. Move specific procedures into skills.
Compact at a natural break	When one part of the work is done, run /compact and say what to keep.

Sources

These public sources were checked on 22 October 2026, and harnesses change often.

Source	Date	Address
Anthropic, Claude Code: Explore the context window		code.claude.com/docs/en/context-window
Anthropic, Claude Code: How Claude Code works		code.claude.com/docs/en/how-claude-code-works
Anthropic, Claude Code: Manage costs effectively		code.claude.com/docs/en/costs
Anthropic, Context windows		platform.claude.com/docs/en/build-with-claude/context-windows
Anthropic, Effective context engineering for AI agents		anthropic.com/engineering/effective-context-engineering-for-ai-agents
GitHub, About context management		docs.github.com/en/copilot/concepts/prompting-and-context/context-management
OpenAI, Codex commands		learn.chatgpt.com/docs/developer-commands
Chroma, Context Rot		trychroma.com/research/context-rot
HumanLayer, Getting AI to Work in Complex Codebases		github.com/humanlayer/advanced-context-engineering-for-coding-agents